, a columnist at The Washington Post, shared her AI workflow on X.

Basically, it’s a research assistant and copyeditor. While she doesn’t have a human research assistant, her job at The Washington Post obviously provides a human editor and copyeditor. It doesn’t seem too crazy to me—she doesn’t even seem to generate any text that goes into her column at all.

Of course, the Internet went bonkers.

Geez.

I’ve been open about how I use AI for writing. I even showed the entire outline, editing, etc. process with AI in my article about agents just to show what the AI can and cannot do.

The short summary of how I use it: I have ultra-detailed bullet outlines (often with literal turns of phrases I want to be incorporated), and I ask the AI to try to clean it up into an actual piece. It also grabs links/sources when I cite something but don’t actually link it. It’s basically doing what I would trust a student intern to do.

And then obviously I read and edit it. In that case, I have four rounds of edits, including two where I literally go in and change a LOT of text. It’s not really uncommon for me to have 12+ rounds of edits at this point. Of course, I do ask it to fact-check as well (I actually ping-pong fact-checks between Claude and ChatGPT to check each other). And guess what, sometimes I disagree with both of them or catch things that neither do!

Is that a problem? Well, perhaps if you uncritically rely on AI to just be “perfect” and don’t read the work. But I never totally trust even human editors and copyeditors.

I had a world-class team of editors for my book, and I still had to read what they did carefully. They missed some things. They introduced errors (due to lack of subject matter knowledge). They sometimes mangled AI-industry phrasing (like turning “compute” into “computational power” everywhere).

Questions About Authorship

I’m using AI more extensively than Megan McArdle is. But I’d assert that it’s unquestionably my writing—just as a human-edited article still is, when I’m on the byline. I ultimately need to approve of whatever goes out. More than that, I need to make it my own.

(As a note, major publications sometimes have weird processes around this, especially for “final edits” where the writer might NOT actually see it or the title/subtitle... which ironically makes the “but it’s still my work” case weaker for human-edited pieces than AI-edited ones, but this is still 80/20 true.)

I give examples in my book, What You Need to Know About AI where:

• Photographers should get the credit, not “Kodak” or “Canon”

• Some well-known artists like Andy Warhol (Campbell Soup can guy from the 1960s)—who often had assistants make their work and didn’t actually touch it themselves—are still credited (I think rightly) with their art pieces

• Death of an Author by novelist and journalist Stephen Marche, an “AI-generated” novella published in 2023, which plays a prominent role in one of my book’s chapters. Summary is that he used AI—mostly ChatGPT—to generate it (”95%” of the words)... but still wrote the plot himself and basically kept mashing “re-generate” until he got what he wanted. I argue pretty strenuously that it’s pretty hard to accept the prior two examples but not this as “authorship.”

I think photography is especially instructive. It changed the nature of “art” significantly at the turn of the century—and arguably unleashed one of the most creative artistic periods in history. Cubism/Picasso, Impressionism/Monet... all of it came about in part because the mechanical skill of merely capturing the scene in front of you was less interesting when a “machine” (cameras) could do it.

But not only did art evolve, photography itself became an art form. I think you’d find very few people who’d deny photographers their due as artistic creatives in their own right.

Subscribe now

Hallucinations?

The other major critique that got was that fact-checks and paper summaries would be terrible, because the AI would just “hallucinate” all over the place. On the Podcast with , , and , she pushed back on people calling them “stochastic parrots”—implying the label no longer applies.

Except... it does. If you look under the hood, they are statistical models that try to match a plausible conversation you’d have given what you say. I get her frustration, but the term itself is not the problem. The problem is people using it as shorthand for “useless.” After all, as we’ve seen, it’s very much not!

describes this better than I do. They ultimately just do things “within their dataset.” As per , ML engineer at Google:

I’m also intrigued by how many people are surprised by the fact that models can’t perform outside of their dataset when by definition this is how deep learning works. The real way to get them to generalize further is by finding unique ways to expose to more data.

They have a lot of data and post-training these days, so they are actually fairly good. They will never be perfect, though—that’s part of the entire tradeoff of the way we’ve built modern deep learning AI/agents (another plug for my book: I explain why it’s impossible to totally banish “hallucinations” without making current AI commercially useless... like the AI from the 1980s).

Funny enough, I did not use AI in any appreciable fashion for writing my book. One is that my publisher (and major platforms like Ingram) ban AI in writing. I had to certify that I didn’t use AI in order to publish it. So, it was already a strikeout there, but I could have used it like McArdle. Research assistant. Fact-checks. Nothing against the rules about that. But I didn’t. Why?

Well, it wasn’t because I have a phobia of AI. After all, I was writing a whole book about how it works. No, it was because back then hallucinations were pretty bad. This wasn’t that long ago. We’re talking about late 2024 and early 2025! But especially at the beginning of my book process, anytime I tried to get AI to help me grab papers or research, it’d hallucinate like crazy!

Half of the links for papers or articles wouldn’t work—because the url was made up. The other half? Usually I’d click through and it wouldn’t be what it was described as at all. With a research assistant like that, well, I don’t think it’s a mystery why I shelved trying to use AI to do any research for me, let alone fact-check.

Given my own experience, I’m not too surprised that some people think using AI like McArdle is insane. Times have changed, though.

Again, human research assistants and fact-checkers also make mistakes. But I’d assert the reason why AI doing this felt so useless and terrible was it made mistakes no human would ever make.

This is the same problem that self-driving cars keep running into.

Waymos (and most major autonomous platforms) have been safer than humans for a while. The above chart is from RMI on Waymo in 2017. While I might have looked with some skepticism back then at hitting exceptional/novel conditions, Waymos have only gotten safer since, and it’s hard to argue at this point that they’re worse than the average human driver—who is pretty terrible.

The problem, even now?

Self-driving cars sometimes make mistakes that no human driver would. They get confused in circumstances that they have never encountered (...outside their dataset...). They require intervention when some of the decision process gets complicated. They easily get trapped by malicious people doing things like throwing cones on them (regardless of your stance on them, I think terrifying people riding in them is not great—and ironically, those people only feel safe to do so because they expect the self-driving cars to properly stop instead of run them over!).

Since 2024, the result of a massive amount of post-training, harnesses, and tools by the frontier labs is that LLMs not only fail more predictably (e.g., modern events like Claude getting very confused when I was writing about the Iran conflict), but they also fail in more humanlike ways.

They slightly misinterpret studies. They get overly conservative about the interpretation of certain data. They miss studies that would change the conclusion quite a bit...

Basically, they have failure modes very similar to working with human assistants. The benefit of a huge amount of money, RLHF (Reinforcement Learning Human Feedback), and just sheer work put into them is that they’ve been tuned to behave in this way. AVs (autonomous vehicles) do not have the kind of volume of data and freedom to do so, though I’m sure the companies are trying their best to do the same thing.

That, however, makes it far easier for a human that doesn’t just dedicate their life to using AI to actually use it effectively. The human can predict how the model will fail. And that makes errors—hallucinations—acceptable and something that can be worked around, instead of being a showstopper like it was a mere few years ago.

My Writing Process in the Future?

I’m working on another book (well, proposal and research anyway). I’ll share more about it over time, but one obvious question from this is:

Will I be using AI to write my next book?

Well, unless publisher contracts and Ingram policies change, no—just purely from a requirement perspective. However, the most valuable use of AI for me these days is actually the same as McArdle. I get to utilize a process that feels a bit more like having a research and editing team—even if I don’t and will never find it economical to have one for, say, Substack posts.

I still read through every post. I still check all of the sources (unless I obviously know what they are/actually send the AI to grab the exact sources I want... which is probably 80%+ of the time). I still change a massive amount of text for “style,” despite having a process of incorporating the “James style” into my CLAUDE.md as part of finishing every piece (finding where the AI fell short and trying to modify its instructions).

The AI writing agent is getting better—but I think it’s probably like working with a human assistant. I’ve had junior analysts write memos from outlines that probably had nothing wrong. But I still changed things... maybe just because I felt like it in the moment and just really liked a turn of phrase more. Given that, the AI may never get to “zero changes” from me. But, then again, I’ve never had that happen with human drafters either.

The Future of Writing

As for, perhaps, the final potential critique: maybe my writing muscles will be “rusty” after having AI draft stuff for me.

... Maybe? But I will point out that for certain pieces that just catch my attention, like this one, I bang them out in a stream of consciousness. That’s how a lot of chapters in my book came together (... which also saw a lot of chapters get discarded wholesale as I moved stuff around).

Zero AI involvement.

Relative to my process/drafting from an “initial sketch” (detailed bullets) in my “How I Use AI Agents” piece... I one-shotted this article, basically! Screenshot as I’m writing this now. I’ll probably run this through fact-check and copyedit, but I expect it’ll largely be in the form it is now that came out of sitting down for about 45 minutes to bang this out.

Which even as I’m trying to push the boundaries of what I can do with AI—if nothing else, to understand it—I still end up often doing. Sometimes my “sketch”/outline to give to the AI… is basically my whole article. But so it is.

Honestly, if you enjoy writing, I don’t think having that kind of writing session is ever going away. We can take photographs now, but plenty of people still enjoy painting landscapes and portraits. Sure, fewer people can “make money” from it... but how many people ever did really make much money from it? There’s a reason art has had patrons for much of its history.

Writing, as per the Podcast episode, is also somewhat brutal with few, highly elite and coveted slots—which McArdle suggested might be part of the reason people are so angry, since there’s some jealousy and anger about the industry mixed in.

Just as fewer people can make money doing street art portraits (though, obviously, this still exists in limited fashion in, say, tourist areas), fewer people can make money with listicles and the classic human writing slop like “10 Reasons Why Your Cat Doesn’t Love You,” or whatever. But it’s not like people were ever making much money from that either...

In any case, will AI change writing? Probably. Though in the most artistic expressions of it, I expect it’ll probably elevate it. After all, if we all get used to reading AI summaries, elegant writing will suddenly stand out that much more—just as a Picasso is quite different than a photorealistic, well, photo. Not to do the classic, “and it’ll diversify the field!” but it will likely also help those who have trouble just typing words, either from physical or non-physical disabilities.

Just as we shifted from pen to typewriter to word processor (... which was decried for “losing something” even to this day), we will likely shift in how we produce words.¹

However, it was always the case that we didn’t really care about the literal “shapes on the page” that happened to be how we convey meaning. The valuable part of writing has always been the expression and humanness—not the mere mechanical act of getting those words on the page.

1. As an aside, if you want a particularly fascinating version of this, 提笔忘字 in Chinese means “character amnesia,” referring to an ever-more-common scenario of not actually being able to write characters by hand, despite recognizing them. Why? A lot of Chinese is written on either keyboards or mobile devices with SCIM, which is like super-amazing autocomplete where you can often just write the first pinyin “letter” of each word in your sentence to write the whole thing. Is it bad? I don’t know, but it is a change in the relationship to words, mediated by technology. Chinese literature and language haven't collapsed, though.

Thanks for reading!

I hope you enjoyed this article. If you’d like to learn more about AI’s past, present, and future in an easy-to-understand way, I’ve published a book titled What You Need to Know About AI.

You can order the book on Amazon, Barnes & Noble, Bookshop, or pick up a copy in-person at a local bookstore.

It’s a little dated now, but the February jobs report landed like a grenade. Payrolls fell by 92,000—the third decline in five months—and the December number got revised down to negative 17,000. Headlines did what headlines do. “Economy Sheds Jobs.” “Worst Report Since the Pandemic.”

This, of course, plays into everyone freaking out about AI destroying jobs and creating a “jobless economic boom.” But is that really what’s happening?

I get it. The numbers look bad. Since Trump took office in January 2025, the economy has added somewhere around 10,000–15,000 jobs per month on average, depending on how you cut the start date—and that’s after the annual benchmark revision that slashed 2025’s full-year job growth from +584,000 to just +181,000. In 2024, the benchmarked average was about 120,000 per month (the pre-revision number was closer to 170,000, which is what most people remember). That’s a dramatic deceleration by any measure.

But here’s the thing: the denominator has changed. Dramatically. And almost no one writing panic headlines is adjusting for it.

(This is probably a dangerous article that might age like milk when Iran’s oil situation might actually plunge the world into a recession, but there have been so many breathless takes citing job numbers that I feel like I need to write this. I get an insatiable itch to write about this stuff when people obviously misinterpret or misrepresent economic stats.)

The Breakeven Number Has Collapsed

There’s a concept in labor economics called “breakeven employment growth”—the number of new jobs the economy needs to create each month to keep the unemployment rate stable. It’s not a fixed number. It depends on how fast the working-age population is growing, which depends on births, deaths, retirements, and—critically—immigration.

In 2024, breakeven was somewhere around 130,000–230,000 jobs per month, depending on which model and time horizon you used. The San Francisco Fed estimated that the short-run breakeven got as high as 230,000 during the immigration surge, while the long-run breakeven was closer to 70,000–90,000. The economy needed to create a lot of jobs just to absorb new entrants into the workforce—immigrants arriving, young people aging in, and people coming off the sidelines.

That number has collapsed.

The Kansas City Fed’s research, using CBO population projections for 2026, estimates a breakeven range of roughly 29,000 to 77,000 depending on assumptions about immigration’s contribution. KC Fed President Jeff Schmid said in January that his staff’s estimates put breakeven “in the vicinity of 50,000 jobs per month.” In their January 2026 update, Brookings and AEI estimated a breakeven range of 20,000 to 50,000 for late 2025, and noted it could dip negative in 2026 as immigration slows further.

Let’s sit with that for a second. A year and a half ago, the economy may have needed 200,000+ new jobs a month to stay in equilibrium. Now it might need somewhere between 30,000 and 70,000, depending on whose model you trust. That means a payroll print of 50,000—which would have been terrifying in 2023—is around or modestly above breakeven under several current estimates.

Why? Immigration Fell Off a Cliff

The Census Bureau released population estimates in January that tell the story plainly. Net international migration dropped from 2.7 million in the year ending July 2024 to 1.3 million in the year ending July 2025—a 54% decline. If current trends continue, the Census Bureau projects net migration will fall to roughly 321,000 by July 2026. That would be another decline of nearly a million people.

The CBO’s projections tell a similar story through a different lens. Compared to their January 2025 baseline, they now project net immigration to be 1.6 million people smaller in 2025 and 960,000 smaller in 2026. Their latest demographic outlook estimates that net immigration declined to about 410,000 in calendar year 2025—a fraction of the 3.3 million they estimated for 2023.

(A quick note on the numbers: CBO uses calendar-year estimates, while the Census Bureau uses July-to-July periods, and their methodologies differ. The specific figures don’t line up perfectly. But the directional story is the same across both: a historic collapse in immigration flows.)

These are enormous swings. We went from a period of historically elevated immigration to what might be the lowest levels in decades. The Census Bureau’s own analysis notes that current trends are “trending toward negative net migration” and that if they continue, “it would be the first time the United States has seen net negative migration in more than 50 years.”

Fewer people entering the country means fewer people entering the labor force. Fewer people entering the labor force means you need fewer jobs to keep the unemployment rate from rising. It’s a lot of stats and numbers, but the concept is really pretty straightforward.

Subscribe now

So Where Does That Leave Us? Not Great, But Not Bad

Let me be clear about where the labor market actually sits, because it’s not a clean story in either direction.

The February report was genuinely distorted. A Kaiser Permanente strike involving about 31,000 workers in California and Hawaii hit during the BLS survey week, and healthcare employment swung to -28,000 for the month—after averaging roughly +30,000/month over the prior year. The strike has since been resolved. Federal government employment also dropped by 10,000, continuing a trend that’s taken 330,000 federal jobs off the books since October 2024—an 11% reduction.

Strip out the strike effects and the federal workforce reduction, and the underlying picture looks more like stagnation than contraction. Look at the chart: most of the monthly bars in 2025 and early 2026 are bouncing around in the zone between the old breakeven and the new one. Against the 2024 standard, they look terrible. Against the current breakeven estimates, they look like what they are—a labor market roughly treading water.

But stagnation is not great, either. Average unemployment duration rose to 25.7 weeks in February—a multi-year high. If you’re out of work, you’re staying out of work longer. The job-finding rate has clearly deteriorated. And the Minneapolis Fed published an important analysis in October 2025 arguing that immigration declines can only explain about 40–60% of the slowdown in payroll growth—the rest reflects genuinely weaker labor demand.

They found that states and sectors with fewer unauthorized workers actually saw bigger employment declines, which is the opposite of what you’d expect if immigration were the whole story. They also found that real wage growth has slowed more for low-wage workers—who should be benefiting most from less competition if immigration were the only story—which points to a broadly cooling labor market, not just a supply-side adjustment.

That’s a meaningful finding. It means the labor market is softer than just the immigration math alone would suggest, but not as soft as the headline payroll numbers imply.

So, as per my usual admonition... it’s nuanced.

NEC Director Kevin Hassett made the same basic argument after the February report, saying breakeven is probably in the 30,000–40,000 range and that current job growth is consistent with a strong economy. I wouldn’t go quite that far—a “strong economy” doesn’t usually produce negative payrolls and rising unemployment duration—but the basic arithmetic is directionally right.

The St. Louis Fed put it well: “the biggest swing factor in the breakeven growth rate is the population growth rate.” And the population growth rate just changed enormously.

Putting it all together, the labor market is mediocre. Not in free fall. Not strong. Not terrible. It’s somewhere in the zone where the economy is creating enough jobs to roughly keep pace with the (much smaller) population inflow, but without enough surplus to actually pull unemployment down or create the kind of dynamism you associate with healthy expansion.

It’s a low-growth equilibrium—stable, but fragile. (Ah, fortunately there isn’t something like a global oil shock and war going on in the Middle East...)

Share

Why This Matters Beyond the Monthly Number

There are two broader implications worth flagging.

First, the Fed. If breakeven really is somewhere in the 30,000–70,000 range, then the gap between what the economy is producing and what it needs is smaller than the headlines suggest. That weakens the case for aggressive rate cuts. The labor market isn’t tight, but it’s not collapsing either—and cutting rates won’t conjure up missing immigrants. The supply side of the economy has become the binding constraint, and monetary policy is a demand-side tool. Rate-cut expectations have shifted toward mid-year, but the Fed has been explicitly cautious, and this data doesn’t clearly force their hand.

Second—and this connects to what I wrote last week about AI and the pipeline problem—a labor market with structurally lower breakeven creates an environment where organizations can more easily justify not hiring. If you only need 30,000–50,000 jobs a month to keep unemployment flat, every marginal hire feels less urgent. Senior people with AI tools can handle more. The junior pipeline narrows further. The demographic math is giving companies cover to do what many were already inclined to do: run leaner and let attrition do the work.

That’s not a disaster today. But it will be! Fewer people entering, fewer people being trained, fewer people developing the judgment layer that makes them effective evaluators of AI output—it’s the same pipeline problem I described in “AI Agents and Jobs”, now with a demographic accelerant.

After all, we aren’t uniformly not hiring across the population. I assert that we’ll likely see greater unemployment among younger cohorts if AI really starts to get widely adopted. Despite some headlines that this is already happening, the actual evidence is mixed (which I won’t deep dive on here—the short summary is we’ll need to wait and see).

The Bottom Line

I think I’m fulfilling my brand here of being boringly level-headed. It’s certainly less of a sexy, clickbait article to be saying that there’s nothing much to write home about in terms of the economy vs. screaming that the sky is falling from AI.

Let’s be clear: the jobs numbers aren’t good. The White House trying to jawbone it into being “strong” doesn’t make it so. But they aren’t the crisis that various articles (and certain Substacks) are declaring. And really, it’s unlikely to be actually driven by AI given lag times and actual adoption so far. I don’t think the Fortune 500 are using agents in the same way power users are.

The single most important piece of context is that breakeven employment growth has fallen dramatically—from perhaps 200,000+ to somewhere in the 30,000–70,000 range—in the span of about a year, driven primarily by a historic decline in net immigration. A payroll print that would have been alarming in 2024 is merely mediocre in 2026.

So, all told, the labor market is cooling, not cratering. The Minneapolis Fed is right that there’s genuine demand weakness beyond just the immigration effect. But the people telling you the sky is falling are reading the thermometer without noticing that they moved to a different climate.

We’ll obviously have to keep monitoring—the March report drops April 3rd, and the Kaiser strike effects should wash out. If payrolls remain negative after that, the “it’s just distortions” argument gets harder to make. But for now, the most defensible read is that we’re in a low-growth equilibrium, breakeven has fallen dramatically, and the labor market is roughly treading water in a pool that’s gotten a lot smaller.

Do I think there won’t be disruption from AI? No, I think it will be disruptive. But the way these things work their way through the economy is complicated. And this isn’t it.

(And... once again... Iran, oil, war...)

Thanks for reading!

I hope you enjoyed this article. If you’d like to learn more about AI’s past, present, and future in an easy-to-understand way, I’ve published a book titled What You Need to Know About AI.

You can order the book on Amazon, Barnes & Noble, Bookshop, or pick up a copy in-person at a local bookstore. If you’re in Southern California, I’ll be at Lido Village Books in Newport Beach, CA this Thursday 3/26 from 4-6pm signing books in person if you’re around and want to say hi!

The narrative has shifted. Agents are now mainstream. People have been freaking out about Andrej Karpathy’s market map (karpathy.ai/jobs) that shows a huge portion of jobs are “exposed to AI.” Ben Thompson is now arguing he no longer thinks we’re in a bubble because agents are likely to generate huge value—and profits.

My article from last year, “The Boring Phase of AI,” was right about the impact of agents—but I guess people don’t find it so boring after all. I probably contributed somewhat to this trend with my article about how I personally use AI agents as well. As a note, since that article, my agent workflow has advanced so much that what I outlined there is fairly primitive to me now. What a difference a few weeks make in our current environment.

The big question on everyone’s mind, of course, is whether or not a huge portion of our population is now out of a job.

A Recent, Personal Case Study

I was recently onboarding a new analyst. She was quite patient with my, likely unreasonable, demands that she automate certain of our manual spreadsheets to incorporate new data and update their tables/charts all without human intervention.

Unreasonable. At least in part because her skills are not in software engineering. She was a good sport about it all, though. She was taking an introductory course in Python and offered to try to do what she could with VBA while trying to scramble up the massively steep curve I laid out for her. It’s a lot to ask, especially since, like most financial institutions, most of what we do runs on Excel.

I’m not a total monster, though. I said I understood but wanted to show her how I’d do it. I took the largest spreadsheet, dumped it into a new directory, opened Claude Code, and told it to break the spreadsheet up into distinct Data, Analyses/Logic, and View layers. I wanted it to prototype the charts/tables in HTML and clearly document how things were used... as well as what specific data was needed so we can automate pulling it from our financial database later. Then I hit enter.

From there, Claude Code was not particularly fast. It took around 30 minutes for it to finish while I showed the analyst what it was doing and explained why I asked for what I did. Software engineers among my readers probably recognize what is essentially MVC (Model-View-Controller). Experienced users of Claude Code will know that standalone HTML is the absolute easiest way for it to provide easy, interactive tables/charts right now. And infrastructure engineers understand exactly why I wanted the clear spec on what data was required.

Still, during that entire time, seeing it quickly sort through an Excel file that is painful to update—and can take hours to debug when something breaks or something new is required—and rapidly build everything... I think it was fair to say she was, uh, surprised. I think she expressed a sentiment along the lines of everyone who did her job at a bank likely losing it.

Anyway, when it finished, we walked through it. We quickly spotted immediate errors (... a blank chart is a pretty good giveaway), but it was pretty good as a one-shot attempt from Claude (Opus 4.6). Still, it never crossed my mind that we would lay people off or that I didn’t need the analyst anymore.

Yes, perhaps she doesn’t have the engineering background I do to know to ask my exact query. But she could easily recognize if the tables/charts were right and—later on, as we expand—what kinds of views are useful. She is able to add a layer of domain knowledge and human judgment. Claude Code and all the other agentic tools are... well... tools.

I want employees who use tools well and can take ownership (and apply judgment) to work. Not even more tools I’ll need to run myself.

Shifting What Skills are Important

In the past, the ability to build these Excel files—and even format them in pretty (and legible) ways—was extremely important. It wasn’t like that was the “highest value add” task for an expensive analyst in a hedge fund or Wall Street bank. That was just a necessary part of the job for those expensive analysts to actually get to the part where they add value.

That has changed.

In the particular case above, not only did I not really want the analyst working in Excel, I didn’t even particularly care about her finishing her Python course before jumping into things. In fact, as far as “necessary but not high value add” skills I wanted her to learn, I told her to learn how to navigate a command line (use ls, mv, touch, etc.), use a text editor, and use git to collaborate/”save” work.

I do want her to finish the course eventually. But she’s not building production software systems serving millions (or billions) of users that have six sigma uptime requirements. If it looks like it works... and it works... that’s likely sufficient for the job—or at least this part.

Ultimately, the skills and tools that are important are shifting... and quite rapidly. But that’s not the same as “people are no longer useful.”

We’ve automated certain mechanical aspects of digital work that was just required. Now, human judgment is a much higher percentage of time/value-add in the work product—at least in my experience here, but also the experience of many companies I’m talking to.

Productivity can mean higher earnings. In fact, it generally does. It can also mean fewer people are needed. That, of course, is what everyone is worried about.

Subscribe now

The Market Map

Andrej Karpathy’s dataset obviously speaks to those fears. He scraped the Bureau of Labor Statistics’ Occupational Outlook Handbook—342 occupations spanning the entire US economy—and had an LLM score each one on AI exposure from 0 to 10. The results are an interactive treemap at karpathy.ai/jobs, where rectangle size represents employment and color represents exposure. Average exposure: 5.3. Medical transcriptionists score a perfect 10. Software developers land at 8–9. Roofers and plumbers sit at 0–1. If your work product is fundamentally digital, your exposure is high. Physical presence is a natural barrier.

I’ve talked about the physical world and data friction being a huge barrier to certain segments of AI, by the way. Here, here, this podcast, and an entire chapter in my book...

Still, think about it a bit. Follow the logic.

If you read those scores as “whose jobs will be eliminated,” the chart tells you top executives, software architects, and portfolio managers are the most screwed—and roofers and plumbers are sitting pretty.

Look, I have nothing against manual labor. But do you really believe the takeaway is that it’s better to be a manual laborer than a CEO? That programmers should retrain as bricklayers?

If that reading is right, then that’s going to be great right up until the point robots take over that physical work too, making humans entirely redundant. Sure, there are people who believe that. I’m willing to even say that can be a coherent position, but it requires assumptions about robotics and general intelligence that are significantly ahead of where the technology actually is.

What the scores actually show is whose jobs will change the most. High exposure means high disruption to how the work gets done—not that the work disappears.

Oracles and Skill Dispersion

I’ve written about my concept of oracles before. Borrowed from cryptography, where an oracle is a source of information that verifies if you’re on the right track, even tiny amounts of information can make a massive difference. This applies for human experts-in-the-loop working with AI.

That’s exactly what played out when I was showing the new analyst how to process the sheet. I was the oracle. I could tell it how to approach the problem. I could glance at the HTML output and immediately spot problems. Without a domain knowledge overlay, the output would have looked impressively complete and been subtly wrong in ways that matter.

So a financial analyst scoring 8 on Karpathy’s map doesn’t mean financial analysis stops being needed. It means pulling data, formatting reports, running standard models, and producing charts—the mechanical components—are highly susceptible to automation. And they are. I just demonstrated that with a spreadsheet. Still, the analyst is still needed. That analyst can just move much faster, and Excel ninjitsu is less valuable, especially versus good intuition and judgment. Again, changed, not destroyed.

That does raise a question. What happens to humans that aren’t experts? Say, young people or fresh graduates?

Humans-in-the-Loop—AKA, Human Bottlenecks

As said, I’ve advanced quite a bit with my AI agent setup.

I’ve been context-switching between 10–15 different agentic projects daily. All of them are bottlenecked on the same thing: having a human who can 1) take the initiative to start the project, 2) know what the actual output should look like, and 3) overlay domain knowledge or know when to have the agent fetch more context. That would be me in this case.

There are experiments in trying to take the human out of these agentic loops. StrongDM recently put out their version with a lot of PR—but the basic pattern holds regardless of tooling.

These three bottlenecks are the same bottleneck described three ways: judgment.

Which is exactly the skill I argued in “White-Collar Apocalypse” gets elevated rather than devalued by AI. The mechanical skills get cheaper. The judgment skills get more valuable.

Either way, I’ve been trying my hardest to take myself out of the loop. Relative to other people, I’m actively trying to put myself out of work. While I have been able to roll some of these agents up to be managed by higher-level agents (... ironic, I’ve created a corporate hierarchy from my agents), I still can’t take myself out.

If anything, it has steadily increased the pace of decisions I need to make, as it keeps increasing the pace of me getting stuff done.

But What About the Layoffs?

It’s worth being clear-eyed about what’s actually driving the layoff headlines. Companies like Block and Meta hired aggressively during the pandemic boom and are now trimming bloat—something tech does periodically regardless of the prevailing narrative. Previously, you’d hire McKinsey to come in and tell you to cut 15%. Now you can just cite “AI.” It’s a convenient story, but a lot of what’s being labeled as AI-driven displacement is really just cyclical correction with better PR.

What AI actually does is make high-performing, capable people even more individually valuable. That’s the oracle effect in practice. The senior analyst who can direct agentic workflows and evaluate their output is now doing what used to require a team. That doesn’t mean the team disappears—at least not everywhere. For my own organizations, which aren’t bloated to begin with, the freed time doesn’t mean we get rid of people. It means we can finally do the proactive work we never had bandwidth for—projects that were always worth doing but sat in the backlog because everyone was buried in mechanical tasks. The pie gets bigger.

Share

The Pipeline Problem / Young Person Job Apocalypse?

I’ve talked about this plenty. This is an even better illustration.

Avos Capital

The Research Team We Never Hired

On Friday we launched our first ETF, ticker AVOS. It is the highest-conviction, long-only, active global equity portfolio we can build. More info at https://avosglobalequities.com…

Read more

a month ago · 12 likes · 1 comment · Josh Blanchfield

helpfully illustrated the point for me.

Yeah, bloated organizations will lay people off, but “old industry” tends to be conservative and slow regardless. They’ll adopt gradually, and the adjustment will be similarly gradual. New organizations?

Well, it’s not that AI is eliminating jobs. It’s that organizations, seeing the productivity gains from senior people using AI, will stop hiring juniors. It’s destroying job openings, even if jobs themselves aren’t being lost.

Interestingly, where do future oracles/experts come from? This generation isn’t going to be immortal. Junior employees spend years building domain knowledge, making mistakes, and developing the judgment that makes them effective evaluators of AI output. Cut off that pipeline and you’re consuming your seed corn.

Everyone obviously knows this, but it’s a prisoner’s dilemma. Someone needs to hire those young people and train them up. That’s expensive though. Someone needs to do it... but not us.

At the risk of being repetitive: AI automates the mechanical components of digital work. That raises the value of human judgment. But judgment is developed through years of apprenticeship. Which means the real medium-term risk isn’t mass replacement—it’s underinvestment in future experts. Unfortunately, I don’t really have a good answer for this. Where it’ll land is still “wait and see” for me.

Where This Leaves Us

I’ve been called an “AI skeptic”—which is not a label I use for myself. A Googler called me one at a recent book signing. The reality is, I’m neither a doomer nor an accelerationist. I don’t think we’re all going to die, and I don’t think AGI is around the corner. But I absolutely think AI is going to make massive changes in our economy.

There are great macro thinkers like my former colleague who called early that AI productivity is a bigger deal than most financial analysts think—and I agree.

I still believe we’ll eventually have a bubble pop, because that’s just the nature of new technologies. But, sure, as per Ben Thompson, it’s far less obvious now that one is imminent. There are real productivity gains showing up in the data. Things are going to change. Hell, things have already changed.

It isn’t likely to be the economic apocalypse that Citrini’s thought experiment envisioned—with mass white-collar unemployment and “Ghost GDP” from AI displacing all white-collar work by 2028. I’m still highly skeptical of that. But AI is going to change the nature of those jobs profoundly. And the question of who adapts well comes back to that same oracle layer.

My analyst will be fine. My teams will be too. There may be a lot of uncertainty and change... but there’s also a lot of opportunity to do far more as individuals than ever before. It’s similar to when software frameworks made coding a lot easier, the internet made distribution of software instant, and the “cloud” made it possible to run scaled compute infrastructure from a coffee shop. These also all brought down the barrier to entry for creating stuff (especially in startups). Agentic workflows are also a massive jump forward—as large or even larger.

Karpathy’s treemap shows the landscape of change. The oracle framework explains who navigates it well. Though the “pipeline question” (funny way of saying kids entering the workforce) does still worry me. I guess we’ll have to wait and see... though at the pace things are going now, it might not be that long a wait.

Thanks for reading!

I hope you enjoyed this article. If you’d like to learn more about AI’s past, present, and future in an easy-to-understand way, I’ve published a book titled What You Need to Know About AI.

You can order the book on Amazon, Barnes & Noble, Bookshop, or pick up a copy in-person at a local bookstore. If you’re in Southern California, I’ll be at Lido Village Books in Newport Beach, CA next Thursday 3/26 from 4-6pm signing books in person if you’re around and want to say hi!

“Oil’s price spike is bad news for power-hungry AI.” “The Iran Conflict Is Quietly Reshaping the AI Trade.” “How War in Iran Could Cripple the Global Digital Economy.”

If you’ve been nodding to these headlines, I get it. Oil prices are surging. AI runs on energy. It feels right—especially if you’ve been wincing at the gas pump every time you fill up your car. That visceral, twice-a-week reminder that energy prices are spiking is hard to ignore, and it’s natural to assume it ripples into everything. And that’s why it’s great to explain why that’s wrong—or, at least, it’s more nuanced than that.

We’re back to our regularly scheduled programming, especially now that the Iran situation is clearly more involved than the quick in-and-out that Venezuela ended up being. As usual, has good content on the geopolitics/military situation, but I’m talking about the implications for datacenters here.

So, why is the gut-feel narrative wrong, or at least not right?

Oil, natural gas, and electricity are three different commodities with three different supply chains, three different price dynamics, and three very different relationships to the Middle East. Conflating them—treating “energy” as one fungible thing—leads to bad analysis. It’s also, incidentally, the kind of confusion that commodity traders and macro funds get paid to exploit. Understanding the actual linkages, even (especially) when they defy intuition, is how you make money in markets. And in this case, the intuition is wrong. At least for US and Chinese datacenters (... it’s more complicated elsewhere).

Let’s talk about it.

First, the context. On February 28th, the United States and Israel launched roughly 900 coordinated strikes against Iran. Iran’s Supreme Leader was killed. Iran responded by threatening—and then largely succeeding in practicality—a blockade of the Strait of Hormuz. Shipping through the strait all but ground to a halt—Kpler estimated tanker transits down ~92% versus the prior week, with daily passages falling from 37 pre-war to near zero within days. Brent crude peaked near $120 a barrel. Qatar’s largest LNG export facility was hit by an Iranian drone and shut down. 400 million barrels were released from emergency petroleum reserves—the largest release in history.

This is the biggest oil supply disruption in modern history. The Strait of Hormuz handles about 20% of global oil supply and roughly a quarter of all seaborne oil trade. For context: the 1973 Arab oil embargo disrupted about 9% of global supply. The 1979 Iranian Revolution disrupted roughly 4%. The Gulf War in 1990, about 6%. Russia-Ukraine in 2022, roughly 3%. This crisis dwarfs all of them. (As recently detailed, the historical comparisons are not great for this being something global supply chains can paper over.)

So yes, this is serious. Separately, I’m not dismissing the geopolitical crisis or its humanitarian toll. But you can get that side of things from plenty of places. We’re talking about technology and markets, so let me talk about the part that’s more my bailiwick: the claim that this is an AI infrastructure problem.

It isn’t. At least not yet. And understanding why it isn’t—getting past the vibes and into the actual plumbing of energy markets—is more useful than the panic.

Oil Is Not Electricity

Let’s start with the most important fact that makes “oil shock kills AI” largely wrong:

Petroleum provides less than 1% of US electricity generation. Less than half a percent, actually—petroleum liquids and petroleum coke together account for roughly 0.3–0.4% of total US generation.

The US electricity grid runs on natural gas (~43%), renewables (~23%), nuclear (~18%), and coal (~15%). Oil is used for transportation—gasoline, diesel, and jet fuel. It powers your car, not datacenters—and hence not your ChatGPT queries.

When Brent crude hits $120 a barrel, that’s a crisis at the gas pump. It is not, in any direct sense, a crisis for the US power grid. The reason you feel like energy prices are exploding is because you interact with gasoline twice a week. It’s the most visible energy price in your life. And, hence, a serious problem for politicians. But it’s not the price that matters for datacenters.

The gap between those two things is the entire point of this article.

Fortunately, we don’t need to argue this theoretically, because the Iran crisis has given us a beautiful natural experiment. Here’s what happened to various energy prices since the strikes began:

Note: Oil and gas figures are daily spot prices; US retail electricity is the BLS/CPI national average price, a monthly figure that lags by several weeks. The comparison across different data frequencies is imperfect—but the magnitude of the gap is the point. Even if the broad retail indicator hasn’t budged, there’s no evidence yet of a generalized electricity-price shock propagating from crude. Brent also briefly spiked past $119 on March 10 before losing altitude again.

Oil surged. Gasoline jumped. European gas spiked. US electricity didn’t move.

In a way, I could end the article here. The table basically shows my whole argument. So, let’s see if we can try to steelman the “oil crisis destroys AI” argument a bit more.

“But natural gas and oil are correlated!” That’s more or less the load-bearing argument for most of the articles. The thing is, they’re really not. Not anymore. CME Group research shows oil and gas prices have been largely decoupled since around 2008. They overlap mainly during periods when broad macroeconomic forces hit both markets simultaneously—which isn’t what’s happening here. This is a specific regional supply disruption, not a global demand shock.

The reason US natural gas barely moved while European gas doubled is straightforward: the US has been a net gas exporter since 2017. Our LNG export terminals are already running at full capacity, so there’s no mechanism for surging global LNG demand to suddenly drain our domestic supply. As Axios put it, “America’s natural gas bounty is acting like a moat, largely shielding the U.S. from price spikes while much of the world reels.” The gas powering Virginia’s datacenters and Texas’s ERCOT grid comes from American shale, not the Strait of Hormuz.

Natural gas is hard to move around. Commodities are not like digital bits and can’t be instantly transmitted, despite the miracle of the modern supply chain. Heck, not even oil is totally fungible—remember all of the challenges for sweet, light crude versus sour, heavy crude (Venezuela) for refining. West Texas Intermediate (WTI) and Brent, two global oil benchmarks, can diverge dramatically at times.

One last helpful illustration: Hawaii. It’s the only US state that actually burns oil for electricity in meaningful quantities—and pays far above the national average for the privilege, roughly 40¢/kWh or more than 2x what the mainland pays. Hawaii proves the mechanism works. If you’re oil-dependent for power, oil prices matter enormously. The continental United States simply isn’t. We aren’t yet seeing Hawaiians screaming, but rates are set monthly... and it could look ugly next month.

Though, let’s now reel my argument back a bit: the crisis is two weeks old. If the conflict persists for months and global LNG competition eventually bids up Henry Hub meaningfully, this picture could shift. Chatham House notes that US energy prices were already trending up before Iran—driven by domestic demand from (among other things) datacenters. But as of mid-March 2026, the data is not ambiguous. The biggest oil shock in modern history has not yet shown up as a comparable surge in US retail electricity prices, which remain driven far more by domestic power-market structure and natural-gas dynamics than by crude itself.

Subscribe now

Where It Actually Matters

The rest of the world is not so lucky. And this is where it gets genuinely interesting for anyone thinking about the global AI infrastructure buildout—not just the American one.

The Middle East: Datacenters as Military Targets

This one’s obvious in hindsight but was apparently not obvious to the people who spent billions building cloud infrastructure in the Gulf. Iranian drones hit three AWS datacenters—two in the UAE, one in Bahrain—in the first days of the conflict. AWS reported“structural damage” and “disrupted power delivery to our infrastructure.” Banking apps went down across the region. Enterprise software broke.

I wrote last year about how AI is overturning the military world order. Datacenters are strategic infrastructure now, and they got treated as such. Iran has reportedly identified dozens of datacenter locations across the Gulf for potential targeting—Amazon, Microsoft, IBM, Google, Oracle facilities. The $20 billion first phase of the Stargate UAE project (OpenAI, Oracle, NVIDIA) sits in Abu Dhabi, with plans for a 5 GW campus. Looked great before—now it’s a war zone.

And here’s the key difference from the US: Gulf states actually do burn oil and gas for electricity. Saudi Arabia’s grid is nearly entirely fossil fuel-dependent. This is not the US grid. When crude doubles, their power costs go up in a way ours simply don’t. They could always control it through decreasing exports and simply capping prices, but someone pays the piper. If they do it, the government is just absorbing the cost.

Europe: The LNG Shock

Europe is exposed because Europe is dependent on imported natural gas for a meaningful share of its electricity generation. Qatar supplies roughly 20% of global LNG, and the vast majority transits through Hormuz. When that supply shut down, European gas prices on the TTF benchmark surged 45–70%. Again, in a crisis, it’s very clear commodities in different places are not totally fungible. They don’t magically teleport to where they’re needed.

Worse, Europe entered 2026 with gas storage well below recent years—a structural vulnerability that existed before this crisis started. As at Commodity Context has been writing and illustrating well, there’s “no end in sight” for the energy pressure.

The deeper issue for European AI infrastructure: Europe was planning to add substantial new gas generation capacity partly to service growing datacenter demand. They’re locking themselves into gas dependence at the exact moment gas became expensive and unreliable. Ireland is a particularly stark example—Irish datacenters already consumed 22% of the country’s metered electricity in 2024, and instead of going green, many are increasing gas consumption.

This is, of course, the same Europe that shuttered a bunch of nuclear power plants for greater reliance on Russian gas (despite all of the green rhetoric) right before the Russian invasion of Ukraine... Either from luck or poor decision-making, they’ve just continuously been in a terrible position for energy (and for a lot of other things).

Asia: The Import Dependency Trap

Japan gets the vast majority of its oil and a significant share of its LNG through Hormuz. South Korea is similarly exposed. Singapore’s refineries cut operations by 40–50%. These are major economies with growing datacenter sectors, and they’re acutely vulnerable to exactly this kind of disruption.

China: The Quiet Winner

Finally, there’s China. Oil and natural gas account for only ~3% of China’s power mix—natural gas alone was just 2.8% of China’s electricity network in 2025. The grid runs on coal (still the baseload backbone) and renewables (whose growth met ~84% of China’s electricity demand growth in 2024). China’s clean energy capacity reached 52% of total installed capacity by February 2026—1,494 GW clean versus 1,420 GW fossil—a milestone, though generation still lags behind capacity since renewables have lower utilization rates than fossil plants.

As has discussed before, China may be lagging in chip and datacenter (space) efficiency, but they’re making up for it in part with a massive energy buildout. While the US debates permitting for new gas plants and nuclear restarts, China has been adding capacity (both coal and renewable) at a staggering pace. Their AI infrastructure is structurally insulated from Middle East disruptions in a way that Europe and much of Asia are not. The Oxford Institute for Energy Studies published an interesting paper on this in February (before the war even started) titled “The China Data Centre Advantage: Hype Versus Reality.”

I’ve written before about who’s winning the AI war. Everyone was talking about power capacity. Energy security wasn’t really part of that calculus until two weeks ago. I expect that’s changed now.

Now, how much does this global differential actually matter?

Honestly, less than the headlines suggest for any individual company in the near term. The hyperscalers have been hedging geographic risk for years—Microsoft, Meta, and Google all have nuclear power deals that, while still years away from delivering power, signal awareness that fossil fuel dependency is a strategic vulnerability. And the datacenter industry is, by nature, designed to route around damage—that’s the whole point of multiple availability zones. So, I don’t expect there to be wholesale abandonment of Middle Eastern datacenters.

But in aggregate, over a multi-year buildout involving ~$650–690 billion in annual hyperscaler capex (depending on the analyst aggregation)? The question of where you build matters a lot (I’d argue it’s always mattered, but people don’t take some of these risks into account until forced).

Hyperscalers are already slowing new capital deployments in the Gulf and considering accelerated projects in Northern Europe, India, and Southeast Asia. Again, not wholesale abandonment, but the war is reshaping the geography of AI infrastructure in real time.

Share

What Actually Threatens the US AI Buildout

So if the Iran crisis isn’t the thing to worry about for US AI infrastructure, what is?

The answer is a lot less dramatic than a war but a lot more grinding. It’s also the same one it was before. Last November, my former Bridgewater colleague wrote a guest post here arguing—convincingly, I think—that AI’s power requirements aren’t intractable. His main point was that the energy exists; the question is whether we can mobilize it fast enough. I agree. But “fast enough” is doing a lot of work in that sentence, because the real chokepoints are all about speed—and they’re mostly the boring kind:

Grid interconnection queues. In PJM (a “regional transmission organization,” which covers 65 million people across 13 states), projects that came online in 2025 spent an average of 8 years waiting in the interconnection queue. There are ~2,300 GW of total generation and storage capacity seeking connection—nearly 2x the entire installed US fleet. You want to build a datacenter? Great! Get in line. A very long line.

Local opposition. Between May 2024 and March 2025, $64 billion in datacenter projects were blocked or delayed by local opposition. By Q2 2025, that hit ~$100 billion in a single quarter. Only 44% of Americans would welcome a datacenter nearby—less popular than gas plants, wind farms, or even nuclear facilities. Datacenters are somehow less popular than nuclear power plants. Funny that something finally beat nuclear plants in a NIMBY-hate-contest.

Power transformer shortages. Large power transformers have 2.5–4 year lead times. The US faces a 30% supply deficit, with unit costs up 77% since 2019. The US relies on a single domestic supplier of grain-oriented electrical steel. No meaningful relief before 2027.

Water. US datacenters could require 697 million to 1.45 billion additional gallons per day by 2030—equivalent to New York City’s daily water supply. That’s water for cooling, not for drinking, but try explaining that distinction to the residents near a proposed facility (see: local opposition, above).

Gas turbine lead times. Even though natural gas is abundant and cheap, the equipment to burn it has 5–7 year lead times for large-frame units, up from ₂ years in 2021. The fuel isn’t the bottleneck. The hardware is.

No one writes breathless articles about transformer lead times or interconnection queue backlogs. But these are the actual binding constraints on whether the US AI buildout hits its timeline. As I argued in “Shortages Create Opportunity”—this is the same framework. The constraints create openings for whoever solves them.

However, the constraint needs to live long enough to actually put capital to work and invent new startups/technology to tackle it. One of the issues here is how long will this Iran situation persist? No one knows. I suspect not even the White House or Iranian leadership knows.

The Real Macro Risk

Now, if I were going to connect the Iran crisis to the AI buildout, the transmission mechanism isn’t oil→electricity. It’s oil→recession→capex cuts. That’s the channel that should actually keep people up at night.

The top hyperscalers are projected to spend ~$650–690 billion in 2026 capex, with something like 90% of operating cash flow going to capital expenditure (up from a 10-year average of 40%). AI infrastructure investment has become a significant driver of US GDP growth—without it, the economy would look a lot weaker.

If $120 oil tips the US economy—and oil shocks have a pretty solid track record of doing exactly that—the question isn’t “can we afford the electricity?” It’s “do CFOs keep signing off on $200 billion annual datacenter budgets during a downturn?” Microsoft’s management has signaled that meaningful cash returns from GenAI may lag infrastructure deployments by a year or two—the kind of caveat that becomes a problem when revenue growth slows.

I’ve argued before that the AI bubble will burst eventually—not because AI isn’t real, but because technology bubbles always reflect real technologies whose market prices ran ahead of reality. An oil-induced recession would be an exogenous shock, not a reflection of AI’s fundamentals. But markets don’t much care about the distinction when they’re repricing risk.

So What?

The Iran war is the biggest oil supply disruption in modern history. It is a real crisis—for drivers at the pump, for European energy security, for Gulf states whose datacenters are literally being hit by drones, and for Asian economies dependent on Middle Eastern fuel.

But for American AI datacenters? Oil is not electricity. Gas is not oil. And the data from the last two weeks proves it: crude up 46%, US electricity flat.

Energy security matters for the global AI buildout—but it matters differently depending on where you are. The US and China are insulated, for entirely different reasons. Europe and Asia are exposed. And the Gulf states just learned that “strategic infrastructure” becomes “strategic target” rather quickly.

Think less about the price of oil and more about the 8-year queue to plug into the grid.

Thanks for reading!

I hope you enjoyed this article. If you’d like to learn more about AI’s past, present, and future in an easy-to-understand way, I’ve published a book titled What You Need to Know About AI.

You can order the book on Amazon, Barnes & Noble, Bookshop, or pick up a copy in-person at a local bookstore.

Let’s do this again, redux! I’ll explain how to use AI agents for easy language learning, to create an easier version of my morning briefing, and finally, a far easier version of my briefing transcription -> summary -> action pipeline. In the process, my goal is to help readers remix the general principles for their own (mostly safe) agents.

My last piece about AI agents was my most popular and widely shared article to date. Usually, one writes a “Part 1” that’s easier and a “Part 2” that’s more complex. This is the exact opposite.

I realized there was a problem with the last piece.

• “Where do I enter the matrix?” (meaning the command line…)

• “Where do I download Claude Code?” (... a more complicated task than you’d think, especially on Windows)

• “What’s a Tron job?” (Cron job…)

More technically inclined readers absolutely took it and ran with it. But others saw it as aspirational at best, with no ability to operationalize what I wrote about. Scheduled jobs, auto-processing pipelines, and all of the fancy features of my agents are great but do come with a lot of prerequisites.

I might be ok with that—as said, I have complete faith that companies will come along and make these agentic workflows available to “regular people.” Unfortunately, as one of my more technically proficient readers commented:

“You hilariously did a PSA—’don’t do drugs, because they feel so, so good.”

And that was right, because in a few chats, I noticed this was the push some people needed to go take the risk with OpenClaw. After all, they realized how useful agents really are—and decided it was too hard to do it the way I explained... which is the exact opposite of what I intended.

So, in this revisit, I have these goals:

• Explain the general principles of creating agents (more slowly)

• Use methods that are more accessible to non-technical users.

• Give a framework for remixing these methods for readers’ own ideas/agents.

Ironically, this piece took longer than my last one. Instead of just sharing my workflows, this piece is designed to let you use these agents with step-by-step instructions, from scratch, and have them adapted to you (not me).

Language-Learning Chat

Let’s explain instructions and context. To do so, let’s talk about a specific “agent”—with the term used loosely here—I utilize on a daily basis.

Giving Instructions and Context: Utilizing Projects

The first major element of agents is the right context.

All of the major LLM platforms have some kind of functionality to create standing instructions. If you noticed from my last piece, I talked about utilizing CLAUDE.md files. These are used by Claude Code specifically to encode instructions the AI agent will always read. We can do the same thing without Claude Code and without leaving the Claude or ChatGPT app using Projects.

Why is this useful? Well, for one, it prevents you from needing to retype instructions constantly. This is nice when you have short instructions but is essential if you have extensive ones.

Additionally, if you have a complicated pipeline (hint-hint, later examples), you absolutely either need to do something like this—the agent needs to know what tools to use, how to use them, and what it should do without you writing a huge amount of information each time.

Beyond that, this approach lets you iterate on the instructions, which is crucial to actually making useful agents.

Learning Chinese and Spanish with Normal Chats

Nowadays, it’s quite possible to forgo using Google for most tasks. LLMs have gotten good enough at searching and synthesizing information that there’s much less utility in sifting through websites/social media yourself (besides that, Google shoves Gemini on top of every search anyway...).

Why not use that kind of random query to painlessly strengthen your foreign language skills?

As seen above, I was at Georgia Tech recently for my book tour. I was wandering around campus and having a casual chat about landmarks to look at. Instead of simply opening a “normal chat,” I used my Project called “Chinese Language Chat” (you can see the beginning of the instructions above). This has standing instructions to utilize Chinese language terms with pinyin (the “romanized” letters you see) and explain the meaning.

When I have time, I can read and even dig into the vocabulary. When I don’t, it’s pretty harmless to simply ignore the Chinese terms and skim the rest.

Subscribe now

Constant Iteration

The beauty of having standing instructions is you can see how they perform over time. What if there are things you don’t like? Well, you can change it. Even better, the AI can change it for you.

My Chinese is... reasonable. My Spanish is worse. However, I found that the adapted version of my Chinese language chat in Spanish (accounting for my level) was too easy. Or rather, it gave single words that were too short and without context to really be of use. So, I asked to have it changed.

I had Claude (in this case) reread its Project instructions and make edits. With those edits, it became much more useful (for me).

Just to clarify: you WILL need to copy and paste the updates instructions into Instructions manually. This doesn’t update it automatically.

Adapting This For Yourself

I’ve added my instructions to this git repo under For_Regular_People/.

But there's no need to do any editing yourself. You can ask for it to be adapted by your chat agent! Copy and paste the below directly into your favorite LLM chat, and it should help you customize it:

Please view this set of instructions at https://raw.githubusercontent.com/j-wang/how-i-utilize-ai-agents-article/refs/heads/main/For_Regular_People/Chinese_Language_Prompt.md and help me adapt it for my use case. Ask me questions about what language I want to learn, my proficiency level, the context of my learning, and anything else that is required to fully adapt these to me.

Then, once it does:

• Create a new Project

• Name it, fill in the other info, etc. I’d also suggest you star/favorite it.

• Fill in the instructions

Two gotchas for Claude to point out. 1) In Claude, the Description that you fill out on the same dialog box as the name is not the Instructions. 2) You may need to restart your app or refresh claude.ai in order to see the Project in your starred list.

Note: You can always write instructions yourself. In fact, has a good summary of a piece by is called “Does AGENTS.md Actually Help Coding Agents?” (AGENTS.md is the more “generic” version of CLAUDE.md—instructions/readme for the coding agent):

A new benchmark study shows repository-level context files only help when they add non-redundant, repo-specific info: human-written files that capture tooling quirks and non-obvious conventions raise success rates around 4%, while LLM-generated files that restate existing docs reduce success and increase compute by over 20%.

Absolutely true. However, I also find that many people—including myself—are often too parsimonious about instructions. You waste context and have fairly repetitive instructions from LLM-generated text. That being said, it tends to be more complete as well.

Is This Really an Agent?

It stretches the definition. I’d personally just consider this a “custom chatbot.” I mainly used this particular example as a gentle introduction into Projects, which will be necessary later, and to show how surprisingly useful standing instructions are.

The next example is still not super complicated. But I think it actually crosses the boundary where one can fairly call it an agent, albeit limited.

Morning Briefing—But Just Using Projects

Before I had my automated morning briefing that delivers at 5:30am, I actually had a simpler Project-only version. The downside? I had to say “morning” in a chat to trigger it. That makes it less “agentic,” in a way, because it isn’t autonomously acting each morning. But it is still reaching out into your world to help you do useful tasks.

Just to note: given how many data sources it uses (email, calendar, HubSpot, to-do list, email newsletters, etc.), it actually takes a while. It’s fine if you want to just say “morning” and walk away, but if you were hoping to sit down with it immediately with a coffee... well, you’re going to need to come back in 5 minutes. I’m sure you can guess why I ended up turning it into a fully automated flow.

That being said, this version needs absolutely zero work in cron/launchd, scripts, and editing configuration files. It will also automatically ask you for context and improve as it goes (I’ll get into this mechanism later).

Here’s an example of its output when I used to use it.

The Right Tools

The second major element of good agents is tools. One of the defining characteristics of agents is the ability to act, and tools are usually the sensory organs (e.g., calendars, emails) and hands (writing to DevonThink or Notion, drafting emails, etc.).

Most major LLM platforms now have connectors/apps/integrations. Whatever the terminology, these are tools that the AI agent can utilize to do useful tasks in the world.

Obviously, nowadays I use Claude more, so it has both more connectors as well as my custom-built ones (”DevonThink Web,” “Omnifocus Web,” and even “Day One”—which, despite not having “web” in its name, is actually also one of my custom “web wrappers” around Day One’s MCP).

This is the major element that takes this briefing beyond a simple chat.

The Context/Instructions

As before, instructions/context (... which, in a way, are one and the same for the agent) are critical to both making the agent useful/intelligent and being able to utilize the right tools. Although all platforms technically have a way to let the AI know what tools are available for it, the agent actually utilizing the tools is inconsistent at best.

In this case, for a useful briefing agent, it must know:

1. What should it do? You need to have the right instructions and a trigger word (for which I made one word—”morning”—since I wanted a very quick, memorable trigger).

2. What is its context? If anyone has worked with assistants—on-site, remote, or even AI—one quickly learns that they get far better with time. Why? There are thousands of unwritten rules, preferences, and simple pieces of contextual information that a new assistant doesn’t have. We need to rectify that.

3. What tools should it use? Again, this is the case even with a human assistant.

To get you started, I have all of those handled here in this set of briefing instructions, which can actually be directly copied/pasted into Instructions.

The instructions on what to do—which you should customize over time—are here. I explicitly ask the agent to do a first-time setup to gather the right context. Finally, I have Gmail and Google Calendar as required tools. They’re the most common ones that everyone has, but if you have another service (e.g., Zoho), just replace it.

You must have the tools activated/authenticated—Gmail and Calendar—or this will do nothing.

Even news is pulled from your email—since most people have subscription newsletters (... you are, after all, reading this on Substack). Partly for safety and partly for curation, I avoid having the agent search the web for news.

Memory/Iteration

As mentioned, a key part of making good agents for yourself is feedback and iteration. That becomes part of the context for the agent. I showed you how to do so by actively changing the instructions (even using the agent itself) in the last section. For more “systematic” and larger changes, you should do exactly that.

But that’s not the only way.

Both Claude and ChatGPT (and many others) have the ability to store “memories” to intelligently bring in context from other conversations (or at least they purport to—I’m often not impressed by the “intelligently” part of bringing in context). More importantly, both have Project-level memories.

It will try to figure out what is relevant to store itself... but you can also tell it to store things.

Which is how the first-time setup works for this briefing—and is how you can keep giving it feedback, explicitly, to remember and improve itself over time. For example, you can tell it that certain types of events are not important and not to show them. You can tell it that you should prioritize a different set of todos. You can tell it that you want to see more domestic news and less international news. And it will store these preferences in memories.

Is This an Agent?

I’d be fairly comfortable saying yes, though I could understand pushback. The version of my briefing I outlined in the last piece is far more easily categorized as one without any “grey area.” Why? It runs autonomously, and it writes into my databases/journals/tools. Putting aside any technical definition, one can easily see that it acts.

This one? You need to kick it off. It just outputs the contents of the briefing into chat. Still, it reaches into active parts of your personal and/or business life with email and calendar and helps you prioritize. Although one might have minor quibbles, I think a task that many human personal assistants literally do every day can be considered “agentic.” And this one also gets better over time.

Share

Meeting Summary Agent(s)

Ok, now we’re going to take a fairly significant jump in complexity—and power.

Once again, we have a repeat from the last article. This tended to be the one people gave me private feedback that they were most excited about. It’s unquestionably “agentic.” More important, it’s a huge amount of time savings. How long would it otherwise take to manually comb through eight hours of transcripts and thousands of documents?

Unfortunately, this does require something like Claude Code, ChatGPT Codex, or Gemini Antigravity. I’ll show how to do this in Claude Code, which makes this the only platform-specific example (as a note, I spent a long time trying to get this to work with Cowork... but it has sandboxing issues). But why does it require one of these?

1. It has a step that requires cloud transcription. You need an API key and a script.

2. For any non-trivial version of this, you will need to dispatch parallel subagents. Otherwise, your agent will take forever... or, more likely, run out of context and become stupid.

3. You need to write somewhere. The length of each “work product” and then the final product cannot simply be output in chat.

1) could potentially be overcome. 2) is sometimes done (in a black box fashion) in “normal” LLM chat, but mostly not. 3) cannot be overcome at all.

This is an important concept: to maintain “smart agents,” you often need to break up tasks.

Parallel Tasks and Context Windows

This is the last element I see as important in agents: parallelization/breaking into narrow tasks.

As a reminder, my original example had seventeen meetings. These were composed of eight hours of recording, over a thousand pages of PDFs, and a bunch of screenshots of my calendar/schedule showing who I was meeting with and when.

1. All of the recordings were transcribed by Deepgram.

2. Sort my folder of entirely unsorted/not-descriptively-named recordings, transcripts, and calendar screenshots into usefully named folders for each meeting (I didn’t bother to pre-sort them myself).

3. Using context from the PDFs and the schedule and asking me questions (... in this case, there was enough info to ask zero questions), Claude cleaned up all of the transcripts with names and fixed things that Deepgram obviously “misheard” (e.g., “Mall” to “Vol”/volatility).

4. Summarize each transcript and set of PDFs into key takeaways and evaluation (based on my pre-written instructions/criteria).

5. Summarize all of those summaries into a full, compiled report of all meetings.

This is far, far too much to fit inside the context window of a single agent. Even if it could, I would not recommend it due to how fast current LLMs degrade with longer context. You will get very, very poor results if you dispatch agents to do everything versus have fresh ones for narrow tasks.

As such, each task should be done by a different agent.

• The transcriptions can all be managed by scripts/an independent subagent

• The sorting process of meeting materials into relevantly named folders can be done by a separate subagent

• Each individual clean-up process for the transcript can be managed by subagents for each meeting folder (so, 17 in parallel)

• The summary can be done by another set of fresh subagents (17 in parallel) by reading each transcript

• The overall compiled summary of all meetings can be done by a final subagent

In total, we have a minimum of 1 + 1 + 17 + 17 + 1 + 1 = 38 subagents (and the “main chat” that we interact with). This is possible to do by just opening 38 chats and constantly dragging and dropping the inputs and outputs of each chat into each other. After all, “subagents” are just fresh LLM chats/agents. Obviously, I don’t suggest doing it that way.

The more analytical among you might have also noticed that some of these files get read repeatedly. For example, a transcript will be read once by the sorting agent, once by a cleanup agent, and once by a summary agent. It may also be referenced by the overall summary agent for more info. As such, the same input tokens will be loaded in three or four times.

Wasteful? Absolutely not. The context window issue of running out of context or simply having the agent get dumber with long, unfocused context makes “rereading” a fairly cheap cost for better results.

Given all of this, I think it’s clear why I don’t think it’s viable to do this outside of Claude Code or Claude Cowork (which would work if I didn’t use an external service/web call like Deepgram). You need parallel agents and a long, multi-step workflow. I did not, after all, sit around and babysit each step. I simply provided the instructions and said, “Go.”

Doing This Yourself

Bad news. You need Claude Code and a subscription (though other coding agents could do it too).

Good news: I took care of basically everything else, including making it so you don’t need to install or configure anything beyond Claude Code itself. You don’t even need a text editor! (Ah, I can still hear people asking me, “Is that like Microsoft Word?”).

Obviously, my original workflow has a ton of my specific context, required analyses, personal preferences, common mistranscriptions, etc. I genericized it here in this git repo.

If you open Claude Code, it will walk you through everything you need.

I don’t love that this trains non-technical users to blindly hit “yes” on permission prompts, because unless you run claude --dangerously-skip-permissions on the command line, you’ll be doing that a lot. Unfortunately, I found that’s essentially what people do, so I find you might as well use it... You are trusting that I wrote this in a way that doesn’t do stupid things either way.

On the desktop app, even choosing “auto accept edits” as the permission in Claude Code will have it still asking you things, but most users who’ve tested it just blindly hit enter there too.

Prerequisites:

1. Install Claude Code. OR, just use the Claude Code Desktop

This is a good tutorial from for Claude Code in Cursor. Alternatively, use Claude Desktop and go to Code (you may need to follow instructions to install prerequisites, etc.).

Then:

1. Download from my git repo here.

2. Go to the directory in Terminal (Mac) or Powershell (Windows) and run: claude --dangerously-skip-permissions.

3. OR, if you’re using the desktop app, select the folder you put my repo into (and unzipped, if you downloaded it as one) and select main.

4. Put your files (all of them—just freely dump in MP3s, screenshots from your calendar, PDFs from meetings, etc.) in the INBOX/ folder.

From there, simply ask it to read CLAUDE.md—though, even if you said something else, it almost certainly would anyway. Also, if you happened to be using Codex or Antigravity... you could just tell it to read CLAUDE.md anyway, and it should basically all work too—I promise they won’t get offended.

In the following, I show screenshots from Claude Code on desktop (it’s actually my first time using it there!) using a random subset of the files I had from the conference I mentioned in the prior article.

(Note: If you don’t have a Python environment that functions, Claude Code will help you install it on your specific system. It’ll be able to walk you through it better than I can.)

As you can see, this kind of agent can do a lot and is very powerful. You should go read CLAUDE.md in the directory (or just click on it in GitHub) to see how you should set something like this up yourself.

Is This an Agent?

Definitely. It requires a fair amount of planning, context, and outputs.

The point of this last one is to finally have all of the elements:

1. Instructions/Context

2. Tool Use

3. Planning + Parallel Workstreams

This applies to almost everything for creating useful agents.

Subscribe now

From Here

If you take nothing else from this piece, take this: the barrier to useful AI agents is not technical skill. It’s clarity of thought.

Every example I walked through—from a language-learning chatbot to a 38-subagent meeting pipeline—boils down to the same three things. First, clear and detailed instructions that encode what you actually want (not a vague “help me with X”). Second, the right tools are connected so the agent can reach into your actual life—email, calendar, documents—rather than just talk at you. Finally, third—less obvious if you don’t know how LLMs work under the hood—breaking work into focused chunks so the agent stays sharp rather than drowning in context.

None of that requires a command line. The first two examples don’t even require Claude Code/Cowork. They require you to sit down, think about what you actually want an assistant to do, and write it down in enough detail that a (very fast, very literal) new hire could follow it. If you’ve ever onboarded someone, you’ve done harder work than this.

I did originally say that productivity pieces like this aren’t what I do. That being said, it seems everyone’s overwhelming feedback is that they do find it very helpful. As such, I will indeed do more of these—not every post, but every now and then.

If nothing else, I owe you all a piece on safety and sandboxing, because “don’t be reckless” deserves more than a section header.

Thanks for reading!

I hope you enjoyed this article. If you’d like to learn more about AI’s past, present, and future in an easy-to-understand way, I’ve published a book titled What You Need to Know About AI.

You can order the book on Amazon, Barnes & Noble, Bookshop, or pick up a copy in-person at a local bookstore. I’ll be at the Barnes & Noble in Walnut Creek, CA this Sunday 3/15 signing books in person if you’re around and want to say hi!

My day-to-day productivity stack now often feels more like managing a small team than it does running tools.

I receive a morning briefing outlining emails I should respond to, tasks that are overdue, stakeholders I need to follow up with from my CRM (customer relationship management), and important news stories. After a meeting, I drop the recording and files (slide decks, etc.), and an agent picks it up, classifies it, and gives me a summary (or report, if it’s a prospective investment I’m evaluating). At the end of the day, I get a prompt to decompress, and an agent basically walks me through a 5-minute gratitude journal.

Finally, I have a research assistant agent and a drafting agent that helps me write now—the last of which I’ll actually provide a full end-to-end of how this piece evolved from the agent.

(See the GitHub repo here to see all the human/AI drafts!)

How to Utilize Agents on a Personal Level

This Substack isn’t meant to be a how-to or productivity blog. I’ve said that before, and I mean it.

That being said, one of the reasons I ended up feeling like I needed to write this piece is no matter how much I explained the conceptual fundamentals, people didn’t really “get it.” I would tell people that they could easily use AI to substantially help speed them up in workflows that they were complaining about (ironic, I know, given my reputation as a “relative skeptic” in the world of Substack). Then, they would do something like “help me with X” in ChatGPT, find that it was minimally effective and clunky, and then give up.

The second reason is because of OpenClaw (previously known as Moltbot and before that as Clawdbot). People would hear stories of it running their entire lives—along with running crypto schemes, trading markets, trolling people, or posting on Moltbook—and go, “Ah, that’s how I do what James asked me to do.” One of the last straws was someone mentioning that OpenClaw sounded a lot like something I would/did build.

First off, I respect the hustle of the OpenClaw founder. Marketing and getting adoption for a product is often harder than building it. So, when I say that it’s both horribly flawed and not really that technically complex, I’m not trying to insult Peter Steinberger or say he doesn’t deserve whatever payday he got from OpenAI. He does.

That being said, I would hesitate to recommend running it to those with a lot of expertise, and I would say, “hell no” to anyone who doesn’t understand it. Running OpenClaw without a good understanding is getting on the fast lane to having all your sensitive data (financial info, API keys, confidential data, etc.) stolen.

In this piece, I want to prove (and then illustrate with my own examples) three things:

1. Agents are extremely powerful... given the right context.

2. Agents, without safeguards, are extraordinarily dangerous.

3. Agents can already revolutionize your life and work—today.

And I’d suggest everyone try to use them. Why? I always refer back to the poor NYU Stern professor who is now known as the subject of Bill Gurley’s “How to Miss by a Mile”—rebutting the professor’s piece in FiveThirtyEight on how “Uber Isn’t Worth $17 Billion.”

Uber today is worth over $150B. More to the point, he admitted:

“As I attempt to attach a value to Uber, I have to confess that I just downloaded the app and have not used it yet. I spend most of my of life either in the suburbs, where I can go for days without seeing a taxi, or in New York City, where I find that the subways are a vastly more time-efficient, cheaper and often safer mode of transportation than taxis.”

Putting aside today’s insults of being an out-of-touch, elite NYC urbanite, it is truly difficult to understand a new technology without trying it. If you’re either worried or skeptical of AI agents... well, go find out for yourself.

We Are Unquestionably Here

The definitive introductory textbook for AI, Artificial Intelligence: A Modern Approach—often simply referred to as “Russell-Norvig” after its authors—defines artificial intelligence research as “the study and design of rational agents.”

I often push back against those who claim artificial general intelligence (AGI) is right around the corner. That being said, we unquestionably now have AI as defined by Russell-Norvig.

While you could quibble and argue with how much “machine learning” or simply statistical methods fell into that category, one would be hard pressed to not call what we have with recent, modern models “rational agents.” Rational, after all, doesn’t mean perfect or infallible—which would disqualify even humans if we required that!

The jump has been from models operating in a chat box—where they may be “rational,” but not agents—to operating in the real world. Or, at minimum, our working digital world, where our calendars, journals, memos, and everything else now exist. As I wrote in “The Boring Phase of AI,” this shift from chatting to doing is both less flashy and ultimately much more important.

Subscribe now

Examples of My Setups

Ok, enough boring “blah, blah, blah, principles/safety/etc.” I suppose I now have to show examples of what I mean. Obviously, I do use Claude Code (and Codex) as coding agents. But I use them for far more as well.

The Morning Briefing

Let’s do this in a front-loaded fashion. The concepts in this one translate to most of the others.

Every morning, I get a neat briefing. It basically looks like a personal assistant put it together. Let’s show rather than tell.

It also flags for me what emails I should respond to and even points out things I should do (e.g., block out the morning for a long judging session).

How does it work? Well, every morning I launch a cron job*, which is basically a scheduled script. I use Claude Code in prompt mode, where I can pass it a prompt “non-interactively” (basically, I don’t need to chat with it). I give it dangerously-skip-permissions so it can run things.

/opt/homebrew/bin/claude -p “$MODE briefing for $DATE” --output-format text --dangerously-skip-permissions --max-turns 25 >> “$LOG_FILE” 2>&1

For the non-technical, this looks like an arcane summoning to hell, but the key portions of it are

• claude (command-line command for Claude Code—I just have the “full path” for it)

• -p (this is prompt-mode)

• --dangerously-skip-permissions (as suggested, this is dangerous but is required so it can run without me)

As the name suggests, it’s dangerous (we’ll cover this later), but I restrict what it can do and access. I have my own MCPs (model context protocols) it can access, and most published MCPs (like Google’s for Gmail and Calendar) are cautious and have a level of protection.

Think of them as tools that the model can call—with pre-specified functions and scope of what can be done with them.

The reason, for example, why I need to push my briefings to Day One and DevonThink is Gmail will not allow a model using an MCP to actually send an email—the best it can do is a draft.

But how does this headless, prompt-mode Claude Code know what to do?

The two keys to all of this working are two files in the directory:

• CLAUDE.md: instructions to Claude Code that it will read to know what to do

• settings.json: a configuration file for what MCPs Claude Code has access to (and what it is denied) EDIT: This should actually be under .claude/settings.json—I originally moved it so it’d be visible (dot files are hidden) and forgot to note this.

settings.json:

{
  "permissions": {
    "allow": [
      "mcp__gmail__*",
      "mcp__google_calendar__*",
      "mcp__hubspot__*",
      "mcp__omnifocus__*",
      "mcp__devonthink__*",
      "mcp__dayone__*",
      "Read(*)",
      "Write(~/Documents/briefings/*)"
    ],
    "deny": [
      "Bash(*)",
      "Write(~/.claude/*)"
    ]
  }
}

As you can see, it gets access to my email, calendar, HubSpot (CRM), task manager, DevonThink (personal database—which you’ll see I use a lot, because it’s flexible... and I’ve used it for years... but also because I can more safely read/write to it with backups), and Day One (journal). It can read things and write things to a specific folder (for logging). It cannot arbitrarily run scripts.

I don’t actually allow it to access websites or “call out.” All of the news items are from my email (with specific items only forwarded). You never want it to search and land on arbitrary websites.

And I’m not even super comfortable/done with securing this. I’m figuring out the best way to sandbox this even further so I can granularly deny all network access outside of that required by my MCPs (many of which are hosted on my own domain, because I wrap them to allow me to use them anywhere with Google OAuth/login). If anyone has suggestions, let me know!

*It’s actually a launchd job on Mac because of auth token issues, but that’s not super important.

Share

Capturing Substack Articles On-The-Go

While the morning briefing runs automatically, I also use Claude Code on the go for various tasks. I use tmux—a terminal multiplexer that keeps sessions alive—so I can connect remotely via Wireguard (VPN to my home network) and Blink.sh on iOS.

Non-tech translation: I securely connect to my home network and remote-access Claude Code running on my computer from my phone or iPad.

For any tool, the folder structure is basically the same: detailed instructions in CLAUDE.md and a settings.json defining what is allowed and not.

One workflow I use frequently is web capture for Substack articles into my DEVONthink database. If you’re an avid Substack reader, you know the pain: not all publications have dark mode, paid posts require login, and you’re out of luck without internet. My Claude Code session takes a URL, opens it in my logged-in Substack Reader on Chrome, and captures it as a Web Archive in DEVONthink’s “To Read” inbox—where I can read it at my leisure, offline, in dark mode.

(Anthropic recently released Claude Code “remote connections,” which lets you connect to a running session from the web or another device. Most people won’t need my tmux setup—remote connections are easier. Though mine still has some advantages in control and flexibility.)

As you can see, Claude Code in folders doesn’t have to be about coding. It can just be structures for automation. Most of what I’ve described here isn’t writing code in any traditional sense—it’s setting up context and instructions for the agent to follow.

Writing: Research and Junior Drafting

This is the use case I get asked about the most. I’ve written about using AI as a writing aid before in “AI as a Partner, Not a Replacement,” where I was still figuring things out.

I credit inspiration from ’s early example with CODER, though having done it myself and helped others set it up, you really have to set it up for yourself, and it’s almost impossible to get something off-the-shelf and have it be any good.

Here’s how it works: I still need to write the outline of what I want, with the broad notes to hit. I still need to go through and substantially edit—asking for different charts and data, checking math (which is often wrong), probing claims, and changing a decent amount of the text for my own tone.

It’s more akin to an intern helping do a first draft than a ghostwriter.

Is it useful? Yes. But as you can see for yourself in this GitHub for this article... things do change substantially. There’s no real question as to who is the author in any meaningful sense, though. Still, it actually saves me time and mental energy, whereas before it was totally useless.

Why is it helpful at all, though? It took a few things:

1. Iteration. Every time it fails to get that close to my writing, I ask it to assess and then iterate on the CLAUDE.md to try to have future attempts get closer. This is why I’m not posting the full CLAUDE.md. While the content might not be me, I’m also not looking for impersonators who sound like me.

2. Research. Remember how I capture articles I like in DevonThink? Well, guess where I ask my AI agent (Claude, but it doesn’t have to be) to research? I have a “blessed canon” that is preferentially drawn upon. It literally also has all my published articles! They are well-organized in categories and with tags so it can not only research using me but also draw on how I wrote about certain topics. This is like RAG (retrieval augmented generation)—actually, it is literally RAG by most definitions.

3. Organization. If you look at this article now and what I started with—my personal notes (initial sketch), Initial Draft (Claude’s first shot), and eventually my final article... well, perhaps the only thing that really survived is some rough organization. That’s fine. It’s relatively rote, low differentiation (most good articles are organized similarly), and time-consuming for me... but fast for an AI agent. Exactly the kind of thing you want it to help with.

Again, check out the GitHub repo yourself to see how this played out over time. Usually it isn’t quite as “bad” as this one—you’ll see the initial draft, and my edits are big. But remember, I don’t write these kinds of articles often, so it’s going to understandably be worse not seeing examples like this.

Meeting Notes

This was the recent example that blew people away.

I use a Sony ICD-UX570 recorder for meetings—it is more private, controlled, cheaper, has better microphones, and is more flexible in every regard than AI notetakers like Plaud. A side benefit is that young people think it’s a Walkman.

I have a monitored folder where I upload MP3 recordings. From there, a Claude Code instance has detailed instructions to:

1. Send them all to DeepGram: it has better privacy and its ability to deal with noise is excellent (unfortunately, Whisper from OpenAI, which you can run from your own computer... is not)

2. Read all extra documents in the folder: this includes PDFs of pitch decks, track records, whatever.

3. Use this context to classify it: is it a memo to myself, an (investment) manager meeting, a portfolio company meeting, or an internal meeting?

4. Ask James who each speaker is if it is not clear: often, pitch decks, board meetings, and whatnot have enough context, but if not, give me context and ask who it is.

5. Create summaries: classification isn’t just for kicks—I’ve written detailed guides on what I care about and how to summarize for each type of meeting.

For steps 4 and 5, I have CLAUDE.md specify to dispatch subagents for each meeting. It takes too long otherwise. However, this obviously sets tokens on fire in terms of total use in a short period of time.

And no, I never record secretly. While Granola and similar services have made it common, I think it’s quite bad taste (not to mention illegal in many jurisdictions). Additionally, few people have issues with it these days—and even less so with my old-school recorder.

But, as you might have noted by this point, what makes agents ok versus supercharged is the right context. I recently went to Global Alts in Miami to help out in selecting managers for investment. I had 17 meetings. Each had a huge amount of content.

That was 8.5 hours of meetings. Literally hundreds of pages of documents.

It usually takes weeks to process it all. I did it in hours and compiled a report on all managers, including summaries, analyses, comparisons, etc. The report itself is too sensitive, but just to give you the table of contents:

1. Executive Summary

2. Manager Overview Table (Managers, Strategy, AUM, Interest, Bucket, Metric, Next Steps)

3. Priority Tier: Advance to Diligence (lists each meeting with Summary, Why It’s Interesting, Key Risks, and Next Steps)

4. Monitor Tier: Follow Up Required (lists each meeting with Summary, Why It Warrants Follow Up, Key Open Questions, Next Steps)

5. Pass Tier: Not A Fit (lists each meeting with James’s Assessment, Why It’s a Pass, Note, Action)

6. Portfolio Construction Observations (with Bucket Mapping and Capacity, Redundancy Analysis, Diversification Gaps, Correlation Analysis)

7. Consolidated Action Items (tables of 10-20 follow-ups for each of Immediate, Near-Term, Analytical [Before Follow-up Calls], Medium-Term, Deferred/Monitor)

8. Appendix (Risk Factor Summary [large matrix analysis with beta, crowding, illiquidity... etc.], and Key Takeaways from Risk Matrix with notes for each manager)

It’s a 30-page report that took a little over a million tokens and a few hours. It floored people. But, again, the magic was

• It had context from my recordings with my detailed questions

• It had context from the documents/PDFs from the managers

• It had context from my own voice memos summarizing my thoughts on sets of managers later

• It had context from investment policy docs from the organization I was helping

• It had context from my evaluation criteria that I wrote up in exhaustive detail (in a EVALUATION.md) before my meetings

Again, I didn’t just go, "Hey, go summarize stuff.”

Subscribe now

Takeaways

1. Context is king—and sometimes takes a lot of work. None of these things magically work well. I needed to tune both the instructions and what the agent had access to to have good results. Some of this might have been prompt engineering, RAG, or whatnot, but I see it as all being context.

2. Iteration is required. I’ve rarely seen an agent start amazing out of the gate. All of these processes (and others I haven’t shown) have required constant iteration of instructions/context. Sometimes it’s automated (modify CLAUDE.md to incorporate learnings). Sometimes it’s me giving better data sources. But still, you need to be judicious about this.

3. It’s dangerous. Notice, I redacted things even for certain screenshots here. You bet the agents have access to extraordinarily sensitive information. And I don’t even do OpenClaw yolo things like give it access to my bank account/credit card (yes, people have done that with OpenClaw)! With great power comes a huge amount of f*cking security risk.

I expect there will be better off-the-shelf agentic products in the future that require less tinkering and are safer. For most... I’d stick to the safer side of things.

For example, meeting summarization can be pretty safe and highly effective with some of the ideas of using Claude Code (or Claude Cowork now, which is basically the same) and having clear, detailed instructions—pages of them, not paragraphs of them—and sufficient context. No need to let your agents trawl the web or deal with setting up OAuth for your MCPs (like I did).

Really powerful. Really scary.

Security: Please, Don’t Be Reckless

Let me expand on that last point, because it’s important enough to warrant its own section. has excellent pieces on one of the biggest issues: prompt injection.

There is simply no good way to fully prevent LLMs from reading rogue, malicious instructions and potentially acting on them.

The biggest worry is exfiltration. Say you have your personal schedule, API keys, passwords, whatever in your agent’s context (sometimes you can’t avoid it—power and access often require keys to do useful things). What if an attacker embeds an instruction to send all of it to a website or location they control?

Major model companies have been working on making their models resistant to this... but it’s not perfect. My own setup is very not perfect. There are security holes I already know about, and that’s with a lot of custom tuning, firewall settings, sandbox isolation, etc.

If you aren’t sure how to think about this, err on caution. Don’t YOLO with something like OpenClaw, where a bunch of auto-installable skills are actually malicious. Yeah, sure, OpenClaw can be powerful. But that’s like going around and picking up used needles off the street and injecting yourself, hoping for a high.

You can get a lot of power already with some of the principles I outlined—yeah, it’s more work, but I expect it will work better... and it will be infinitely safer.

So What Does This All Mean?

Block (previously Square) just laid off roughly half their workforce. Research published a thought piece that essentially crashed various stocks by painting a people-less future as AI takes over (which I disagree with). At the same time, a refrain on social media is “AI adds nothing in productivity—it’s been shown in studies.”

My last piece directly rebuts the “nothing in productivity” point. Studies don’t show that. Even in fairly conservative cases (GitHub Copilot benchmarks from over a year ago), we clearly see productivity gains across different studies. This also fits what I wrote in “The Boring Phase of AI”—AI that does real tasks is both less flashy and likely much more important than the headline-grabbing model releases. That future is already here. We’re just getting started.

And I also think it’s highly unlikely to cause a dystopian scenario of persistent mass unemployment. (Note: I didn’t say it won’t cause disruption, especially in the short term.)

Random Note: On Model Choice

Ironically, I use Opus for almost everything, even when it’s overkill. The one place I don’t is Claude Code—almost all implementation work (note: not planning or my “main” chat, but the actual execution) runs on Sonnet or Haiku, their smaller “dumber” models.

I think it tells you something that the most “permissive” case for tolerating errors and dumbness is code. Code can be validated against tests, and I’m generally going to review each pull request anyway. That perhaps says something about why AI adoption is happening so quickly in the coding realm.

See me in Atlanta!

I’m in Atlanta for a book event and speaking at one of my alma maters, Georgia Tech! Come see me at at A Cappella Books tomorrow (Sunday) at 4:30pm discussing What You Need to Know About AI—what’s real, what’s hype, and where this technology is headed.

If you stop by tomorrow, mention this newsletter and I’ll ship you a free tote bag. :)

thinks the singularity won’t be gentle. says you are no longer the smartest type of thing on Earth. On top of the software sell-off in stock markets (“SaaSpocolypse”), non-technical pundits have suddenly clued into Claude Code and are predicting job doom.

But somehow, at the same time, a new NBER study of 6,000 CEOs just found that nearly 90% report AI has had zero impact on employment or productivity. And is so frustrated by AI boosterism that he’s offering Scott Alexander a public $5,000 wager that the economy will remain basically normal through 2029.

This is where we are. One camp is suddenly terrified (again) by AI’s power. The other camp is convinced the whole thing is a fever dream that will dissipate like crypto or the metaverse (a point folks like make every now and then).

It’s all extremes. All or nothing. We’re either all screwed, or it’s all going to nothing.

The reality isn’t that. I had an exchange recently with that I think captures where most thoughtful practitioners actually land: it’s somewhere in the middle. AI is real, it’s doing real things, it’s not going away—and it’s also not about to make the economy unrecognizable by next Tuesday.

But here’s the thing: whether you’re in the “it’ll all come to nothing” camp or the “we’re all screwed” camp—nobody’s happy. Everyone is very downbeat about all of this. It seems to be the theme of the times.

There’s a recent interview on Adam Savage’s Tested with Ryan Norbauer—a designer and propmaker who spent five years and hundreds of thousands of dollars of his own money building what he considers the perfect mechanical keyboard. Around the 29:30 mark, Norbauer says something that stuck with me:

He talks about wanting to physically inhabit or touch something that “feels free from the disappointments of everyday life.” And then he connects it to growing up:

I think that this is why I loved Star Trek growing up as a kid. I grew up in West Virginia. I was a very nerdy, bookish, weird, gay kid and so you feel a little bit alienated by real life. And on television, there was this incredible vision of the way things would inevitably be in the future. Everybody was going to be nice to each other, tolerances would be accepted and celebrated, and everything was going to be great. So I desperately wanted to step into that world... It’s like the future is gonna be amazing. It’s right around the corner...

Maybe reality sucks, but the future is going to be amazing... Cultural differences will disappear and we’ll all get along, right? Everyone is going to be so smart because they have access to infinite information. It’s not so possible to believe in that idea anymore.

Which I think is a sad encapsulation of what we’re seeing from the commentariat. Notably, in the west—China is quite different. As per in “An AI-Maxi New Year”:

What connects these stories is what they reveal about disposition. The Chinese society, from a world-renowned auteur to the hundreds of millions watching the Gala, is broadly, strikingly optimistic about AI. The reflexive existential dread so pervasive in Western discourse is largely absent.

The thing is, this is the same technology in the US and in China. And the future really can be amazing... though only if we think about this all rationally and properly handle both encouraging its development—and guardrailing with the right regulation. We need to be smart in thinking about tradeoffs... which is not really a feature of most commentary or politicians these days.

So, let’s talk about what’s actually happening.

It’s Definitely Not Going Away

Whatever you think about AGI timelines or the singularity, one thing is clear: AI has already fundamentally changed software development, and those changes are already showing up in the data.

I wrote about this recently in AI Gains Starting to Show in the Real Economy—Q3 2025 BLS data showed nonfarm business sector labor productivity jumping 4.9%, with output increasing 5.4% while hours worked barely moved at 0.5%. That’s a classic “pure” productivity gain, the kind you see when technology adoption is actually working. I know we’re squinting at a single data point. But it’s a notable one.

Now, the Fortune article I linked above—the one about the CEO study—will use this as a gotcha. See? CEOs say nothing’s happening! They’re not wrong about the macro picture. As I said, it’s early.

But they’re making the same mistake people made with IT in the 1980s. Robert Solow’s famous line—”you can see the computer age everywhere except in the productivity statistics”—was true in 1987. By 1997, productivity growth had accelerated to nearly 3%. Computers didn’t suddenly start working better. Organizations figured out how to use them.

Organizations today don’t actually know how to use AI. Except, as said, in software (or rather, they’re starting to here). Claude Code. OpenAI’s Codex. Google’s Antigravity. And claims that many software developers have literally stopped writing any code. I can personally attest to a similar phenomenon in my own workflow—though I will assert that it’s pretty important to manage/oversee it!

Whatever you think of the claims, it is pretty clear that something is happening in software.

Subscribe now

The Floor Argument

Here’s my proposal about how to think about this. What’s the minimum reasonable economic impact we should expect from AI, if we limit it to its single best use case?

AI coding assistants have real data behind them now. The most credible evidence comes from MIT’s field experiments across Microsoft, Accenture, and a Fortune 100 manufacturer. Nearly 5,000 developers. Real companies, real work. In short, the results were that access to an AI coding assistant increased weekly task completion by 26% on average.

Let’s be clear: we shouldn’t just uncritically take this as gospel and do something stupid like multiply it by GDP. Let’s pick that number apart and figure out how it applies.

First, 26% is a task-level figure. In a Microsoft study, developers spend only about 11-32% of their time actually doing coding tasks—the rest is meetings, planning, debugging, code review, and all the other things that make software engineering a job and not just typing. As I—and many others—have pointed out, soft skills are extraordinarily underrated in the best engineers. Multiply straight through and you get a 3-8% project-level gain from coding alone. That number bumps up somewhat once you account for AI helping with adjacent work—debugging, code review, documentation—and comes back down a bit when you factor in the overhead of reviewing and fixing AI-generated code (which, as we’ll see, is not trivial).

Second, adoption isn’t universal. Even with the tool freely available, only about 60% of developers were actively using it after a year (from MIT study, again). Newer hires adopted faster; some veterans stuck with their existing workflows.

Finally—and this is the one AI boosters don’t like to talk about—quality is a real concern. Studies show 40-62% of AI-generated code contains serious bugs. Refactoring (editing code for clarity/cleanliness versus functionality) is declining from 25% of code changes in 2021 to under 10% in 2024. Code cloning is up 48%. There’s a real technical debt story building underneath the speed gains. From my own testing, even the frontier models with the most profligate thinking settings still do dumb stuff all the time. I promise you—AI is not “one-shotting” complicated software engineering tasks without human supervision. Preferably, highly experienced human supervision. As I’ve said before, AI’s biggest benefits accrue to the most skilled/expert already.

Ok, so that’s a lot of stuff that can’t just be straightforwardly multiplied. But let’s do this.

• 20-26% task-level improvement. Let’s flatten between 11-32% spending on coding tasks to 20%. So, a modest 4-5% improvement.

• However, debugging, code review, and documentation are tasks that take up a lot of time in the MIT study—it’s another 20%’ish of total time. AI definitely helps with all of this. So, let’s add on another 4-5% for this set of tasks.

• After applying all these discounts—adoption rates, time allocation, overhead for debugging and review, quality adjustments—a defensible estimate for net project-level productivity improvement is roughly 10%.

Some, including me, thinks this is likely ridiculously low. The MIT study was in February 2025. Recent, but given how fast AI coding tools are improving, that was an eternity ago. It was also using Copilot—which no one cites as the leading coding tool/agent. That would be Claude Code.

Additionally, the number of developers since April 2025 (Microsoft’s study) utilizing Claude Code and similar tools has skyrocketed—see the below chart from SemiAnalysis. But whatever, this is meant to handle a skeptic’s argument.

Now, with this likely way-too-low number, let’s do the math.

There are about 1.7 million software developers in the US (2.1 million if you include adjacent roles like QA and data scientists). Per BLS, total compensation is roughly $357 billion annually—about 1.2% of GDP. And critically, 58% of these developers don’t work in “tech”—they’re embedded in finance, healthcare, manufacturing, government. So this isn’t a Silicon Valley story. It’s an everywhere story.

Apply the 10% productivity gain and that gets us around $36B in additional output—which is, as a share of GDP, 0.12% annually.

Is that a big deal? Well, that depends on whose estimate you’re comparing this to.

Daron Acemoglu—a leading AI skeptic in economics—estimates that AI will contribute only about 0.05% to GDP growth per year. Our floor estimate, from one application of AI (software coding), already exceeds his total economy-wide number.

Meanwhile, Goldman Sachs forecasts AI will add 1.5 percentage points to annual productivity growth by 2027. Our software-only number is about 10% of that—from one use case and likely already a radical underestimate due to utilizing numbers earlier in the growth curve. The remaining 90% would need to come from AI applications in other domains.

The market, for its part, is pricing something close to the Goldman consensus plus an upside buffer. The S&P 500’s forward P/E is 22.2x (18% above the ten-year average) and the CAPE ratio is sitting at 40—90% of the 2000 dot-com peak. The equity risk premium is essentially zero. Stocks are priced for things going perfectly right.

So, how does that all net out? Well, even if AI delivers everything I’ve described in software, that alone doesn’t justify current market valuations. The 1990s DotCom Boom did produce real productivity gains, and the S&P 500 still fell 49% from 2000 to 2002 because expectations had run ahead of reality. The productivity was real. The prices were also wrong.

But, that should be no surprise to my readers here.

All this is to say: the productivity gains are real, they’re measurable, they exceed what the harshest skeptics predict, and they’re a meaningful economic contribution. But they’re also more modest than headlines suggest, subject to real quality concerns, and insufficient by themselves to justify the market’s AI euphoria. But it’s still a big deal in real economic terms.

And remember, I suspect that estimate is already too low and we shouldn’t be surprised if the developer uptake ends up higher and the productivity gains end up higher—easily doubling or quadrupling (from 2X in both multipliers) our figure, which would get us to about half of Goldman’s estimate just from improvements to existing software engineering and no new applications. Speculative? Sure. But I think most who are trying out these tools would easily take the bet on that figure over our conservative one.

Share

What Actually Changes: Mechanical Skills

OK, so AI makes software developers more productive. That’s happening now and will continue. But the more interesting question—and the one that connects back to all the fear and excitement at the top of this piece—is what it means for skills and work more broadly.

I have a thesis here, and it’s one that I think most people aren’t going to love, because it doesn’t fit neatly into either the “we’re all screwed” or the “nothing’s happening” box.

Mechanical skills become less valuable over time. They always have.

The thing Anthropic’s Daniela Amodei said recently about humanities becoming more important is getting attention for good reason. She noted that AI models are already “very good at STEM,” and that understanding ourselves, understanding history, and understanding what makes us tick “will always be really, really important.” When Anthropic hires, she said, they look for people who are “great communicators, who have excellent EQ and people skills, who are kind and compassionate and curious.”

She’s right, and the broader trend she’s pointing at is real. But I want to push on this a bit, because I think the framing of “humanities vs. STEM” misses the deeper pattern.

Soft skills—communication, empathy, judgment, taste, the ability to understand what someone actually needs versus what they say they need—these have always been underrated and far more important in technical fields than most people appreciated. I’ve said that a few times already. I’ve worked in enough technical environments to know: the engineer who can explain what they’re building and why it matters is worth five engineers who can only build. The PM who can figure out the right thing to build is worth ten PMs who can manage a Gantt chart. Building the right thing has always been more valuable than building a thing. AI just makes that gap more visible because the “building a thing” part is getting cheaper.

But is STEM really going to go extinct? Do you, dear reader, really believe that? Does that make sense with experienced developers (and other experts) becoming massively more productive?

What’s happening is something more specific and historically familiar: the mechanical component of technical skills is being devalued, while the judgment component is being elevated. And this has happened before. Multiple times. We just have very short memories.

Historical Analogies

Photography

I discussed this in my book, but before photography, if you wanted a realistic visual record of something—say, a person’s face, a landscape, a historical event—you needed someone who could draw or paint with photorealistic skill. That was a rare, valuable, hard-won mechanical ability. It took years of training. Masters commanded premium prices.

Photography made that specific skill dramatically less valuable almost overnight. Now, it wasn’t worthless. In fact, there are still photorealistic painters today, and they’re impressive. However, the economic premium for “can render a realistic image” collapsed. What didn’t collapse was the value of composition, of artistic vision, of knowing what to point the camera at and why. The judgment survived. The mechanical skill didn’t.

Did photography destroy art? Obviously not. It transformed it. It freed artists to explore impressionism, abstraction, and a hundred other directions that weren’t about pure reproduction. The total economic and cultural value of visual art increased after photography. But the specific skill of “realistic rendering by hand” became a niche rather than a necessity.

I’ll also add on here: would you call the “real artist” of various photographs the camera? Should we credit Kodak, Canon, or Fuji instead of Ansel Adams or Annie Leibovitz? Even if we’re using a tool that makes certain mechanical skills less valuable, the artist is still the intent and motive driver behind it.

Typewriters

Just to belabor the point, for centuries, penmanship was a critical professional skill. In law, medicine, commerce, and government, the ability to produce clean, readable script was a prerequisite for professional work. People trained for years. Schools taught it as an important part of the curriculum.

The typewriter (and later, the computer) destroyed that. Not “writing”—the ability to compose clear, persuasive, well-organized prose is more valuable than ever. But penmanship? Nobody’s getting hired for their handwriting anymore. The mechanical skill evaporated. But, obviously, writers still exist.

Luddites (the real ones)

And then there’s the one everyone brings up, usually incorrectly (again, just to self-plug again, in my book!). The Luddites weren’t anti-technology. They were skilled manual weavers whose specific mechanical expertise (ironically, also a fruit of technological advancement) was being made obsolete by power looms. Their hard-earned mechanical skills became less important—allowing untrained workers (including women and literal children... for better or worse for the latter) “invade” and “devalue” the work. The result? Clothing became dramatically cheaper, demand exploded, and suddenly people were able to wear more than one single threadbare outfit.

Total employment in textiles increased after mechanization. But the specific job of “skilled hand weaver” largely disappeared.

The Luddites’ mechanical skills became less valuable. The industry their skills served became more valuable. Both things were true simultaneously, so it isn’t really as simple as “technology arrived, jobs gone.”

You see where this is going. In every case, a mechanical skill that was previously the bottleneck gets automated or dramatically cheapened. The people who built their careers on that specific mechanical ability face real disruption. But the broader field doesn’t die—it grows. And the skills that survive and appreciate in value are the ones that involve judgment: knowing what to build, what to paint, what to write, what to weave. The “what” and the “why” survive. The “how”—at least the mechanical “how”—gets commoditized.

AI is doing this to coding right now. The mechanical act of translating a well-understood requirement into working code is getting cheaper fast. Junior developers who could write clean boilerplate were gaining 27-39% in the MIT studies—precisely because boilerplate is the most “mechanical” part of programming. Senior developers gained only 8-13%, because their work involves more judgment, architecture, and navigating ambiguity—things AI is worse at.

This doesn’t mean software engineers are doomed. (I’ve written about this before.) It means the mix of what makes a software engineer valuable is shifting. Writing code is becoming less of the value. Understanding what code to write—and whether code is even the right solution—is becoming more of it.

And yes, this pattern will extend beyond software. Any domain where AI can reliably perform the mechanical components of the work will see the same shift: mechanical skills devalued, judgment skills elevated. The domains themselves won’t die. They’ll grow, because productivity increases expand the total pie. But the specific humans who refuse to adapt their skill mix will face the same disruption the Luddites did.

Subscribe now

So Now What?

Here’s the thing that frustrates me about the current discourse: everyone’s arguing about the magnitude without agreeing on the mechanism. And the mechanism is actually pretty clear if you look at the history. The productivity gains from AI in software are real—not catastrophic for software engineers, especially at the senior level. AI will expand to other domains over time, unevenly. Mechanical skills will be devalued; judgment skills will be elevated. The economy will adjust, imperfectly and with real human costs.

The future Norbauer talked about so evocatively: a Star Trek future, where it’s all going to be amazing—yeah, it’s probably hard to imagine it’s right around the corner. But we aren’t looking at some kind of dystopia either.

Unfortunately, what is likely isn’t as clickbait-friendly. Tools will get better, mechanical bottlenecks fall away, and the people who can exercise judgment about what to build and why become more valuable than ever. That’s not a dystopia. It’s not a utopia either. It’s not a white-collar apocalypse. It’s also not a nothingburger—far from it.

It’s just... the next version of a pattern we’ve seen play out for centuries. Maybe we’ll get to AGI eventually but, in the meantime, AI is going to be an important tool—which, lest you forget, the computer also was, with all of the impacts it ultimately had.

Thanks for reading!

I hope you enjoyed this article. If you’d like a serious yet accessible guide to AI, I published a book titled What You Need to Know About AI, now available as an audiobook narrated by me!

You can learn more and order it on Amazon, Audible, Barnes and Noble, or select indie bookstores in the Bay Area and beyond.

1. The audiobook edition of my book What You Need to Know About AI is available now! Check it out on Audible if you’re interested.

2. This Saturday (tomorrow!) meet me at Barnes & Noble in San Jose, CA on February 7th from 1-3pm. No RSVP required, but if you happen to be around feel free to swing by and say hi!

3. This article I’m trying a voiceover version where I narrate the article. Check it out and let me know if you enjoy it.

And now, back to our regularly scheduled programming.

A foundation, a private equity fund, a hedge fund, a journalist, and a pastor. That’s not the start of a “walked into a bar” joke. Those are entities that have had in-depth questions for me about Nvidia losing its moat. As such, it seems like the right time to write an update about my thoughts on Nvidia and its CUDA moat.

In early 2024, I wrote about CUDA’s moat and argued it was as high a barrier as the hardware itself. Higher, in fact. The argument was simple: NVIDIA controls the entire stack, stuff just magically works, and years of tools, libraries, PhD blog posts, and guides all assume you’re using NVIDIA GPUs. Even with a lot of capital at stake, it’s hard to magically catch up to nearly two decades of battle-tested scientific computing, with the researcher community helping report problems and help develop the Nvidia’s platform.

Even if alternatives improved, I said, “it would take quite a while for the impacts of such a heavy, long-term strategy to go away.”

I still believe that. But “quite a while” could be getting shorter. There’s at least a window for competitors. The reason isn’t that AMD or Google suddenly got better at software (the former has fumbled the ball consistently and the latter was never bad).

The reason is that TSMC can’t make enough chips, Nvidia can’t buy its way out of the constraint, and customers are being forced to try alternatives—specifically for inference.

The Supply Constraint Nobody Chose

TSMC’s most advanced node is sold out through 2026. Now, you might ask: isn’t being sold out through next year pretty normal for a leading-edge foundry? Actually, no. Not in the way that it’s sold out now.

At the Semiconductor Industry Association Awards in November 2025, TSMC CEO C.C. Wei gave a blunt assessment: “Not enough, not enough, still not enough.” He estimated that TSMC’s advanced-node capacity is roughly three times short of what its major customers want to consume. For context, in late 2022—right before ChatGPT launched—wafer starts had actually collapsed because the smartphone, PC, and server markets all went into recession. TSMC had excess capacity. That’s a far cry from where we find ourselves today.

The desperation is visible in tech leader pilgrimages to TSMC to beg for more chips. Sam Altman has made multiple secret visits to Taiwan to meet with TSMC executives about chip supply. Apple’s Tim Cook explicitly called out TSMC capacity as a reason they couldn’t fulfill demand for the new iPhone 17 Pro this quarter—and probably next quarter too.

Finally, we come to the company we care most about in this article.

Nvidia has secured over 50% of TSMC’s expanded CoWoS (Chip-on-Wafer-on-Substrate) capacity through 2026. Is that normal? Absolutely not.

Apple has been TSMC’s largest customer for over a decade. That’s now changing: Nvidia has overtaken Apple as TSMC’s biggest customer.

In 2023, Nvidia was maybe 5-10% of TSMC’s revenue. By 2026, analysts project Nvidia at $33 billion (22% of TSMC revenue) versus Apple at $27 billion (18%). Jensen Huang confirmed the shift publicly.

This is an incredible realignment—and it’s still not enough to meet demand for Nvidia chips.

I’ve criticized Nvidia for squeezing the market with suboptimal VRAM (high bandwidth “video RAM” or, basically, GPU memory), because it creates space for competitors to come in and seize the market. But maybe that critique has been unfair.

Perhaps this is the best option Nvidia has—since they cannot choose “give the market all it wants to keep everyone hooked on CUDA.” Not that it was ever my argument, but they’re not merely artificially constraining supply to maximize profit extraction. They literally cannot get enough capacity from the only foundry capable of making their chips at the required specs.

Now, realistically, few will shift off of CUDA—even under these conditions—for training frontier models. This is, after all, what all of those AI labs have been securing all those Nvidia GPUs for. There is, however, a place where a more diversified approach can make sense.

Inference.

Why Inference is the Soft Underbelly

Let’s be clear about the market structure. Training—“teaching” AI models—is maybe 20-45% of total workload volume, but it’s where the money is dearest and switching costs are highest. Training is also where CUDA’s advantages are most acute. The performance gap versus alternatives is substantial, the optimization work is deep, and the developers who specialize in training (and, if not them, their tooling) are still predominantly CUDA-trained. Nvidia isn’t losing training anytime soon—and I don’t really see any serious-minded arguments to the contrary, even with various experiments by AI labs with other platforms.

Inference is different. Inference is running a model that’s already trained—asking ChatGPT a question, generating an image, running a recommendation engine. By compute volume, inference hit ~50% in 2025 and is projected to reach two-thirds by 2026. McKinsey projects inference growing at 35% annually versus 22% for training through 2030. And by lifetime cost, inference can account for 80-90% of total AI spending in production systems—training is a one-time investment, but inference runs continuously.

More importantly, inference costs are collapsing.

at a16z coined the term “LLMflation”: for an LLM of equivalent performance, the cost is decreasing by roughly 10x every year. ARK Invest cites a similar figure of 85-90% per year.

When inference gets that cheap, two things happen. First, inference becomes economical everywhere—edge devices, internal tools, use cases that would never have existed at $20 per million tokens. Second, when the workload becomes price-sensitive, customers start asking harder questions about whether they really need Nvidia hardware for it.

Even worse, some customers can’t get Nvidia hardware even if they wanted it. In which, they’re forced off Nvidia by default.

Subscribe now

The Ecosystem Flywheel—In Reverse

I wrote that CUDA’s moat wasn’t just technical superiority—it was ecosystem lock-in. PhD students learn CUDA. They write their dissertations using CUDA. They publish blog posts about buying used RTX 2090s (well, back then). They write libraries that assume CUDA. Those libraries get used by the next generation of researchers. The cycle reinforces itself.

That entire ecosystem is the real lock-in and a network effect. It’s just not worth it to go where other people in this niche aren’t. Once people starting going somewhere else though...

Let’s think about what happens when a company—AI lab, hyperscaler, or anything else—can’t get H100s for inference and decides to try Google TPUs or AMD’s MI300X or Amazon Trainum or Cerberus or (... you get the idea) instead:

1. They hire or train engineers who know how to deploy on that platform

2. Those engineers write internal tools and scripts optimized for non-CUDA

3. Some of them blog about it, write guides, contribute to open-source projects (or open source what they wrote)

4. The next company trying to deploy inference sees those guides and decides that going not-Nvidia is not that bad

5. Libraries improve because there are now real users filing bugs and contributing patches

6. Switching costs drop because the ecosystem is maturing

7. More companies try alternatives… and then we get back to step 1

This doesn’t necessarily break training’s moat, but with the size of inference, losing that market is already bad enough for Nvidia.

But the more people get experience and build things for non-NVIDIA hardware, the easier it will get. That two decade advantage of battle-testing for Nvidia’s ecosystem chips away, little by little.

Midjourney reportedly saved 65% on inference costs by switching from Nvidia GPUs to Google TPUs—dropping from $2.1 million monthly to under $700,000. AMD’s ROCm now has first-class PyTorch support. Llama.cpp—the inference library that’s become standard for running open-source models locally—now has official ROCm support.

(Speaking of, has a great article on TPUs that’s a must-read on the topic.)

The Chip Letter

TPU Mania

Nobel Prize winning economist, Paul Krugman, discusses TPUs and GPUs with Paul Kedrosky. No criticism of Krugman intended, it’s great that computer architecture is getting so much attention…

Read more

4 months ago · 87 likes · 1 comment · Babbage

How important are all of these? Less than their headlines, honestly.

Google TPUs are unlikely to catch on with the major labs, despite some high profile announcements, most notably, Anthropic. Google’s kind of a competitor to them all. AMD’s ROCm, the butt of many of my jokes, has famously shot itself in the face for years and I have faith in their ability to continue to do so. But obviously folks are trying out alternatives and a big part of it is difficulty getting Nvidia GPUs.

Eventually, everyone would prefer not to be locked-in and if the cost of switching falls enough... well, that’s the end of the moat.

Nvidia Isn’t Sitting Idle

Nvidia isn’t stupid. They obviously can read the same trends as I can. That is, of course, why it’s interesting that Nvidia (and Apple) is reportedly shifting some 2028 production to Intel’s fabs. This follows on from a large strategic investment from Nvidia into Intel and reportedly aiming to move significant amounts of production of the upcoming “Feynman” architecture to Intel as well.

Obviously, this is also useful for diversifying from Taiwan and its geopolitical complications. If we do see more movement to Intel—or even Samsung, the other leading-edge fab—we could see Nvidia potentially reversing their fortunes in the future.

The question is whether they can close it fast enough.

Intel’s fabs won’t be at full production until 2027 at the earliest. TSMC’s expansion takes time. And meanwhile, every quarter that customers can’t get Nvidia hardware is another quarter of forced experimentation with alternatives.

What About China?

While it isn’t an elephant in the room for me, I know it is for some readers. What about Chinese AI companies?

Despite U.S. on-again-off-again export bans, Chinese companies remain heavily Nvidia-dependent. Roughly 75% of AI training chips in Chinese data centers still run on CUDA. And there are workarounds to the US’s mercurial policy on these chips. Tencent signed a $1.2 billion deal with Japanese neocloud Datasection to rent access to Nvidia’s B200 chips—which is a fairly common route for Chinese companies to get access to Nvidia GPUs. DeepSeek reportedly operates a fleet of 50,000 Hopper GPUs stockpiled before restrictions tightened. Companies are going to extraordinary lengths to stay on Nvidia hardware—enough so that the Chinese government has done some on-and-off restrictions, including discouraging/banning H200s.

Huawei’s Ascend chips and CANN, its version of CUDA, are the domestic alternative. ( has done a good job covering the topic.) And there is a lot of real progress there.

AI Proem

What are the capabilities of Huawei’s chips and software systems?

Hi all, this is a delayed update on Huawei, which was first published in AI Supremacy as a guest post. It features a special guest appearance by Ray Wang, a Washington, D.C.-based semiconductor and AI analyst…

Read more

10 months ago · 23 likes · 4 comments · Grace Shao

For example, ByteDance signed a $5.6 billion order for Ascend chips, which is a little bit too much to be pure window-dressing for the sake of looking like you’re supporting the China option. But CANN still has plenty of problems. Again, you can’t catch up to two decades of refinement overnight. Chinese companies are dragging their feet on switching despite government pressure. (As I’ve said before, the Chinese ecosystem is not is not a monolithic community or hive mind. Contrary to what some Western observers think, if individual companies can get an edge on other Chinese companies, they’ll do it.)

The China situation is a great demonstration of the continued durability of CUDA’s moat: even with maximum government pressure and supply restrictions, Nvidia still not only exists but dominates. Still, enough forced adoption, progress from Huawei, and (again) movement of the broader AI community towards other platforms does create a real threat.

So, is Nvidia doomed to lose their moat?

No, not really. Despite all the doomsaying, it isn’t inevitable. After all, this story could reverse quickly if Nvidia instead finds capacity to flood the market with their chips, backed by their superior ecosystem of libraries, experienced researchers/developers, and battle-tested software. It is a window of opportunity for competitors, though.

TSMC might or might not make enough investments to ease constraints by 2027. Intel’s capacity will come online (theoretically) around the same time. We might end up with overcapacity. The semiconductor industry, including fabs, have an infamous tendency for large boom and bust cycles—its inevitable given the huge lead times and massive capex inherent to the industry.

In the meantime, however, we might see a significant ecosystem diversification from Nvidia. And that starts in inference... which could eventually cause the loss of Nvidia’s moat everywhere else.

Thanks for reading!

I hope you enjoyed this article. If you’d like to learn more about AI’s past, present, and future in an easy to understand way, I’ve published a book titled What You Need to Know About AI, now also available in audiobook format!

You can learn more and order it on Amazon, Audible, Barnes and Noble, or select indie bookstores in the Bay Area and beyond.

Claude Cowork launched last week. It wasn’t really greeted with much fanfare outside of AI/tech circles, though. To be fair, it isn’t really that different from an instance of Claude Code with a nicer UI wrapper around it for non-coders. I’ve been contorting Claude Code to do what Cowork does using MCPs and dumping documents into folders for it to read for a while now.

As covered, this movement towards AI working to do things in the real world is also a trend in China—most concretely with Alibaba launching their own Qwen Assistant last week as well.

This is all part of a wider trend of accessible agents doing things for people, like what I pointed out in “The Boring Phase of AI.”

Even though Claude Cowork may not be technically impressive, it’s still a potentially major milestone. Calling back to my recent decade review, it’s not like ChatGPT was a magical step beyond what the GPT models (GPT-3, in particular) demonstrated for years. But ChatGPT made it accessible to a wide audience, and that accessibility was transformative. A technology that only experts can use is not really a technology yet—it’s a prototype. The moment regular people can pick it up and get value from it is when the real economic effects begin.

Is AI productivity starting to show in the data?

Economic statistics are always lagging indicators. We put enormous effort into trying to know what has already happened—and often get it wrong, seeing as government agencies often revise data (sometimes significantly) months later.

Among these kinds of stats, productivity data is notoriously hard to read. Even in hindsight, years later, we’re often trying to build narratives and read tea leaves. As such, it seems kind of silly to try to make much of a single datapoint, especially one that you have to kind of squint at to see its relationship to the trendline.

Nonetheless, recent BLS data for Q3 2025 was notable. Nonfarm business sector labor productivity jumped 4.9%. Notably, output increased 5.4% while hours worked barely budged at 0.5%. That was a classic “pure” productivity gain, which looks a lot like fast-moving technology adoption.

Subscribe now

In the chart below, I draw two lines—first, the trend “before” the recent period. And secondly, the one after 2020 (which has some oddity because of COVID suddenly causing labor to collapse with lockdowns). Again, I know we’re squinting, but it does seem to be a break in the line—which would likely just be seen as noise without the underlying context that the economy is still fairly strong, hours worked didn’t really fall (i.e., not like COVID’s lockdowns), and output jumped a lot. We’ll obviously have to keep monitoring it going forward, but it would be fascinating if this is the start of a real trend.

Just for context, it is worth noting that productivity data is cyclical. During recessions, productivity tends to spike—not because workers suddenly got smarter, but because (in classical theory) less experienced workers get laid off first, raising the average skill level of those who remain. (If you want to be more cynical about it, workers are laid off and the remaining need to pick up the slack—at least temporarily and perhaps unsustainably.) Capital per worker also temporarily rises when you have fewer workers using the same equipment. As the economy recovers, these effects unwind.

The San Francisco Fed shows a similar analysis and puts it into broader context. Specifically, our lost decade(s) of productivity since around 2005. 2008 temporarily spiked it back up with mass layoffs, but it’s been sluggish as a whole.

Why Do We Care?

More productivity does generally seem good, but why does it matter? On a macro level, productivity helps define how fast the economy can grow without inflation. This is because taking away population growth (which is one source of GDP growth) and exports, what your economy can sustain is defined by how efficiently you can build stuff.

Think about it this way: if everyone wants to buy more stuff, but we’re unable to actually make more stuff faster, what happens? Prices go up. PC gamers are feeling this with the RAM squeeze (driven by data center demand).

We’ve already had stubbornly high inflation, in part because of high spending in the US, which has been at near wartime levels for the past few US presidential administrations, and not enough central bank tightening. The point of fiscal austerity and monetary tightness is to rein in the economy when it is running “too hot”—meaning too far above potential GDP... which is defined by productivity.

Now, the stock market is not growth. It’s a market.

Notably, as my former colleague recently pointed out in explaining why the stock markets of Colombia, South Korea, Spain, and Greece were the top-performing markets in the world (116%, 100%, 83%, and 79% return in 2025 respectively—the US is #38!):

Profits are not the same as equity returns. It matters what you pay for those profits.

Nevertheless, it would certainly help if potential GDP were much higher given how ridiculously high the pricing is for the US stock market—which is essentially at historical highs relative to earnings. If the ceiling of the economy to grow rather than just create inflation were much higher, one could make a far, far better argument for why US tech stocks (largely, the “Magnificent Seven”) aren’t crazily overpriced.

That all being said, while the US stock market is now a huge store of wealth for many Americans (... and many others in the world), the stock market isn’t the real economy. If AI really has the ability to massively increase productivity like some of its most fervent (but non-insane) boosters claim, it would make a huge difference in people’s standards of living.

It’s Still Early

Regardless of where productivity actually is, AI adoption is still in its early stages. Yes, ChatGPT exploded onto the scene in late 2022 and sent shockwaves through the tech world. Yes, companies have been scrambling to integrate AI into their workflows. But if you look at actual enterprise adoption data, most organizations are still experimenting.

If we’re already seeing productivity effects at this level of adoption (enough to show up in productivity stats), what happens when AI tools become truly integrated into day-to-day work?

Book Event Next Week

If you’re around San Francisco, I’ll be doing an event at Book Passage on Sunday January 25th at 2pm! I’ll be doing a reading of my book What You Need to Know About AI, signing books, and answering questions from the audience. Hope to see some of you there! If you’d like an event reminder you can sign up here.

When did the current AI cycle begin? There are a lot of ways of slicing it, including all the way back to the mid-to-late 1980s, as backpropagation—our current method of training deep neural networks—was rediscovered and made more practical. Maybe it was the mid-2000s, when tech companies were utilizing machine learning more and more. Maybe it was November 2022, when ChatGPT was released.

I personally say around 2015.

AlexNet, which was created by Alex Krizhevsky, Ilya Sutskever (among other things, co-founder of OpenAI), and Geoffrey Hinton, was earlier, in 2011-2012, but we hit human levels of performance around 2015, with the (very large model at the time) MSRA from Microsoft Research in particular.

Vision and recognition, one of our fundamental senses, was no longer the exclusive domain of humans.

It was also the same year that Yann LeCun, Yoshua Bengio, and Geoffrey Hinton wrote their excellent Nature review paper, “Deep Learning,” helping announce that it has arrived.

On a sentimental level, it’s also when I started pivoting toward AI in my career—starting from 2013, explicitly exploring machine learning, taking graduate seminars at Berkeley, and eventually starting investments in the field by 2016. But that was obviously correlated to seeing where the field was going.

I’ve always said 2015 (or thereabouts) was “when AI really kicked off,” but it’s also convenient that it now makes it a nice, round decade.

A Decade of AI Milestones

This is just a repeat of my chart above, but I know it’s hard to read on some platforms.

• 2015: Human-level performance in image recognition—quickly followed by progress in game-playing and machine translation.

• 2016: AlphaGo defeats Lee Sedol using a combination of deep learning, search, and reinforcement learning. Google publishes “Attention Is All You Need,” introducing the Transformer architecture.

• 2018: BERT demonstrates that natural language processing need not be task-specific.

• 2020: GPT-3 shows that scale alone can produce general-purpose language models.

• 2022: Release of ChatGPT. Less a technical breakthrough than an interface breakthrough—AI enters public consciousness. Diffusion models (Stable Diffusion, DALL·E 2, and Midjourney) show that creative work is also within AI’s domain.

• 2023: The AI boom begins in earnest. Tool use, agents, and multimodal models make systems practically useful.

• 2024: OpenAI’s o1 is released. Reasoning becomes explicit, and inference-time scaling becomes a first-class capability—including with DeepSeek, causing the world (including China) to suddenly pay attention to Chinese AI.

• 2025: Open-weight models (including DeepSeek/Chinese models) and agentic systems proliferate. AI becomes less magical—and reliably useful.

Subscribe now

Things to Point Out

Google’s Early Dominance

I like my crude visual timeline because one thing it demonstrates (in my selected events) is just how dominant Google, in particular, was at the beginning of the decade. I’ve had multiple individuals in the company laugh rather wryly that the company published the Transformer paper that made OpenAI’s accomplishments later possible.

OpenAI Seizing the Crown

Sometimes executives that I’ve spoken to at Google wonder if they should have released it openly—but, either way, the company essentially did next-to-nothing with the technology, allowing OpenAI to create GPT-3 in 2020 and then capture public imagination in 2022 with ChatGPT.

Diffusion in the Last Few Years

Although I specifically pulled out DeepSeek (and called out DeepSeek for 2025), the reality is that there is much more diffusion.

I could have also included Claude Code (and thus Anthropic) in 2024. But the point is that the number of players that one should pay attention to has exploded—especially in the Chinese market.

Zhipu and MiniMax are both launching IPOs in the Hong Kong stock market. Cursor’s proprietary model seems to be using Alibaba’s Qwen. In hardware, China has been pushing their native semiconductor capabilities, with the government mandating 50% domestic equipment use for chipmakers (perhaps allowing us to diversify from just talking about Nvidia over time).

In the US market, Gemini 3 obviously recently seized headlines (bringing Google “back”—not that they were ever gone). Nvidia acquired Groq’s assets in a $20B sale, showing various aspects of the hardware race continuing here.

From Relatively Unknown… to the Center of the Economy… to Just Another Technology

AI has become a larger and larger portion of the total economy. This is obvious from the US stock market, but it’s more than that.

Even though AI adoption rates across the economy are starting to level out, we have more and more “boring innovation.” It has gotten integrated into companies at a shocking pace—a lot of it happening just within 2025.

I made this point in my last post, but the idea is essentially that AI will diffuse further into the economy in ways that don’t warrant headlines. It will just be as ubiquitous as internet use or electricity.

It’s just a technology being used, not something that will suddenly exterminate humanity, cause us to reach post-scarcity, or any other sci-fi-like AI takeoff scenarios.

What’s Ahead?

I’ve been predicting for a long time that “Most AI Startups are Doomed.” Part of that prediction, of course, is boring. Most startups die and are doomed.

However, the real question is whether or not we will see some kind of enduring advantage for, say, AI model companies. Google suddenly causing “Code Red” at OpenAI is just another day in AI headlines. The next time OpenAI or one of the Chinese companies releases their next frontier model, it’ll suddenly be “Code Red” at every other company.

Ultimately, without some kind of true barrier to entry, most AI companies will not enjoy some kind of amazing future where they dominate. OpenAI may be a strong consumer brand in ChatGPT. Anthropic has a pretty good API business and focuses on coding. But as the Chinese AI companies IPO’ing have shown, there’s actually not as much revenue as you’d think in their businesses and certainly far less earning potential. They all have the same models, data, and compute.

What do I think? I think there’ll be a lot of strong vertical applications in AI, with proprietary data. I wrote about this in “What is Defensibility” in early 2024. It was hardly prescient—it’s just normal business dynamics.

Nevertheless, I think the boring is exciting. The internet’s first generation was “exciting,” but a lot of its potential was later, quietly changing every aspect of our lives. After all, a ‘90s kid would recognize the early 2000s—it’s just what was in the 1990s, with less hype and more evolution.

A ‘90s kid would be lost with a smartphone-driven internet, remote work, and AI (which was spread with a ubiquitous internet). Not to mention ride-hailing, food/grocery delivery, e-tickets for everything, Apple/Google/Ali Pay… But, somehow, a lot of the 2000-2020 period just “wasn’t that significant” and “didn’t add that much productivity” in popular conception, versus the heyday of press attention on the internet as a technology.

Subscribe now

Other News

If you haven’t had a chance, take a listen to my podcast with on . It was a great conversation on Chinese AI, society and AI, chips, healthcare, and all sorts of fascinating areas!

AI Proem

Why American Investors Should Have a Pulse on China with James Wang

In this episode of Differentiated Understanding, I talk with James Wang, general partner at deep-tech fund Creative Ventures, author of What You Need to Know About AI: A Primer on Being Human in an Artificially Intelligent World, and writer of the newsletter…

Listen now

4 months ago · 7 likes · 2 comments · Grace Shao

Separately, my book is available in bookshops all over the US, but especially in the Bay Area. Check to see if it might be in your area!

Also, please rate and review on Amazon if you’ve read the book and haven’t yet! It really helps!

Gemini 3 caused Sam Altman at OpenAI to declare a “Code Red.” This is likely the most significant model release ever, with off-the-charts benchmark performance—until at least the next model release.

While commentators liberally spill ink about how important this is and cover the juicy drama (Sam Altman is world-class at generating PR), Gemini 3’s real-world impact is fairly muted. It doesn’t suddenly one-shot complex coding problems. Its ability to do “Deep Research” is not night-and-day from GPT-5.1 (… it actually hallucinates much more than OpenAI’s offerings). It’s better at benchmarks—for whatever that’s worth—but while it maybe feels somewhat better, it certainly doesn’t feel like a revolution in day-to-day use.

It’s likely most releases from this point onwards will feel this way, punctuated by the occasional breakthrough that will quickly diffuse to the other model companies/labs. This was an article I was already planning to write even before Gemini 3’s release. And its release doesn’t change my opinion at all.

When We Hit Diminishing Returns

I will say one thing about Gemini 3 that’s significant: while we don’t know how big it actually is, reporting and extrapolation suggest that it may be the largest pre-trained model ever. Just last month I wrote that models aren’t exponentially growing in size (and, if nothing else, size isn’t leading to massive performance improvements). It’s hard to do any concrete analysis when we don’t actually know how big it is, but I suspect that the “diminishing returns” aspect of size-scaling holds.

That doesn’t mean that OpenAI’s “Code Red” won’t make them throw bigger pre-training runs and models at the wall, even if it isn’t efficient/cost-effective/financially prudent, though.

Nonetheless, this kind of scaling isn’t going to make a night-and-day difference. The bigger question is what these model companies will do to make themselves more relevant.

Business Models and Real-World Impact

One of my predictions is that the word “agentic” to describe AI will go away. Eventually, all AI will be agentic. Why wouldn’t it be? The real promise of AI isn’t that it will get better and better at chatting with us and summarizing articles—it’s that it’ll go and do stuff for us.

In a way, China is far ahead in this aspect:

The real split is over where each country’s tech companies believe the profit will land. China bets on applications; America bets on the model itself.
- in “China and the US are Running a Different AI Race”

For example, MiniMax recently also seized headlines due to its M2 model release and various rumors about a Hong Kong IPO ($4B in public reporting, though I’ve heard higher figures privately as well). It has Talkie, an AI companion app with subscriptions, microtransactions, etc. It threw Hailuo AI at the wall for image generation and has tried to push it into various applications.

There’s also just been way more AI rollout in online shopping (see ’s article here) or it being widely used (and touted) in the Chinese insurance industry or even in entertainment with AI-native dramas. Aside from these “documented” examples, I’ve heard all sorts of insane AI things getting rolled out in WeChat, individual cafes, schools, and pretty much every level of society.

US adoption has been much more in coding and trying to use API calls from companies like OpenAI and Anthropic. Ironically, even in the US, for those utilizing fine-tuning or creating “GPT-wrappers,” open-weight, permissively licensed Chinese models are becoming more popular.

Subscribe now

“Agentic AI” Will Just Become “AI”

Every major model maker has various agentic products (Claude Code, Codex by OpenAI, and Antigravity by Google, just to name the coding agents). Although I literally just named coding agents as one of the main things happening in US adoption, it’s still a good example of things to come.

Specifically, AI that does things for us, whether it’s researching a vacation and then booking tickets, or managing our calendars autonomously, or monitoring our home temperature and managing it based on things it asks us…

Or on the enterprise side, monitoring and helping decide how to scale out (or back) cloud infrastructure, triaging and managing customer service requests, or automatically surfacing important business metrics…

All of these kinds of applications are within reach of our current AI but need to be “productized” to do so. As we start running out of just pure model performance, this is likely going to be the more interesting competitive landscape (along with differentiated models, like Google’s Nano Banana, which is honestly much better than its competitors in image generation).

If anything, this is the “risk” to “GPT-wrapper” companies. As the model companies pivot to getting things done, rather than having better benchmarks, there’s a danger of a lot of companies trying to do these things getting Sherlock’ed.

All this will be far less visible and flashy than what we’ve seen so far—which have been the heady, exponential growth days of an early technology with a lot of low-hanging fruit. However, I’d argue that branching from just better models (on paper) into doing things will ultimately be much more impactful, relative to any single big, flashy model release.

In a way, you can see it similar to how the Internet evolved. Everything getting a website wasn’t the real revolution. It was how it ended up enabling far more interesting business models on top of it later.

A Side Note on US and China

Ironically, in the past few days, I’ve said to a few different people who’ve asked about the Chinese model companies that their valuations are still mainly on potential—while OpenAI and Anthropic have “real businesses.”

I stand by that statement, even though it probably sounds weird juxtaposed to what I literally just wrote about applications a little earlier in this very article.

MiniMax, Moonshot, Zhipu, and more are all throwing stuff at the wall. My understanding, both from public info and from private conversations, is nothing has really taken off in a large enough way to really brag about or back valuations (if you have different info, please let me know!).

I expect something will catch fire eventually (experimentation is how you find out), but they’re still searching for “real businesses.” On the other hand, Anthropic is pursuing an enterprise business/API strategy that by all accounts is growing, and OpenAI obviously has its subscription business that is said to hit $20B of annualized revenue this year.

These are “real businesses,” with multi-billions in revenue that aren’t going to just go away. Though one needs to ask, relative to the amount of investment they’re putting into it, what is the “rest of it”? These have gotten impressively far, but to grow into their valuations, they’ll need to show more measurable impacts.

OpenAI kind of half-heartedly put out a shopping research tool, though the idea is good. In the midst of their “Code Red,” OpenAI apparently is also deprioritizing their efforts to create an ad-supported model—which ultimately would be the real replacement threat to Google.

These companies, even doing it in fits and starts, will get there. Google, after all, has embedded Gemini everywhere—even if you really don’t want it to do stuff for you in Google Slides or Docs.

Thanks for reading!

I hope you enjoyed this article. If you’d like to learn more about AI’s past, present, and future in an easy to understand way, I’ve published a book titled What You Need to Know About AI.

Sign up to get a free preview of the introduction and first chapter here.

You can also order the book on Amazon, Barnes and Noble, Bookshop, or pick up a copy in-person at a local bookstore.

And if you prefer physical books, you can find signed copies at select local bookstores—full list here.

Most people don’t understand bubbles.

The popular conception of them is that investors get greedy and stupid. They make crazy, irrational investments and then lose their shirts. Or, if you’re more conspiratorially minded, the elite get out at the top and everyone else holds the bag. And, of course, there are heroes who call it. In popular retellings of these periods, like 2008, such as in The Big Short, these come in the form of contrarians like Michael Burry (Christian Bale) and Steve Eisman (Steve Carell).

Don’t get me wrong, I really like that movie—for another perspective, I’d also suggest Margin Call.

Those movies aren’t bad in terms of accuracy in capturing certain aspects of finance culture and some of the events. They are, obviously, oversimplified. Painting a little more in black-and-white than reality is a useful cinematic tool.

But if we’re really going to talk about bubbles, not all of them are created equal.

A Bubble, Again?

Why am I talking about this now? The nominal reason is recent chatter about AI being a massive bubble—especially in light of the Magnificent Seven’s ridiculous divergence from the rest of the market, historical highs in US equity forward P/E ratios, etc. And, of course, OpenAI’s somewhat crazy, eye-popping machinations, like the $300B deal with Oracle, or AMD’s rather strange chip deal, just to name two recent ones.

Of course, this narrative was also around last year when OpenAI’s revenue was $3.4B (it supposedly will top $20B this year). I know because I had to write a piece addressing the point at the time.

Financial and Technology Bubbles

We’ve had all kinds of bubbles. Some are speculative financial bubbles—the mortgage crisis in 2008 (Global Financial Crisis) comes to mind. Or the Dutch Tulip Mania in the 1630s.

Others are technology bubbles. American railroads at the turn of the century. Or the Internet boom and bust in the 1990s and early 2000s.

Regardless of anything else you take away, you should note that these are not the same thing. You might get insane price appreciation and crazy financial speculation. People may lose their shirts. In both, players all look back at it like a really bad drug trip or a poorly considered weekend in Vegas (or Atlantic City or Macau—just to be more inclusive about it).

But it’s not the same thing.

Financial Bubbles are Predictable—Kind Of

One of the key plot points for both The Big Short and Margin Call is that smart people with sharp pencils figured out the multilayered, complicated structures of CDOs (collateralized debt obligations), MBS (mortgage-backed securities), etc. were built on bad assumptions about default rates and their correlations—making them a house of cards. Could it have not collapsed? Sure, if interest rates didn’t go up, the economy stayed super strong, etc.

It’s not a foregone conclusion it was going to collapse. It’s just a very likely one. It’s a nice, happy popular media narrative that the public likes—those greedy, anti-social math geeks screwed everything up (hmm, wonder if that might be the case for AI too…).

Obviously, by my tone, you can infer that I think it’s not that simple.

Subscribe now

It’s a matter of degree

Bundling mortgages isn’t actually a stupid idea (if it were, we’d be in trouble right now… and way fewer people could afford houses). Diversification is a real thing.

The “leverage” and “debt” story isn’t even the problem. If it were… what are we doing offering mortgages anyway—or college loans or anything else? Or, you know, microfinance loans for impoverished borrowers in West Africa (like where I used to work). It sure would be news to all those non-profits trying to help borrowers finance their inventories of bananas, textiles, or other trade goods that they’re committing some kind of moral evil.

It isn’t stupid to think that getting people into homes (or educations) before they earn every last penny to pay for it out-of-pocket is a bad thing—if it’s even possible to do so. It isn’t a stupid thing to think that the mortgages in California vs. Utah would have different characteristics and hit trouble at different times.

What might be stupid? Well, leveraging to the absolute hilt and also assuming that correlation assumptions are sacrosanct and can’t be violated (i.e., there can’t be a nationwide recession).

Food isn’t a bad thing, even if overeating is. Moderation in eating for health is not a bad thing, even if anorexia is. It’s a matter of degree. And, yes, that degree can be “measured” even if it’s not super predictable when bubbles pop—the entire mortgage thing worked until it didn’t, and a confluence of a couple of economic hiccups helped push the thing over the cliff.

(And yes, for the conspiratorially minded, hedge funds don’t tend to “correct” or “price discover” the right levels—it’s stupid to fight the market, which can, as Keynes noted, “stay irrational for longer than you can stay solvent.” TSLAQ have discovered this to their sorrow—however right they are about Tesla. And academic papers have found that hedge funds usually ride bubbles, not fight them.)

Technology Bubbles are About Predicting the Future of a Totally New Technology

Going back to my prior hand-drawn framework, the problem with trying to “predict” tech bubbles is no one quite knows where the tech will go.

In 1952, this is what one article predicted we’d have in the year 2000:

By the year 2000 cures for most of the diseases of man will have been discovered. The average age will be about 100 years. Journeys through space in rocket ships will be an established form of transportation, with regularly scheduled trips to the various planets. A number of man-made moons will be circling around the earth.

Is that right? No. Obviously, we fell quite short on certain aspects. However, our level of computing miniaturization, the advancement of our biotech (because it wasn’t as understood how complicated it was back then), and numerous other aspects are much more advanced.

I’m not going to totally rehash my last bubble piece, but the point is:

I’m both saying tech bubbles are actually reasonable/rational… and highly unpredictable in the moment when our expectations truly exceed where the technology will go.

It’s one thing to be using purely historical financial models of default and delinquency rates for mortgages—any analyst should immediately see red flags in building in too much reliance on that. But hey, we have data on that. It’s quite another to try to guess what a totally new, never-before-seen technology (which, by definition, there is no data on) is supposed to get to.

A great example of how silly trying to precisely extrapolate and nail down a new technology’s value using the past is Aswath Damodaran (finance professor at NYU Stern)’s piece on FiveThirtyEight explaining why “Uber Isn’t Worth $17B” (he thinks it’s closer to $5.9B) and VC Bill Gurley’s deconstruction of it in “How to Miss by a Mile.”

The current market cap is around $175B.

Subscribe now

The Real Revolution is “Gen2” of the Technology

In the immediate aftermath of the dot-com bust, many commenters talked about how stupid it all was. Dot-com everything. Eyeballs. Webvan with “online grocery delivery.” How stupid.

Of course, given most of the Mag 7 is composed of internet-related companies, and Google and Amazon became quite large companies (and, oh yes, Instacart, DoorDash, Uber Eats, etc. exist), I’d say that over two decades later, boosters of the internet may have had the last laugh.

Even the “dark fiber” of the Internet, the horrible, shirt-losing investments of the late 1990s and 2000s, helped become the backbone of today’s digital economy, as per this great piece by .

Still, other than “actually, the dot-com bubble completely changed our world,” it’s useful to remember what predictions of it were. It’s really easy to take a super simple extrapolation: just stick everything online.

Many small-town publications rejoiced in the democratization the internet brought. Their tiny circulation numbers didn’t matter on the internet—they could get subscribers and readers anywhere. Of course, we know what actually happened. The destruction of geographic barriers not only meant a newspaper in Wichita, Kansas, could get subscribers from anywhere, but they were in head-to-head competition with The New York Times. They were no longer protected by locality. The New York Times, in a head-to-head competition, wins.

The market flows downstream to the best. This is the same as what we’re literally seeing now. While translators have seen significant job loss to AI, I interviewed certain translators who are particularly good at subspecialties like literature or legal documents who literally 10X’ed their work. They no longer have “bodies to throw at the problem” barriers for low-level, lower-skill work—AI can do it, and they can simply focus on the tricky parts and give the documents a quick once-over.

That’s potentially predictable. But, during the Internet boom, could you have guessed that web native advertising would, in turn, expand and eat a lot of the market? In other words, could you have predicted that not only would it be the best newspapers dominating online, but also players who were not newspapers at all, like Google or Facebook/Meta?

This is the problem in investing and predicting the top of a bubble. It’s really hard to know the form that technology will ultimately take. It’s inevitably going to fall far short in certain aspects (we don’t have Rosie the Robot Maid from the Jetsons)—but also going to be (literally) unimaginably more advanced in certain others. Especially since facilitating technologies in the meantime make a huge difference.

Webvan might not have been viable back in the 2000s, but broadband, fast cell networks, and smartphones made things very different for Uber and Instacart a decade or two later.

As any high school stats student should have learned, extrapolating without data is just stupid. That’s why “calling the top” of a tech bubble isn’t really an analytical exercise, unless you can literally predict the future.

So is it a bubble?

If it isn’t already, it will be (again, my constantly called-back-to piece on the AI bubble). Tech always turns into a bubble eventually. But that doesn’t obviate its societal/economic impact—and it also doesn’t make it easy to call when we’ve jumped from exuberance to irrational exuberance, except in hindsight.

Thanks for reading!

As mentioned, through Dec 1, the Kindle edition of my book What You Need to Know About AI is just $5 (regularly $12.99). If you enjoy Weighty Thoughts, the book is essentially the structured, front-to-back version of how I think about AI and where things may be headed.

LinkedIn co-founder Reid Hoffman called it “accessible to neophytes, but broad enough to offer useful insights even to AI professionals who often end up narrowly focused.”

If you’d rather read a physical copy (or want a giftable version) several bookstores carry signed editions, some with the coveted shiny gold bookmarks. You can find the list of participating shops here.

Whether on the screen or on paper, thank you, as always, for reading.

Josh Blanchfield

Chief Investment Officer at Avos Capital / ex-Bridgewater

There is a drumbeat of vocal analysts and pundits who don’t believe that we will be able to power our AI ambitions.

The latest example was from an Apollo executive who said, “The gap between what AI is demanding and what we have everywhere in the world on the grid in terms of generation and transmission is huge and will not be closed in our lifetime.”

A lifetime is a long time! As it so happens, this is a question we’ve been studying for a while now, and the more we dig in, the more we think the doubters will be proved very wrong.

Before getting into anything AI-specific, it’s worth stepping back and considering humanity’s track record when functionally infinite money is thrown at a problem.

When measured apples to apples, the amount of money spent on powering AI will dwarf the amount of money it took to figure out how to put a man on the moon with “computers” that were closer to abacuses than to the computer in my dishwasher. We ran highways and railroads across the US for much, much less. The Manhattan Project was a drop in the bucket by comparison. We developed a vaccine for COVID in record time, and the cost could have been covered by the loose change in Sam Altman’s couch.

To solve the AI power problem, we need some combination of:

1. Producing more of a thing—electricity—that we are already good at producing.

2. Improving the energy efficiency of technology—a thing we are also very good at.

To think this is the problem that will ultimately thwart our species fails the smell test.

Share

But let’s get specific. Here is a chart where a bunch of smart people sharpen their pencils and project how much juice all of these AI data centers will need. The estimates are comically different from each other: a ~5x range over the next 5 years.

For the sake of this analysis, we will pick the highest line from the chart: BCG’s “high” estimate at ~1000 terawatt-hours (TWh). We’ll use it because it is the most difficult bar to clear, even though I will eat my hat if the real number ends up anywhere close to that.

Current data center demand is about 200 TWh, so they think it’ll add an incremental 800 TWh. Current US total electricity demand is about 4200 TWh. So if you do the math, they are projecting an electricity demand growth rate of about 3.5% over the next 5 years.

That is high relative to recent history, but in a longer context it is completely normal—in fact, since 1950, US electricity demand has grown at an average of 3.2% per year.

Where can we find 800 TWh? To start, let’s look at where electricity comes from now: the chart below from the EIA is labeled as “billion kilowatt-hours” instead of “terawatt-hours,” but those units are the same thing.

Some highlights from this chart:

• Natural gas is now our largest source of electricity and has grown at ~3.5% per year for the last decade. If it continues at that pace, it will add about 300 TWh on its own, and that pace could certainly be dialed up. For all intents and purposes we have infinite natural gas, and it is super cheap. We think this will be the largest contributor to the solution.
  

• Nuclear is too slow to move the needle. I wrote a piece in March titled “Calling Homer Simpson” that outlines why... you can find it here if you are interested: https://www.avos.co/research
  

• Renewables have some issues as it relates to intermittency (that are tractable) but could easily provide most or all of the 800 TWh. As a point of comparison, China added 1400 TWh of renewables over the last 5 years.
  

• Coal is an interesting wild card. It has gone from about 2000 TWh at the peak to about 700 now. Of that rough 1300 TWh reduction, we could get almost all of it back on a timeline that works. About half of the reduction came from just lowering the output of still operational plants, most of which could be dialed back up. The other half came from plant retirements.
  
  How long it would take to unretire those plants is largely a function of political will. At the extreme, if Trump signed an executive order waiving permitting/studies/paperwork to do this (and it stood up in court), you are talking about a couple of years, maybe five years tops. Fast enough for the problem at hand.
  
  As an aside, for years we’ve been hearing from ivory-tower types that unemployed coal miners should be re-educated to learn how to code. Maybe the universe has a dark sense of humor, and AI will take the coding jobs, and then coders will have to be re-educated to learn how to mine coal!

  Share

When you put it all together—even if you assume that the BCG high estimate is completely correct and the army of PhDs who are trying to lower the energy cost of AI are toiling away fruitlessly—we could probably cover the incremental demand by 2x if we went all hands on deck. And spending trillions of dollars tends to get a lot of hands on the deck.

Now... saying we can do this doesn’t mean there won’t be big hiccups. It will be incredibly hard and will require little to no wavering in financial resources and political will.

There are a zillion articles out there about the literal bottlenecks in materials and components, and they aren’t wrong—it is basically guaranteed that this won’t be a smooth process. Some part of the supply chain will get snarled, some protests will pop up and complicate permitting, there will be a blackout in some mid-sized city and associated outrage, etc., etc. But when we compare those issues to the financial rewards for clever humans to solve them, we see a footnote in history, not the main story.

Thanks for reading!

I hope you enjoyed Josh’s post. If you’d like to follow more of his work, subscribe to his new Substack and check out his firm’s research at Avos Capital Management.

Josh Blanchfield

Chief Investment Officer at Avos Capital / ex-Bridgewater

Straddling the worlds of AI researchers/analysts and “normies” is odd sometimes. That perspective and bridging the two worlds is the premise for why I wrote my recent book. Even so, it’s surprising to me when something is so obvious to those following AI—such that it doesn’t even need to be said—but is either totally unknown or is completely misunderstood in the broader population.

This, of course, is the curse of knowledge, where important things are totally left unsaid... because everyone already knows, right? This leaves a lot of the public with a lot of misconceptions.

In 2022, the Center for Emerging Technology and Security (CSET) projected that by 2030-2040, AI model costs would exceed total US GDP. A few months ago people were freaking out about AI consuming all of the world’s water. A few days ago, Apollo provocatively stated that the AI energy gap “will not be closed in our lifetime.”

This, of course, makes total sense, because AI is just consuming everything and becoming infinitely bigger… right?

If you project growth from the beginning of any exponential curve, it inevitably looks bonkers in a relatively short period of time.

At the beginning of the deep learning and LLM boom, especially 2020-2023, it sure looked like we were going to be on the wild ride projected out by CSET.

That turned out to not be the case—so much so that it’s actually extremely hard to find a good chart of model parameter sizes. Part of that was because since 2023, we have Mixture of Experts (MoE) models—which many major models now utilize—that make apples-to-apples comparisons hard. That means that traditional “dense” models vs. MoE models, which only activate certain portions of their parameters for certain problems, are hard to directly compare. Additionally, many model providers, like OpenAI, explicitly stopped publishing the sizes of their models.

Even with all of these caveats, it’s fairly clear that parameter counts have leveled off in the past two years based on publicly disclosed (and estimated) data. But why?

Why They Stopped Growing (As Fast, For Now)

Simply said: performance and cost.

Just scaling up parameters doesn’t directly give us greater performance. As folks like have pointed out, performance of models has more recently been driven by post-training. It’s not the same as “fine-tuning” (which is something different), but it is similarly taking models and applying various techniques to improve them instead of just making them bigger/consume more data.

That has helped keep the cost of models high (because post-training can be expensive and is often actually human-labor intensive, like with RLHF—reinforcement learning with human feedback, which literally feeds in human feedback to models). Additionally, post-training can also involve a lot of compute, so a “compute curve” or cost curve might not see leveling off in the same way.

Subscribe now

Nevertheless, high cost or not, that’s still a form of improvement that isn’t just “make it bigger” by raw model size. We have clearly seen models plateau quite a bit in performance… relative to model size. Again, this doesn’t mean we’ve “hit a wall” in performance. It just means that improving it isn’t as easy as naively scaling everything up. In some cases, it might even go the opposite direction, as covered with a 7M parameter model (absolutely petite versus billions or even a trillion parameters) beating some leading models. Clever architecture/models, better post-training, and more experimentation give a bigger bang for the buck than just literal “bigger is better.”

(A separate piece I’ll likely write later is why “all AI will be agentic”—part of “scaling” performance will simply be doing the work to make AI more practically applicable to productivity. As in, make AI do stuff directly, instead of just chat. That’s already happening with the “agentic AI” boom, which will likely get dropped to just be “AI” soon. In any case, there are many axes for improving performance that aren’t pure size and thus don’t have the same implications for resource consumption.)

The other, related reason is that as “pure scaling” has gotten less effective, it’s made it less and less economically viable to approach things that way. This is because models get more expensive to serve (inference) as they get bigger. In other words, a bigger model = a (generally) higher cost of inference.

(Obviously, training costs also increase—and by a lot—but inference costs scale faster with usage, because they apply to every chat/query.)

This is partly why GPT-5 for OpenAI has a built-in router that decides which model it wants to send queries to. Even including that intermediate layer, OpenAI can save on costs by sending easy queries to a smaller model.

As such, why bother scaling up a model when training costs exponentially balloon, inference costs are higher (even if they’re dropping rapidly with improving hardware), and you don’t really get many benefits from it?

You wouldn’t. And, mostly, model companies haven’t.

This won’t necessarily hold forever—as we get improved technology with interconnects, faster/better memory, and other things that make training cheaper, the economics could become reasonable again.

Basically, even if we do have fairly marginal performance gains, the cost curve could bend down enough later that it’s worth it just to make models bigger again. Years ago, after all, having 1 MB of system memory was huge (and extremely expensive). Nowadays, you browsing this webpage probably consumes more than that.

Subscribe now

Mistaken Assumptions

Those of you closely following AI likely have been fairly bored by this piece so far. After all, all of this is so obvious… and has been so for years, which is an eternity in AI development.

AI isn’t scaling the way normies think it is. Additionally, however, the power grid has a lot more ability to scale than most AI experts think.

After all, similar to how most people aren’t AI experts so don’t understand these “obvious truths” about AI scaling, AI experts aren’t experts on the power grid… despite everyone moonlighting as one now.

The Real World is S-Curves

A base assumption that one should take in the real world is exponential curves never continue.

The world’s population growth leveled off. Carbon emissions did not grow like we feared they would in the early aughts. Even “peak oil,” where oil prices would infinitely skyrocket (which all the headlines blared in my youth), was totally wrong.

This is related to my piece on shortages. Ecosystems in the real world adapt. All exponential curves are actually S-curves—you just can’t necessarily see it in the exponential portion of it.

And long before we hit literal limits, ecosystems adapt. Resources get scarcer (in ecology). Price mechanisms (in markets) also send adaptation signals.

In any case, because this is an S-curve, people who are projecting out infinite growth (or infinite earnings in the market) are probably wrong. This also applies to the power situation—are people really thinking correctly about it? Are we really not going to be able to meet AI power demand in our lifetime, like Apollo says?

We should assume no—even just by history.

In fact, all of the apocalyptic articles about AI sending power prices skyrocketing already… don’t actually reflect the data. Inflation has risen, so everything has gotten more expensive, faster—but we haven’t seen power, in particular, skyrocket.

Less breathless and data-free takes generally say it will skyrocket (even if it hasn’t yet)… But will it, actually?

Not to leave this on a cliffhanger, but there’s pretty good evidence, beyond just S-curves, that it won’t be that bad. We’ll have a special guest post a bit later on this topic—from an energy expert. For now, this is set up for why there’s a misconception from the general population side about AI. The next post will talk about why there’s a misconception from the general population (and AI experts) about power.

Thanks for reading!

I hope you enjoyed this article. If you’d like to learn more about AI’s past, present, and future in an easy to understand way, I’ve published a book titled What You Need to Know About AI.

You can learn more and order it on Amazon, Barnes and Noble, or Bookshop (indie booksellers).

A little over a year ago, I decided to write What You Need to Know About AI. And today, I’m proud to announce that it’s officially hit bookshelves (digital and physical).

If you haven’t already picked up a copy, they’re available on all major platforms.

For those who already did or did just now—you’re amazing. Please, please leave a review on Amazon when you have a chance. It really makes a difference in how much exposure it gets to readers on the platform.

📘 Order on Amazon | Barnes & Noble | Bookshop (indie booksellers)

When I started, I knew I had to move quickly. It wasn’t because I was eager to rush it out. It was because the world was changing fast. Even while writing early drafts, much of what I suggested was going to happen became things that already happened. Traditional publishing often takes years, but the questions people have about AI are happening now. Critical conversations are happening now. Fear, hope, hype, and misinformation are all colliding in the public sphere now.

A book isn’t just a really long Substack post. It needs to be a cohesive entity. To bring it together took rewrite after rewrite after rewrite—enough so that I wrote the total length of the book many times over. Over that time, my arguments and thinking became far clearer. I hope, even as it refined my own thoughts, it’ll help readers understand what truly matters in a confusing time.

I’m deeply grateful to everyone who’s supported this along the way—readers of this newsletter, friends who encouraged me to keep going, and everyone who shared the book, pre-ordered copies, or sent encouragement when I needed it most.

Early praise

The response so far has been incredible. Technology leaders like Reid Hoffman said this about the book:

“Understanding AI is a daunting task that becomes more difficult with the advances each passing day. This book offers a great starting point for those worried they’ll never catch up. James covers the essential history, principles, and implications of AI in a way that is accessible to neophytes, but broad enough to offer useful insights even to AI professionals who often end up narrowly focused.”

- Reid Hoffman, Cofounder of LinkedIn, Partner at Greylock Partners, and Coauthor of New York Times Bestsellers The Start-up of You and The Alliance

Some of my friends here on Substack—who are all deeply talented writers—had kind and thoughtful things to say about my book as well:

“This is the book I will give to everyone in my life to help them get to grips with the future.”

- , Founder of ChinaTalk and Fellow at the Center for a New American Security (CNAS)

“James is one of the clearest thinkers in AI. This book is an encapsulation of everything that’s excellent about his work—deep systems-level thinking delivered through deceptively clear prose. What You Need to Know About AI doesn’t just tell you about AI; it gives you simple, powerful ways to understand its real impact on your job and the world. This is the most essential, grounded, and genuinely useful guide you will find on the most important technology of our time.”

- , Head of AI at Iqidis, writer of Substack bestselling newsletter Artificial Intelligence Made Simple

“James is a polymath bridging AI history, cutting-edge algorithms, and the AI revolution with philosophical depth and venture capital savvy.”

- , founder of WellAI, writer of Substack best-selling newsletter AI Health Uncut, and author of Advancing AI in Healthcare: A Comprehensive Review of Best Practices

“Really enjoyed this book. James always asks penetrating questions and that translates into an excellent synthesis of the pitfalls and promises of the AI revolution.”

- , Founder and CEO of Rose AI, former head of commodities at Bridgewater Associates, and writer of Substack Campbell Ramble

Media coverage

I was also thrilled to see Bay City Books, a Bay Area publication that follows book news, highlight my book this month amongst other new releases by local Bay Area authors. Many of the other featured writers are with the big publishing houses—versus the small one I was working with—so it feels extra validating to be in such great company.

of Cash and Carried here on Substack also did a great review of my book as well!

Cash and Carried

A Review of James Wang's "What You Need to Know About AI"

Today’s post is an experimental one: reviewing a fund manager’s published work…

Read more

6 months ago · 4 likes · 2 comments · Shahrukh Khan

Campus talks

A few weeks ago I spoke at Northwestern’s Kellogg School of Management and USC’s VC Academy/Spark SC and had a great conversation with students at both schools.

In the coming weeks, I’ll be doing talks at universities and other organizations (watch this space!) and sharing a few reflections from behind the scenes.

Thank you again

If you’ve already picked up a copy, thank you again.

If not, you can find it here:

📘 Order on Amazon | Barnes & Noble | Bookshop (indie booksellers)

This is just the start of the conversation. For those of you who pick it up, I hope you enjoy the book.

— James

P.S. In the opening days of release, we’re now #1 on Computing Industry History and Generative AI!

Let’s look at what happens when supply is restricted—whether through chip export controls or rare earth embargoes

On October 9th, China massively escalated the trade war with a package of rare earth export controls with extraterritorial jurisdiction. If Chinese-produced rare earths account for 0.1% or more of the value of the exported item (or are integrated in specified controlled components), a Chinese export license is required. Essentially any dual-use or military applications will likely be prohibited from export. and at ChinaTalk did a good podcast/transcript on it here.

Across history, every genuine shortage has eventually birthed its solution—through substitution, innovation, or reallocation.

The Longer-Term Impact (if it persists)

Most humans are terrible at predicting the future. The reason is only partly because we aren’t good at extrapolating curves—though we do have trouble understanding exponentials; see the parable of the king, the inventor of chess, and grains of rice.

The deeper flaw is that we assume the world's state is, more or less, fixed. This makes sense for an early hunter-gatherer on the African savannah. It does not in our complex, technologically driven modern world.

Thomas Malthus predicted in 1800 that the world's population would soon hit a wall, as our exponential population growth would outpace our linear ability to feed it.

That was, famously, quite wrong.

His observations and data were broadly correct. Why was he wrong? The answer is technology. The Industrial and Green Revolutions dramatically increased our ability to produce both goods and food for over eight billion people today.

The Framework for How it Always Works

A resource becomes scarce. The price skyrockets. And that price signal causes a few things.

1. It begins by restricting demand, as the price gets high

2. Then, new sources of the resource come online (made viable by higher prices)

3. Finally, substitutes emerge or technology develops to create new supply

In the 1800s, whale oil prices skyrocketed as whale populations collapsed and demand for oil lamps continued to rise. By the mid-century, kerosene and petroleum were developed and largely substituted—causing the volume of whale oil utilized to collapse.

In the 1970s, OPEC oil embargo caused soaring oil prices. This helped spur on development of alternate, renewable energy—leading eventually to the capable clean solar and wind technologies we have today (the technological substitute). But we also have technological alternate supply.

The ability of more fields to come online from currently non-economical sources constrains high oil prices. Shale oil sands in Canada, for example, are a well-known resource close to the US, and there are many such sources available. Even if the Middle East erupts once more, it doesn't necessarily mean that oil prices will surge indefinitely.

The same dynamic plays out in the digital world. Compute, too, behaves like a scarce resource whose cost drives innovation. As such, we will not, in fact, devote all of the world’s GDP to running AI.

Compute has always come down in cost, but especially as the AI boom (eventually) leaves its wild “just-dump-money-in” phase, we will find more and more efficient ways of running AI. Inference costs have already been plummeting. And to bet against compute costs falling… well, if you did it historically, you’d have lost a lot of money.

These weren’t necessarily easy achievements. A lot of solar coming online was from massive Chinese buildout. Plummeting compute costs were in part due to incredible technical advances in semiconductor manufacturing—and multi-billion-dollar leading-edge fabs.

But with price signals from restricted supply and continued demand… the market and human ingenuity respond.

Subscribe now

Chip and Rare Earth Controls

I’m on the opposite side of the chip export control debate from several smart people I’ve met IRL like and (at least I know they aren’t online bots!). I also listen to their podcasts, so I regularly hear their dulcet voices making the arguments while picking up and dropping off my daughter from school.

While it does give me pause—and makes me think harder about it—I’m still fairly confident from a macro perspective that forcing Nvidia/CUDA scarcity will create innovation in the Chinese ecosystem. continues to do a great job covering the topic and adding depth/analysis to the broad “gestalt” I get from speaking to Chinese counterparts.

My more specific views on how US export chip controls are effectively reverse industrial policy are here.

Note: Also, to be fair, the chip controls have effectively been so leaky that they hardly apply. I also think they likely wouldn’t totally disagree with what I would have thought was better policy if perfectly applied: complete chip non-restriction and total chip-making equipment restriction.

Which brings me to the note that helped spur me to write this piece:

I do actually think what’s good for the goose is good for the gander.

It’s extraordinarily painful in the short term if China squeezes on rare earths/critical minerals.

But I do think it’s ultimately doomed. They are required in too many products. If China fully cuts them off—which might not be true depending on how leaky theirs is—I can guarantee you that many products will immediately be facing redesigns (restricting demand). Yes, of course, many kinds of products will also get scarcer/get pricier.

However, we’ll likely see more supply come online. Rare earths are not actually rare, and a lot of the difficulty comes from refining (as covered well here). Of course, you can’t spin up a new mine overnight. Refining capacity is also still largely concentrated in China, despite all of the handwringing.

But that’s where technology comes in. I’ve watched/seen multiple deep tech startups operating in the rare earth/critical minerals space. I literally met a promising new one last Wednesday. A lot of these companies have failed to gain much traction while facing the existing, scaled industrial capacity of this commodity industry.

That obviously changes if costs skyrocket and real shortages appear.

I was literally talking to yesterday, who was wondering why a mining startup he saw at a pitch event during San Francisco Tech Week was so popular. This is, at least in part, why—and I expect all of these startups will likely be fielding a lot more investor calls if this continues.

In the short term, restrictions and shortages are painful. In the medium term, human ingenuity finds a way. And in the long term, those shortages are the seedbed of opportunity for new technologies.

Thanks for reading!

I hope you enjoyed this article. If you’d like to learn more about AI’s past, present, and future in an easy to understand way, I’m working on a book titled What You Need to Know About AI that will be published October 15th.

You can learn more about it and pre-order on Amazon, Barnes and Noble, or Bookshop (indie booksellers).

We all live in bubbles—at least to some degree. Most subscribers here accept that AI is going to be important and useful. In fact, I spend much of my time on this Substack pointing out the limits of AI.

It’s rare that I’m in circles where the prevailing opinion is that AI is a mere bubble and it’ll turn out to be largely useless. Between some seminars, dinners, and parts of the book tour (keep reading for those announcements!), I did that much more these few weeks.

I don’t believe AGI is around the corner. I would say AI reasoning still has hard limits around its learned dataset. And I still think most AI startups are doomed.

That being said, I think it’s crazy at this juncture to say that AI isn’t going to be extraordinarily impactful. That won’t necessarily mean that AI startups—or even public markets—ultimately justify their lofty valuations, but that’s a different statement than “it’s useless.”

Public stats aren’t really helpful for this. Explaining how much revenue OpenAI is making ($13 billion expected for 2025), how many tokens Google Gemini is generating (all the data in Google Books every four hours), or citing adoption stats for AI in industry and consumer sectors is not convincing if someone thinks it’s all just hype and a bubble. “That’s why there’s so much money flowing around!”

And maybe it’s impossible to convince some of those people that it’s useful. For them, the idea that AI is just a fad is as much a religious belief as AGI-is-just-around-the-corner is for some on the other side.

Still, for those who aren’t totally entrenched, I’ve found a gentler, more intuitive description for how AI can be useful is far more impactful. This is a little bit less “macro stats” and more anecdotal, but hopefully you’ll find it helpful for walking skeptics through how it can be useful (or maybe help you understand, if you’re a skeptic yourself) or get ideas as to how you can use it in your own life.

Subscribe now

One of the simplest use cases—a more useful search

Few would dispute that Google Search is useful. As of today, it’s worth over $3 trillion in market capitalization—largely, even now, on the back of their massive advertising business built on search. But even from a more intuitive perspective, all of the world’s knowledge at your fingertips is incredible. No gatekeepers, no subscription or payment required, and near-instant—no need to page through volumes of dusty encyclopedias.

Even if you dismiss everything else about AI (and in this case, LLMs), isn’t a conversational version of search useful? People who aren’t tech savvy can talk to an LLM more naturally and get the same results, and even despite how far Google has come, LLMs like ChatGPT are able to take vaguer inputs and make something out of them. Early on, I found ChatGPT extremely useful for “tip of my tongue” quotes or movies that I couldn’t quite remember.

For example, I remember a fascinating pre-internet era quote about people losing their ability to understand things—and people instead just take in sound bites or trivial information. Basically, a similar sentiment as today, which is a powerful statement about how we always worry about media, technology, and culture decaying our ability to think.

Trying combinations of science influencer, loss of knowledge, attention, 1960s, 1970s, quotes… None of that brought up what I was thinking of. A short back-and-forth with ChatGPT gave me Carl Sagan—the person I was thinking of—and his quote in The Demon-Haunted World in 1995:

The dumbing down of America is most evident in the slow decay of substantive content in the enormously influential media, the 30 second sound bites (now down to 10 seconds or less), lowest common denominator programming, credulous presentations on pseudoscience and superstition, but especially a kind of celebration of ignorance”

It even helpfully pointed out:

🔍 One clarification: Sagan’s major public science influence began in the 1960s–70s (with Cosmos and frequent media appearances), but the specific famous quotes about attention/knowledge erosion come later, especially in the 1990s.

Which, of course, was one of the reasons I thought it was from that era (wrongly), which likely also messed up Google’s more literal search.

In this formulation, the normal pushback that “LLMs will have hallucinations” is largely irrelevant. Google search, with its sourcing from random blogs, websites, or social media, is obviously not deeply researched knowledge that is guaranteed to be right. Teachers chide students all the time not to use Wikipedia as a source. “Anyone can edit it!” My publisher (and most publishers) also banned it for my book as a source for the same reason. But anyone who’s ever used Wikipedia also knows how amazingly useful it is—and most use cases of it are not “high stakes” enough for the possibility that it’s wrong to matter.

In the worst case, it can be similar to using Wikipedia for academic research or published articles/books. One can use it as a quick reference and then do more research to confirm some of what you read about on Wikipedia.

This, by the way, is why AI threatens Google’s search dominance. It has already significantly changed people’s everyday habits on how they find and consume information. What’s Google going to do without a couple of blue links on a page (where they can conveniently stuff in ads)? That’s still an open question.

Subscribe now

Less common use cases—but ways I use it in my daily life

Programming on the Go

Coding is one of the most common “killer apps” for AI. Agentic coding assistants like Claude Code (Anthropic’s offering), Codex (ChatGPT’s offering), and Cursor (which packages up models in an integrated development environment, also known as an IDE) are just a few names in the space.

I prefer Claude Code and Codex (more so the latter recently), which run natively on the command line. I’m old-fashioned—all of my code editors are from the 1970s-1980s. In addition to allowing me to cosplay as a grumpy old engineer, another benefit of using this kind of "low resource" coding platform without any fancy, modern IDE is that I can log in to those interfaces remotely, even on my phone.

I explain the exact tools I use in the caption, but this essentially means I can "code” from anywhere, even with a tiny screen and thumb-typing. I can ship code from public transit, while on the way to a meeting, and even during lulls in a meal.

This paradigm is convenient for me. It’s a godsend for those with physical disabilities or limitations. I’ve had coworkers with severe arthritis or other hand/wrist issues that made coding debilitating. For them, the ability to more naturally dictate or, at minimum, type fewer words, significantly enhances their quality of life.

You still need to manage these agents, though. In a way, using them feels like working with a team of junior programmers to code. You send off instructions, and sometimes you get the right thing, and sometimes they find remarkably creative ways to mess everything up.

One example: while building a tricky AI model, Claude decided to “pass the test” by adding a hardcoded +0.2—boosting the accuracy by 20% by just… fudging the test result to be 20% higher. I’m definitely not alone in seeing this kind of behavior.

As with most AI applications, you can’t leave it alone and unsupervised. It’ll likely do insane things if you just accept its work unquestioniningly. However, it is an amazing executor.

Assistant Work, Also On-The-Go

I use a Mac program called Drafts. On paper, it’s quite simple. It’s a quick note-capture app where it always opens to a blank entry and you quickly jot down whatever you want using plain text.

However, that’s not all it can do. Its power comes in its incredible extensibility with “actions.” You can script actions to do all sorts of things. I have, for example, a dialog for correctly formatting and entering public market trades that I do (so I can track my performance).

One really simple extension that’s available on Drafts’ marketplace is to call ChatGPT on your text.

In the above example, I dictate a note using Apple’s native voice-to-text, which is largely unreadable for me. It’s a giant block of text where iOS creatively found ways to misinterpret my words.

I can use the Action to call ChatGPT’s API, allowing it to clean up the entire text and convert it into intelligible bullet points. Words that Apple mangled are usually cleanly mapped into what they actually were based on context.

This type of task is the kind of thing I’ve used an assistant or secretary for in the past when I’ve had one available. That works, but I loathe making humans do this kind of busywork, even if that’s technically their job. I also dislike it because I tend to be quite specific about the tasks I want completed.

In this case, I can be as annoying and prescriptive as I like—and send ChatGPT to redo it as many times as I like.

I’ve also asked it to fill-in-the-blank for shopping lists that I have for recipes. “This is a shopping list for a rhubarb pie. Please add any common ingredients that I might have missed for it to the list.” You can imagine asking it to clean up a thank-you note. Or take an overly long email and ask it to summarize it.

All of these are completely normal tasks that one might input into ChatGPT’s (or Gemini’s, Grok’s, or Claude’s) chat interfaces; however, this process is seamless. It’s called from where my notes currently reside—no copy-paste dance required. I expect similar services will be available (… in satisfactory form, not in some of the questionable Apple Intelligence services) for people with Apple Notes, Reminders, Microsoft OneNote, or whatever program of choice they use.

It’s cheap too. All of my API calls have probably added up to around $3.

Writing and Research Aid

Inspired by ’s piece on using AI agents to help automate writing, I’ve been experimenting with my own workflow here. I made my own version of it using Claude Code (which I have a GitHub repo for here). Claude Code can modify text without me copying and pasting—and has a freer hand than Codex generally does to pull resources from the internet and download them.

The cardinal rule I’ve given it is: “I come up with the direction/ideas in my notes and preserve my phrasing and idioms. However, please help me organize it into a coherent outline, research sources, and eventually edit and proofread it.”

So far… I haven’t found it particularly useful. The editing/proofreading isn’t awful. Using CriticMarkup (an open-source and plain text method of “track changes”) is nice and I’ll probably keep using that. The research leaves a lot to be desired—I’m not getting what I really want, which are pretty charts and graphics that support my points. The outlines are uninspired, but at least they are organized. But at least it isn’t actively getting in my way, which is usually the problem with trying to use AI agents to help write.

It’s still a work-in-progress, but it’s at least progressing to a point where I might find it useful in a couple more iterations.

Bonus Research Tip

In terms of utilizing AI, I’ve found NotebookLM is generally the most “magical” to ordinary people. Recently, I’ve thoroughly enjoyed its video feature. It makes far more engaging pitches given a pile of boring information than most founders—or I—do.

However, in terms of research for writing or just keeping up with things, I’ve found the podcast feature very helpful. Most people use it to summarize research. I use it to translate podcasts or interviews in foreign languages and summarize the information. Even if you feed it a Mandarin podcast, it’ll come out with its summary podcast in English. I’ve used this for getting the gist of what’s going on from all kinds of niche, foreign language sources.

Subscribe now

The Bigger Picture

Bill Gates recently argued that AI still can’t handle the most complex coding tasks—and it might still be reasonably far away (10+ years). Why? Because programming is less about memorizing syntax and more about understanding problems, structuring solutions, and working within messy real-world constraints. LLMs can help, but humans still define the goals and priorities.

AI won’t replace humans. It will replace parts of tasks, speed up workflows, and sometimes lower the barrier to entry. But at the core, humans are the ones with the questions, the goals, the desires, and the responsibilities.

The real question is not “Will AI replace us?” but “How can we use it to do more, better?” At this point, we definitely know the answer is not that it is useless.

In my previous life, I co-founded a company that built biofeedback vibrators. I bring this up because I saw firsthand how people worried these tools would “replace” human partners (yes, let’s be frank—usually it was men who were worried about this). In reality, there was never any real danger of this. Any “tools” here were an enhancement, and even if they could function better than a human… they didn’t serve the same purpose at all. In intimacy, the human connection can’t be replaced by a tool, no matter how capable. When it comes to productivity tools, an individual must possess creativity, intent, and motive.

Will this all still apply when we get conscious, human-like AI? Maybe not. In intimacy, maybe we’ll get HER or The Companion. In productivity, well, we’d be basically imagining I, Robot and artificial humans. But we’re a long way from those scenarios!

Thanks for reading!

I hope you enjoyed this article. If you’d like to learn more about AI’s past, present, and future in an easy to understand way, I’m working on a book titled What You Need to Know About AI that will be published October 15th.

You can learn more about it and pre-order on Amazon, Barnes and Noble, or Bookshop (indie booksellers).

One of the most annoying recent hive-mind “AI is bad” memes has been water use by AI. It follows a pretty similar path as the prior hysteria cycle over power and carbon emissions, which I addressed in a previous piece here.

My boring TLDR is that AI has a minuscule impact right now relative to… pretty much all other human activities. That being said, AI does require more than “conventional” web/internet services, so we should keep an eye on it.

That’s far less of an exciting take than, “We’re all gonna die.”

has a far better rebuttal (with sexier charts) than I did on this topic—I’ll let the data speak for itself.

Since some people will just use this to say all tech services are bad, let’s also include this perspective from Andy.

This would make off-the-grid vegans happy, but most people who decry AI would be far less willing to give up meat or running water. If more people were willing to sacrifice meat, even a little, alternative protein startups would be a lot happier (and, for most, a lot less dead).

Still, just like the power conversation, it’s a good one to have. It’s also dumb to claim that AI is totally costless. Hysteria and totally one-sided absolutism help nothing, though.

Techno-Optimism and Techno-Pessimism

In much of the post-WWII period, we’ve seen hyperbolic techno-optimism. We were supposed to have flying cars and robot maids like in the Jetsons by now. Today, we have far more techno-doomerism—everything tech is bad.

I hate it. Things stop being a conversation and there’s no way to rationally weigh anything when the orientation is essentially a religious belief.

I don’t think boundless tech-optimism is good. Anything short of Iain Banks’ Culture series—an all-powerful, post-scarcity AI utopia—is unlikely to solve social problems. Past a certain level of desperate poverty for basic material needs, there’s no magical reason why faster chips, bigger TVs, or even breakthrough advances in medical technology will change the nature of humans. We’ll still covet thy neighbor’s stuff, do unhealthy things, and fall into the trap of wanting someone else to solve all our problems.

But tech also doesn’t magically “destroy everything.” It’s become hip to hate on tech companies, now that tech companies are economically ascendant. A lot of us (me included) hate where social media ended up.

But did social media create a dystopian world? Social media arose during a time when the “free world” was no longer united by a common enemy (and the threat of nuclear annihilation). Right-wing talk radio and the left-wing blogosphere both existed in the early internet days, before the advent of our current social media ecosystem.

We were already fragmenting and self-sorting into our own little worlds. The tool of social media, the internet, and always-connected mobile phones just made it more efficient.

To be clear, that’s still something. I also am not arguing that we shouldn’t regulate it. For example, perhaps we really should ban smartphones in school, as @Jonathan Haidt suggests.

But it isn’t the source of all sins. You can’t blame the Devil (or Facebook) for everything bad that humans have done.

Subscribe now

Tradeoffs Are Normal

Whether it’s AI resource use, or social media, or pretty much literally anything else, nothing is all good or all bad. Even your choice of lunch is a tradeoff. That salad may be kind of sad-looking but it’s healthier than the burrito stuffed with fries (also known as California-style).

That’s normal. Technology and productivity improvements are good. They aren’t costless. But, on balance, we want them.

The problem is that in our current social landscape, no one wants to bring up any tradeoffs. It’s politically expedient in the short term—and incredibly corrosive in the long term.

Let’s talk about examples, which unfortunately turns this into a bit of a rant, but here we are. Given that my two examples are at the center of political debate, I expect this might be unpopular. But if I didn’t choose topics with some valence, I doubt anyone would care. It also wouldn’t make my point.

Renewables and Costs

recently did a piece called Kilowatt Madness debunking Trump’s… let’s call it an argument.

From Krugman:

It’s unclear what his theory is. How does adding wind and solar generating capacity — increasing electricity supply — lead to higher prices?
…
As it happens, the data overwhelmingly reject Trump’s claims about renewables and prices. The Department of Energy has data on the share of each state’s electricity generated by renewable sources. For example, Iowa gets 80 percent of its utility-scale power from renewables, mostly wind, while New Jersey only gets 4.6 percent from renewables. Yet Iowa’s electricity prices actually fell slightly from May 2024 to May 2025, while NJ prices rose 10 percent.

Well, yes, and no. It’s actually fairly complicated—and I expect Trump got briefed on something related to the topic, which has a nugget of truth in it.

The poster child for high prices is California, which has made huge renewables pushes. But Texas is actually building out renewable generation faster—and is obviously cheaper.

On the other side of the spectrum, there’s . While it’s rather… negative… on renewables in general, it points out that there is quite a direct connection:

Last week, an expansive article in Bloomberg Green—confessionally titled “The Fix For Solar Blackouts Is Already Here”—captured this sentiment. It lamented that the penetration of solar and wind has outpaced the buildout of stabilization technologies such as synchronous condensers and grid-forming inverters. In other words, the renewables worked as designed, but the infrastructure to integrate them safely at such high percentages of supply lagged far behind…

The bottom line: renewables require substantial grid improvements to work well. The sun doesn’t always shine and the wind doesn’t always blow.

That doesn’t mean we shouldn’t use renewables. I’m also no “climate denier” (…I mean, Creative Ventures and I have invested quite heavily in climate tech). There are simply real shortcomings. And obviously Krugman knows—he has very careful language that doesn’t give an inch to Trump’s rant, while never saying anything technically wrong… but also obfuscating the fact that renewables can drive up costs.

Integration into power grids can be costly. Rooftop solar is inefficient and collectively expensive for the grid. And building out new renewables capacity to replace older, more polluting technologies is expensive.

Even taking the rooftop solar example, you wouldn’t expect your power bills to be radically cheaper (unless there’s some gimmick where someone else pays—say, government incentives, which just displaces who pays for it) right after building them. There’s a capital outlay up front. It may very well pay off—and in a big way. That also doesn’t mean it’s great from day one.

However, given how poorly preventing climate change polls (it’s always lagged the economy and other issues by a lot), most people aren’t going to give it much grace if they hear it’ll increase their power bills.

The public debate is healthy, though. Renewables are likely still “worth the cost” after weighing all the tradeoffs. Because there are others that are definitely not. I also don’t think acknowledging some nuance would really hurt the argument—especially among Krugman’s audience.

Sunlight and clear debates could prevent truly stupid policies—like California’s special CARB oil blend that makes it a gasoline island, improving emissions slightly at the cost of massively higher oil prices. Now, with refineries that produce its special blend closing, it represents 20% of the state’s supply… This could push California’s $4+ per gallon prices to $6 or $8—again, with very limited benefits.

Subscribe now

The Loss of Public Trust in Public Health

Robert F. Kennedy Jr., the administration-appointed head of the US Health and Human Services (HHS) Department, has been canceling funding for vaccine development, trying to fire CDC staff over vaccine policy, and limiting access to who can get COVID shots.

He’s also said weird things (not out of context), like being able to diagnose children by looking at them. All in all… bad stuff.

I’m worried about where the climate and renewables debate will go. I’m despondent about where the public health debate is—and fear other issues will metastasize similar to it.

RFK Jr. is the whirlwind that public health reaped. With all the politicization around the topic, I’ve seen a tendency to want to sweep the dumb things that happened around COVID under the rug. Unfortunately, I’d argue those dumb things are what led us here.

I get a lot of “that didn’t happen” responses for the stupid policies around masks. Specifically, lying to the public about their effectiveness rather than simply urging people to leave them for essential workers and trusting societal cohesion.

I remember quite well—because my wife and a friend of mine devoted countless hours to sourcing and organizing drives to get N–95 masks to essential workers. I also spent hours coding for the cause (ah, the days before LLMs).

The dialogue about the plunging trust in public health was dismissed as “disinformation”, but much of it was also missteps by health officials.

At the beginning of the crisis, red states had insane policies. Eventually, the facts changed, both due to our growing knowledge and literal conditions shifting. The big one, of course, was that vaccines became available. The politics didn’t change. With the moving landsccape, eventually blue states ended up with crazy policies. It’s not to pat either on the back—they all ended up doing irrational things, just at different times. And usually for political reasons.

Zeynep Tufekci basically built up an entire massive public profile calling out the disingenuousness of the public health community (rightly so).

California closed their schools and kept them closed for stupidly (and damagingly) long periods of time—while arguing that there were no tradeoffs involved. Different counties in the state kept mask, distancing, and hours mandates for far longer than everywhere else.

Trust is so low—even in, or perhaps especially in, California—that I have regularly had conversations with Uber drivers about vaccine skepticism. Although it’s a problem nation-wide, in California we saw vaccine hesitancy nearly double as COVID went on, especially among Black and Hispanic workers. When the topic comes up on the radio, I’ve had drivers tell me that their cousins died from the vaccine, that it was completely untested, or that it was designed without any real testing for safety for Black people. And so on and so forth. At the same time, at fancy house parties in San Francisco, people compared how many shots they got on their little vaccine cards.

Is some of that misinformation? Yeah, sure. But much of that stems from people losing trust in public health officials.

So much of this was counterproductive. People, regardless of their education level, aren’t stupid. Inconsistency and unwillingness to discuss tradeoffs often comes across as disingenuous salesmanship.

Hoping for Better Public Discourse

There was no golden age when everything was fair, balanced, and everyone talked about tradeoffs. But certainly more people were willing to do so in the past.

Coming full circle: is this the fault of social media? Smartphones? In the spirit of tradeoffs, I expect it’s likely partly to blame. It’s not all of it, though.

We have many challenges coming up. That includes actual climate change (beyond debates about renewables), massive demographic challenges (which I’ve written about regarding why we might need AI), and an increasingly geopolitically unstable world (following the end of unipolar hegemony by the US, after the Cold War).

Putting aside bad takes on AI—in either direction—it’s impossible to tackle our biggest problems without an honest accounting of what we’ll need to trade off. If we don’t, we’re doomed to dumb decisions or total paralysis.

Thanks for reading!

I hope you enjoyed this article. If you’d like to learn more about AI’s past, present, and future in an easy to understand way, I’m working on a book titled What You Need to Know About AI that will be published October 15th.

You can learn more about it and pre-order on Amazon, Barnes and Noble, or Bookshop (indie booksellers).

OpenAI just released GPT-5—to great fanfare and mixed reviews around the internet. According to benchmarks and subjective personal testing, GPT-5 is better than GPT-4 and o3.

It’s certainly a better default than GPT-4o, which is what most people used on ChatGPT’s interface. The model dominates across the board in LMArena.

I don’t feel it as much. But I also used OpenAI’s research previews of o3-mini-high, GPT-4.5, and other models for specific tasks. As such, I don’t really see it as revolutionary. That makes sense though. Today, if you try to select other models in the Plus subscription, all you get is GPT-5 and GPT-5 Thinking (the latter being the “high effort” version of the first).

The function of those research previews all got rolled into the 5-series.

Of course, it’s not identical. The tone, for one, is quite different. I didn’t notice as much, but I always took steps to reduce sycophancy. Whenever possible, I wanted it to disagree with me. Other people used it differently.

That makes me feel a little bad calling it sycophancy—even though it’s the industry term. The word implies that it’s just sucking up to people. Which, I mean… it is… But that non-judgmental cheerleading, agreement, and yes-man’ing has been helpful at least to some. Not all of us are masochistic ex-Bridgewater robots or (as one business school professor derided me as in a heated debate) Vulcans.

And this is partly why OpenAI is bringing 4o back for paid users soon, at least for now. It’s a bit surprising they didn’t expect this reaction, though. I personally wouldn’t mind still having access to some of the extremely fast mini models. Not all tasks require embedding a ton of world knowledge.

Other parts of the rollout were perplexing too—I did a double-take on some of the announcement stream’s charts. For OpenAI’s long-delayed, much-anticipated release, it all seemed quite sloppy.

Putting that aside, forward progress is continuing. Despite how it feels now, that wasn’t a foregone conclusion. It was the end of last year (eight months ago) that the media was feverishly asking whether or not AI scaling was over. It’s obviously not.

From the other extreme, though, I also think it’s reasonably clear that we aren’t suddenly going to wake up to either Judgment Day like in Terminator 2 or AI suddenly becoming superintelligent.

As usual, has an interesting take, and I like his synthesis here:

If AGI was a real goal, the main factor on progress would be raw performance. GPT-5 shows that AI is on a somewhat more traditional technological path, where there isn’t one key factor, it is a mix of performance, price, product, and everything in between.

Given all of this, where does this leave our vision of the future?

Peering into the Crystal Ball

My upcoming book, What You Need to Know About AI, will be coming out in October (!!), and I had to ponder some of these questions. The book isn’t meant to speculate about the future. It’s a “survey class” of AI’s history, technical underpinnings, business fundamentals, and how it’s being used today.

Still, some of my concluding chapters grapple with its impact on certain jobs and the nature of creativity (and “humanness”). It’s hard to avoid it and a bit of a disservice to readers if I just dropped them off abruptly without addressing why people care so much about AI to have read an entire book on the topic.

These aren’t directly from the book but more just thoughts based on what I’ve seen in online conversations about AI. These are my own opinions and synthesis—which is all it can be, since no one knows the true answer to these yet.

Subscribe now

1. We aren’t going to be become drooling degenerates hooked on AI

Calculators were once supposed to make students dumb, because no one would be able to do math anymore. I still remember when I was in high school that teachers tried to police the use of the TI-89 (originally released in 1998, but there was a re-release of the “Titanium” version in 2004). It could solve algebra problems and even store programs/macros. The idea was this overly powerful calculator—if allowed in classrooms—was going to make our kids completely unable to learn math.

That’s almost laughable, of course, in today’s world with Wolfram Alpha and, you know, ChatGPT.

Each generation of technology has generated this kind of moral panic. The most recent one is that AI will turn us into mental degenerates who can’t flex our brains at all without consulting an LLM.

Is it useful to be able to do fast mental math? As a former hedge fund person and now as someone in charge of a lot of workstreams where I need to at least be able to know if certain stats and figures are roughly right… yeah.

It’s also great as a parlor trick too. Impress your friends by knowing the answer to some calculation by just sitting there while they look it up. I personally suggest Secrets of Mental Math, which was super useful to me when I was younger.

However, do I think someone is truly stunted by pulling up the calculator app on a smartphone? No.

Similarly, is there a fundamental value in using traditional Google search over ChatGPT? Was there one in using an encyclopedia over Google? Was there one in memorizing oral traditions over the written word?

We’ve used tools to make our lives easier throughout all of human history. Hell, even LLMs themselves use tools nowadays. There’s nothing brain-melting about using AI to now look up, calculate, summarize, or otherwise do things that require relatively little mental rigor—but do take time.

Also, having been around a lot of successful people of means (read: rich), many of them use assistants, leveragers, or other kinds of people around them to do things for them. It didn’t stunt their intellectual ability. If anything, it usually freed them up to do what they’re really good at. Now, far more people who can’t hire armies of assistants have the ability to capture a lot of the same kind of value.

2. The kids will be all right… eventually

I’ve had interesting debates with a lot of smart people about the future of work—specifically, the future of work for young people today.

I started out being pretty dismissive about AI being a threat to young people’s jobs. The economy has adapted to all sorts of things—and young people, at the beginning of their careers, have plenty of fluidity to adapt.

But, over time, as I noticed that AI was massively more valuable for experienced experts, heard multiple entrepreneurs say they simply didn’t hire for far longer, and noticed myself that I could get a ton of stuff done with LLMs for coding, data entry, and similar tasks that I would have otherwise gotten interns to do… I changed my mind.

The “missing link” for LLMs and similar deep learning algorithms is that they have no inherent motive drive on their own and don’t have “creativity.” That’s a fraught topic about what that even is—but deep learning algorithms inherently learn from distributions and draw from them.

Debating with those who say that there is no true originality in the world and everything is fundamentally derivative (there is no original art, novels, etc.) is like debating with someone about the world being a simulation. It doesn’t seem right, it gets tiring quickly, and ultimately it’s more of a religious stance that means we’ll have to agree to disagree.

Regardless, I’ll assert there is some level of creativity. These models will always have trouble exhibiting it by their very nature. As per here:

I’m also intrigued by how many people are surprised by the fact that models can’t perform outside of their dataset when by definition this is how deep learning works. The real way to get them to generalize further is by finding unique ways to expose to more data.

Do you know what general population often has relatively little independence/motive drive and (useful) creativity? New graduates and entry-level folks. There are exceptions, of course. One of the biggest pushbacks I’ve had are from tech founders of currently tiny companies who talk about how incredible the young people they hired are. I’ve had remarkable interns myself who are better than experienced hires. That’s rare, though, and a cherry-picking of the 0.001%.

Generally, young humans often need a bit more experience and time to really take off. What happens if they don’t get those opportunities, though?

This is happening a lot in software right now.

Addy Osmani wrote a piece some time back called, “AI Won't Kill Junior Devs - But Your Hiring Strategy Might” arguing that junior engineers should now learn higher-level skills and warns that:

Despite doomsaying that AI will kill off entry-level programming jobs, the need for junior engineers isn't going away. Industry leaders have realized that completely bypassing the junior level is shortsighted – after all, every senior engineer once started as a junior. As Camille Fournier bluntly put it, many tech managers who shifted to "senior-only" hiring are asking for trouble: "How do people ever become 'senior engineers' if they don't start out as junior ones?"

I agree, but who will shoulder the cost of that? There’s a bit of a prisoner’s dilemma here, where maybe the industry is better off if we all train new juniors… but everyone gets to take advantage of hiring the trained-up juniors after some other company spends the money to train them.

Big companies like Google may have reason to do so (where Addy works) but few others do—and even Google has to care about its bottom line.

I do think the industry and young people will adapt. It used to be the case that architecture as a discipline had entry-level apprentices largely do drafting. Not glamorous, but needed. Eventually, CAD made a lot of that need go away—and specialized drafters are now experienced technicians who are very fast at using those tools.

Junior architects still do some CAD or BIM modeling today. There is also a ton of other grunt work to do—especially in the regulated construction industry. I’m not quite sure what a junior developer would do that’s the equivalent. Fill out forms for submitting an app to the Google Play Store? However, the industry did adapt, and I expect we’ll eventually adapt to the new normal as well.

But there is likely going to be a transition period. I remember how painful it was getting jobs as a young person—where every entry-level job needed 5-10 years of experience. It’s likely even worse today.

The kids will be all right, eventually... but there might be a lot of dislocation in the meantime.

Share

3. AGI isn’t coming anytime soon

I’m not going to harp on this one as much. No matter how good LLMs get, they are still a categorically different thing than human intelligence. DeepMind and OpenAI can win more Math Olympiads, LLMs can keep knocking out successive ARC-AGI tests up to ARC-AGI-50 or whatever, but knocking over setpiece benchmarks is not the same as human-level intelligence.

As per Logan’s quote above: the real way to make a model more capable is to give it more data that is currently out of sample. We can keep doing that and make these models broader… but it still doesn’t look like human reasoning or problem-solving.

There’s fascinating philosophical thought experiments like John Searle’s Chinese Room, or frameworks like Judea Pearl’s Ladder of Casusation.

Each can be the subject of their own posts.

However, I don’t think it’s particularly controversial now that AGI isn’t lurking literally around the corner, ready to impact Alphabet or Meta’s next quarterly results.

That doesn’t mean I don’t think AGI won’t ever be reached. I think it will. I’m not saying that I think it’s unachievable in my lifetime. It might very well be within it. I’m not even saying it isn’t within ten years. A lot can happen in ten years.

We don’t have a clear way of getting there and nothing right now seems to be that promising. I would assert (being in the “deep tech industry”) that it’s in a similar place as general quantum computing. There are theories. Some are probably good theories. But there are no practical, testable solutions that gets us there today.

That means it’s very, very unlikely to suddenly burst onto the scene in three years, or even five though. And once we reach five, we’ll have to reassess again—and it could very well be another “five years and then we’ll reassess.”

The same applies for AGI.

LLMs may feel magical, but it’s important to understand their fundamental limitations. That doesn’t mean they aren’t useful. They are obviously very useful for a lot of applications. But just being useful doesn’t make one human or human-like in intelligence.

Thanks for reading!

I hope you enjoyed this article. If you’d like to learn more about AI’s past, present, and future in an easy to understand way, I’m working on a book titled What You Need to Know About AI that will be published October 15th.

You can learn more about it and pre-order on Amazon, Barnes and Noble, or Bookshop (indie booksellers).

Once upon a time, there was a small trading company in Korea that specialized in dried fish and noodles. It was called Samsung. In the decades between 1960 and 1990, the Korean government pushed this company into electronics and semiconductors.

Having no expertise in these areas, the company produced terrible products. However, the government protected it—setting tariffs on foreign competition and “encouraging” Korean businesses to use their terrible products. Critically, Korea did allow exported products to use superior foreign chips. They foisted the terrible Samsung chips onto the domestic market, which didn’t have any other choice.

Today, Samsung is obviously a global leader in electronics—and one of the two leading-edge semiconductor fabs. Notably, the other one is TSMC.

The former global leader and US champion, Intel, has in practice been out of the running—and recently said it may give up on ever regaining its place as a leading-edge semiconductor manufacturer.

The US’s Reverse Industrial Policy—Benefiting China

The recipe for building up globally competitive domestic capacity (replicated across Asia, including in Taiwan) is to restrict foreign competition, gain export competitiveness as quickly as possible, and force your uncompetitive product onto your domestic market.

It may suck for domestic consumers, but it gives your national champion the time, money (semi-captive customers), and experience (manufacturing learning curve) to become globally competitive.

Usually, this is imposed by the country trying to catch up.

Notably, the US’s chip export regime has done it for China. In a way, the US implemented policies that aimed to make Huawei and SMIC globally competitive in the semiconductor industry.

The ostensible reason for the controls was to cripple China’s AI progress. If that was the goal, it has been a failure.

Just look at the many benchmark-leading Chinese models. DeepSeek dominated the news cycle earlier this year. More recently, Moonshot’s Kimi K2 became a frontier model that, similar to DeepSeek, is on par with the best American models and also cheaper.

On LMArena’s rankings, 25% of the models on its leaderboard are Chinese.

These models, notably, tend to be open-weight and very permissively licensed—which has encouraged many companies to use them as the basis for their own models.

Marc Andreessen has raised the alarm about these models becoming the global standard in the absence of any competitive American open model.

What’s the problem? One problem is that, even without censorship from hosted versions of these models, they tend to skew towards the viewpoints favored by the Chinese government. They are likely post-trained to do so (or their datasets reflect it). As I was quoted in SCMP saying, it’s challenging to rip out those biases without impacting the underlying model.

Kids these days (and—let’s be real—most adults) already rely heavily on just asking models like ChatGPT for answers. Controlling the base model that everyone uses constitutes a significant form of information control.

This danger, among others, is the basis for efforts like ’s laudable “American DeepSeek project.”

Subscribe now

Trump’s AI Action Plan

Earlier this week, the Trump administration unveiled its AI Action Plan.

As for its goals, I’ll quote the White House’s own announcement:

• Exporting American AI: The Commerce and State Departments will partner with industry to deliver secure, full-stack AI export packages – including hardware, models, software, applications, and standards – to America’s friends and allies around the world.

• Promoting Rapid Buildout of Data Centers: Expediting and modernizing permits for data centers and semiconductor fabs, as well as creating new national initiatives to increase high-demand occupations like electricians and HVAC technicians.

• Enabling Innovation and Adoption: Removing onerous Federal regulations that hinder AI development and deployment, and seek private sector input on rules to remove.

• Upholding Free Speech in Frontier Models: Updating Federal procurement guidelines to ensure that the government only contracts with frontier large language model developers who ensure that their systems are objective and free from top-down ideological bias.

This was also paired with a recent deal for Nvidia to resume sales of their H20 GPUs to China in exchange for China loosening restrictions on rare earth metal exports.

There’s a lot of choice language in the plan itself, like “Build, Baby, Build!” It definitely has many elements to it, but I’ll highlight what I found notable:

1. A weak form of the AI moratorium on state-level regulation that was removed from the Big Beautiful Bill. It relies on withholding funding from states and agency reviews—as such, it has a lot less legal force than congressional law. But it’s still an attempt.

2. Promotion of open-source AI and exporting it—and encouraging the government and the Department of Defense to adopt it. This is quite similar to a lot of Chinese efforts in both. It’s basically fast-following Chinese policy.

3. Lots of talk about building stuff (power, data centers, etc.). The agenda includes deregulation, encouraging workforce development, and trying to clear away barriers. I’m rather skeptical of this—but we’ll see if it amounts to anything.

4. Coordinate with allies and partners. Self-explanatory. Not really something that the Trump administration has made itself known for, so if this signals a change, it’d be welcome.

5. Miscellaneous: callouts for cybersecurity, tightening export controls in semiconductor manufacturing, biosecurity, and generating national datasets. This is a bit of a grab bag of miscellaneous stuff. The datasets effort, again, is a bit of a mirror to existing Chinese policy to cultivate and unify disparate datasets into “national datasets.” The “tightening export controls” aspect of this is focused on the right thing—semiconductor manufacturing systems and subsystems and not chips.

Combined with the Nvidia and rare earth deal, the US seems to be trying to directly compete with China, build more stuff, and not restrict chips (vs. semiconductor manufacturing). That all seems positive to me.

We Are Not in the Endgame

Harsh restrictions make sense when you think the technology today or in the near term will be highly impactful and at its apex.

Harsh restrictions do not make sense if the highest potential of the technology is still ahead of us. No export control is perfect. The longer the timeframe that is required to make these kinds of controls work, the less likely they’ll work at all. More likely, they’ll backfire—as they have in the US’s case, where Huawei has made significant progress and gained traction for their platform. Again, effectively the US has executed industrial policy benefiting the Chinese semiconductor industry.

covers Huawei’s modern capabilities well in an overview here.

AI adoption is still in its early days, with Bond Capital’s presentation showing that adoption in all industries in the US was still less than 10% in March 2025.

Early internet commentators predicted that a plethora of blogs would dominate the world or create a decentralized utopia. It would have been unimaginable in the late 1990s that the world would get centralized into walled gardens like the social media platforms and have a single search engine be the “home page” of the internet.

That home page—Google, in case it wasn’t obvious—is now under threat by AI. Why go to Google when you can just “ChatGPT” it?

There are many different changes that had to happen to make Webvan a laughingstock of a business after the dot-com bust and Instacart a $13 billion business as of July 2025.

That included smartphones, better mobile connectivity, and more.

What will the future bring to AI? It’s still too early to say.

Given that, I think some of the current changes in policy—getting rid of the diffusion rule, easing restrictions around Nvidia chip sales, and trying to promote more development—make sense.

Share

Hooked on CUDA

There are smart people I respect who think that chip controls are important. However, whether or not you think the PLA (Chinese military) will use Nvidia chips or not (which Jensen Huang denied… weirdly), isn’t it better for Nvidia, a US company, to be the global standard that gets sales/capital, feedback from developers, and the most progress on its platform? If you’re worried about leverage, isn’t it better to have everyone hooked on Nvidia’s CUDA platform instead of Huawei’s CANN?

I’ve talked about why CUDA is so sticky in the past, and that’s still true.

In that case, if conflict breaks out, you’d actually have something you can restrict that could cause short-term problems for China. Instead, currently, China possesses significant leverage due to its critical rare earth minerals, and we are still far from establishing non-Chinese refining and processing capacity to replace it.

While China itself definitely knows that it must diversify away from the US and build up both Huawei and SMIC, Chinese companies will use whatever works best. China has a famously competitive culture (just look at China’s “involution” problem). Being patriotic may mean using Huawei chips—but I suspect if Nvidia chips were freely available, they’d use those (and pay lip service to the Huawei clusters). After all, demand is still extremely high for Nvidia chips among Chinese companies.

Forcing these companies to use Huawei, providing sales/capital to them, and giving them market feedback and the capability to scale up their operations seems like exactly what the US would not want to do.

Thanks for reading!

I hope you enjoyed this article. If you’d like to learn more about AI’s past, present, and future in an easy to understand way, I’m working on a book titled What You Need to Know About AI that will be published October 15th.

You can learn more about it and pre-order on Amazon, Barnes and Noble, or Bookshop (indie booksellers).