News Roundup: May 15, 2024

Open AI announcements, DeepSeek-v2, and TSMC Arizona

We’re trying something new. Instead of just deep articles every few weeks, I’ll do a round-up of interesting news/articles and add some commentary on my side. Some articles will have paywalls for their full versions (unavoidable nowadays, given the best analysis is often paid), though I’ll always give enough context that further reading from linked content is helpful, but not required.

Please let us know if this is interesting or useful!

OpenAI announces GPT-4o, and not a search engine

On Friday, multiple news networks reported that OpenAI was announcing a Google Search competitor on Monday (May 13th). Instead, OpenAI unveiled GPT-4o (I’d suggest watching the announcement yourself for full details).

The big, constantly emphasized factor of GPT-4o was its “omnimodel” (multimodal, to use the terminology everyone else does). Vision, screenshots, and, especially, voice. Voice interaction with GPT-4o was constantly used during the demo, which was obviously their biggest highlight.

This isn’t precisely a search engine, but one can see where they can argue that it’s meant to replace voice assistants, which (in theory) can replace search.

The big demo points were:

• Voice generation (including with numerous available tones of voice for GPT-4o)

• Basic math (a simple linear equation was explained in the demo)

• Reading and explaining graphs

• Explaining code and what it does (at least simple code)

There’s nothing breakthrough here. I explained previously in my article about Q* (which, interestingly, disappeared from the public eye after the Sam Altman news cycle went away), math is hard for LLMs. However, the demos were basic enough that perhaps the work supposedly in Q* was rolled in, or it might still be inconsistent.

Explaining graphs, code, and the environment from pictures is not new and are features of Humane and Rabbit—which might be in trouble, or could just roll in OpenAI’s likely much-superior model for that purpose as part of their backend.

The interesting thing here is what has always been OpenAI’s strength: taking what exists and cleaning it up with a lot of careful engineering to look impressive to the broader public.

China’s DeepSeek-v2

AI Current had a fun review of China’s DeepSeek-v2, where he tested some politically sensitive questions with fascinating results:

DeepSeek-V2 is a decent LLM. It is also a propaganda tool priced way below the actual cost, likely in the hope of disseminating CCP worldviews globally.

It’s good, much cheaper than its competitors, and highly political. I can’t find it right now, but I saw one commentator say it essentially comes with a Chinese political commissar/censor.

My main take is we’re finally seeing decent models from China (it was bound to happen eventually). I commented on the challenges in ChinaTalk some time back, and it seems like this mostly reflects that challenge: whether it’s explicitly stated, Chinese LLMs need to notgenerate content that runs afoul of Chinese officials.

Given LLMs—and probably all models subsequently that have interesting/useful levels of generalization and power—are not deterministic, this is always going to be a huge pain for Chinese companies and something that slows them down relative to their global counterparts.

TSMC in the Desert (of Arizona)

Viola Zhou has a great feature in Rest of World and a fascinating interview on ChinaTalk about it.

There’s all the expected challenges in terms of culture clashes, misunderstanding between Taiwanese engineers/bosses and American workers, complaints about TSMC military-like demands/hours vs. complaints about lazy Americans that leave work at 5pm, and more. It is really worth a read through to dive into the human drama of it.

However, it’s worth noting: it’s still moving forward.

This is in contrast with the old Foxconn site in Wisconsin that was announced by Trump, which Biden now announced will be a Microsoft AI data center. Both were “industrial” projects being on-shored by Taiwanese companies in response to tensions with China. Only one was actually serious and is still continuing.

Thinking about why, it’s clear: TSMC and semiconductors are much more strategically important than Foxconn and LCD screens.

Foxconn might have gotten tax incentives and brownie points, but it needed the factory to actually be reasonable in terms of economics. It’s too early to tell whether the Arizona fab is actually going to be economically net-positive to TSMC (as in, profitable). But, even if it isn’t, the US is going to throw projects at it and TSMC will be willing to take a loss to keep itself in the US’s good graces (and the US will be willing to offer good graces, given its importance).

Ben Thompson at Stratechery had called this in late 2022:

What is interesting about this Arizona project is that I wouldn’t be surprised if it represents TSMC adopting a bit of the Intel playbook: Apple is TSMC’s largest customer, and Nvidia and AMD are almost certainly in the top 5. Moreover, their combined volume far exceeds what TSMC is building in Arizona, which may be a feature, not a bug: I suspect these Arizona foundries are being built specifically for those three companies, and while I doubt that TSMC is going to the same extremes as Intel’s Copy Exactly! approach, my assumption is that these fabs are going to be designed for mass production of a limited number of chips, not the flexibility that is at the core of the company’s Taiwan operations…

What is more important to the powers-that-be, though, is that President Biden gets to make a big speech about a big dollar figure and Tim Cook gets to talk about “Made in America” without having to discuss the company’s continued dependence on China. TSMC will have to bear the cost, but if that is the price of shoring up U.S. support for Taiwan, well, that is the best possible insurance policy the company could buy for its operations that truly matter, which are intrinsically tied to Taiwan.

In general, TSMC’s Taiwan operations are still irreplaceable, but this is still an important project (for optics, if nothing else) that everyone will still be happy to sink money into.