Silicon Valley is certainly trying to set the bar high for our expectations. The youngest and most impacted among us aren't impressed. Americans without wealth to participate in the stock market are going to start to see LLMs as a threat to their well-being.
I’m slowly working on a piece regarding the economics of it.
I think Silicon Valley has a failure of imagination on how the economy works and wage/job improvements will be significantly more widespread than the industry is messaging. That’s a good thing but the tech industry fundamentally has both ego and myopia on areas outside of itself.
This is persuasive, AND: 1. It is easy to design tests that are obvious but almost all humans get wrong too - optical and cognitive illusions are examples. Complete failure at a category of tasks may reveal only a quirk of cognitive structure - a usually-beneficial shortcut - rather than a deep lack. Or something in between. 2. Reasoning about the necessity of world models should include the Helen Kellers and quadriplegics-since-birth of the world - people whose sensory input was severely restricted (sometimes to just symbolic language and/or limited proprioception) yet developed understandings of basic physics. I don't know what they tell us: is the inherent problem difficulty less than argued here? Is more basic physics wired into our brains than we realize? But articles like this seem to assume successful functioning in the world should be impossible for such outliers.
After all, the history of AI, though trapped in the IF-THEN world of expert systems, was about working with so-called "world models". They are not new.
Bayesian belief networks are about world models, and they are not trapped by the idea that all you need is "attention" while running rampant around word vectors, which are by no means world models of anything.
Perhaps a truly important, but overlooked set of experiments were those conducted by Douglas Lenat, starting with his PhD project called AM - for Automated Mathematician. It was, in fact, an expert system, but not one trained on world models, but instead, with rules on "how to think". AM was just given some random collections of information, and, new to it but not to society, it discovered prime numbers, and other oddities.
He did his postdoc with Herb Simon and came out with an upgraded AM called Eurisko, which, unlike AM, was given real-world things to think about. It's first exercise was that of designing a navy capable of winning a game (Traveller's) which had the winning rule: last boat in the water wins. Eurisko discovered a heuristic which said that it's frequently valuable to explore both extremes of some continuum, but back up just a little. So, it's opening game was a truly massive, heavily armored ship, and a tiny unarmed speedboat. All players saw that and surrendered. They changed some rules and Eurisko came back with a ship which had the ability to immediately replace damaged equipment and sailors from below deck - thus, not loosing fleet agility capabilities, and a flying boat. Again, surrender.
The next task was to answer the question: how can we save one transistor per gate in VLSI design. Eurisko decided that the fundamentally 2-dimensional nature of chip design was ripe for a third dimension; doing so saved one transistor per gate. On a billion-gate chip, that's a lot. Carver Mead built one to show it could work.
So, through my lens, "returning" to world models is a truly reasonable way to move forward.
They were used in DeepMind's protein folding platform to enormous advantage.
Silicon Valley is certainly trying to set the bar high for our expectations. The youngest and most impacted among us aren't impressed. Americans without wealth to participate in the stock market are going to start to see LLMs as a threat to their well-being.
I’m slowly working on a piece regarding the economics of it.
I think Silicon Valley has a failure of imagination on how the economy works and wage/job improvements will be significantly more widespread than the industry is messaging. That’s a good thing but the tech industry fundamentally has both ego and myopia on areas outside of itself.
Congratulations on your book award, James! That's excellent. And well deserved.
This is persuasive, AND: 1. It is easy to design tests that are obvious but almost all humans get wrong too - optical and cognitive illusions are examples. Complete failure at a category of tasks may reveal only a quirk of cognitive structure - a usually-beneficial shortcut - rather than a deep lack. Or something in between. 2. Reasoning about the necessity of world models should include the Helen Kellers and quadriplegics-since-birth of the world - people whose sensory input was severely restricted (sometimes to just symbolic language and/or limited proprioception) yet developed understandings of basic physics. I don't know what they tell us: is the inherent problem difficulty less than argued here? Is more basic physics wired into our brains than we realize? But articles like this seem to assume successful functioning in the world should be impossible for such outliers.
I am not so sure it's a "big if".
After all, the history of AI, though trapped in the IF-THEN world of expert systems, was about working with so-called "world models". They are not new.
Bayesian belief networks are about world models, and they are not trapped by the idea that all you need is "attention" while running rampant around word vectors, which are by no means world models of anything.
Perhaps a truly important, but overlooked set of experiments were those conducted by Douglas Lenat, starting with his PhD project called AM - for Automated Mathematician. It was, in fact, an expert system, but not one trained on world models, but instead, with rules on "how to think". AM was just given some random collections of information, and, new to it but not to society, it discovered prime numbers, and other oddities.
He did his postdoc with Herb Simon and came out with an upgraded AM called Eurisko, which, unlike AM, was given real-world things to think about. It's first exercise was that of designing a navy capable of winning a game (Traveller's) which had the winning rule: last boat in the water wins. Eurisko discovered a heuristic which said that it's frequently valuable to explore both extremes of some continuum, but back up just a little. So, it's opening game was a truly massive, heavily armored ship, and a tiny unarmed speedboat. All players saw that and surrendered. They changed some rules and Eurisko came back with a ship which had the ability to immediately replace damaged equipment and sailors from below deck - thus, not loosing fleet agility capabilities, and a flying boat. Again, surrender.
The next task was to answer the question: how can we save one transistor per gate in VLSI design. Eurisko decided that the fundamentally 2-dimensional nature of chip design was ripe for a third dimension; doing so saved one transistor per gate. On a billion-gate chip, that's a lot. Carver Mead built one to show it could work.
So, through my lens, "returning" to world models is a truly reasonable way to move forward.
They were used in DeepMind's protein folding platform to enormous advantage.