4 Comments

Perhaps asking “What is AI?” and the hunt for “real” intelligence is the wrong question. Perhaps Collet is right about what intelligence really means, but misses the real question, which is “What does AI actually do?”

/

I can’t say I agree. Chollet is concerned with the claims of many prominent researchers in the field that what we have today can be called intelligent. He pushes for a more intellectual honest debate and with ARC wants to stimulate open research, in a time where almost everything has become closed-source.

Expand full comment

Great article! Although the claim that “Reasoning is a type of learning, and learning is a type of reasoning” seems a bit suspect to me as I can’t see how something that is a subset (a “type”) of something can also be a superset of it. But then I am neither a mathematician or a logician, so maybe I’m missing something. Like, a Cadillac is a type of car, but a car is clearly not a type of Cadillac.

Expand full comment

You aren't missing anything. I thought I had replied to this prior, but in any case, since that sentence is confusing and inaccurate, I have revised it to " Reasoning and learning are similar; both augment understanding by operating on information priors. "

Expand full comment

Great summary. I'm never sure whether Chollet is bringing clarity or mud to the situation, but at least the ARC prize is motivating a lot of excellent work.

> LLMs are not intelligent to Chollet because intelligence is about understanding novel situations and LLMs cannot go outside their training data

Well funnily enough, Chollet has supplied a solution in his own definition: train on the novel situation! That's test-time training as you mentioned here and last week.

Test-time training at first sounded to me impossible or cheating -- you can't train on a result you don't have?? -- but MIT's test-time training is showing something simpler and more in line with human reasoning: take the problem you want to solve, and learn from each part of it; test your understanding on what you think you know about the problem by creating sub-problems and verifying each one, learning from your successes and failures. Learning, learning, learning.

I've long believed a form of training has to take place during novel situations (rather than just inference) and seeing the test-time training results you referenced was a real jolt of excitement; this is the path forward!

(But does a solved ARC-AGI alone mean AGI? I'm doubtful. This paper https://arxiv.org/pdf/2410.06405 demonstrated that an improved Vision Transformer (2d visual tokens, combining absolute, relative and object positional info) greatly improves results. Combine this with the fact that LLMs have known poor perceptual reasoning (https://arxiv.org/pdf/2410.07391) and we might reasonably conclude that one reason the ARC benchmark has held up is because Chollet happened to pick an area where AI models are "systematically" poorly trained, hence not really a good AGI test.)

Expand full comment