Why on earth would you tell the model the task in english? The tasks have extremely simple JSON representations that would make for a way, way more natural input. The colorful grids are just to make it easier (for humans) to visualize the symmetries needed. Eg: Here is an (abbreviated) train/test example for task 007bbfb7 from https://github.com/fchollet/ARC/blob/master/data/training/00...
I would argue maybe that the author's premise depends on some faulty prompts.
With some gentle rephrasing, ChatGPT correctly interprets and manipulates the stack problem. I've observed ChatGPT may have some issues with non-standard sentence structure involving, especially with colons.
> Imagine a stack of items (from bottom to top) with a cat, laptop, television, and apple. The apple is moved between the laptop and the television. Which item is on top of the stack afterwards?
The author seems also to have misinterpreted that Stable Diffusion is the same sequence-sensitive model as Transformer LLMs, which it is not. While it is generative it has minimal capacity for abstraction, since the focus would rather be minimizing the "kernel" representation of each word in the weights.
DALL-E has a bit more luck, but requires much more verbose description to get anything like a screen on top of a cat, but that's presumably more due to the bias in the data set of things being displayed on screens, or cats getting on top of things.
Something like
> A cartoon image of a television above a cartoon image of a cat
The author's prompts are all perfectly understandable to humans and expressed in a direct and clear way, and the AIs can't understand them.
This comment essentially says "the prompt is faulty because the AI didn't answer it correctly". This bakes the correctness of the AI into the tautology that it always answers "good" prompts, which are defined as those it answers correctly.
Well, humans don't know how to determine if a proposition is correct.
So it's no surprise we didn't teach AIs yet!
Edit: I don't understand why experts say it failed when they "asked it to reason". It doesn't have pictures of TVs on cats to produce variations of, right? How does "reason" enter into it?
There are different perspectives on how much intelligence these foundation models have. Some people think they just elaborate on or mash up their training data. Others think they have internal representations that allow for more human-like creativity. Maybe they learn from the examples what cat and TV and on top of mean, and can put those together in a way not seen in the training data.
Right but the author's premise is that the AI fails to reason or extend ideas, not that the AI fails to reason or extend ideas in the presence of extraneous, conflicting, or malformed instructions.
If you're prompting ChatGPT (which is already loaded with prompt-engineering cheez-whiz mind you) to reason as a computer, why use words like "imagine"? why phrase things as a question? I think the fragility the author is observing here is extraneous to the actual underlying technology, and a factor of the UI that's been stacked-on and purpose-built for softer-purpose use, not hard academic research such as this.
I mean unless you've got a better heuristic, but my conclusion is primarily structural with regard to Transformer units, basically that they're Turing-complete -- if there's anything more to reasoning and abstraction, I don't know myself, perhaps Scott Aaronson is the guy to talk to.
The author mentions that Chollet's ARC tasks contain 200 unreleased "hidden" problems. This set off a light bulb for me: hidden problems are going to be increasingly important for benchmarking AI meaningfully. With RLHF, ChatGPT is memorizing the answer to every problem thrown at it. Eventually it may become impossible to test with any public question bank.
We are going to go through a period of fooling ourselves, as we always do, then we're going to hit a wall where models are preforming perfectly on benchmarks while still clearly lacking in genuine intelligence. And then we are going to need to test with a concern for security as serious as that used for the gaokao or SAT. Benchmarking will be a paid service performed under highly controlled conditions.
"Consider, for example, the “simple” concept on top of."
This is from fifty years ago:
Person: Pick up a big red block.
Computer: OK.
Person: Grasp the pyramid.
Computer: I DON'T UNDERSTAND WHICH PYRAMID YOU MEAN.
Person (changing their mind): Find a block which is taller than the one you are holding and put it into the box.
Computer: BY "IT", I ASSUME YOU MEAN THE BLOCK WHICH IS TALLER THAN THE ONE I AM HOLDING.
Computer: OK.
Person: What does the box contain?
Computer: THE BLUE PYRAMID AND THE BLUE BLOCK.
Person: What is the pyramid supported by?
Computer: THE BOX.
Person: How many blocks are not in the box?
Computer: FOUR OF THEM.
Person: Is at least one of them narrower than the one which i told you to pick up?
Computer: YES, THE RED CUBE.
Person: Is it supported?
Computer: YES, BY THE TABLE.
Person: Can the table pick up blocks?
Computer: NO.
Person: Can a pyramid be supported by a block?
Computer: YES.
Person: Can a pyramid support a pyramid?
Computer: I DON'T KNOW.
Person: Stack up two pyramids.
Computer: I CAN'T.
Person: Will you please stack up both of the red blocks and either a green cube or a pyramid?
Computer: OK.
Person: Which cube is sitting on the table?
Computer: THE LARGE GREEN ONE WHICH SUPPORTS THE RED PYRAMID.
Person: Is there a large block behind a pyramid?
Computer: YES, THREE OF THEM: A LARGE RED ONE, A LARGE GREEN CUBE, AND THE BLUE ONE.
P.S. "Any sufficiently advanced technology is indistinguishable from a rigged demo"
16 comments
[ 3.8 ms ] story [ 52.8 ms ] threadARC 2 — https://news.ycombinator.com/item?id=32740353
The Measure of Intelligence — https://news.ycombinator.com/item?id=21547958
Neural Abstract Reasoner — https://news.ycombinator.com/item?id=25167182
A first lesson in meta-rationality (Bongard problems) — https://news.ycombinator.com/item?id=27411960
A HN dicussion about ARC — https://news.ycombinator.com/item?id=29867200
That was the a competition. Nobody scored well.
With some gentle rephrasing, ChatGPT correctly interprets and manipulates the stack problem. I've observed ChatGPT may have some issues with non-standard sentence structure involving, especially with colons.
> Imagine a stack of items (from bottom to top) with a cat, laptop, television, and apple. The apple is moved between the laptop and the television. Which item is on top of the stack afterwards?
The author seems also to have misinterpreted that Stable Diffusion is the same sequence-sensitive model as Transformer LLMs, which it is not. While it is generative it has minimal capacity for abstraction, since the focus would rather be minimizing the "kernel" representation of each word in the weights.
DALL-E has a bit more luck, but requires much more verbose description to get anything like a screen on top of a cat, but that's presumably more due to the bias in the data set of things being displayed on screens, or cats getting on top of things.
Something like
> A cartoon image of a television above a cartoon image of a cat
https://imgur.com/q9NMY0W Illustrates that it's preference rather than capability.
This comment essentially says "the prompt is faulty because the AI didn't answer it correctly". This bakes the correctness of the AI into the tautology that it always answers "good" prompts, which are defined as those it answers correctly.
So it's no surprise we didn't teach AIs yet!
Edit: I don't understand why experts say it failed when they "asked it to reason". It doesn't have pictures of TVs on cats to produce variations of, right? How does "reason" enter into it?
If you're prompting ChatGPT (which is already loaded with prompt-engineering cheez-whiz mind you) to reason as a computer, why use words like "imagine"? why phrase things as a question? I think the fragility the author is observing here is extraneous to the actual underlying technology, and a factor of the UI that's been stacked-on and purpose-built for softer-purpose use, not hard academic research such as this.
Case study, even GPT2 can reason about the stack https://transformer.huggingface.co/share/bQXbmeyBZD
"Can reason" is a generous interpretation of "got the right answer in one instance".
We are going to go through a period of fooling ourselves, as we always do, then we're going to hit a wall where models are preforming perfectly on benchmarks while still clearly lacking in genuine intelligence. And then we are going to need to test with a concern for security as serious as that used for the gaokao or SAT. Benchmarking will be a paid service performed under highly controlled conditions.
More like “and then, as we always do, we’re going to redefine ‘intelligence’ to exclude whatever we don’t want to accept as intelligent.”
This is from fifty years ago:
P.S. "Any sufficiently advanced technology is indistinguishable from a rigged demo"