One thing I have not seen commented on is that ARC-AGI is a visual benchmark but LLMs are primarily text. For instance when I see one of the ARC-AGI puzzles, I have a visual representation in my brain and apply some sort of visual reasoning solve it. I can "see" in my mind's eye the solution to the puzzle. If I didn't have that capability, I don't think I could reason through words how to go about solving it - it would certainly be much more difficult.
I hypothesize that something similar is going on here. OpenAI has not published (or I have not seen) the number of reasoning tokens it took to solve these - we do know that each tasks was thoussands of dollars. If "a picture is worth a thousand words", could we make AI systems that can reason visually with much better performance?
Yeah this part is what makes the high performance even more surprising to me. The fact that LLMs are able to do so well on visual tasks (also seen with their ability to draw an image purely using textual output https://simonwillison.net/2024/Oct/25/pelicans-on-a-bicycle/) implies that not only do they actually have some "world model" but that this is in spite of the disadvantage given by having to fit a round peg in a square hole. It's like trying to map out the entire world using the orderly left-brain, without a more holistic spatial right-brain.
I wonder if anyone has experimented with having some sort of "visual" scratchpad instead of the "text-based" scratchpad that CoT uses.
A file is a stream of symbols encoded by bits according to some format. It’s pretty much 1D. It would be susprising that LLM couldn’t extract information from a file or a data stream.
I hypothesize that something similar is going on here. OpenAI has not published (or I have not seen) the number of reasoning tokens it took to solve these - we do know that each tasks was thoussands of dollars. If "a picture is worth a thousand words", could we make AI systems that can reason visually with much better performance?