The point of ARC is NOT to compare humans vs AI, but to probe the current boundary of AIs weaknesses. AI has been beating us at specific tasks like handwriting recognition for decades. Rather, it's when we can no longer readily find these "easy for human, hard for AI" reasoning tasks that we must stop and consider.

If you look at the ARC tasks failed by o3, they're really not well suited to humans. They lack the living context humans thrive on, and have relatively simple, analytical outcomes that are readily processed by simple structures. We're unlikely to see AI as "smart" until it can be asked to accomplish useful units of productive professional work at a "seasoned apprentice" level. Right now they're consuming ungodly amounts of power just to pass some irritating, sterile SAT questions. Train a human for a few hours a day over a couple weeks and they'll ace this no problem.

o3 low and high are the same model. Difference is in how long was it allowed to think.

It works the same with humans. If they spend more time on the puzzle they are more likely to solve it.

But does it matter if it "really, really" reasons in the human sense, if it's able to prove some famous math theorem or come up with a novel result in theoretical physics?

While beyond current motels, that would be the final test of AGI capability.

If it's gaming the system, then it's much less likely to reliably come up with novel proofs or useful new theoretical ideas.

That would be important, but as far as I know it hasn’t happened (despite how often it’s intimated that we’re on the verge of it happening).

I've seen one Twitter thread from a mathematician who used an llm to come up with a new math result. Both coming up with the theorem statement and a unique proof,iirc.

Though to be clear, this wasn't a one shot thing - it was iirc a few months of back and forth chats with plenty of wrong turns too.

Then he used it as a random text generator, LLM is by far the most configurable and best random test generators we have. You can use that to generate random theorem noise and then try to work with that to find actual theorems, still doesn't replace mathematicians though.

I think we should let the professional mathematician who says the llm helped him be the judge of how and why it helped.

Found the thread: https://x.com/robertghrist/status/1841462507543949581?s=46&t...

From the thread:

> AI assisted in the initial conjectures, some of the proofs, and most of the applications it was truly a collaborative effort

> i went back and forth between outrageous optimism and frustration through this process. i believe that the current models can reason – however you want to interpret that. i also believe that there is a long way to go before we get to true depth of mathematical results.

Yeah, it really does matter if something was reasoned, or whether it appears if you metaphorically shake the magic 8 ball.

How would gaming the system work here? Is there some flaw in the way the tasks are generated?

AI models have historically found lots of ways to game systems. My favorite example is exploiting bugs in simulator physics to "cheat" at games of computer tag. Another is a model for radiology tasks finding biases in diagnostic results using dates on the images. And of course whenever people discuss a benchmark publicly it leaks the benchmark into the training set, so the benchmark becomes a worse measure.

I am not expert in llm reasoning but I think because of RL. You cannot use AlphaZero to play other games.

Nope. AlphaZero taught itself to play games like chess, shogi, and Go through self-play, starting from random moves. It was not given any strategies or human gameplay data but was provided with the basic rules of each game to guide its learning process.

Yes its reinforcement learning, but need to create policy and each policy is specialized for specific tasks.

I thought that AlphaZero could play three games? Go, Chess and Shogi?

Think I mean Catan :)

Humans and AIs are different, the next benchmark would be build so that it emphasize the weak points of current AI models where a human is expected to perform better, but I guess you can also make a benchmark that is the opposite, where humans struggle and o3 has an easy time.

I think you've hit the nail on the head there. If these systems of reasoning are truly general then they should be able to perform consistently in the same way a human does across similar tasks, baring some variance.

AGI wouldn't necessarily entail any autonomy or goals though. In principle there could be a superintelligent AI that's completely indifferent to such outcomes, with no particular goals beyond correctly answering question or what not.

AGI is a spectrum, not a binary quality.

Not sure why I am being downvoted. Why would a sufficiently advanced intelligence reveal its full capabilities knowing fully well that it would then be subjected to a range of constraints and restraints?

If you disagree with me, state why instead of opting to downvote me