[swaraj]

You should try the arc agi puzzles yourself, and then tell me you think these things aren't intelligent

https://arcprize.org/blog/openai-o1-results-arc-prize

I wouldn't say it's full agi or anything yet, but these things can definitely think in a very broad sense of the word

[gessha]

[Back in the 1980s]

You should try to play chess yourself, and then tell me you think these things aren't intelligent.

[jhbadger]

While I agree that we should be skeptical about the reasoning capabilities of LLMs, comparing them to chess programs misses the point. Chess programs were specifically created to play chess. That's all they could do. They couldn't generalize and play other board games, even related games like Shogi and Xiangqi, the Japanese and Chinese versions of chess. LLMs are amazing at being able to do things they never were programmed to do simply by accident.

[gessha]

Are they though? They’ve been shown to generalize poorly to tasks where you switch up some of the content.

[jhbadger]

Here's an example. I'm interested in obscure conlangs like Volapük. I can feed a LLM (which had no idea what Volapük was), a English-language grammar of Volapük and suddenly it can translate to and from the language. That couldn't work with a chess program. I couldn't give it a rule book of Shogi and have it play that.

[refulgentis]

That's not true

[gessha]

They don’t generalize well on logic puzzles

https://huggingface.co/blog/yuchenlin/zebra-logic

[refulgentis]

Apologies, I was a bit curt because this is a well-worn interaction pattern.

I don't mean anything by the following either, other than, the goalposts have moved:

- This doesn't say anything about generalization, nor does it claim to.

- The occurrences of the prefix general* refer to "Can fine-tuning with synthetic logical reasoning tasks improve the general abilities of LLMs?"

- This specific suggestion was accomplished publicly to some acclaim in September

- To wit, the benchmark the article is centered around hasn't been updated since since September, because the preview of the large model accomplishing that blew it out of the water, 33% on all at the time, 71%: https://huggingface.co/spaces/allenai/ZebraLogic

- these aren't supposed to be easy, they're constraint satisfaction problems, which they point out are used on the LSAT

- The major other form of this argument is the Apple paper, which shows a 5 point drop from 87% to 82% on a home-cooked model

[daveguy]

LLMs don't do too well on those ARC-AGI problems. Even though they're pretty easy for a person.

[cubefox]

https://arcprize.org/blog/oai-o3-pub-breakthrough

[daveguy]

Let me know when they can perform that well without a 300-shot. Or that well on unseen ARC-AGI-2.

[aoeusnth1]

Two years give or take 6 months.

[mrbungie]

Give a group of "average human" two years, give or take 6 months, and they will also saturate the benchmark and probably some humans would beat the SOTA LLM/RLM.

People tend to do so all the time, with games for example.

[aoeusnth1]

Average humans cannot be copy-pasted.

[refulgentis]

Done (link says 6 samples?)

[daveguy]

> OpenAI shared they trained the o3 we tested on 75% of the Public Training set.

I'm talking transfer learning and generalization. A human who has never seen the problem set can be told the rules of the problem domain and then get 85+% on the rest. o3 high compute requires 300 examples using SFT to perform similarly. An impressive feat, but obviously not enough to just give an agent instructions and let it go. 300 examples for human level performance on the specific task, but that's still impressive compared to SOTA 2 years ago. It will be interesting to see performance on ARC-AGI-2.