Live data from Hacker News

OpenAI O3 breakthrough high score on ARC-AGI-PUB

arcprize.org

291–300 of 1001 posts

Re: OpenAI O3 breakthrough high score on ARC-AGI-PUB

#291
Very cool. I recommend scrolling down to look at the example problem that O3 still can’t solve. It’s clear what goes on in the human brain to solve this problem: we look at one example, hypothesize a simple rule that explains it, and then check that hypothesis against the other examples. It doesn’t quite work, so we zoom into an example that we got wrong and refine the hypothesis so that it solves that sample. We keep iterating in this fashion until we have the simplest hypothesis that satisfies all the examples. In other words, how humans do science - iteratively formulating, rejecting and refining hypotheses against collected data.

From this it makes sense why the original models did poorly and why iterative chain of thought is required - the challenge is designed to be inherently iterative such that a zero shot model, no matter how big, is extremely unlikely to get it right on the first try. Of course, it also requires a broad set of human-like priors about what hypotheses are “simple”, based on things like object permanence, directionality and cardinality. But as the author says, these basic world models were already encoded in the GPT 3/4 line by simply training a gigantic model on a gigantic dataset. What was missing was iterative hypothesis generation and testing against contradictory examples. My guess is that O3 does something like this:

1. Prompt the model to produce a simple rule to explain the nth example (randomly chosen)

2. Choose a different example, ask the model to check whether the hypothesis explains this case as well. If yes, keep going. If no, ask the model to revise the hypothesis in the simplest possible way that also explains this example.

3. Keep iterating over examples like this until the hypothesis explains all cases. Occasionally, new revisions will invalidate already solved examples. That’s fine, just keep iterating.

4. Induce randomness in the process (through next-word sampling noise, example ordering, etc) to run this process a large number of times, resulting in say 1,000 hypotheses which all explain all examples. Due to path dependency, anchoring and consistency effects, some of these paths will end in awful hypotheses - super convoluted and involving a large number of arbitrary rules. But some will be simple.

5. Ask the model to select among the valid hypotheses (meaning those that satisfy all examples) and choose the one that it views as the simplest for a human to discover.

Re: OpenAI O3 breakthrough high score on ARC-AGI-PUB

#292
post #237

Complete aside here: I used to do work with amputees and prosthetics. There is a standardized test (and I just cannot remember the name) that fits in a briefcase. It's used for measuring the level of damage to the upper limbs and for prosthetic grading. Basically, it's got the dumbest and simplest things in it. Stuff like a lock and key, a glass of water and jug, common units of currency, a zipper, etc. It tests if y…

There's a lot of truth in this. I sometimes joke that robot benchmarks should focus on common household chores. Given a basket of mixed laundry, sort and fold everything into organized piles. Load a dishwasher given a sink and counters overflowing with dishes piled up haphazardly. Clean a bedroom that kids have trashed. We do these tasks almost without thinking, but the unstructured nature presents challenges for rob…

I maintain that whoever invents a robust laundry folding robot will be a trillionaire. In that, I dump jumbled clean clothes straight from a dryer at it and out comes folded and sorted clothes (and those loner socks). I know we're getting close, but I also know we're not there yet.

Re: OpenAI O3 breakthrough high score on ARC-AGI-PUB

#293

Complete aside here: I used to do work with amputees and prosthetics. There is a standardized test (and I just cannot remember the name) that fits in a briefcase. It's used for measuring the level of damage to the upper limbs and for prosthetic grading. Basically, it's got the dumbest and simplest things in it. Stuff like a lock and key, a glass of water and jug, common units of currency, a zipper, etc. It tests if y…

> We had hand prosthetics that could play Mozart at 5x speed on a baby grand, but could not pick up a silver dollar or zip a jacket even a little bit. " I must be missing something, how can they be able to play Mozart at 5x speed with their prosthetics but not zip a jacket? They could press keys but not do tasks requiring feedback? Or did you mean they used to play Mozart at 5x speed before they became amputees?

I'm far from a piano player, but I can definitely push piano buttons quite quickly while zipping up my jacket when it's cold and/or wet outside is really difficult.

Even more so for picking up coins from a flat surface.

For robotics, it's kind of obvious, speed is rarely an issue, so the "5x" part is almost trivial. And you can program the sequence quite easily, so that's also doable. Piano keys are big and obvious and an ergonomically designed interface meant to be relatively easy to press, ergo easy even for a prosthetic. A small coin on a flat surface is far from ergonomic.

Re: OpenAI O3 breakthrough high score on ARC-AGI-PUB

#294
We should NOT give up on scaling pretraining just yet!

I believe that we should explore pretraining video completion models that explicitly have no text pairings. Why? We can train unsupervised like they did for GPT series on the text-internet but instead on YouTube lol. Labeling or augmenting the frames limits scaling the training data.

Imagine using the initial frames or audio to prompt the video completion model. For example, use the initial frames to write out a problem on a white board then watch in output generate the next frames the solution being worked out.

I fear text pairings with CLIP or OCR constrain a model too much and confuse

Re: OpenAI O3 breakthrough high score on ARC-AGI-PUB

#295

Congratulations to Francois Chollet on making the most interesting and challenging LLM benchmark so far. A lot of people have criticized ARC as not being relevant or indicative of true reasoning, but I think it was exactly the right thing. The fact that scaled reasoning models are finally showing progress on ARC proves that what it measures really is relevant and important for reasoning. It's obvious to everyone that…

> making the most interesting and challenging LLM benchmark so far. This[1] is currently the most challenging benchmark. I would like to see how O3 handles it, as O1 solved only 1%. 1. https://epoch.ai/frontiermath/the-benchmark

Apparently o3 scored about 25%

https://youtu.be/SKBG1sqdyIU?t=4m40s

Re: OpenAI O3 breakthrough high score on ARC-AGI-PUB

#296
post #22

Earlier quoted context omitted.

Are there any single-step non-reasoner models that do well on this benchmark? I wonder how well the latest Claude 3.5 Sonnet does on this benchmark and if it's near o1.

Here are the results for base models[1]: o3 (coming soon) 75.7% 82.8% o1-preview 18% 21% Claude 3.5 Sonnet 14% 21% GPT-4o 5% 9% Gemini 1.5 4.5% 8% Score (semi-private eval) / Score (public eval) [1]: https://arcprize.org/2024-results

I'd love to know how Claude 3.5 Sonnet does so well despite (presumably) not having the same tricks as the o-series models.

Re: OpenAI O3 breakthrough high score on ARC-AGI-PUB

#297

Efficiency is now key. ~=$3400 per single task to meet human performance on this benchmark is a lot. Also it shows the bullets as "ARC-AGI-TUNED", which makes me think they did some undisclosed amount of fine-tuning (eg. via the API they showed off last week), so even more compute went into this task. We can compare this roughly to a human doing ARC-AGI puzzles, where a human will take (high variance in my subjective…

some other imporant quotes: "Average human off the street: 70-80%. STEM college grad: >95%. Panel of 10 random humans: 99-100%" -@fchollet on X So, considering that the $3400/task system isn't able to compete with STEM college grad yet, we still have some room (but it is shrinking, i expect even more compute will be thrown and we'll see these barriers broken in coming years) Also, some other back of envelope calculat…

I don't follow how 10 random humans can beat the average STEM college grad and average humans in that tweet. I suspect it's really "a panel of 10 randomly chosen experts in the space" or something?

I agree the most interesting thing to watch will be cost for a given score more than maximum possible score achieved (not that the latter won't be interesting by any means).

Re: OpenAI O3 breakthrough high score on ARC-AGI-PUB

#298
post #8

How much longer can I get paid $150k to write code ?

I think they will have to figure out how to get around context limits before that happens. I also wouldn't be surprised if the future models that can actually replace workers are sold at such an exorbitant price that only larger companies will be able to afford it. Everyone else gets access to less capable models that still require someone with knowledge to get to an end result.

Re: OpenAI O3 breakthrough high score on ARC-AGI-PUB

#299
post #42

Human performance is 85% [1]. o3 high gets 87.5%. This means we have an algorithm to get to human level performance on this task. If you think this task is an eval of general reasoning ability, we have an algorithm for that now. There's a lot of work ahead to generalize o3 performance to all domains. I think this explains why many researchers feel AGI is within reach, now that we have an algorithm that works. Congrat…

It actually beats the human average by a wide margin: - 64.2% for humans vs. 82.8%+ for o3. ... Private Eval: - 85%: threshold for winning the prize [1] Semi-Private Eval: - 87.5%: o3 (unlimited compute) [2] - 75.7%: o3 (limited compute) [2] Public Eval: - 91.5%: o3 (unlimited compute) [2] - 82.8%: o3 (limited compute) [2] - 64.2%: human average (Mechanical Turk) [1] [3] Public Training: - 76.2%: human average (Mecha…

Super human isn't beating rando mech turk.

Their post has stem grad at nearly 100%

Re: OpenAI O3 breakthrough high score on ARC-AGI-PUB

#300
post #19

My initial impression: it's very impressive and very exciting. My skeptical impression: it's complete hubris to conflate ARC or any benchmark with truly general intelligence. I know my skepticism here is identical to moving goalposts. More and more I am shifting my personal understanding of general intelligence as a phenomenon we will only ever be able to identify with the benefit of substantial retrospect. As it is…

It doesn't need to be general intelligence or perfectly map to human intelligence. All it needs to be is useful. Reading constant comments about LLMs can't be general intelligence or lack reasoning etc, to me seems like people witnessing the airplane and complaining that it isn't "real flying" because it isn't a bird flapping its wings (a large portion of the population held that point of view back then). It doesn't…

On the contrary, the pushback is critical because many employers are buying the hype from AI companies that AGI is imminent, that LLMs can replace professional humans, and that computers are about to eliminate all work (except VCs and CEOs apparently).

Every person that believes that LLMs are near sentient or actually do a good job at reasoning is one more person handing over their responsibilities to a zero-accountability highly flawed robot. We've already seen LLMs generate bad legal documents, bad academic papers, and extremely bad code. Similar technology is making bad decisions about who to arrest, who to give loans to, who to hire, who to bomb, and who to refuse heart surgery for. Overconfident humans employing this tech for these purposes have been bamboozled by the lies from OpenAI, Microsoft, Google, et al. It's crucial to call out overstatement and overhype about this tech wherever it crops up.

Post reply on HN