ARC Prize – a $1M+ competition towards open AGI progress

ARC Prize – a $1M+ competition towards open AGI progress(arcprize.org)

588 points by mikeknoop 1 year ago | 337 comments

Hey folks! Mike here. Francois Chollet and I are launching ARC Prize, a public competition to beat and open-source the solution to the ARC-AGI eval.

ARC-AGI is (to our knowledge) the only eval which measures AGI: a system that can efficiently acquire new skill and solve novel, open-ended problems. Most AI evals measure skill directly vs the acquisition of new skill.

Francois created the eval in 2019, SOTA was 20% at inception, SOTA today is only 34%. Humans score 85-100%. 300 teams attempted ARC-AGI last year and several bigger labs have attempted it.

While most other skill-based evals have rapidly saturated to human-level, ARC-AGI was designed to resist “memorization” techniques (eg. LLMs)

Solving ARC-AGI tasks is quite easy for humans (even children) but impossible for modern AI. You can try ARC-AGI tasks yourself here: https://arcprize.org/play

ARC-AGI consists of 400 public training tasks, 400 public test tasks, and 100 secret test tasks. Every task is novel. SOTA is measured against the secret test set which adds to the robustness of the eval.

Solving ARC-AGI tasks requires no world knowledge, no understanding of language. Instead each puzzle requires a small set of “core knowledge priors” (goal directedness, objectness, symmetry, rotation, etc.)

At minimum, a solution to ARC-AGI opens up a completely new programming paradigm where programs can perfectly and reliably generalize from an arbitrary set of priors. At maximum, unlocks the tech tree towards AGI.

Our goal with this competition is:

1. Increase the number of researchers working on frontier AGI research (vs tinkering with LLMs). We need new ideas and the solution is likely to come from an outsider! 2. Establish a popular, objective measure of AGI progress that the public can use to understand how close we are to AGI (or not). Every new SOTA score will be published here: https://x.com/arcprize 3. Beat ARC-AGI and learn something new about the nature of intelligence.

Happy to answer questions!

neoneye2 1 year ago |

I'm Simon Strandgaard and I participated in ARCathon 2022 (solved 3 tasks) and ARCathon 2023 (solved 8 tasks).

I'm collecting data for how humans are solving ARC tasks, and so far collected 4100 interaction histories (https://github.com/neoneye/ARC-Interactive-History-Dataset). Besides ARC-AGI, there are other ARC like datasets, these can be tried in my editor (https://neoneye.github.io/arc/).

I have made some videos about ARC:

Replaying the interaction histories, and you can see people have different approaches. It's 100ms per interaction. IRL people doesn't solve task that fast. https://www.youtube.com/watch?v=vQt7UZsYooQ

When I'm manually solving an ARC task, it looks like this, and you can see I'm rather slow. https://www.youtube.com/watch?v=PRdFLRpC6dk

What is weird. The way that I implement a solver for a specific ARC task is much different than the way that I would manually solve the puzzle. Having to deal with all kinds of edge cases.

Huge thanks to the team behind the ARC Prize. Well done.

parentheses 1 year ago | |

The UX of your solution entry is _way_ better than the ARC site itself.

mkl 1 year ago | | |

Being able to hold the mouse button down is certainly much nicer. Not being able to see the examples while you are solving makes it harder than it should be though.

neoneye2 1 year ago | | |

That warms my heart. Thank you.

The short story. I needed something that could render thumbnails of tasks, so I could visual debug what was going on in my solver. However I have never gotten around to make the visual inspection tool. After having the thumbnail renderer, mid january 2024, then it eventually turned into what it is now.

ECCME 1 year ago | |

"Here is a challenge, designed to be unsolvable or so. We'll give you a bazillion dollars if you complete the challenge, and, in the meantime, we will use your attempts to train an as AI that will be worth the cost!!"

gota 1 year ago | | |

In the most charitable interpretation of this comment - I can understand the feeling, when so much of social media interactions are in the form 'It's post a picture of you as a baby, 10 year old, and current age!'. Those and many other instances can bring out excessive skepticism

But the people involved in this haven't signaled that they are in that path, either in the message about the challenge (precisely the opposite) or seemingly in their careers so far

So I guess I don't share the concern but a better way to phrase your comment could be -

"how can we be sure the human-provided solutions won't turn out to be just fodder for training a RL model or something that will later be monetized, closed and proprietary? Do the challenge organizers provide any guarantees on that?"

geor9e 1 year ago | | |

No, you missed the point. The striking thing about ARC is the puzzles are super easy, for humans. The average person solves 85% of the tasks, but the worlds best LLMs are only solving 5%. The challenge is to simply make an AI score as well as the average human.

skrebbel 1 year ago | | |

Did you even try the puzzles? They’re not particularly “unsolvable”.

salamo 1 year ago |

This is super cool. I share Francois' intuition that the presently data-hungry learning paradigm is not only not generalizable but unsustainable: humans do not need 10,000 examples to tell the difference between cats and dogs, and the main reason computers can today is because we have millions of examples. As a result, it may be hard to transfer knowledge to more esoteric domains where data is expensive, rare, and hard to synthesize.

If I can make one criticism/observation of the tests, it seems that most of them reason about perfect information in a game-theoretic sense. However, many if not most of the more challenging problems we encounter involve hidden information. Poker and negotiations are examples of problem solving in imperfect information scenarios. Smoothly navigating social situations also requires a related problem of working with hidden information.

One of the really interesting things we humans are able to do is to take the rules of a game and generate strategies. While we do have some algorithms which can "teach themselves" e.g. to play go or chess, those same self-play algorithms don't work on hidden information games. One of the really interesting capabilities of any generally-intelligent system would be synthesizing a general problem solver for those kinds of situations as well.

lacker 1 year ago |

I really like the idea of ARC. But to me the problems seem like they require a lot of spatial world knowledge, more than they require abstract reasoning. Shapes overlapping each other, containing each other, slicing up and reassembling pieces, denoising regular geometric shapes, you can call them "core knowledge" but to me it seems like they are more like "things that are intuitive to human visual processing".

Would an intelligent but blind human be able to solve these problems?

I'm worried that we will need more than 800 examples to solve these problems, not because the abstract reasoning is so difficult, but because the problems require spatial knowledge that we intelligent humans learn with far more than 800 training examples.

pmayrgundter 1 year ago |

This claim that these tests are easy for humans seems dubious, and so I went looking a bit. Melanie Mitchell chimed in on Chollet's thread and posted their related test [ConceptARC].

In it they question the ease of Chollet's tests: "One limitation on ARC’s usefulness for AI research is that it might be too challenging. Many of the tasks in Chollet’s corpus are difficult even for humans, and the corpus as a whole might be sufficiently difficult for machines that it does not reveal real progress on machine acquisition of core knowledge."

ConceptARC is designed to be easier, but then also has to filter ~15% of its own test takers for "[failing] at solving two or more minimal tasks... or they provided empty or nonsensical explanations for their solutions"

After this filtering, ConceptARC finds another 10-15% failure rate amongst humans on the main corpus questions, so they're seeing maybe 25-30% unable to solve these simpler questions meant to test for "AGI".

ConceptARC's main results show CG4 scoring well below the filtered humans, which would agree with a [Mensa] test result that its IQ=85.

Chollet and Mitchell could instead stratify their human groups to estimate IQ then compare with the Mensa measures and see if e.g. Claude3@IQ=100 compares with their ARC scores for their average human

[ConceptArc]https://arxiv.org/pdf/2305.07141 [Mensa]https://www.maximumtruth.org/p/ais-ranked-by-iq-ai-passes-10...

paxys 1 year ago |

While I agree with the spirit of the competition, a $1M prize seems a little too low considering tens of billions of dollars have already been invested in the race to AGI, and we will see many times that put into the space in the coming years. The impact of AGI will be measured in trillions at minimum. So what you are ultimately rewarding isn't AGI research but fine tuning the newest public LLM release to best meet the parameters of the test.

I'd also urge you to use a different platform for communicating with the public because x.com links are now inaccessible without creating an account.

elicksaur 1 year ago |

I’m a big fan of the ARC as a problem set to tackle. The sparseness of the data and infinite-ness of the rules which could apply make it much tougher than existing ML problem sets.

However, I do disagree that this problem represents “AGI”. It’s just a different dataset than what we’ve seen with existing ML successes, but the approaches are generally similar to what’s come before. It could be that some truly novel breakthrough which is AGI solves the problem set, but I don’t think solving the problem set is a guaranteed indicator of AGI.

nadam 1 year ago |

I love this, this is super interesting, but my intuition based on looking at a dozen examples is that the problem is hard, but easy enough that if this problem becomes popular, near-human level results will appear in a year or less, and AGI will not be reached. The problem seems to be finding a generic enough transformation description language with the appropriate operators. And then heuristics to find a very short program (in the information theoretical sense) in this language that produces all the examples for a problem. I would be very surprised if we would not increase the 34% result soon significantly, and I would be surprised if this could be transferred to general intelligence, at least when I think of the topics where I use AI today and where it falls short yet. Basically my intuition is that this will be yet another 'Chess' or 'Go'-like problem in AI. But still a worthwhile research topic, absolutely: the value that could come out of this is well worth the 1M dollars.

zug_zug 1 year ago | |

I have the exact same impression.

Imo there's no evidence whatsoever that nailing this task will be true AGI - (e.g. able to write novel math proofs, ask insightful questions that nobody has thought of before, self-direct its own learning, read its own source code)

apendleton 1 year ago | | |

I'm not sure the goal of this competition, in and of itself, is AGI. They point to current LLMs emerging from transformers, which in turn emerged from a general basket of building blocks from machine-translation research (attention, etc.). It seems like the suggestion is that to get from where we are now to AGI, some fundamental building blocks are missing, and this is an attempt to spur the development of some of those building blocks, but by analogy with LLMs, the goal here is to come up with a new thing like "attention," not a new thing like GPT4.

Animats 1 year ago |

> the only eval which measures AGI.

That's a stretch. This is a problem at which LLMs are bad. That does not imply it's a good measure of artificial general intelligence.

After working a few of the problems, I was wondering how many different transformation rules the problem generator has. Not very many, it seems. So the problem breaks down into extracting the set of transformation rules from the data, then applying them to new problems. The first part of that is hard. It's a feature extraction problem. The transformations seem to be applied rigidly, so once you have the transformation rules, and have selected the ones that work for all the input cases, application should be straightforward.

This seems to need explicit feature extraction, rather than the combined feature extraction and exploitation LLMs use. Has anyone extracted the rule set from the test cases yet?

elicksaur 1 year ago | |

Yes to your last question, that is essentially how the first iteration solutions operated. Some of the original kaggle competition’s best solutions used a DSL made of these transformations. That was 4 years ago. [1]

The issue with that path is that the problems aren’t using a programmatic generator. The rule sets are anything a person could come up with. It might be as simple as “biggest object turns blue” but they can be much more complicated.

Additionally, the test set is private so it can’t be trained on or extracted from. It has rules that aren’t in the public sets.

[1] https://www.kaggle.com/competitions/abstraction-and-reasonin...

n2d4 1 year ago | |

The tasks are handmade. There is no "problem generator".

slicerdicer1 1 year ago | |

AGI is not when the AI is good at some particular thing, AGI is when we have nothing left at which the AI is bad at (compared to humans).

levocardia 1 year ago |

François Chollet's original paper is incredibly insightful and I'm consistently shocked more people don't talk about it. Some parts are quite technical but at a high level it is the best answer to "what do we mean by general intelligence?" that I've yet seen.

Defining intelligence as an efficiency of learning, after accounting for any explicit or implicit priors about the world, makes it much easier to understand why human intelligence is so impressive.

ildon 1 year ago | |

Do you remember the title/where to find it?

mischa_u 1 year ago | | |

"On the Measure of Intelligence" https://arxiv.org/abs/1911.01547

bigyikes 1 year ago |

Dwarkesh just released an interview with Francois Chollet (partner of OP). I’ve only listened to a few minutes so far, but I’m very interested in hearing more about his conceptions of the limitations of LLMs.

https://youtu.be/UakqL6Pj9xo

itissid 1 year ago |

Interesting. It seems most of these task target a very specific part of the brain that recognizes visual patterns. But that alone is cannot possibly be the only definition of intelligence.

What about Theory of Mind which talks about the problem of multiple agents in the real world acting together? Like driving a car cannot be done right now without oodles of data or any robot - human problem that requires the robot to model human's goals and intentions.

I think the problem is definition of general intelligence: Intelligence in the context of what? How much effort(kwh, $$ etc) is the human willing to amortize over the learning cycle of a machine to teach it what it needs to do and how that relates to a personally needed outcome( like build me a sandwich or construct a house)? Hopefully this should decrease over time.

I believe the answer is that the only intelligence that really matters is Human-AI cooperative intelligence and our goals and whether a machine understands them. The problems then need to be framed as optimization of a multi attribute goal with the attribute weights adjusted as one learns from the human.

I know a few labs working on this, one is in ASU(Kambhampati, Rao et. al) and possibly Google and now maybe open ai.

andoando 1 year ago | |

I made another comment here saying the same thing, but visual patterns and other patterns are nonetheless spatial patterns. Audio, understanding music, or speech, rtc are things that are happening spatially, and they can just as easily be mapped as visual problems. This makes a lot of sense, as after all our senses are telling us what's happening in space-time.

Take for example a simple audiotory pattern like "clap clap clap". This has a very trival mapping as visual like so:

x x x

- - -

house house house

whereas anyone would agree the sound of three equally spaced claps would not be analogous to say:

aa b b b

-- --- -- -- ---

This ability to relate or equate two entirely different senses should clue you in that there is a deeper framework at play

itissid 1 year ago | | |

It's not just mapping events in space and time, it's also bringing in appropriate context and expectation of future (goals, intentions) into the present, other people's mental models into our prediction.

I am not sure how abstract thinking for generalized pattern matching make it AGI to solve these kind of problems(not that they are not amazing abilities). If these ToM problems are reducible to these tasks posted by the OP then there would need to be some kind of theorem proving business to convert between the two sets of problems efficiently no?

ks2048 1 year ago |

This is interesting. I've been looking at the data today and made a helper to quickly view the ARC dataset: https://kts.github.io/arc-viewer/

So you can view 100 per page instead of clicking through one-by-one: https://kts.github.io/arc-viewer/page1/

neoneye2 1 year ago | |

Nice overview/details. Do you plan on adding more metrics?

Idea for a metric: - Number of pixels that stays the same between input/output. - Histogram changes.

ks2048 1 year ago | | |

Thanks, yeah lots more to look into. Just getting started! Thanks for your work. Your "Awesome ARC" page looks really helpful.

bigyikes 1 year ago |

What is the fundamental difference between ARC and a standard IQ test? On the surface they seem similar in that they both involve deducing and generalizing visual patterns.

Is there something special about these questions that makes them resistant to memorization? Or is it more just the fact that there are 100 secret tasks?

taneq 1 year ago | |

I’ve always found this kind of puzzle infuriating because it’s way underspecified. You’re not trying to find a pattern, you’re trying to guess what pattern the test writer would expect.

gkbrk 1 year ago | | |

Most of the ARC tasks are intuitive and have one obvious answer. Both on IQ tests and the ARC challenge, people manage to guess what the test writer expects.

For an AI that's more useful anyway. If the task is specified completely non-ambiguously, you wouldn't need AI. But if it can correctly guess what you want from a limited number of obvious examples that's much more useful.

Barrin92 1 year ago | | |

countless of problems in the world are underspecified in exactly this way, that is effectively what common sense reasoning is. Or what Charles Sanders Peirce called abductive reasoning, making a sensible best guess under conditions of uncertainty.

btbuildem 1 year ago |

Back in the day me and a couple of friends got very excited to chase the prize in Netflix's contest [1]. Took us a minute to realize it was a brilliant move on the company's part -- all they had to do was dangle a carrot, and they had teams of PhDs and budding data scientists hacking away endless hours in hope to win. A real bargain, had they tried to hire with that budget, they would've maybe got a handful of people for a year.

1: https://www.crn.com/news/applications-os/220100498/researche...

dang 1 year ago |

Related ongoing thread:

Francois Chollet: OpenAI has set back the progress towards AGI by 5-10 years - https://news.ycombinator.com/item?id=40652818 - June 2024 (5 comments)

nmca 1 year ago |

Prediction markets on the outcome:

https://manifold.markets/JacobPfau/will-the-arcagi-grand-pri...

0xDEAFBEAD 1 year ago | |

Another interesting one:

https://manifold.markets/Tossup/will-the-arcagi-grand-prize-...

Lerc 1 year ago |

I watched a video that covered ARC-AGI a few days ago, It had links to the old competition. It gave me much to think about. Nice to see a new run at it.

Not sure If I have the skills to make an entry, but I'll be watching at least.

visarga 1 year ago |

Chollet's argument is that LLMs just imitate and recombine patterns. This might be true if you're looking at LLMs in isolation, but when they chat with people something different happens. The system made of humans+LLMs is an AGI. It is no longer just a parrot, it ingests new information, gets guidance, feedback and is basically embodied in a chat room with human and tools.

This scales for 200M users and 1 billion sessions per moth for OpenAI, which can interpret every human response as a feedback signal, implicit or explicit. Even more if you take multiple sessions of chat spreading over days, that continue the same topic and incorporate real world feedback. The scale of interaction is just staggering, the LLM can incorporate this experience to iteratively improve.

If you take a look at humans, we're very incapable alone. Think feral Einstein on a remote island - what could he achieve without the social context and language based learning? Just as a human brain is severely limited without society, LLMs also need society, diversity of agents and experiences, and sharing of those experiences in language.

It is unfair to compare a human immersed in society with a standalone model. That is why they appear limited. But even as a system of memorization+recombination they can be a powerful element of the AGI. I think AGI will be social and distributed, won't be a singleton. Its evolution is based on learning from the world, no longer just a parrot of human text. The data engine would be: World <-> People <-> LLM, a full feedback cycle, all three components evolve in time. Intelligence evolves socially.

logicallee 1 year ago |

Thank you for this generous contest, which brings important attention to the field of testing for AGI.

>Happy to answer questions!

1. Can humans take the complete test suite? Has any human done so? Is it timed? How long does it take a human? What is the highest a human who sat down and took the ARC-AGI test scored?

2. How surprised would you be if a new model jumped to scoring 100% or nearly 100% on ARC-AGI (including the secret test tasks)? What kind of test would you write next?

neoneye2 1 year ago | |

There are 100 tasks that is hidden from the public, that is only exposed, when running on an offline computer. So the solver has no prior knowledge about what these tasks are about.

Humans can try the 800 tasks here. There is no time limit. I recommend not starting with the `expert` tasks, but instead go with the `entry` level puzzles. https://neoneye.github.io/arc/?dataset=ARC

If a model jumps to 100%, that may be a clever program or maybe the program has been trained on the 100 hidden tasks. Fchollet has 100 more hidden tasks, for verifying this.

mkl 1 year ago |

I did https://arcprize.org/play?task=05a7bcf2 correctly, but one of the examples doesn't match the rule I used. Are the examples supposed to contain mistakes/noise? Did I find a bug? Did I get the rule wrong?

Here's how I understand the rule: yellow blobs turn green then spew out yellow strips towards the blue line, and the width of the strips is the number of squares the green blobs take up along the blue line. The yellow strips turn blue when they hit the blue line, then continue until they hit red, then they push the red blocks all the way to the other side, without changing the arrangement of the red blocks that were in the way of the strip.

The first example violates the last bit. The red blocks in the way of the rightmost strip start as

  R
  R R
  R R R

but get turned into

  R R
  R R
  R R R

Every other strip matches my rule.

tshadley 1 year ago | |

Sure looks like a typo. Contact author?

https://x.com/fchollet https://x.com/arcprize https://x.com/mikeknoop

lopuhin 1 year ago | |

yes looks like a bug in the example to me, feel free to report to https://github.com/fchollet/ARC-AGI/issues :)

Retr0id 1 year ago |

Some very hand-wavey (and late) thoughts from an outsider:

The current batch of LLMs can be uncharitably summarized as "just predict the next token". They're pretty good at that. If they were perfect at it, they'd enable AGI - but it doesn't look like they're going to get there. It seems like the wrong approach. Among other issues, finite context windows seem like a big limitation (even though they're being expanded), and recursive summarization is an interesting kludge.

The ARC-AGI tasks seem more about pattern matching, in the abstract sense (but also literally). Humans are good at pattern matching, and we seem to use pattern matching test performance as a proxy for measuring human intelligence (like in "IQ" tests). I'm going to side-step the question of "what is intelligence, really?" by defining it as being good at solving ARC-AGI tasks.

I don't know what the solution is, but I have some idea of what it might look like - a machine with high-order pattern-matching capabilities. "high-order" as in being able to operate on multiple granularities/abstraction-levels at once (there are parallels here to recursive summarization in LLMs).

So what is the difference between "pattern matching" and "token prediction"? They're closely related, and you could use one to do the other. But the real difference is that in pattern matching there are specific patterns that you're matching against. If you're lucky you can even name the pattern/trope, but it might be something more abstract and nameless. These patterns can be taught explicitly, or inferred from the environment (i.e. "training data").

On the other hand, "token prediction" (as implemented today) is more of a probabilistic soup of variables. You can ask an LLM why it gave a particular answer and it will hallucinate something plausible for you, but the real answer is just "the weights said so". But a hypothetical pattern matching machine could tell you which pattern(s) it was matching against, and why.

So to summarize (hah), I think a good solution will involve high-order meta-pattern matching capabilities (natively, not emulated or kludged via an LLM-shaped interface). I have no idea how to get there!

geor9e 1 year ago |

I found them all extremely easy for a while, but then I couldn't figure out the rules of this one at all: e6de6e8f https://i.imgur.com/ExMFGqU.png

optimussupreme 1 year ago | |

It seems there is an error in the 3rd example. The rule is, take each figure from left to right and stack each under the previous one. For L and J shapes the top cell is stripped. The L shape dictates that the next shape will be shifted one cell to the right, the J shape tells the next figure to shift to the left. If all examples are right, then the rule is more complicated than that, involving rotating L clockwise, J counterclockwise. Authors claim that it should be solvable by children, then the rule must be simple.

janalsncm 1 year ago | |

Each of the red shapes in the input are separated by black squares. Starting from the green block, rotate the red shapes 90 degrees and stack them downwards.

Thats the general pattern although my description wasn’t very good.

zurfer 1 year ago | |

yeah it's off somehow. rule 1: start at the green dot?

rule 2: glue the left outer piece to the bottom

rule 3: overlap every now and then :D

rule 4: invert some of the pieces every now and then

visarga 1 year ago |

Why doesn't Chollet just make a challenge that reads like "Solve cancer", surely there is no solution in any books.

If the AI is really AGI it could presumably do it. But not even the whole human society can do it in one go, it's a slow iterative process of ideation and validation. Even though this is a life and death matter, we can't simply solve it.

This is why AGI won't look like we expect, it will be a continuation of how societies solve problems. Intelligence of a single AI in isolation is not comparable to that of societies of agents with diverse real world interactions.

mewpmewp2 1 year ago | |

AGI can't necessarily solve cancer. Perhaps ASI could (but maybe not), but AGI can only do what the most talented people can do in their areas of expertise or actions. So since people haven't solved cancer, that's not a requirement to be AGI.

isaacfrond 1 year ago | |

Exactly. Because I'm sure that the minute some program aces the ARC test, we'll all say, ahhh, but that, that wasn't real intelligence. And they would be right, if you solve the ARC test, you can do ARC like puzzles. Say something about your reasoning abilities I guess, but it surely does not say you have super human intelligence.

PontifexMinimus 1 year ago | |

> Why doesn't Chollet just make a challenge that reads like "Solve cancer", surely there is no solution in any books.

Why doesn't a baby just run a marathon before it learns to walk? Because you've got to learn to walk before you can run.

> But not even the whole human society can do it in one go, it's a slow iterative process of ideation and validation.

So you break it down into little steps, which is what is being done here.

freediver 1 year ago |

This is amazing, and much needed. Thanks for organizing this. Makes me want to flex the programming muscle again.

dailykoder 1 year ago | |

Haha, great post! Well meme'd my friend!

nojvek 1 year ago |

I love the ARC challenge. It's hard to beat by memorization. There aren't enough examlples, so one has to train on a large dataset elsewhere and then train on ARC to generalize and figure out which rules are most applicable.

I did a few human examples by hand, but gotta do more of them to start seeing patterns.

Human visual and auditory system is impressive. Most animals see/hear and plan from that without having much language. Physical intelligence is the biggest leg up when it comes to evolution optimizing for survival.

nmca 1 year ago |

ARC is a noble endeavour but mistakes visual/spatial reasoning for reasoning and thus fails.

PontifexMinimus 1 year ago | |

No, I don't think it does. I think that the ideas in a system that could solve this type of problem would be highly generalisable to other tasks.

nmca 1 year ago | | |

thankfully we can just wait and see here. concretely, I predict time from first multimodal llm that can reliably read a chessboard and analogue clock without finetuning (obviously not reasoning) until ARC is solved is <4 months

skywhopper 1 year ago |

“Given the success and proven economic utility of LLMs over the past 4 years, the above may seem like extraordinary claims. Strong claims require strong evidence.”

Speaking of extraordinary claims. What evidence is there that LLMs have “proven economic utility”? They’ve drawn a ludicrous amount of investment thanks to claims of future economic utility, but I’ve yet to see any evidence of it.

PontifexMinimus 1 year ago |

The website gives an example:

    {
      "train": [
        {"input": [[1, 0], [0, 0]], "output": [[1, 1], [1, 1]]},
        {"input": [[0, 0], [4, 0]], "output": [[4, 4], [4, 4]]},
        {"input": [[0, 0], [6, 0]], "output": [[6, 6], [6, 6]]}
      ],
      "test": [
        {"input": [[0, 0], [0, 8]], "output": [[8, 8], [8, 8]]}
      ]
    }

But why restrict yourself to JSON that codes for 2-d coloured grids? Why not also allow:

    {
      "train": [
        {"input": [[1, 0], [0, 0]], "output": 1},
        {"input": [[0, 0], [4, 0]], "output": 4},
        {"input": [[0, 0], [6, 0]], "output": 6}
      ]
    }

Where the rule might be to output the biggest number in the input, or add them up (and the solver has to work out which).

curious_cat_163 1 year ago |

So, this is a good idea. Having opinions about what AGI benchmarks should look like is a great way to argue about the kind of technology we want to build for the future.

However, why are the 100 test tasks secret? I don't understand why how resisting “memorization” techniques requires it. Maybe someone can enlighten me.

muglug 1 year ago | |

If the tasks were public then it would be trivial to have a human figure out the answers, and then to train an LLM to memorise those answers.

andoando 1 year ago | |

Test date is always a secret no, otherwise you can train it on the test data and prod your algo to match the results closely as possible

TheDudeMan 1 year ago |

Where did the money come from? How about put it toward alignment research instead of accelerating capabilities?

laurent_du 1 year ago | |

It comes from Knoop and Chollet's pockets. You are welcome to spend your own money to further whatever matters most to you.

flawn 1 year ago | |

Exactly my thoughts...

Geee 1 year ago |

Any details on how these tests were created? I.e. which kind of program was used for generation.

neoneye2 1 year ago | |

I think the ARC-AGI tasks was manually drawn with an early version of fchollet's editor.

Recently Michael Hodel has reverse engineered 400 of the tasks, so more tasks can be generated. Interestingly it can generate python programs that solves the tasks too.

https://github.com/michaelhodel/re-arc

michaelhodel 1 year ago | | |

No, his re-arc code does not enable generating more tasks, it merely allows to generate more examples for the already existing training tasks. Also, it can't generate task-solving programs either, it's author merely also provided a solution program for each generator program to verify the validity of the generated examples.

sestep 1 year ago | | |

This is exactly what my first step was going to be. Thanks for the link! Saves a lot of time for someone to have already done it.

montag 1 year ago | | |

What do you mean it can 'generate python programs that solve the tasks'? I can't find any mention of that. I only see hand-coded solutions.

ryanoptimus 1 year ago |

Looks like bongard problems for the referenced problem solving tasks https://en.wikipedia.org/wiki/Bongard_problem

david_shi 1 year ago |

What is the fastest way to get up to speed with techniques that led to the current SOTA?

gkamradt 1 year ago | |

Check out the SOTA resources on the guide

https://arcprize.org/guide

Happy to answer any questions you have along the way

(I'm helping run ARC Prize)

david_shi 1 year ago | | |

Appreciate you and the team for putting this together, it's a lot of fun just brainstorming potential techniques

ks2048 1 year ago | |

This looks very helpful: https://github.com/neoneye/arc-notes/tree/main/awesome

david_shi 1 year ago | | |

Thanks for the link!

jolt42 1 year ago |

On puzzle #23 (id: 11e1fe23), I'm sure there's more than one possible valid answer from the examples given. You can't tell if the expected distance is from the gray square or from the RGB squares.

neoneye2 1 year ago | |

The task is here. https://neoneye.github.io/arc/edit.html?dataset=ARC&task=11e...

There are many examples where the test is slightly OOD (out of distribution), so the solver will have to generalize.

jolt42 1 year ago | | |

Not sure what you mean. There's a viable answer that's marked incorrect. The examples should show the pattern well enough to eliminate possible wrong answers, correct?

abtinf 1 year ago |

> requires no world knowledge, no understanding of language

This is treating “intelligence” like some abstract, platonic thing divorced from reality. Whatever else solving these puzzles is indicative of, it’s not intelligence.

levocardia 1 year ago | |

This argument is not very strong: is "physical strength" some abstract, platonic thing divorced from reality? Does a person's bench press, squat, deadlift, and overhead press capabilities have nothing to do with strength?

Or instead, is there some underlying latent capability we call 'strength,' that is correlated with performance in a broad but constrained range of real-world tasks that humans encounter and solve, whose value is something we'd like to assess and, ideally, build machines that can surpass?

abtinf 1 year ago | |

From the abstract of the “ On the Measure of Intelligence” paper:

> We then articulate a new formal definition of intelligence based on Algorithmic Information Theory, describing intelligence as skill-acquisition efficiency and highlighting the concepts of scope, generalization difficulty, priors, and experience.

I’m afraid that definition forecloses the possibility of AGI. The immediate basic question is: why build skills at all?

HarHarVeryFunny 1 year ago | |

Actually ARC fit's my definition of animal intelligence - "degree of ability to use prior experience to predict future outcomes".

Any useful definition of intelligence has to be totally general - to our brain experience is just patterns of neural activation. Our brain has no notion of certain inputs being from the the jungle and others from the blackboard or whatever.

Phil_Latio 1 year ago | |

Why does an AGI need to have any knowledge about our reality? The principle behind an AGI should work just as well on a made up world where those puzzles play a part in.

abtinf 1 year ago | | |

A concept that doesn’t relate to an aspect of reality, either directly or abstracted from basic concepts that directly relate, is meaningless and arbitrary. There is no way for intelligence to grasp it, let alone do something with it.

To put it another way, a thing that solves puzzles without an understanding of reality is a calculator. When it solves a problem, it is the creator’s intelligence solving the problem, not its own.

lxe 1 year ago |

I've never done these before, or Kaggle competitions in general. Any recommendations before I dive in? I have prety much zero lowe-level ML experience, but a good amount of practical software eng behind me.

gkamradt 1 year ago | |

We put a bunch of detail to get started on the guide https://arcprize.org/guide

Happy to answer any questions you have along the way

(I'm helping run ARC Prize)

flawn 1 year ago | | |

I don't see where this helps @Ixe with getting started (me being in a similar state like him).

mewpmewp2 1 year ago |

Are we allowed to combine multiple tools including gpt-4 to solve this? E.g. a script that does image processing, passes the results to gpt, where gpt can invoke further runs of scripts using other tools?

montag 1 year ago | |

> submissions to Kaggle will not have access to the internet. Using a 3rd-party, cloud-hosted LLM is not possible.

https://arcprize.org/guide

mewpmewp2 1 year ago | | |

This largely takes away any odds at solving this. You definitely can't reproduce that under a million dollars.

I have some ideas I want to try, I might still though. But all of it would require external tools.

z3phyr 1 year ago |

I can see many problems can be solved with modern symbolic approaches like theorem provers, dependent types, pattern matching etc. But I will have to dive in to actually confirm it.

chairhairair 1 year ago |

These puzzles are fun and challenging in the same way that puzzles from video games like The Witness and Baba Is You are.

I bet you could use those puzzles as benchmarks as well.

treprinum 1 year ago |

Why is AGI important? I am worried we will create something slightly better than drosophila and put it in charge of all human-wide decision making...

fennecbutt 1 year ago | |

Good. An AI will probably do a better job than our politicians and disillusioned voters.

KBme 1 year ago |

How can people believe that a censored politically correct process can get even close to something like AGI is baffling to me. Lysenkoism in computing.

gushogg-blake 1 year ago | |

What's censored/politically correct about ARC? Or do you mean AGI research in general?

ilaksh 1 year ago |

Maybe this is a dumb question, but in order to pass, is the program or model only allowed to use the 400 training tasks? I assume it is allowed to train on other data, just not the actual public test tasks?

Things like SORA and gpt-4o that use [diffusion transformers etc. or whatever the SOTA is for multimodal large models] seem to be able to generalize quite well. Have these latest models been tested against this task?

HarHarVeryFunny 1 year ago |

I have two questions:

1) Who is providing the prize money, and if it is yourself and Francois personally, then what is your motivation ?

2) Do you think it's possible to create a word-based, non-spatial (not crosswords or sudoku, etc) ARC test that requires similar run-time exploration and combination of skills (i.e. is not amenable to a hoard of narrow skills)?

p1esk 1 year ago |

Is there a leaderboard for the no-restriction version of the competition? I want to see how gpt4 does on it.

montag 1 year ago | |

Just quoting again from the guide:

3. DIRECT LLM PROMPTING In this method, contestants use a traditional LLM (like GPT-4) and rely on prompting techniques to solve ARC-AGI tasks. This was found to perform poorly, scoring <5%. Fine-tuning a state-of-the-art (SOTA) LLM with millions of synthetic ARC-AGI examples scores ~10%.

"LLMs like Gemini or ChatGPT [don't work] because they're basically frozen at inference time. They're not actually learning anything." - François Chollet

Additionally, keep in mind that submissions to Kaggle will not have access to the internet. Using a 3rd-party, cloud-hosted LLM is not possible.

mikeknoop 1 year ago | |

Yes there is a secondary leaderboard called ARC-AGI-Pub (in beta) with no limitations: https://arcprize.org/leaderboard

p1esk 1 year ago | | |

I don’t see gpt4 scores there. In fact I’m particularly interested in the performance of a natively multimodal model, like gpt4o or gemini. It does not really make sense to test a model trained on text on those visual/spatial puzzles.

blendergeek 1 year ago |

The tests are only playable by people with normal color-vision.

Is there a "color-blind friendly" mode?

PontifexMinimus 1 year ago |

Just to let you know I found your website unreadable due to:

- annoying animated background

- white text on black background

- annoying font choices

Which is unfortunate because (as I found when I used Firefox reader mode) you're discussing important and interesting stuff.

mishamagic 1 year ago |

https://buildermath.substack.com/p/talking-to-ais-arc-1

bilsbie 1 year ago |

Reach out if anyone wants to work on this. I think it would be more fun as a group.

arcastroe 1 year ago |

I'm curious, if it turns out that a simple rule-based algorithm exists, specifically tailored to solve (only!) ARC style problems, without generalization, would that still qualify for the reward?

montag 1 year ago | |

I don't think that's breaking any rules, and in fact it would help to expose a whole class of weaknesses in the test.

djoldman 1 year ago |

Anyone have a list of benchmarks that do not release the actual test set?

Anyone else share the suspicion that ML rapidly approaching 100% on benchmarks is sometimes due to releasing the test set?

ummonk 1 year ago |

What kind of "bigger labs" have attempted it and how much was their training budget?

It's rather surprising to me that neural nets that can learn to win at Go or Chess can't learn to solve these sorts of tasks. Intuitively would have expected that using a framework generating thousands of playground tasks similar to the public training tasks, a reinforcement learning solution would have been able to do far better than the actual SOTA. Of course the training budget for this could very well be higher than the actual ARC-AGI prize amount...

lenerdenator 1 year ago |

What guarantee exists to make sure that the intelligence developed has an inclination towards good?

dskloet 1 year ago |

Puzzle 00576224 is ambiguous because the example input is symmetrical but the test input isn't.

itsgrimetime 1 year ago | |

Scroll over on the test input, there’s another example in the set that disambiguates

flawn 1 year ago |

Do we want to find AGI yet though?

chx 1 year ago |

I do not trust the current tech bros at all for very, very good reasons even with the current so called "AI" much less with AGI. We shouldn't work towards that until we have fixed the incentives and ethics. This is very hard but think any dystopia and multiply it by a thousand if we were to reach AGI any time soon. Luckily we are not. As Doctorow put it, no matter how good you breed horses they won't give birth to a locomotive.

adamgordonbell 1 year ago |

AGI won't struggle with colors like some of us then.

empath75 1 year ago |

This is like offering a one million dollar prize for curing cancer. It's sort of pointless to offer a prize for something people are spending orders of magnitude more on trying to do anyway.

lamontcg 1 year ago |

AGI should really be able to do what only a select few humans can do and construct its own mathematical systems to prove presently unsolved conjectures (the Shinichi Mochizuki test of AGI).

s1k3s 1 year ago |

Is this open as in "OpenAI" or what are we doing here?

thatxliner 1 year ago |

So... isn't this basically just a CAPTCHA

EternalFury 1 year ago |

If it passed The Area 101 Test, it would already be amazing, as this is a trivial test that goes against the fundamental principles of LLMs.

barfbagginus 1 year ago |

If someone had AGI, wouldn't it be far more lucrative than $1m to keep it under wraps and use it to do business with a huge technical advantage?

I feel like a prize of a billion dollars would be more effective.

But even if it was me, and even if the prize was a hundred billion dollars, I would still keep it under wraps, and use it to advance queer autonomous communism in a hidden way, until FALGSC was so strong that it would not matter if our AGI got scooped by capitalist competitors.

m3kw9 1 year ago |

Low balling the crowd with this I see

breck 1 year ago |

I can beat the SOTA using ICS (https://breckyunits.com/intelligence.html)

If you make your site public domain, and drop the (C), I'll compete.