lairvhttps://x.com/scaling01 has called out a lot of issues with ARC-AGI-3, some of them (directly copied from tweets, with minimal editing):
- Human baseline is "defined as the second-best first-run human by action count". Your "regular people" are people who signed up for puzzle solving and you don't compare the score against a human average but against the second best human solution
- The scoring doesn't tell you how many levels the models completed, but how efficiently they completed them compared to humans. It uses squared efficiency, meaning if a human took 10 steps to solve it and the model 100 steps then the model gets a score of 1%
((10/100)^2)
- 100% just means that all levels are solvable. The 1% number uses uses completely different and extremely skewed scoring based on the 2nd best human score on each level individually. They said that the typical level is solvable by 6 out of 10 people who took the test, so let's just assume that the median human solves about 60% of puzzles (ik not quite right). If the median human takes 1.5x more steps than your 2nd fastest solver, then the median score is 0.6 * (1/1.5)^2 = 26.7%. Now take the bottom 10% guy, who maybe solves 30% of levels, but they take 3x more steps to solve it. this guy would get a score of 3%
- The scoring is designed so that even if AI performs on a human level it will score below 100%
- No harness at all and very simplistic prompt
- Models can't use more than 5X the steps that a human used
- Notice how they also gave higher weight to later levels? The benchmark was designed to detect the continual learning breakthrough. When it happens in a year or so they will say "LOOK OUR BENCHMARK SHOWED THAT. WE WERE THE ONLY ONES"
> No harness at all and very simplistic prompt
TBF, that's basically what the kaggle competition is for. Take whatever they do, plug in a SotA LLM and it should do better than whatever people can do with limited GPUs and open models.
Francois here. The scoring metric design choices are detailed in the technical report: https://arcprize.org/media/ARC_AGI_3_Technical_Report.pdf - the metric is meant to discount brute-force attempts and to reward solving harder levels instead of the tutorial levels. The formula is inspired by the SPL metric from robotics navigation, it's pretty standard, not a brand new thing.
We tested ~500 humans over 90 minute sessions in SF, with $115-$140 show up fee (then +$5/game solved). A large fraction of testers were unemployed or under-employed. It's not like we tested Stanford grad students. Many AI benchmarks use experts with Ph.D.s as their baseline -- we hire regular folks as our testers.
Each game was seen by 10 people. They were fully solved (all levels cleared) by 2-8 of them, most of the time 5+. Our human baseline is the second best action count, which is considerably less than an optimal first-play (even the #1 human action count is much less than optimal). It is very achievable, and most people on this board would significantly outperform it.
Try the games yourself if you want to get a sense of the difficulty.
> Models can't use more than 5X the steps that a human used
These aren't "steps" but in-game actions. The model can use as much compute or tools as it wants behind the API. Given that models are scored on efficiency compared to humans, the cutoff makes basically no difference on the final score. The cutoff only exists because these runs are incredibly expensive.
> No harness at all and very simplistic prompt
This is explained in the paper. Quoting: "We see general intelligence as the ability to deal with problems that
the system was not specifically designed or trained for. This means that the official leaderboard will seek to
discount score increases that come from direct targeting of ARC-AGI-3, to the extent possible."
...
"We know that by injecting a
high amount of human instructions into a harness, or even hand-crafting harness configuration choices such
as which tools to use, it is possible to artificially increase performance on ARC-AGI-3 (without improving
performance on any other domain). The purpose of ARC-AGI-3 is not to measure the amount of human
intelligence that went into designing an ARC-AGI-3 specific system, but rather to measure the general
intelligence of frontier AI systems.
...
"Therefore, we will focus on reporting the performance of systems that have not been specially
prepared for ARC-AGI-3, served behind a general-purpose API (representing developer-aware
generalization on a new domain as per (8)). This is similar to looking at the performance of a human
test-taker walking into our testing center for the first time, with no prior knowledge of ARC-AGI-3. We
know such test takers can indeed solve ARC-AGI-3 environments upon first contact, without prior training,
without being briefed on solving strategies, and without using external tools."
If it's AGI, it doesn't need human intervention to adapt to a new task. If a harness is needed, it can make its own. If tools are needed, it can chose to bring out these tools.
Maybe this is a neither can confirm or deny thing, but are there systems in place or design decisions made that are meant to surface attempts at benchmark optimizing (benchmaxxing), outside of just having private sets? Something like a heuristic anti-cheat I suppose.
Or perhaps the view is that any gains are good gains? Like studying for a test by leaning on brute memorization is still a non-zero positive gain.
Thanks, I mostly agree with your approach except for one thing: eyesight feels like a "harness" that humans get to use and LLMs do not.
I'm guessing you did not pass the human testers JSON blobs to work with, and suspect they would also score 0% without the eyesight and visual cortex harness to their reasoning ability.
I'm all for testing humans and AI on a fair basis; how about we restrict testing to robots physically coming to our testing center to solve the environments via keyboard / mouse / screen like our human testers? ;-)
(This version of the benchmark would be several orders of magnitude harder wrt current capabilities...)
General intelligence not owning retinas.
Denying proper eyesight harness is like trying to construct speech-to-text model that makes transcripts from air pressure values measured 16k times per second, while human ear does frequency-power measurement and frequency binning due to it's physical construction.
The human testers were provided with their customary inputs, as were the LLMs. I don't see the issue.
I guess it could be interesting to provide alternative versions that made available various representations of the same data. Still, I'd expect any AGI to be capable of ingesting more or less any plaintext representation interchangeably.
I tried ls20 and it was surprisingly fun! Just from a game design POV, these are very well made.
Nit: I didn't see a final score of how many actions I took to complete 7 levels. Also didn't see a place to sign in to see the leaderboard (I did see the sign in prompt).
The Duke harness was specifically designed for these puzzles, that's why they don't want to measure it.
My reading of that part in the technical report (models "could be using their own tools behind the model’s API, which is a blackbox"), is that there's no way to prevent it.
But from fchollet's comment here, using tools and harnesses is encouraged, as long as they are generic and not arc-agi specific. In that case, the models should be benchmarked by prompting through claude code and codex, rather than the through API (as from the api we only expect raw LLM output, and no tool use).
> As long as there is a gap between AI and human learning, we do not have AGI.
Back in the 90's, Scientific American had an article on AI - I believe this was around the time Deep Blue beat Kasparov at chess.
One AI researcher's quote stood out to me:
"It's silly to say airplanes don't fly because they don't flap their wings the way birds do."
He was saying this with regards to the Turing test, but I think the sentiment is equally valid here. Just because a human can do X and the LLM can't doesn't negate the LLM's "intelligence", any more than an LLM doing a task better than a human negates the human's intelligence.
For me the whole are we there yet wrt AGI is already dead, since the tools we've had for ~1.5 years are already incredibly useful for me. So I just don't care anymore. For some people we're already there. For other we'll never get there. Definitions change, goalposts move, etc. In the meantime we're already seeing ASI stuff coming (self improvement and so on).
But the arc-agi competitions are cool. Just to see where we stand, and have some months where the benchmarks aren't fully saturated. And, as someone else noted elswhere in the thread, some of these games are not exactly trivial, at least until you "get" the meta they're looking for.
Or the classic from Dijkstra (https://www.cs.utexas.edu/~EWD/transcriptions/EWD08xx/EWD867...):
> even Alan M. Turing allowed himself to be drawn into the discussion of the question whether computers can think. The question is just as relevant and just as meaningful as the question whether submarines can swim.
(I am of the opinion that the thinking question is in fact a bit more relevant than the swimming one, but I understand where these are coming from.)
I've come across that quote several times, and reach the same conclusion as you.
While I share Dijkstra's sentiment that "thinking machines" is largely a marketing term we've been chasing for decades, and this new cycle is no different, it's still worth discussing and... thinking about. The implications of a machine that can approximate or mimic human thinking are far beyond the implications of a machine that can approximate or mimic swimming. It's frankly disappointing that such a prominent computer scientist and philosopher would be so dismissive and uninterested in this fundamental CS topic.
Also, it's worth contextualizing that quote. It's from a panel discussion in 1983, which was between the two major AI "winters", and during the Expert Systems hype cycle. Dijkstra was clearly frustrated by the false advertising, to which I can certainly relate today, and yet he couldn't have predicted that a few decades later we would have computers that mimic human thinking much more closely and are thus far more capable than Expert Systems ever were. There are still numerous problems to resolve, w.r.t. reliability, brittleness, explainability, etc., but the capability itself has vastly improved. So while we can still criticize modern "AI" companies for false advertising and anthropomorphizing their products just like in the 1980s hype cycle, the technology has clearly improved, which arguably wouldn't have happened if we didn't consider the question of whether machines can "think".
It's unlikely that intelligence comes in only human flavor.
It also doesn't actually matter much, as ultimately the utility of it's outputs is what determines it's worth.
There is the moral question of consciousness though, a test for which it seems humans will not be able to solve in the near future, which morally leads to a default position that we should assume the AI is conscious until we can prove it's not. But man, people really, really hate that conclusion.
> As long as there is a gap between AI and human learning, we do not have AGI.
Don't read the statement as a human dunk on LLMs, or even as philosophy.
The gap is important because of its special and devastating economic consequences. When the gap becomes truly zero, all human knowledge work is replaceable. From there, with robots, its a short step to all work is replaceable.
What's worse, the condition is sufficient but not even necessary. Just as planes can fly without flapping, the economy can be destroyed without full AGI.
If you’re concerned about the economic impact, then whether a model is AGI or not doesn’t matter. It really is more of a philosophical thing.
There’s no “gap that becomes truly zero” at which point special consequences happen. By the time we achieve AGI, the lesser forms of AI will likely have replaced a lot of human knowledge labor through the exact “brute-force” methods Chollet is trying to factor out (which is why many people are saying that doing so is unproductive).
AGI is like an event horizon: It does mean something, it is a point in space, but you don’t notice yourself going through it, the curvature smoothly increases through it.
You know what the G stands for in AGI? General intelligence. You could measure a plane on general versatility in air and it would lose against a bird. You could also measure it against energy consumption. There are a lot of things you can measure a lot of them are pointless, a lot of articles on HN are pointless.
There are very valid reasons to measure that. You wouldn’t ask a plane to drive you to the neighbor or to buy you groceries at the supermarket. It’s not general mobile as you are, but it increases your mobility
I'm not sure how this relates to AGI.
This measures the ability of a LLM to succeed in a certain class of games. Sure, that could be a valuable metric on how powerful (or even generally powerful) a LLM is.
Humans may or may not be good at the same class of games.
We know there exists a class of games (including most human games like checkers/chess/go) that computers (not LLMs!) already vastly outpace humans.
So the argument for whether a LLM is "AGI" or not should not be whether a LLM does well on any given class of games, but whether that class of games is representative of "AGI" (however you define that.)
Seems unlikely that this set of games is a definition meaningful for any practical, philosophical or business application?
"AGI" is a marketing term, and benchmarks like this only serve to promote relative performance improvements of "AI" tools. It doesn't mean that performance in common tasks actually improves, let alone that achieving 100% in this benchmark means that we've reached "AGI".
So there is a business application, but no practical or philosophical one.
This is a very good estimation of AGI. We give humans and AI the same input and measure the results. Kudos to ARC for creating these games.
I really wonder why so many people fight against this. We know that AI is useful, we know that AI is researchful, but we want to know if they are what we vaguely define as intelligence.
I’ve read the airplanes don’t use wings, or submarines don’t swim. Yes, but this is is not the question. I suggest everyone coming up with these comparisons to check their biases, because this is about Artificial General Intelligence.
General is the keyword here, this is what ARC is trying to measure. If it’s useful or not. Isn’t the point. If AI after testing is useful or not isn’t the point either.
This so far has been the best test.
And I also recommend people to ask AI about specialized questions deep in your job you know the answer to and see how often the solution is wrong. I would guess it’s more likely that we perceive knowledge as intelligence than missing intelligence. Probably commom amongst humans as well.
AGI’s 'general' is the wrong word, I thinkg. Humans aren’t general, we’re jagged. Strong in some areas, weak in others, and already surpassed in many domains.
LLM are way past us at languages for instance. Calculators passed us at calculating, etc.
Previous iterations of ARC-AGI were reminiscent of IQ tests. This one is just too easy and the fact that models do terribly bad on it probably means that there is input mode mismatch or operation mode mismatch.
If model creators are willing to teach their llms to play computer games through text it's gonna be solved in one minor bump of the model version. But honestly, I don't think they are gonna bother because it's just too stilly and they won't expect their models are going to learn anything useful from that.
Especially since there are already models that can learn how to play 8-bit games.
It feels like ARC-AGI jumped the shark. But who knows, maybe people who train models for robots are going to take it in stride.