I feel like I'm the only one who isn't convinced getting a high score on the ARC eval test means we have AGI. It's mostly about pattern matching (and some of it ambiguous even for humans what the actual true response aught to be). It's like how in humans there's lots of different 'types' of intelligence, and just overfitting on IQ tests doesn't in my mind convince me a person is actually that smart.
Getting a high score on ARC doesn't mean we have AGI and Chollet has always said as much AFAIK, it's meant to push the AI research space in a positive direction. Being able to solve ARC problems is probably a pre-requisite to AGI. It's a directional push into the fog of war, with the claim being that we should explore that area because we expect it's relevant to building AGI.
We don't really have a true test that means "if we pass this test we have AGI" but we have a variety of tests (like ARC) that we believe any true AGI would be able to pass. It's a "necessary but not sufficient" situation. Also ties directly to the challenge in defining what AGI really means. You see a lot of discussions of "moving the goal posts" around AGI, but as I see it we've never had goal posts, we've just got a bunch of lines we'd expect to cross before reaching them.
> Getting a high score on ARC doesn't mean we have AGI and Chollet has always said as much AFAIK
He only seems to say this recently, since OpenAI cracked the ARC-AGI benchmark. But in the original 2019 abstract he said this:
> We argue that ARC can be used to measure a human-like form of general fluid intelligence and that it enables fair general intelligence comparisons between AI systems and humans.
Now he seems to backtrack, with the release of harder ARC-like benchmarks, implying that the first one didn't actually test for really general human-like intelligence.
This sounds a bit like saying that a machine beating chess would require general intelligence -- but then adding, after Deep Blue beats chess, that chess doesn't actually count as a test for AGI, and that Go is the real AGI benchmark. And after a narrow system beats Go, moving the goalpost to beating Atari, and then to beating StarCraft II, then to MineCraft, etc.
At some point, intuitively real "AGI" will be necessary to beat one of these increasingly difficult benchmarks, but only because otherwise yet another benchmark would have been invented. Which makes these benchmarks mostly post hoc rationalizations.
A better approach would be to question what went wrong with coming up with the very first benchmark, and why a similar thing wouldn't occur with the second.
ARC is definitely about achieving AGI and it doesn't matter whether we "have" it or not right now. That is the goal:
> where he introduced the "Abstract and Reasoning Corpus for Artificial General Intelligence" (ARC-AGI) benchmark to measure intelligence
So, a high enough score is a threshold to claim AGI. And, if you use an LLM to work these types of problems, it becomes pretty clear that passing more tests indicates a level of "awareness" that goes beyond rational algorithms.
I thought I had seen everything until I started working on some of the problems with agents. I'm still sorta in awe about how the reasoning manifests. (And don't get me wrong, LLMs like Claude still go completely off the rails where even a less intelligent human would know better.)
"Being able to solve ARC problems is probably a pre-requisite to AGI." - is it? Humans have general intelligence and most can't solve the harder ARC problems.
My problem with AGI is the lack of a simple, concrete definition.
Can we formalize it as giving out a task expressible in, say, n^m bytes of information that encodes a task of n^(m+q) real algorithmic and verification complexity -- then solving that task within a certain time, compute, and attempt bounds?
Something that captures "the AI was able to unwind the underlying unspoken complexity of the novel problem".
I feel like one could map a variety of easy human "brain teaser" type tasks to heuristics that fit within some mathematical framework and then grow the formalism from there.
I'm all for benchmarks that push the field forward, but ARC problems seem to be difficult for reasons having less to do with intelligence and more about having a text system that works reliably with rasterized pixel data presented line by line. Most people would score 0 on it if they were shown the data the way an LLM sees it, these problems only seem easy to us because there are visualizers slapped on top.
In the video, François Chollet, creator of the ARC benchmarks, says that beating ARC does not equate to AGI. He specifically says they will be able to be beaten without AGI.
I think the people behind the ARC Prize agree that getting a high score doesn't mean we have AGI. (They already updated the benchmark once to make it harder.) But an AGI should get a similarly high score as humans do. So current models that get very low scores are definitely not AGI, and likely quite far away from it.
> I think the people behind the ARC Prize agree that getting a high score doesn't mean we have AGI
The benchmark was literally called ARC-AGI. Only after OpenAI cracked it, they started backtracking and saying that it doesn't test for true AGI. Which undermines the whole premise of a benchmark.
But on a serious note, I don't think Chollet would disagree. ARC is a necessary but not sufficient condition, and he says that, despite the unfortunate attention-grabbing name choice of the benchmark. I like Chollet's view that we will know that AGI is here when we can't come up with new benchmarks that separate humans from AI.
I agree with you but I'll go a step further - these benchmarks are a good example of how far we are from AGI.
A good base test would be to give a manager a mixed team of remote workers, half being human and half being AI, and seeing if the manager or any of the coworkers would be able to tell the difference. We wouldn't be able to say that AI that passed that test would necessarily be AGI, since we would have to test it in other situations. But we could say that AI that couldn't pass that test wouldn't qualify, since it wouldn't be able to successfully accomplish some tasks that humans are able to.
But of course, current AI is nowhere near that level yet. We're left with benchmarks, because we all know how far away we are from actual AGI.
The AGI test I think makes sense is to put it in a robot body and let it navigate the world. Can I take the robot to my back yard and have it weed my vegetable garden? Can I show it how to fold my laundry? Can I take it to the grocery store and tell it "go pick up 4 yellow bananas and two avocados that will be ready to eat in the next day or two, and then meet me in dairy"? Can I ask it to dice an onion for me during meal prep?
These are all things my kids would do when they were pretty young.
The problem with "spot the difference" tests, imho, is that I would expect an AGI to be easily spotted. There's going to be a speed of calculation difference, at the very least. If nothing else, typing speed would be completely different unless the AGI is supposed to be deceptive. Who knows what it's personality would be like. I'd say it's a simple enough test just to see if an AGI could be hired as, for example, an entry level software developer and keep it's job based on the same criteria base-level humans have to meet.
I agree that current AI is nowhere near that level yet. If AI isn't even trying to extract meaning from the words it smiths or the pictures it diffuses then it's nothing more than a cute (albeit useful) parlor trick.
Give the AI tools and let it do real stuff in the world:
"FounderBench": Ask the AI to build a successful business, whatever that business may be - the AI decides. Maybe try to get funded by YC - hiring a human presenter for Demo Day is allowed. They will be graded on profit / loss, and valuation.
Testing plain LLM on whiteboard-style question is meaningless now. Going forward, it will all be multi-agent systems with computer use, long-term memory & goals, and delegation.
This sounds like a terrible idea to me, you're training intelligent computer to aim for power. It's fine as long as they're bad but if they get good then we have a problem
I find the "what is intelligence?" discussion a little pointless if I'm honest. It's similar to asking a question like does it mean to be a "good person" and would we know whether an AI or person is really "good"?
While understanding why a person or AI is doing what it's doing can be important (perhaps specifically in safety contexts) at the end of the day all that's really going to matter to most people is the outcomes.
So if an AI can use what appears to be intelligence to solve general problems and can act in ways that are broadly good for society, whether or not it meets some philosophical definition of "intelligent" or "good" doesn't matter much – at least in most contexts.
That said, my own opinion on this is that the truth is likely in between. LLMs today seem extremely good at being glorified auto-completes, and I suspect most (95%+) of what they do is just recalling patterns in their weights. But unlike traditional auto-completes they do seem to have some ability to reason and solve truly novel problems. As it stands I'd argue that ability is fairly poor, but this might only represent 1-2% of what we use intelligence for.
If I were to guess why this is I suspect it's not that LLM architecture today is completely wrong, but that the way LLMs are trained means that in general knowledge recall is rewarded more than reasoning. This is similar to the trade-off we humans have with education – do you prioritise the acquisition of knowledge or critical thinking? Maybe believe critical thinking is more important and should be prioritised more, but I suspect for the vast majority of tasks we're interested in solving knowledge storage and recall is actually more important.
Roughly speaking, the job of a medical doctor is to diagnose the patient - and then, after the diagnosis is made, to apply the healing from the book, corresponding to the diagnosis.
The diagnosis is pattern matching (again, roughly). It kinda suggests that a lot of "intelligent" problems are focused on pattern matching, and (relatively straightforward) application of "previous experience". So, pattern matching can bring us a great deal towards AGI.
Pattern matching is instinct. (Or at least, instinct is a kind of pattern matching. And once you learn the patterns, pattern matching can become almost instinctual). And that's fine, for things that fit the pattern. But a human-level intelligence can also deal with problems for which there is no pattern. (I mean, not always successfully - finding a correct solution to a novel problem is difficult. But it is within the capability of at least some humans.)
You're not the only one. ARC-AGI is a laudable effort, but its fundamental premise is indeed debatable:
"We argue that human cognition follows strictly the same pattern as human physical capabilities: both emerged as evolutionary solutions to specific problems in specific evironments" (from page 22 of On the Measure of Intelligence)
But I believe that because of this "even edge" thing which people call of AI weakenesses being not necessarily same of humans, once we run out of these tests which AI is worse than humans it will actually in effect be very much superhuman. My main evidence for this is leela-zero the Go AI who struggled with ladders and some other aspects of Go play well into the superhuman regime (in go it's easier to see when it's superhuman bc you can have elos and play win-rates etc and there's less room for debates)
Perhaps it's because the representations are fractured. The link above is to the transcript of an episode of Machine Learning Street Talk with Kenneth O. Stanleyabout The Fractured Entangled Representation Hypothesis[1]
If you can write code to solve ARC by "overfitting," then give it a shot! There's prize money to be won, as long as your model does a good job on the hidden test set. Zuckerberg is said to be throwing around 8-figure signing bonuses for talent like that.
But then, I guess it wouldn't be "overfitting" after all, would it?
He’s playing the game. You have to say AGI is your goal to get attention. It’s just like the YouTube thumbnail game. You can hate it, but you still have to play if you want people to pay attention.
Much like other forms of psychometry, especially related to so called intelligence, it's mainly about stratification and discrimination for ideological purposes.
I understand Chollet is transparent that the "branding" of the ARC-AGI-n suites is meant to be suggestive of its purpose, than substantial.
However, it does rub me the wrong way - as someone who's cynical of how branding can enable breathless AI hype by bad journalism. A hypothetical comparison would be labelling SHRDLU's (1968) performance on Block World planning tasks as "ARC-AGI-(-1)".[0]
A less loaded name like (bad strawman option) "ARC-VeryToughSymbolicReasoning" should capture how the ARC-AGI-n suite is genuinely and intrinsically very hard for current AIs, and what progress satisfactory performance on the benchmark suite would represent. Which Chollet has done, and has grounded him throughout! [1]
I've said this somewhere else, but we have the perfect test for AGI in the form of any open world game. Give the instructions to the AGI that it should finish the game and how to control it. Give the frames as input and wait. When I think of the latest Zelda games and especially how the Shrine chanllenges are desgined they especially feel like the perfect environement for an AGI test.
And if someone makes a machine that does all that and another person says
"That's not really AGI because xyz"
What then? The difficulty in coming up with a test for AGI is coming up with something that people will accept a passing grade as AGI.
In many respects I feel like all of the claims that models don't really understand or have internal representation or whatever tend to lean on nebulous or circular definitions of the properties in question. Trying to pin the arguments down usually end up with dualism and/or religion.
Doing what Chollet has done is infinitely better, if a person can easily do something and a model cannot then there is clearly something significant missing
It doesn't matter what the property is or what it is called. Such tests might even help us see what those properties are.
Anyone who wants to claim the fundamental inability of these models should be able to provide a task that it is clearly possible to tell when it has been succeeded, and to show that humans can do it (if that's the bar we are claiming can't be met). If they are right, then no future model should be able to solve that class of problems.
Well for most, the next steps are probably towards removing the highly deterministic and discrete characteristics of current approaches (we certainly don't think in lock steps). Those have no measures. Even the creative aspect is undermined by those characteristics.
> I feel like I'm the only one who isn't convinced getting a high score on the ARC eval test means we have AGI.
Wait, what? Approximately nobody is claiming that "getting a high score on the ARC eval test means we have AGI". It's a useful eval for measuring progress along the way, but I don't think anybody considers it the final word.
Today’s llms are fancy autocomplete but lack test time self learning or persistent drive.
By contrast, an AGI would require:
– A goal-generation mechanism (G) that can propose objectives without external prompts
– A utility function (U) and policy π(a│s) enabling action selection and hierarchy formation over extended horizons
– Stateful memory (M) + feedback integration to evaluate outcomes, revise plans, and execute real-world interventions autonomously
Without G, U, π, and M operating llms remain reactive statistical predictors, not human level intelligence.
Looking at the human side, it takes a while to actually learn something. If you've recently read something it remains in your "context window". You need to dream about it, to think about, to revisit and repeat until you actually learn it and "update your internal model". We need a mechanism for continuous weight updating.
Goal-generation is pretty much covered by your body constantly drip-feeding your brain various hormones "ongoing input prompts".
In fact, there is no technical threshold anymore. As long as the theory is in place, you can see such AGI at most half a year. It will even be more energy efficient than the current dense models.
The first highlight from this video is getting to see a preview of the next ARC dataset. Otherwise it feels like most of what Chollet says here has already been repeated in his other podcast appearances and videos. It's a good video if you're not familiarized with his work, but if you've seen some of his recent interviews then you can probably skip the first 20 minutes.
The second highlight from this video is the section from 29 minutes onward, where he talks about designing systems that can build up rich libraries of abstractions which can be applied to new problems. I wish he had lingered more on exploring and explaining this approach, but maybe they're trying to keep a bit of secret sauce because it's what his company is actively working on.
One of the major points which seems to be emerging from recent AI discourse is that the ability to integrate continuous learning seems like it'll be a key element in building AGI. Context is fine for short tasks, but if lessons are never preserved you're severely capped with how far the system can go.
There are dozens of ready-made, well-designed, and very creative games there. All are tile-based and solved with only arrow keys and a single action button. Maybe someone should make a PuzzleScript AGI benchmark?
Maybe someone can make an MCP connection for the AIs to practice. But I think the idea of the benchmark is to reserve some puzzles for private evaluation, so that they're not in the training data.
I've been thinking lately about how AGI runs up against the No Free Lunch Theorem. This is what irritates me: science is not determining the narrative. Money is. I highly recommend mathematician David Wolpert's work on the topic. I think he inadvertently proved that ASI is physically impossible. Certainly he proved that AOI (artificial omniscient intelligence) is impossible.
One thing he showed is that you can't have a universe with two omniscient intelligences (as it would be intractable for them to predict the other's behavior.)
It's also very questionable whether "humanlike" intelligence is truly general in the first place. I think cognitive neurobiologists would agree that we have a specific "cognitive niche", and while this symbolic niche seems sufficiently general for a lot of problems, there are animals that make us look stupid in other respects. This whole idea that there is some secret sauce special algorithm for universal intelligence is extremely suspect. We flatter ourselves and have committed to a fundamental anthropomorphic fallacy that seems almost cartoonishly elementary for all the money behind it.
TIL there actually is something called "no free lunch in search and optimization"[1].
See however that the theorem is quite weak. Requires eg the assumption that the search space has no structure. They even have the example of quadratic problems. It's mostly a useless saying, it appears to me.
AGI can't be defined, because it's the means by which definitions are created. You can only measure it contemporaneously by some consensus method such as ARC.
You can't define AGI, any more than you can define ASA (artificial sports ability). Intelligence, like athleticism changes both quantitively and qualitatively. The Greek Olympic champions of 2K yrs ago wouldn't qualify for high school championships today, however, they were once regarded as great athletes.
Ironically, you missed the point. The person you're responding to is saying the opposite: the narrative (that there is something unique about how our brain solves problems) determines the money (invested).
I think intelligence is search. Search is exploration + learning. So intelligence is not in the model or in the environment, but in their mutual dance. A river is not the banks, nor the water, but their relation. ARC is just a frozen snapshot of the banks, not the dynamic environment we have.
I agree strongly with this take but find it hard to convince others of it. Instead, people keep thinking there is a magic bullet to discover resulting in a lot of wasted resources and money.
I dislike the term AGI, as intelligence (of any type) always involves tradeoffs. Being exceptional at solving 2D grid-based pattern tasks is just one skill. Humans have a strong visual bias, while some hypothetical superintelligent slime molds might value entirely different problems. I know smart people (PhDs in STEM fields at major universities) who struggle with geometric puzzles, yet excel at linguistic or algebraic ones.
Getting a perfect ARC-AGI-n score isn't a smoking gun indicator of general intelligence. Rather, it simply means we're now able to solve a class of problems previously beyond AI capabilities (which is exciting in itself!).
I view ARC-AGI primarily as a benchmark (similar in spirit to Raven's matrices) that makes memorization substantially harder. Compare this with vocabulary-focused IQ tests, where cognitive skills certainly matter, but results depend heavily on exposure to a particular language.
It's obvious why having human visual intelligence in a machine is desirable.
But if slime mold symbolic space is better suited for something like understanding of biology or abstract math, that's a good damn reason to go for the slime mold route too.
> As long as it's easy to come up with tasks that any one of you can do, that are easy for humans, but that AI cannot figure out … we don't have AGI yet. And you will know you are close to having AGI when it becomes increasingly difficult to come up with such tasks.
Man I don't know. A random 10 year old has general intelligence and ain't gonna do too well on these tests. AGI is not consciousness, I feel like that also gets confused, and general intelligence is not superhuman intelligence.
I wonder how much slow progress on ARC can be explained by their visual properties making them easy for humans but hard for LLMs.
My impression is that models are pretty bad at interpreting grids of characters. Yesterday, I was trying to get Claude to convert a message into a cipher where it converted a 98-character string into 7x14 grid where the sequential letters moved 2-right and 1-down (i.e., like a knight it chess). Claude seriously struggled.
Yet, Francois always pumps up the "fluid intelligence" component of this test and emphasizes how easy these are for humans. Yet, humans would presumably be terrible at the tasks if they looked at it character-by-character
This feels like a somewhat similar (intuition-lie?) case as the Apple paper showing how reasoning model's can't do tower of hanoi past 10+ disks. Readers will intuitively think about how they themselves could tediously do an infinitely long tower of hanoi, which is what the paper is trying to allude to. However, the more appropriate analogy would be writing out all >1000 moves on a piece of paper at once and being 100% correct, which is obviously much harder
There are some major hints that this is indeed the case.
I've seen a simple ARC-AGI test that took the open set, and doubled every image in it. Every pixel became a 2x2 block of pixels.
If LLMs were bottlenecked solely by reasoning or logic capabilities, this wouldn't change their performance all that much, because the solution doesn't change all that much.
Instead, the performance dropped sharply - which hints that perception is the bottleneck.
I thought so too back when the test was first released, but now that we have multimodal models which can take images directly as input, shouldn't this point be moot?
I think the top performer afaik (ChatGPT o3) is still treating ARC as a series of characters. I imagine complex reasoning in multimodal processing wouldn't be nearly as advanced so treating it as characters is still better
Even the very best multimodal LLMs still suffer from a harsh perception bottleneck. They're impressive, but nowhere near as good as human visual cortex.
You would think we would have to take a statistical approach to AGI.
Look how we learned physics. Aristotelian physics was "An object in motion tends to come to a stop." That looked right most of the time a bowling ball on sand, grass, or even dirt comes to a stop pretty fast. But once you have a nice smooth marble floor the ball goes a lot further.
Newtonian physics solved that and several other issues and works fine, most of the time, but has corner cases when going very fast or getting near a high gravity location. Then relativity and the rest.
We need to build a system that we can teach like we do children that lets them reason that something is true under certain circumstances but may not hold generally so have to update what true is. And that looks like statistics.
The Cyc project basically achieved what you're talking about, even without approaching AGI. They manually programmed concepts and relationships between things into a huge knowledge graph. Then they had heuristics for choosing the appropriate version of facts for a given context (e.g. level of rigor). It was arguably able to use a library of abstractions similarly to what Chollet is talking about, but couldn't learn new ones automatically through exploration or play.
I think intelligence is search. Search is exploration and learning. So intelligence is not in the model, or in the environment, but in their mutual dance. A river is not the banks, nor the water, but their relation.
From the talk: “Intelligence is the conversion ratio between past experience and potential operating area [where the potential operating area of interest is the area of problems that system has no experience of].”
So we might say, “General Intelligence is the ability to do the things we haven’t yet thought of.”
“Like what?”
“Well, as soon as I name something it stops counting.”
Gödellian - I like it. Does that mean a constructive definition of General Intelligence is uncomputable?
Edit: Also, the more competent models (Opus/ Sonnet to a lesser degree) are good at very complex subtask delegation that it can blow through and attempt and then verify in seconds, so not sure hand crafted regex examples are the best counter examples here
New code patch models that I didn't even take seriously are actually really impressive and pretty new
I would have thought/considered AGI to be something that is constantly aware, a biological brain is always on. An LLM is on briefly while it's inferring.
A biological brain constantly updates itself adds memories of things. Those memories generally stick around.
This may be a silly question, I'm no expert. But why not simply define as AGI any system that can answer a question that no human can. So for example, ask AGI to find out, from current knowledge, how to reconcile gravity and qed.
Computers can already do a lot of things that no human can though. They can reliably find the best chess or go move better than a human.
It's conceivable (though not likely) that given training enough training in symbolic mathematics and some experimental data, an LLM-style AI could figure out a neat reconciliation of the two theories. I wouldn't say that makes it AGI though. You could achieve that unification with an AI that was limted to mathematics rather than being something that can function in many domains like a human can.
Wouldn't this unification need to be backed by empirical data? Let's say the AI discovers the two theories can be unified using let's say some configuration 8 spatial dimensions and 2 time dimensions. Neat trick, but how do we know the world actually has those dimensions?
But consider: technically AlphaTensor found new algorithms no human did before (https://en.wikipedia.org/wiki/Matrix_multiplication_algorith...). So isn't it AGI by your definition of answering a question no human could before: how to do 4x4 matrix multiplication in 47 steps?
Aside from other objections already mentioned, your example would require feasible experiments for verification, and likely the process of finding a successful theory of quantum gravity requires a back and forth between experimenters and theorists.
Current AI systems don't have a great ability to take instructions or information about the state of the world and produce new output based upon that. Benchmarks that emphasize this ability help greatly in progress toward AGI.
When I watched the video I had the subtitles on. The automatic transcript is pretty good. "Test-time" which is used frequently gets translated as "Tesla" so watch out for that.
Has Chollet ever talked about his change of heart regarding AGI? It wasn't that long ago when he was one of the loudest voices decrying even the concept of AGI, let alone us being on the path to creating it. Now he's an advocate and has his own prize dataset? Seems rather convenient to change your tune once hundreds of billions are being thrown at AGI (not that I would blame him).
People are allowed to evolve opinions. It seems to me he believes that a combination of transformer and program synthesis are key. The big unknown at the moment is how to do program search.
Absolutely. Presumably there is some specific considerations or evidence that helped him evolve his opinion. I would be interested in seeing a writeup about it. With him having been a very public advocate against AGI, a writeup of his evolution seems appropriate and would be very edifying for a lot of people.
He has recently co-founded NDEA company, so he has to align himself for that. Same kind of vibe change feels for Joscha Bach after having some position in Liquid AI company. Communication is not so relaxed anymore.
That said, I'd still listen these two guys (+ Schmidhuber) more than any other AI-guy.
I think you're basically saying that ARC-AGI doesn't achieve a goal that _it didn't set_. The point of ARC-AGI is not to benchmark LLMs specifically. The point is to measure fluid intelligence in a way which supports comparisons between models and between models and humans. It's not the obligation of the test to be tailored to the form of model that's most popular now.
>The point is to measure fluid intelligence in a way which supports comparisons between models and between models and humans. It's not the obligation of the test to be tailored to the form of model that's most popular now.
The problem is that the test may not be giving an accurate comparison because the test is problematic when used to assess LLMs, which are the kind of model that people are most interested in assessing for general capabilities.
Let's not. Seriously. I absolutely love François and have used his work extensively. But looking around me at the social impact of AI I am really not convinced that this is what the world needs right now and that if we can stave off the turning point for another decade or two that humanity will likely benefit from that. The last thing we need is to inject yet another instability into a planet that is already fighting existential crisis on a number of fronts.
It doesn't matter what should or should not happen. Technology will continue to race forward at breakneck speed while everyone involved pats each other on the back for making a bunch of money before the consequences hit
There is some kind of massive brigading happening on this thread. Lots of thoughtful comments are downmodded or flagged (including mine, which I thought was pretty thoughtful. I even said poop instead of shit.).
I addressed the entire concept of the talk, and made other relevant points. The correct response to "let me tell you something I can't possibly know" isn't to argue the points within that frame.
If you see a talk like: "How we will develop diplomacy with the rat-people of TRAPPIST-5." you don't have to make some argument about super-earths and gravity and the rocket equation. You can just point out it's absurd to pretend to know something like whether there are rat-people there.
I've noticed that Hacker News has become increasingly strict with tone lately; almost to the point of aggressive "tone policing." It feels like you’re expected to maintain a very diplomatic tone: praise the author, acknowledge the strengths of the article or video, and only then, very carefully, offer any criticism. You can't just bash it, and not get downvoted :)
The Arc prize/benchmark is a terrible judge of whether we got to AGI.
If we assume that humans have "general intelligence", we would assume all humans could ace Arc... but they can't. Try asking your average person, i.e. supermarket workers, gas station attendants etc to do the Arc puzzles, they will do poorly, especially on the newer ones, but AI has to do perfectly to prove they have general intelligence? (not trying to throw shade here but the reality is this test is more like an IQ test than an AGI test).
Arc is a great example of AI researchers moving the goal posts for what we consider intelligent.
Let's get real, Claude Opus is smarter than 99% of people right now, and I would trust its decision making over 99% of people I know in most situations, except perhaps emotion driven ones.
Arc agi benchmark is just a gimmick. Also, since it's a visual test and the current models are text based it's actually a rigged (against the AI models) test anyway, since their datasets were completely text based.
Basically, it's a test of some kind, but it doesn't mean quite as much as Chollet thinks it means.
He said in the video that they tested regular people (uber driver, etc.) on arc-agi2 and at least 2 people were able to solve each task (an average of 9-10 people saw each task). Also this quote from the paper: None of the self-reported demographic factors recorded for all participants—including occupation, industry,
technical experience, programming proficiency, mathematical background, puzzle-solving aptitude, and var-
ious other measured attributes—demonstrated clear, statistically significant relationships with performance
outcomes. This finding suggests that ARC-AGI-2 tasks assess general problem-solving capabilities rather
than domain-specific knowledge or specialized skills acquired through particular professional or educational
experiences.
It is not a judge of whether we got to AGI. And literally no one except straw-manning critics are trying to claim it is. The point is, an AGI should easily be able to pass it. But it can obviously be passed without getting to AGI (as . It's a necessary but not sufficient criteria. If something can't pass a test as simple as AGI (which no AI currently can) then it's definitely not AGI. Anyone claiming AGI should be able to point their AI at the problem and have an 80+% solution rate. Current attempts on the second ARC are less than 10% with zero shot attempts even worse. Even the better performing LLMs on the first ARC couldn't do well without significant pre-training. In short, the G in AGI stands for general.
So do you agree that a human that CANNOT solve ARC doesn't have general intelligence?
If we think humans have "GI" then I think we have AIs right now with "GI" too. Just like humans do, AIs spike in various directions. They are amazing at some things and weak at visual/IQ test type problems like ARC.
This is what is called "spikey" intelligence, where a model might be able to crack phd physics problems and solve byzantine pattern matching games at the 90th percentile, but also can't figure out how to look up a company and copy their address on the "customer" line of an invoice.
Maybe it is a cultural difference aspect, but I feel that "supermarket workers, gas station attendants" (in an Asian country) that I know of should be quite capable of most ARC tasks.
Out of 100 of evals, ARC is a very distinct and unique eval, most frontier models are also visual now, don't see the harm in having this instead of another text eval.
By both definitions of intelligence in the presentation we should be saying "how we got to AGI" in the past tense. We're already there. AI's can deal with situations they weren't prepared for in any sense that a human can. They might not do well, but they'll have a crack at it. We can trivially build systems that collect data and do a bit more offline training if that is what someone wants to see, but there doesn't really seem to be a commercial need for that right now. Similarly, AIs can whip most humans at most domains that require intelligence.
I think the debate hqas been flat-footed by the speed all this happened. We're not talking AGI any more, we're talking about how to build superintelligences hitherto unseen in nature.
According to this presentation at least, ARC-AGI-2 shows that there is a big meaningful gap in fluid intelligence between normal non-genius humans and the best models currently, which seems to indicate we are not "already there".
If there is a gap in intelligence between two humans, does that mean to you that one of them is necessarily not a general intelligence? The current crop of AIs get some of the questions right by reasoning through them. That means they are already intelligent in the way ARC-AGI-2 measures intelligence. They just aren't very capable ones.
If AI at least equal humans in all intellectual fields then they are super-intelligences, because there are already fields where they dominate humans so outrageously there isn't a competition (nearly all fields, these days). Before they are superintelligences there is a phase where they are just AGIs, we've been in that phase for a while now. Artificial superintelligence is very exciting, but Artificial non-super Intelligence or AGI is here with us in the present.
There's already a big meaningful gap between the things AIs can do which humans can't, so why do you only count as "meaningful" the things humans can do which AIs can't?
I enjoy seeing people repeatedly move the goalposts for "intelligence" as AIs simply get smarter and smarter every week. Soon AI will have to beat Einstein in Physics, Usain Bolt in running, and Steve Jobs in marketing to be considered AGI...
Current robots perform very badly on my patented and highly scientific ROACH-AGI benchmark - "is this thing smarter at navigating unfamiliar 3D spaces than a cockroach?"
I feel like I'm the only one who isn't convinced getting a high score on the ARC eval test means we have AGI. It's mostly about pattern matching (and some of it ambiguous even for humans what the actual true response aught to be). It's like how in humans there's lots of different 'types' of intelligence, and just overfitting on IQ tests doesn't in my mind convince me a person is actually that smart.
Getting a high score on ARC doesn't mean we have AGI and Chollet has always said as much AFAIK, it's meant to push the AI research space in a positive direction. Being able to solve ARC problems is probably a pre-requisite to AGI. It's a directional push into the fog of war, with the claim being that we should explore that area because we expect it's relevant to building AGI.
We don't really have a true test that means "if we pass this test we have AGI" but we have a variety of tests (like ARC) that we believe any true AGI would be able to pass. It's a "necessary but not sufficient" situation. Also ties directly to the challenge in defining what AGI really means. You see a lot of discussions of "moving the goal posts" around AGI, but as I see it we've never had goal posts, we've just got a bunch of lines we'd expect to cross before reaching them.
20 replies →
> Getting a high score on ARC doesn't mean we have AGI and Chollet has always said as much AFAIK
He only seems to say this recently, since OpenAI cracked the ARC-AGI benchmark. But in the original 2019 abstract he said this:
> We argue that ARC can be used to measure a human-like form of general fluid intelligence and that it enables fair general intelligence comparisons between AI systems and humans.
https://arxiv.org/abs/1911.01547
Now he seems to backtrack, with the release of harder ARC-like benchmarks, implying that the first one didn't actually test for really general human-like intelligence.
This sounds a bit like saying that a machine beating chess would require general intelligence -- but then adding, after Deep Blue beats chess, that chess doesn't actually count as a test for AGI, and that Go is the real AGI benchmark. And after a narrow system beats Go, moving the goalpost to beating Atari, and then to beating StarCraft II, then to MineCraft, etc.
At some point, intuitively real "AGI" will be necessary to beat one of these increasingly difficult benchmarks, but only because otherwise yet another benchmark would have been invented. Which makes these benchmarks mostly post hoc rationalizations.
A better approach would be to question what went wrong with coming up with the very first benchmark, and why a similar thing wouldn't occur with the second.
4 replies →
ARC is definitely about achieving AGI and it doesn't matter whether we "have" it or not right now. That is the goal:
> where he introduced the "Abstract and Reasoning Corpus for Artificial General Intelligence" (ARC-AGI) benchmark to measure intelligence
So, a high enough score is a threshold to claim AGI. And, if you use an LLM to work these types of problems, it becomes pretty clear that passing more tests indicates a level of "awareness" that goes beyond rational algorithms.
I thought I had seen everything until I started working on some of the problems with agents. I'm still sorta in awe about how the reasoning manifests. (And don't get me wrong, LLMs like Claude still go completely off the rails where even a less intelligent human would know better.)
2 replies →
"Being able to solve ARC problems is probably a pre-requisite to AGI." - is it? Humans have general intelligence and most can't solve the harder ARC problems.
6 replies →
My problem with AGI is the lack of a simple, concrete definition.
Can we formalize it as giving out a task expressible in, say, n^m bytes of information that encodes a task of n^(m+q) real algorithmic and verification complexity -- then solving that task within a certain time, compute, and attempt bounds?
Something that captures "the AI was able to unwind the underlying unspoken complexity of the novel problem".
I feel like one could map a variety of easy human "brain teaser" type tasks to heuristics that fit within some mathematical framework and then grow the formalism from there.
3 replies →
I'm all for benchmarks that push the field forward, but ARC problems seem to be difficult for reasons having less to do with intelligence and more about having a text system that works reliably with rasterized pixel data presented line by line. Most people would score 0 on it if they were shown the data the way an LLM sees it, these problems only seem easy to us because there are visualizers slapped on top.
1 reply →
[dead]
[flagged]
1 reply →
In the video, François Chollet, creator of the ARC benchmarks, says that beating ARC does not equate to AGI. He specifically says they will be able to be beaten without AGI.
He only says this because otherwise he would have to say that
- OpenAI's o3 counts as "AGI" when it did unexpectedly beat the ARC-AGI benchmark or
- Explicitly admit that he was wrong when assuming that ARC-AGI would test for AGI
4 replies →
I think the people behind the ARC Prize agree that getting a high score doesn't mean we have AGI. (They already updated the benchmark once to make it harder.) But an AGI should get a similarly high score as humans do. So current models that get very low scores are definitely not AGI, and likely quite far away from it.
> I think the people behind the ARC Prize agree that getting a high score doesn't mean we have AGI
The benchmark was literally called ARC-AGI. Only after OpenAI cracked it, they started backtracking and saying that it doesn't test for true AGI. Which undermines the whole premise of a benchmark.
2 replies →
https://en.m.wikipedia.org/wiki/AI_effect
But on a serious note, I don't think Chollet would disagree. ARC is a necessary but not sufficient condition, and he says that, despite the unfortunate attention-grabbing name choice of the benchmark. I like Chollet's view that we will know that AGI is here when we can't come up with new benchmarks that separate humans from AI.
I agree with you but I'll go a step further - these benchmarks are a good example of how far we are from AGI.
A good base test would be to give a manager a mixed team of remote workers, half being human and half being AI, and seeing if the manager or any of the coworkers would be able to tell the difference. We wouldn't be able to say that AI that passed that test would necessarily be AGI, since we would have to test it in other situations. But we could say that AI that couldn't pass that test wouldn't qualify, since it wouldn't be able to successfully accomplish some tasks that humans are able to.
But of course, current AI is nowhere near that level yet. We're left with benchmarks, because we all know how far away we are from actual AGI.
The AGI test I think makes sense is to put it in a robot body and let it navigate the world. Can I take the robot to my back yard and have it weed my vegetable garden? Can I show it how to fold my laundry? Can I take it to the grocery store and tell it "go pick up 4 yellow bananas and two avocados that will be ready to eat in the next day or two, and then meet me in dairy"? Can I ask it to dice an onion for me during meal prep?
These are all things my kids would do when they were pretty young.
3 replies →
The problem with "spot the difference" tests, imho, is that I would expect an AGI to be easily spotted. There's going to be a speed of calculation difference, at the very least. If nothing else, typing speed would be completely different unless the AGI is supposed to be deceptive. Who knows what it's personality would be like. I'd say it's a simple enough test just to see if an AGI could be hired as, for example, an entry level software developer and keep it's job based on the same criteria base-level humans have to meet.
I agree that current AI is nowhere near that level yet. If AI isn't even trying to extract meaning from the words it smiths or the pictures it diffuses then it's nothing more than a cute (albeit useful) parlor trick.
1 reply →
Why even bother with the people in the mix? Just tell the AGI: make as much money as you can in 6 months. Preferably without breaking any laws.
I think next year's AI benchmarks are going to be like this project: https://www.anthropic.com/research/project-vend-1
Give the AI tools and let it do real stuff in the world:
"FounderBench": Ask the AI to build a successful business, whatever that business may be - the AI decides. Maybe try to get funded by YC - hiring a human presenter for Demo Day is allowed. They will be graded on profit / loss, and valuation.
Testing plain LLM on whiteboard-style question is meaningless now. Going forward, it will all be multi-agent systems with computer use, long-term memory & goals, and delegation.
This sounds like a terrible idea to me, you're training intelligent computer to aim for power. It's fine as long as they're bad but if they get good then we have a problem
You're not alone in this; I expect us to have not yet enumerated all the things that we ourselves mean by "intelligence".
But conversely, not passing this test is a proof of not being as general as a human's intelligence.
I find the "what is intelligence?" discussion a little pointless if I'm honest. It's similar to asking a question like does it mean to be a "good person" and would we know whether an AI or person is really "good"?
While understanding why a person or AI is doing what it's doing can be important (perhaps specifically in safety contexts) at the end of the day all that's really going to matter to most people is the outcomes.
So if an AI can use what appears to be intelligence to solve general problems and can act in ways that are broadly good for society, whether or not it meets some philosophical definition of "intelligent" or "good" doesn't matter much – at least in most contexts.
That said, my own opinion on this is that the truth is likely in between. LLMs today seem extremely good at being glorified auto-completes, and I suspect most (95%+) of what they do is just recalling patterns in their weights. But unlike traditional auto-completes they do seem to have some ability to reason and solve truly novel problems. As it stands I'd argue that ability is fairly poor, but this might only represent 1-2% of what we use intelligence for.
If I were to guess why this is I suspect it's not that LLM architecture today is completely wrong, but that the way LLMs are trained means that in general knowledge recall is rewarded more than reasoning. This is similar to the trade-off we humans have with education – do you prioritise the acquisition of knowledge or critical thinking? Maybe believe critical thinking is more important and should be prioritised more, but I suspect for the vast majority of tasks we're interested in solving knowledge storage and recall is actually more important.
4 replies →
Unfortunately, we did it. All that is left is to assemble the parts.
https://news.ycombinator.com/item?id=44488126
The point is not that having a high score -> AGI, their ideas are more of having a low score -> we don't have AGI yet.
AGI isn't defined anywhere, so it can be anything you want.
Oh, it's defined in lots of places. The problem is.. it's defined in lots of places!
Yes. And a lot of humans also don't pass for having AGI.
Roughly speaking, the job of a medical doctor is to diagnose the patient - and then, after the diagnosis is made, to apply the healing from the book, corresponding to the diagnosis.
The diagnosis is pattern matching (again, roughly). It kinda suggests that a lot of "intelligent" problems are focused on pattern matching, and (relatively straightforward) application of "previous experience". So, pattern matching can bring us a great deal towards AGI.
Pattern matching is instinct. (Or at least, instinct is a kind of pattern matching. And once you learn the patterns, pattern matching can become almost instinctual). And that's fine, for things that fit the pattern. But a human-level intelligence can also deal with problems for which there is no pattern. (I mean, not always successfully - finding a correct solution to a novel problem is difficult. But it is within the capability of at least some humans.)
You're not the only one. ARC-AGI is a laudable effort, but its fundamental premise is indeed debatable:
"We argue that human cognition follows strictly the same pattern as human physical capabilities: both emerged as evolutionary solutions to specific problems in specific evironments" (from page 22 of On the Measure of Intelligence)
https://arxiv.org/pdf/1911.01547
But I believe that because of this "even edge" thing which people call of AI weakenesses being not necessarily same of humans, once we run out of these tests which AI is worse than humans it will actually in effect be very much superhuman. My main evidence for this is leela-zero the Go AI who struggled with ladders and some other aspects of Go play well into the superhuman regime (in go it's easier to see when it's superhuman bc you can have elos and play win-rates etc and there's less room for debates)
> I feel like I'm the only one who isn't convinced getting a high score on the ARC eval test means we have AGI
Francois explicitly says that's not how ARC is supposed to be interpreted.
[1]https://app.rescript.info/public/share/W_T7E1OC2Wj49ccqlIOOz...
Perhaps it's because the representations are fractured. The link above is to the transcript of an episode of Machine Learning Street Talk with Kenneth O. Stanleyabout The Fractured Entangled Representation Hypothesis[1]
> It's mostly about pattern matching...
For all we know, human intelligence is just an emergent property of really good pattern matching.
If you can write code to solve ARC by "overfitting," then give it a shot! There's prize money to be won, as long as your model does a good job on the hidden test set. Zuckerberg is said to be throwing around 8-figure signing bonuses for talent like that.
But then, I guess it wouldn't be "overfitting" after all, would it?
Who says intelligence is anything more than "pattern matching"? Everything is patterns
He’s playing the game. You have to say AGI is your goal to get attention. It’s just like the YouTube thumbnail game. You can hate it, but you still have to play if you want people to pay attention.
Much like other forms of psychometry, especially related to so called intelligence, it's mainly about stratification and discrimination for ideological purposes.
I understand Chollet is transparent that the "branding" of the ARC-AGI-n suites is meant to be suggestive of its purpose, than substantial.
However, it does rub me the wrong way - as someone who's cynical of how branding can enable breathless AI hype by bad journalism. A hypothetical comparison would be labelling SHRDLU's (1968) performance on Block World planning tasks as "ARC-AGI-(-1)".[0]
A less loaded name like (bad strawman option) "ARC-VeryToughSymbolicReasoning" should capture how the ARC-AGI-n suite is genuinely and intrinsically very hard for current AIs, and what progress satisfactory performance on the benchmark suite would represent. Which Chollet has done, and has grounded him throughout! [1]
[0] https://en.wikipedia.org/wiki/SHRDLU [1] https://arxiv.org/abs/1911.01547
I get what you're saying about perception being reality and that ARC-AGI suggests beating it means AGI has been achieved.
In practice when I have seen ARC brought up, it has more nuance than any of the other benchmarks.
Unlike, Humanity's Last Exam, which is the most egregious example I have seen in naming and when it is referenced in terms of an LLMs capability.
I've said this somewhere else, but we have the perfect test for AGI in the form of any open world game. Give the instructions to the AGI that it should finish the game and how to control it. Give the frames as input and wait. When I think of the latest Zelda games and especially how the Shrine chanllenges are desgined they especially feel like the perfect environement for an AGI test.
And if someone makes a machine that does all that and another person says
"That's not really AGI because xyz"
What then? The difficulty in coming up with a test for AGI is coming up with something that people will accept a passing grade as AGI.
In many respects I feel like all of the claims that models don't really understand or have internal representation or whatever tend to lean on nebulous or circular definitions of the properties in question. Trying to pin the arguments down usually end up with dualism and/or religion.
Doing what Chollet has done is infinitely better, if a person can easily do something and a model cannot then there is clearly something significant missing
It doesn't matter what the property is or what it is called. Such tests might even help us see what those properties are.
Anyone who wants to claim the fundamental inability of these models should be able to provide a task that it is clearly possible to tell when it has been succeeded, and to show that humans can do it (if that's the bar we are claiming can't be met). If they are right, then no future model should be able to solve that class of problems.
6 replies →
Well for most, the next steps are probably towards removing the highly deterministic and discrete characteristics of current approaches (we certainly don't think in lock steps). Those have no measures. Even the creative aspect is undermined by those characteristics.
It is a necessary condition, but not a sufficient one.
You're not alone in this, no.
My definition of AGI is the one I was brought up with, not an ever moving goal post (to the "easier" side).
And no, I also don't buy that we are just stochastic parrots.
But whatever. I've seen many hypes and if I don't die and the world doesn't go to shit, I'll see a few more in the next couple of decades
The author of the test agrees with you.
> I feel like I'm the only one who isn't convinced getting a high score on the ARC eval test means we have AGI.
Wait, what? Approximately nobody is claiming that "getting a high score on the ARC eval test means we have AGI". It's a useful eval for measuring progress along the way, but I don't think anybody considers it the final word.
Today’s llms are fancy autocomplete but lack test time self learning or persistent drive. By contrast, an AGI would require: – A goal-generation mechanism (G) that can propose objectives without external prompts – A utility function (U) and policy π(a│s) enabling action selection and hierarchy formation over extended horizons – Stateful memory (M) + feedback integration to evaluate outcomes, revise plans, and execute real-world interventions autonomously Without G, U, π, and M operating llms remain reactive statistical predictors, not human level intelligence.
I'd say we're not far off.
Looking at the human side, it takes a while to actually learn something. If you've recently read something it remains in your "context window". You need to dream about it, to think about, to revisit and repeat until you actually learn it and "update your internal model". We need a mechanism for continuous weight updating.
Goal-generation is pretty much covered by your body constantly drip-feeding your brain various hormones "ongoing input prompts".
13 replies →
we're closer than you think...
In fact, there is no technical threshold anymore. As long as the theory is in place, you can see such AGI at most half a year. It will even be more energy efficient than the current dense models.
https://dmf-archive.github.io/docs/posts/beyond-snn-plausibl...
[dead]
To pass Arc, you need a living model with sentient abilities, not the dead frog now.
https://news.ycombinator.com/item?id=44488126
[dead]
The first highlight from this video is getting to see a preview of the next ARC dataset. Otherwise it feels like most of what Chollet says here has already been repeated in his other podcast appearances and videos. It's a good video if you're not familiarized with his work, but if you've seen some of his recent interviews then you can probably skip the first 20 minutes.
The second highlight from this video is the section from 29 minutes onward, where he talks about designing systems that can build up rich libraries of abstractions which can be applied to new problems. I wish he had lingered more on exploring and explaining this approach, but maybe they're trying to keep a bit of secret sauce because it's what his company is actively working on.
One of the major points which seems to be emerging from recent AI discourse is that the ability to integrate continuous learning seems like it'll be a key element in building AGI. Context is fine for short tasks, but if lessons are never preserved you're severely capped with how far the system can go.
ARC-AGI 3 remindes me of PuzzleScript games: https://www.puzzlescript.net/Gallery/index.html
There are dozens of ready-made, well-designed, and very creative games there. All are tile-based and solved with only arrow keys and a single action button. Maybe someone should make a PuzzleScript AGI benchmark?
This game is great!
https://nebu-soku.itch.io/golfshall-we-golf
Maybe someone can make an MCP connection for the AIs to practice. But I think the idea of the benchmark is to reserve some puzzles for private evaluation, so that they're not in the training data.
I've been thinking lately about how AGI runs up against the No Free Lunch Theorem. This is what irritates me: science is not determining the narrative. Money is. I highly recommend mathematician David Wolpert's work on the topic. I think he inadvertently proved that ASI is physically impossible. Certainly he proved that AOI (artificial omniscient intelligence) is impossible.
One thing he showed is that you can't have a universe with two omniscient intelligences (as it would be intractable for them to predict the other's behavior.)
It's also very questionable whether "humanlike" intelligence is truly general in the first place. I think cognitive neurobiologists would agree that we have a specific "cognitive niche", and while this symbolic niche seems sufficiently general for a lot of problems, there are animals that make us look stupid in other respects. This whole idea that there is some secret sauce special algorithm for universal intelligence is extremely suspect. We flatter ourselves and have committed to a fundamental anthropomorphic fallacy that seems almost cartoonishly elementary for all the money behind it.
TIL there actually is something called "no free lunch in search and optimization"[1].
See however that the theorem is quite weak. Requires eg the assumption that the search space has no structure. They even have the example of quadratic problems. It's mostly a useless saying, it appears to me.
[1] https://en.m.wikipedia.org/wiki/No_free_lunch_in_search_and_...
AGI can't be defined, because it's the means by which definitions are created. You can only measure it contemporaneously by some consensus method such as ARC.
You can't define AGI, any more than you can define ASA (artificial sports ability). Intelligence, like athleticism changes both quantitively and qualitatively. The Greek Olympic champions of 2K yrs ago wouldn't qualify for high school championships today, however, they were once regarded as great athletes.
ASI is as different from AOI as BB(8) is from infinity. The impossibility of AOI says bubkis about ASI.
When hasn't money determined the narrative?
Ironically, you missed the point. The person you're responding to is saying the opposite: the narrative (that there is something unique about how our brain solves problems) determines the money (invested).
I think intelligence is search. Search is exploration + learning. So intelligence is not in the model or in the environment, but in their mutual dance. A river is not the banks, nor the water, but their relation. ARC is just a frozen snapshot of the banks, not the dynamic environment we have.
I agree strongly with this take but find it hard to convince others of it. Instead, people keep thinking there is a magic bullet to discover resulting in a lot of wasted resources and money.
I dislike the term AGI, as intelligence (of any type) always involves tradeoffs. Being exceptional at solving 2D grid-based pattern tasks is just one skill. Humans have a strong visual bias, while some hypothetical superintelligent slime molds might value entirely different problems. I know smart people (PhDs in STEM fields at major universities) who struggle with geometric puzzles, yet excel at linguistic or algebraic ones.
Getting a perfect ARC-AGI-n score isn't a smoking gun indicator of general intelligence. Rather, it simply means we're now able to solve a class of problems previously beyond AI capabilities (which is exciting in itself!).
I view ARC-AGI primarily as a benchmark (similar in spirit to Raven's matrices) that makes memorization substantially harder. Compare this with vocabulary-focused IQ tests, where cognitive skills certainly matter, but results depend heavily on exposure to a particular language.
Call me crazy but we should be optimizing for human visual intelligence rather than slime mold symbolic space
It's obvious why having human visual intelligence in a machine is desirable.
But if slime mold symbolic space is better suited for something like understanding of biology or abstract math, that's a good damn reason to go for the slime mold route too.
> As long as it's easy to come up with tasks that any one of you can do, that are easy for humans, but that AI cannot figure out … we don't have AGI yet. And you will know you are close to having AGI when it becomes increasingly difficult to come up with such tasks.
Man I don't know. A random 10 year old has general intelligence and ain't gonna do too well on these tests. AGI is not consciousness, I feel like that also gets confused, and general intelligence is not superhuman intelligence.
I wonder how much slow progress on ARC can be explained by their visual properties making them easy for humans but hard for LLMs.
My impression is that models are pretty bad at interpreting grids of characters. Yesterday, I was trying to get Claude to convert a message into a cipher where it converted a 98-character string into 7x14 grid where the sequential letters moved 2-right and 1-down (i.e., like a knight it chess). Claude seriously struggled.
Yet, Francois always pumps up the "fluid intelligence" component of this test and emphasizes how easy these are for humans. Yet, humans would presumably be terrible at the tasks if they looked at it character-by-character
This feels like a somewhat similar (intuition-lie?) case as the Apple paper showing how reasoning model's can't do tower of hanoi past 10+ disks. Readers will intuitively think about how they themselves could tediously do an infinitely long tower of hanoi, which is what the paper is trying to allude to. However, the more appropriate analogy would be writing out all >1000 moves on a piece of paper at once and being 100% correct, which is obviously much harder
There are some major hints that this is indeed the case.
I've seen a simple ARC-AGI test that took the open set, and doubled every image in it. Every pixel became a 2x2 block of pixels.
If LLMs were bottlenecked solely by reasoning or logic capabilities, this wouldn't change their performance all that much, because the solution doesn't change all that much.
Instead, the performance dropped sharply - which hints that perception is the bottleneck.
I thought so too back when the test was first released, but now that we have multimodal models which can take images directly as input, shouldn't this point be moot?
I think the top performer afaik (ChatGPT o3) is still treating ARC as a series of characters. I imagine complex reasoning in multimodal processing wouldn't be nearly as advanced so treating it as characters is still better
2 replies →
Even the very best multimodal LLMs still suffer from a harsh perception bottleneck. They're impressive, but nowhere near as good as human visual cortex.
You would think we would have to take a statistical approach to AGI.
Look how we learned physics. Aristotelian physics was "An object in motion tends to come to a stop." That looked right most of the time a bowling ball on sand, grass, or even dirt comes to a stop pretty fast. But once you have a nice smooth marble floor the ball goes a lot further.
Newtonian physics solved that and several other issues and works fine, most of the time, but has corner cases when going very fast or getting near a high gravity location. Then relativity and the rest.
We need to build a system that we can teach like we do children that lets them reason that something is true under certain circumstances but may not hold generally so have to update what true is. And that looks like statistics.
The Cyc project basically achieved what you're talking about, even without approaching AGI. They manually programmed concepts and relationships between things into a huge knowledge graph. Then they had heuristics for choosing the appropriate version of facts for a given context (e.g. level of rigor). It was arguably able to use a library of abstractions similarly to what Chollet is talking about, but couldn't learn new ones automatically through exploration or play.
I think intelligence is search. Search is exploration and learning. So intelligence is not in the model, or in the environment, but in their mutual dance. A river is not the banks, nor the water, but their relation.
From the talk: “Intelligence is the conversion ratio between past experience and potential operating area [where the potential operating area of interest is the area of problems that system has no experience of].”
So we might say, “General Intelligence is the ability to do the things we haven’t yet thought of.”
“Like what?”
“Well, as soon as I name something it stops counting.”
Gödellian - I like it. Does that mean a constructive definition of General Intelligence is uncomputable?
Today we asked Gemini CLI to check if any of the files in /home/user/dirname contained a word.
It churned for >5 minutes and didn't solve the problem.
grep, of course, solved the problem in under a second.
AGI is going to be a while, and your jobs are safe.
Try it with Claude Code on Opus
Gemini CLI isn't very good
Edit: Also, the more competent models (Opus/ Sonnet to a lesser degree) are good at very complex subtask delegation that it can blow through and attempt and then verify in seconds, so not sure hand crafted regex examples are the best counter examples here
New code patch models that I didn't even take seriously are actually really impressive and pretty new
How do we define AGI?
I would have thought/considered AGI to be something that is constantly aware, a biological brain is always on. An LLM is on briefly while it's inferring.
A biological brain constantly updates itself adds memories of things. Those memories generally stick around.
Did anybody already let o3 crack Bongard Problems? o1 did not convince: https://arxiv.org/abs/2410.19546
This may be a silly question, I'm no expert. But why not simply define as AGI any system that can answer a question that no human can. So for example, ask AGI to find out, from current knowledge, how to reconcile gravity and qed.
Computers can already do a lot of things that no human can though. They can reliably find the best chess or go move better than a human.
It's conceivable (though not likely) that given training enough training in symbolic mathematics and some experimental data, an LLM-style AI could figure out a neat reconciliation of the two theories. I wouldn't say that makes it AGI though. You could achieve that unification with an AI that was limted to mathematics rather than being something that can function in many domains like a human can.
Wouldn't this unification need to be backed by empirical data? Let's say the AI discovers the two theories can be unified using let's say some configuration 8 spatial dimensions and 2 time dimensions. Neat trick, but how do we know the world actually has those dimensions?
1 reply →
That would be ASI I think.
But consider: technically AlphaTensor found new algorithms no human did before (https://en.wikipedia.org/wiki/Matrix_multiplication_algorith...). So isn't it AGI by your definition of answering a question no human could before: how to do 4x4 matrix multiplication in 47 steps?
Aside from other objections already mentioned, your example would require feasible experiments for verification, and likely the process of finding a successful theory of quantum gravity requires a back and forth between experimenters and theorists.
"What is the meaning of life, the universe, and everything?"
42
This quest for an ill defined AGI is going to create a million of Cpt Ahab
Current AI systems don't have a great ability to take instructions or information about the state of the world and produce new output based upon that. Benchmarks that emphasize this ability help greatly in progress toward AGI.
Is the text available for those who don't hear so well?
At the very least, YouTube provides a transcript and a "Show Transcript" button in the video description, which you can click on to follow along.
When I watched the video I had the subtitles on. The automatic transcript is pretty good. "Test-time" which is used frequently gets translated as "Tesla" so watch out for that.
Has Chollet ever talked about his change of heart regarding AGI? It wasn't that long ago when he was one of the loudest voices decrying even the concept of AGI, let alone us being on the path to creating it. Now he's an advocate and has his own prize dataset? Seems rather convenient to change your tune once hundreds of billions are being thrown at AGI (not that I would blame him).
People are allowed to evolve opinions. It seems to me he believes that a combination of transformer and program synthesis are key. The big unknown at the moment is how to do program search.
Absolutely. Presumably there is some specific considerations or evidence that helped him evolve his opinion. I would be interested in seeing a writeup about it. With him having been a very public advocate against AGI, a writeup of his evolution seems appropriate and would be very edifying for a lot of people.
5 replies →
ARC-AGI was introduced in 2019:
https://arxiv.org/abs/1911.01547
GPT-3 didn't come out until 2020.
In my view that just makes his evolution more interesting as it wasn't just a matter of being wow'ed by what ChatGPT could do.
He has recently co-founded NDEA company, so he has to align himself for that. Same kind of vibe change feels for Joscha Bach after having some position in Liquid AI company. Communication is not so relaxed anymore.
That said, I'd still listen these two guys (+ Schmidhuber) more than any other AI-guy.
[dead]
[dead]
I think you're basically saying that ARC-AGI doesn't achieve a goal that _it didn't set_. The point of ARC-AGI is not to benchmark LLMs specifically. The point is to measure fluid intelligence in a way which supports comparisons between models and between models and humans. It's not the obligation of the test to be tailored to the form of model that's most popular now.
Right, that's exactly what I'm saying.
>The point is to measure fluid intelligence in a way which supports comparisons between models and between models and humans. It's not the obligation of the test to be tailored to the form of model that's most popular now.
The problem is that the test may not be giving an accurate comparison because the test is problematic when used to assess LLMs, which are the kind of model that people are most interested in assessing for general capabilities.
[flagged]
Let's not. Seriously. I absolutely love François and have used his work extensively. But looking around me at the social impact of AI I am really not convinced that this is what the world needs right now and that if we can stave off the turning point for another decade or two that humanity will likely benefit from that. The last thing we need is to inject yet another instability into a planet that is already fighting existential crisis on a number of fronts.
It doesn't matter what should or should not happen. Technology will continue to race forward at breakneck speed while everyone involved pats each other on the back for making a bunch of money before the consequences hit
The Iranian nuclear program disagrees. There could be treaties and the analogue of the IAEA for "AI".
technology doesn't just advance itself
5 replies →
[flagged]
There is some kind of massive brigading happening on this thread. Lots of thoughtful comments are downmodded or flagged (including mine, which I thought was pretty thoughtful. I even said poop instead of shit.).
https://news.ycombinator.com/item?id=44492241
My comment was basically instantly flagged. I see at least 3 other flagged comments that I can't imagine deserve to be flagged.
You didn’t address anything from the actual talk.
I addressed the entire concept of the talk, and made other relevant points. The correct response to "let me tell you something I can't possibly know" isn't to argue the points within that frame.
If you see a talk like: "How we will develop diplomacy with the rat-people of TRAPPIST-5." you don't have to make some argument about super-earths and gravity and the rocket equation. You can just point out it's absurd to pretend to know something like whether there are rat-people there.
Either way, it isn't flag-able!
2 replies →
I've noticed that Hacker News has become increasingly strict with tone lately; almost to the point of aggressive "tone policing." It feels like you’re expected to maintain a very diplomatic tone: praise the author, acknowledge the strengths of the article or video, and only then, very carefully, offer any criticism. You can't just bash it, and not get downvoted :)
That this is downvoted to the point of being greyed out proves what it says is true.
1 reply →
The Arc prize/benchmark is a terrible judge of whether we got to AGI.
If we assume that humans have "general intelligence", we would assume all humans could ace Arc... but they can't. Try asking your average person, i.e. supermarket workers, gas station attendants etc to do the Arc puzzles, they will do poorly, especially on the newer ones, but AI has to do perfectly to prove they have general intelligence? (not trying to throw shade here but the reality is this test is more like an IQ test than an AGI test).
Arc is a great example of AI researchers moving the goal posts for what we consider intelligent.
Let's get real, Claude Opus is smarter than 99% of people right now, and I would trust its decision making over 99% of people I know in most situations, except perhaps emotion driven ones.
Arc agi benchmark is just a gimmick. Also, since it's a visual test and the current models are text based it's actually a rigged (against the AI models) test anyway, since their datasets were completely text based.
Basically, it's a test of some kind, but it doesn't mean quite as much as Chollet thinks it means.
He said in the video that they tested regular people (uber driver, etc.) on arc-agi2 and at least 2 people were able to solve each task (an average of 9-10 people saw each task). Also this quote from the paper: None of the self-reported demographic factors recorded for all participants—including occupation, industry, technical experience, programming proficiency, mathematical background, puzzle-solving aptitude, and var- ious other measured attributes—demonstrated clear, statistically significant relationships with performance outcomes. This finding suggests that ARC-AGI-2 tasks assess general problem-solving capabilities rather than domain-specific knowledge or specialized skills acquired through particular professional or educational experiences.
It is not a judge of whether we got to AGI. And literally no one except straw-manning critics are trying to claim it is. The point is, an AGI should easily be able to pass it. But it can obviously be passed without getting to AGI (as . It's a necessary but not sufficient criteria. If something can't pass a test as simple as AGI (which no AI currently can) then it's definitely not AGI. Anyone claiming AGI should be able to point their AI at the problem and have an 80+% solution rate. Current attempts on the second ARC are less than 10% with zero shot attempts even worse. Even the better performing LLMs on the first ARC couldn't do well without significant pre-training. In short, the G in AGI stands for general.
So do you agree that a human that CANNOT solve ARC doesn't have general intelligence?
If we think humans have "GI" then I think we have AIs right now with "GI" too. Just like humans do, AIs spike in various directions. They are amazing at some things and weak at visual/IQ test type problems like ARC.
1 reply →
This is what is called "spikey" intelligence, where a model might be able to crack phd physics problems and solve byzantine pattern matching games at the 90th percentile, but also can't figure out how to look up a company and copy their address on the "customer" line of an invoice.
Maybe it is a cultural difference aspect, but I feel that "supermarket workers, gas station attendants" (in an Asian country) that I know of should be quite capable of most ARC tasks.
Out of 100 of evals, ARC is a very distinct and unique eval, most frontier models are also visual now, don't see the harm in having this instead of another text eval.
By both definitions of intelligence in the presentation we should be saying "how we got to AGI" in the past tense. We're already there. AI's can deal with situations they weren't prepared for in any sense that a human can. They might not do well, but they'll have a crack at it. We can trivially build systems that collect data and do a bit more offline training if that is what someone wants to see, but there doesn't really seem to be a commercial need for that right now. Similarly, AIs can whip most humans at most domains that require intelligence.
I think the debate hqas been flat-footed by the speed all this happened. We're not talking AGI any more, we're talking about how to build superintelligences hitherto unseen in nature.
According to this presentation at least, ARC-AGI-2 shows that there is a big meaningful gap in fluid intelligence between normal non-genius humans and the best models currently, which seems to indicate we are not "already there".
If there is a gap in intelligence between two humans, does that mean to you that one of them is necessarily not a general intelligence? The current crop of AIs get some of the questions right by reasoning through them. That means they are already intelligent in the way ARC-AGI-2 measures intelligence. They just aren't very capable ones.
If AI at least equal humans in all intellectual fields then they are super-intelligences, because there are already fields where they dominate humans so outrageously there isn't a competition (nearly all fields, these days). Before they are superintelligences there is a phase where they are just AGIs, we've been in that phase for a while now. Artificial superintelligence is very exciting, but Artificial non-super Intelligence or AGI is here with us in the present.
2 replies →
There's already a big meaningful gap between the things AIs can do which humans can't, so why do you only count as "meaningful" the things humans can do which AIs can't?
I enjoy seeing people repeatedly move the goalposts for "intelligence" as AIs simply get smarter and smarter every week. Soon AI will have to beat Einstein in Physics, Usain Bolt in running, and Steve Jobs in marketing to be considered AGI...
1 reply →
Well, there is also robotics, active inference, online learning, etc. Things animals can do well.
Current robots perform very badly on my patented and highly scientific ROACH-AGI benchmark - "is this thing smarter at navigating unfamiliar 3D spaces than a cockroach?"