> Only GPT-6 Astra solved anything: 2 of the 68 problems. It disproved problem 74 by finding a counterexample, at a cost of $218 and 15 hours of working time, and it proved problem 126, at a cost of $247 and 16 hours
> Across all attempts, GPT-6 Astra solved 5 of the 68 problems at least once: the two above, plus problem 1, which it disproved, and problem 548 and problem 571, which it proved. Most of the remaining problems were attempted between two and five times in total (172 attempts), and none was solved. Reaching these five solutions took over $220,000 of compute across all attempts, compared with roughly $20,000 for the benchmark run itself.
Which implies a genuine improvement in capability, but there's a still a very long tail ahead that models will continue to need to improve to capture.
"Problems solved before a model's training cutoff can be filtered out, and all models compared on the remaining problems" means that the problems an older model actually solved are the ones that get filtered out, while the remaining problems are the ones it already tried and failed on. So older models end up with 0s on the filtered set and you can't really use this to compare new models to older ones.
Also, since these are known public problems, you can't stop people from spending far more than your arbitrary time and $ limits on them. So the number of clean problems will go down over time.
A very long tail of problems that weren't solved by humans? Sure.
It's a sarcastic take and I understand that you are probably talking about "spiky intelligence", but you've chosen unsolved problems as a measure of the progress yourself.
Unlike traditional benchmarks, it's difficult to overfit your models to produce flattering results to unsolved problems. The open problems very likely do not have published solutions (the initial batches were merely models surfacing data that wasn't published in obvious places, but we're past that now). New models will exhibit something novel by adding solutions. And the matter of solution is interesting as well (contradiction versus a positive proof).
I would venture to predict it'll take years to decades to get to 0 open problems. But I'd be very happy to have this comment look foolish in retrospect as models continue to improve
I mean that it's hard to determine retroactively how much time it would have taken humanity to solve an open problem that was solved by AI. This is a measure that can make ASI look mundane because we don't know how long it would have taken mathematicians to solve a subset of the Erdős problems.
Yeah, I wouldn't purport to use this as a measure of intelligence as it applies to humans. I'd leave that to the philosophers, but my intuition is that models are still a long way off of true human-like intelligence (though benefit from certain unfair advantages).
I'm most interested in these problems as a relative measure of performance for successive model generations. If the prior generation couldn't solve a problem but the current one can, that's useful information, especially when we take into account what the proofs look like.
It's not a perfect benchmark, but I prefer it to many others that I see floating around.
but i feel like its purpose is to measure superintelligence in math. So to be a good measure of it, it can't saturate easily/has to be somewhat mundane at even insanely good levels. (though i do expect that once ais are across all areas/approaches superhuman at math at least 30% will be solved--then probably long-term (like after 2 years) less than 30% will remain unsolved.
It's pretty close to how we measure IQ. The standard test is basically a series of spatial puzzles.
I know there's a lot of people who complain that we're moving goalposts, but I think that the progress in LLMs really just shows that we don't know how to really measure intelligence in the first place, if we understand it as "human-like agency / ingenuity / adaptability". For decades, we saw the Turing test as the proxy for AGI, but then early LLMs could easily pass for a human in a casual conversation while clearly not matching human performance on most other tasks.
Since then, every benchmark we come up with, it turns out that an LLM can be fine-tuned to solve it while still clearly lacking something. They make very non-human mistakes, are easily tricked because they have a pretty tenuous grasp of reality, etc. But I think this just shows that AGI is a meaningless marketing term. We could as well be arguing if they have souls.
When I was 18, my high school girlfriend took me to the local Mensa chapter’s New Year’s party because her mother was a member and she was used to hanging out there.
It was a useful lesson that whatever IQ tests measure, it is completely devoid of value or interest to me.
At the risk of sounding like one of those people at Mensa that annoyed you...
The people at Mensa aren't a valid sample of people who score high on IQ tests, because there is a such a strong selection effect for people with certain personality traits, such as wanting to join a club based on your IQ.
By looking at interactions of people in Mensa I feel like 2% of top IQ is actually too low bar to notice anything particular about "high" IQ people. We are still pretty random in our interests and performance.
I suspect my IQ isn’t quite high enough to join Mensa but I’ve always flirted with the idea of trying to join and getting in just to see what a group of Mensa people are like.
We are basically random. 2% of population by top IQ test result has pretty much very similar distribution in all aspects to 100% of the population. It's not a high bar to be fastest processing one among 50 people.
Is it a serious question? IQ tries to quantify the positive correlation between the results of all intellectual tasks a person takes (AKA positive manifold).
I wanted to highlight that we're actually looking at a trio of concepts: intelligence, IQ, and value. Strongly correlated concepts, yes, but also meaningfully distinct!
Whatever the people who came up with IQ intended isn’t really relevant to the question of whether IQ measures intellect. At best is is loosely correlated. Very loosely.
IQ strongly correlates with many real world outcomes that people tend to associate with being more "intellectual".
IQ has precise definition. It's what the tests measure. Intellect has only fuzzy handwavey definition. Correlation between IQ and intellect is about as loose as the definition of the intellect. The way people make the correlation even looser is by defining intellect in even more fuzzy and nebulous manner.
IQ tests are incredibly good at what they're designed for, which is discriminating relatively higher intelligence humans from lower intelligence humans. Also discriminating within a single human - they are routinely and reliably used to track cognitive decline.
For these purposes they are highly reliable (repeatable, internally consistent) and valid (correlate with ~everything to about the degree one would reasonably expect).
They were never designed for machines or non-human animals.
Nor were they designed for rare ranges of intelligence - these are by definition hard to create tests for, since it's hard to gather the sample sizes you need. So they work well for the middle ~98% of humans but can't discriminate well among the most profoundly intellectually disabled nor among true geniuses.
It might, but if there's one thing that hasn't changed since 2022, it's that the models tend to ace the tasks where you have gobs of training data and where verification loops are fast and cheap... and they are not nearly as amazing elsewhere. If it's close enough, they can generalize, e.g. translate one programming language to another. But there's a pretty steep cliff past a certain distance.
Case in point: you had hundreds of millions of JPEGs to vacuum up and bitmap image generation is amazing. But if you ask them to recreate the same scene as vector art, they will struggle to generate a decent SVG. Like, kindergarten-style pelicans on bicycles are the state of the art. It should generalize seamlessly, but somehow, doesn't?
I think it will happen, just like self-driving cars are happening, but it will probably be a slow process.
Nit: Turing’s actual imitation game is a party game (like Werewolf/Mafia) and nobody’s even trying to win at that. The LLM’s will just tell you they’re an AI.
This is an artifact of how we deliberately craft these models though. We could just as easily fine tune a model that will believe it is not an AI or will attempt to deceive users asking about it
There's a lot more to it. For example, its writing style would also have to improve so it doesn't immediately give itself away.
Also, the skill of the human opponents matters. You'd want to test it against people who have practiced playing the game. Otherwise, it's like the difference between building a chess bot that can win against random undergrads who don't normally play, versus winning against grandmasters. And it's not like there's a pool of skilled human players of the imitation game.
Their input and output interfaces are too different from human's and they're not nearly as smart to take our IQ tests, but both dolphins and octopuses can solve complex puzzles tailored for their environment. Those puzzles are the whole reason scientists know that dolphins and octopuses are more intelligent than other animals.
But we do know they're "intelligent" and also smart in an important capacity. So how gives we don't measure them by IQ? Because the IQ is not a good measure of intelligence or smarts.
> No, it is just because they have difficulties at the bench.
I'll put it in another way. A "gifted kid" can be measured incredibly well on an IQ test, but fail miserably at incredibly normal but very difficult tasks such as consoling someone for their loss and managing family crisis. This is a clear example where an IQ measure doesn't translate to a person being capable of meaningfully changing their environments for good which is one way we define intelligence.
On the other hand saying "the gifted person is highly intelligent/smart just not good at some things" really diminishes the other tasks, because they really are very difficult tasks but are not measured by an IQ test.
I don’t follow your logic. “The gifted person is highly intelligent, just not good at deadlifting 500kg” does not diminish the 500kg deadlift and I wouldn’t expect an IQ test to measure it.
Deadlifting 500kg is not a form of intelligence (or it could be under certain scenarios). I carefully chose specific tasks for my comment, because those tasks do reflect a kind of intelligence that's not measured by IQ.
I think the key here is to be wary of measurements that promise to capture the whole of what we consider intelligence (ie what people think of with IQs).
That would be odd framing: it's a good test for a form of intelligence and other forms of intelligence still require good tests. It remains a good metric - for its specific thing; those who believe it to be the whole metric are naïve. It is almost necessary though not really sufficient.
There's currently a big market for figuring out ways to measure intelligence. With a particular interest in ways that humans can score much higher than LLMs. If you have some ideas please do share!
Why? Seems like benchmarks that closely mirror the tasks you'd want an LLM to help with would be a lot more useful than some general intelligence benchmark.
I think children having learning abilities exceeding LLM test-time learning (currently only happens in-context). But it's unethical to determine the true baseline of a child age 6 spending 6 years learning a radically new skill to mastery--and besides if you apply RL pressure to the AIs it would be able to surpass it. I guess I still believe future AIs should have some form of continual learning at test-time.
Give away access to the model and go ask people from time to time if the model was of use to the person and if they were able to make the model work with them.
This is like arguing about whether a hot dog is a sandwich (of course it is) or whether the chicken or the egg was first (obviously the egg since all chickens come from eggs). Intelligence is just problem solving in the context of self-awareness. Machines don't have it and never will but they can simulate the process given inputs. You can argue whether humans and animals truly possess self-awareness and in what degree, but the definition of intelligence is as simple as the hot dog debate.
It was defined in Animal Intelligence by George John Ramones in 1882 as "intelligence is the capacity to do the right thing at the right time. It is the ability to respond to the opportunities and challenges presented by a context"
You should read more on the ARC prize, it actually has a pretty long history. We're on the 3rd iteration because they keep getting saturated. If you look at the score history over time on ARC AGI 1, 2 and 3 it's pretty impressive.
No, but figuring out that you're playing a snake-like puzzle game at all in an extremely general input domain and then solving it in the least number of moves definitely feels like evidence of intelligence.
You forget the benchmark. The human subjects were told they were being timed. If you believe the lowest time is the primary metric you will absolutely trial and error at speed instead of meticulously plan out your moves to minimize that metric.
LLMs are not timed and given that it costs tens of thousands of dollars to run this test they're not optimizing for speed.
So you've got a deceptive test, with one metric being told to humans and not applied to LLM and a hidden metric humans aren't aware of but LLMs are as the test.
This is flawed from the get go. It almost seems like this was deliberately setup to be able to claim AGI and superiority of LLMs
Not fully relevant: timing is crucial in all-pass tests, not crucial in pass-or-fail tests. I.e.: first of all, they have to be able to reach the goal, and that is already an achievement. Then - and in parallel - the problem solving must also be optimized for efficiency. But "solving" and "efficiency" are non coincident dimensions.
I'm also unclear as to how basic inferential logic puzzles spells out intelligence
I think if you summed up measures of intelligence as 'can it do basic symbolic logic in a chain with memory' then yes, you've now achieved the intelligence of an e. coli colony [0], congratulations
so give it an IQ test and call it AGI. these weird little puzzles are grounded in no research with no replication or mechanistic chain to practical use
"For a cost comparison, during our controlled testing, human participants were paid $115 per 90-minute session, plus $5 per game completed. Participants attempted approximately nine games per session, roughly $12.78 per attempted game before bonuses.
Most of this fee pays for the participant’s time and willingness to take the test, rather than the energy their brain uses (a closer proxy to compare with AI). If we look at only the brain’s energy, and price it as electricity, the estimate drops to about 0.6 cents per session, or 0.067 cents per game attempted."
Well I dont know about all of you, but I am celebrating meat based humans...
I think raw brain energy is not a fair comparison. Humans are not willing and able to serve requests at identical competence all hours of the day. You have to invest considerable resources to get a person to even do so for part of the day.
$360 per puzzle. When they tested people it took about 10 minutes per puzzle. If price/performance keeps falling at the same rate it has been, this will cost less than US minimum wage humans within two years. Three for Phillipines minimum wage.
Once it figures out a puzzle it could probably be instructed to design a specialized harness for Luna to be able to solve other instances of the same puzzle. Minimum wage workers are not solving novel problems.
> Astra’s progress helps clarify which AI capabilities are out of reach and which questions remain open.
Okay. But I don't think this entire article at all explained which AI capabilities remain out of reach. Did I miss something? Other than "oh I guess it could still get even more superhuman on ARC-AGI-3 than it is?"
Since low scored much lower than none, and none scored ~ around medium, could none default to medium in the API? I don't think the new models can even have "instant" via API, unless they train them for that (there was one gpt5 variant called instant or something).
Was OpenAI able to run ARC-AGI-3 tests previously so that they could build a custom harness for the specific tests in the set? Even with the standard harness, if they knew the problems ahead of them they could have used supervised reinforcement learning to teach the model how to solve these specific tests.
Actually I can answer my own question: we know that they have had previous access to the tests because they’ve run older models against the same benchmark.
I wouldn’t put it past a company like OpenAI with a long history of lying and being deceptive to record the tests and benchmaxx ARC. They have trillions of dollars of incentive to cheat any way they can.
“AGI” never made sense to me. It’s a purely marketing term right?
I’ve ignored it thinking it would go away, but it keeps coming up.
I get that consciousness differs from intelligence and that our waking awareness of life is a complete mystery.
Knowledge and thus intelligence however I consider as actively being solved by these large ML models. That is, with the right combination of machinery and know-how, you’ll get it.
But you’d be no nearer to solving consciousness.
Given this thought trajectory - what is AGI supposed to be?
AGI is the term invented because arguments about what AI meant had gotten annoying. Originally there was no distinction and people thought "AI" would mean human level intelligence. Chess and conversations and robotics and math all in one package. Then games and classification and some robotics got solved and called AI, but that didn't solve math or conversation or online learning or a host of other things, so AGI was coined to refer to most of the whole package, virtually all the capabilities you'd need to replace humans intellectually. Now we're quibbling about whether AGI includes robotics or full real-world physical agents or something less.
The consensus now seems to be that once you've got human-level intelligence and planning and executive function then you get recursive self-improvement that can eventually autonomously solve the robotics and world-modeling and other portions of human-equivalence.
Not sure why you are bringing up consciousness, that’s largely orthogonal to intelligence. AGI is usually taken to mean the capability to match or surpass human intelligence across all conceivable cognitive tasks, as opposed to being limited to certain kinds of tasks, or to not matching the general level of human intelligence in some respect.
Intelligence, and hence AGI, doesn’t require consciousness or emotions or sentience.
I think the distinction between these options probably doesn’t matter alll that much if the threshold you’re using is either “random person” or higher? Maybe bump it up to “random educated person”?
They are still different concepts of course, but I imagine that once one is achieved, the others aren’t far off.
Ok, but a random educated person will not perform well on vast majority of specialized tasks where professionals operate. A model like Astra probably will beat random educated person performance on majority of specialized tasks. It’s getting close to the level of professionals in many domains, and to genius level on some (e.g. math).
I’m just trying to understand the implications of the current frontier model capabilities.
We cannot in a declarative sense define what is economically valuable work even now let alone into the future.
People take what they can get for pay. Very few individuals can demand a wage. The value of employment is obviously designed around that, not some arbitrary definition of “valuable”.
Of course an AI will accept $0/hr, it doesn’t mean it does the job.
Anyone who could accurately define the value of work would be wildly successful without having to try.
That is not a useful definition for me unfortunately.
I like recursive self improvement instead. It seems like something that is actually quantifiable and kinda “the point” of why consciousness is important to humans.
So basically, being able to set it free on some long running goal and it sort of “lives” and autonomously does its own tasks?
I wonder at what point consciousness is necessary… that is, if you can have anything like that without it.
To the point that solving consciousness (and combining it with intelligence) is what gives you the autonomous, recursive, self-improving thing otherwise it can only drive in the dark and make big mistakes.
To your point I think - it’s why we don’t see too many non-conscious advanced biology (it rarely survives against those with it).
Right, you are interested in “AGI” and presuming none of that requires consciousness right?
For example, how do you know that “feeling pain” is not a functional prerequisite for a task. And that consciousness is a prerequisite for feeling pain
Because there is no feeling of pain involved during the reasoning about "How to improve the balance of power in the Indo-Pacific" for the human reasoner, hence there is no need to have any experience of pain for any reasoner.
99.9% with the right harness? Ok, we're at AGI then.
Prediction:
We will now see the goalposts moved towards "well, a human costs less / is more efficient" - that will prevail for a few months until they come up with some other test that humans can do easily but is hard for the bots. This cycle will continue for ever and in 25 years, despite having hyper intelligent embodied robots or whatever, we'll still be arguing about if the singularity is here and if we're at AGI for the rest of my life most likely.
TLDR: The official ARC harness throws away old context and reasoning. No real-world harness is this bad, the model has to re-learn the game repeatedly. OpenAI basically just added standard compaction. Their harness is still "general".
24% sounds correct to me. Many of my friend who completed phd are currently unemployed .... And, some of them started doing random job like uber to make a living.
If government don't step up and regulate outsourcing like 100% tax, big problems are coming up.
24.9% is damn close to the 24.8% average over the ast 25 years, for the thing it's measuring-- jobless, people working part-time or involuntarily, and workers earning less than $26,000. Since 2000, it's been over 30% for more years than it's been under 24%.
I agree, like the average human isn't generally intelligent.
IMO, AGI is literally no different from ASI, though people think it is. Like, Imagine you have 1,000 generally intelligent humans working for you (which nobody is really) and you were to point them at your pet project. That would be amazing!
What's the metric on "average human" from a global population of > 8 billion people .. and why would they score mid on a standardised IQ test skewed toward western education / culture?
Doesn't "saturated" mean that essentially there won't be any more progress in the benchmarch? Also of note is that two of your points only mean something on an occidental capitalist system.
Perhaps, but I think a bigger problem than lack of compute is the cost of rewards. Games like Chess and Go were solved long before self-driving, partly because it's incredibly cheap to acquire the reward of a bad board game decision, relatively to how expensive it is to acquire the cost of a bad driving decision. With driving, acquiring the reward can cost you $20/hr for human supervisors to generate disengagements, or $100k if you crash, or $30B if you crash the car into a person in a way that causes your company to collapse (e.g., Cruise).
yeah but I think you may be underestimating the amount of capital available for compute. if AGI is possible through some 5 trillion of expenditure on computers, there will be money for it.
also, you are underestimating how short a 10 year time frame is. we are close to self driving, the first neural net image model was in 2013. 13 years is a blink of an eye
You mean any repeatable benchmark will be saturated.
The problem is that there is a huge perverse incentive. The intelligence is in the training layer not in the model parameters, but the intelligence is really good at remembering things, so if you let it take the test, it can RL it.
So, you’re telling me I need to start a benchmark as a side gig to get a bunch of free compute.
Astra please create a benchmark that’s favorable to your reasoning skills with a human interface but don’t make the score too attainable add some small issues that keep you below 100% to look sensible and to keep my evaluation metric side gig going.
You missed the hard part getting on HN front page , ie. Getting the acceptance of the community / zeitgeist .
There is no incentive for OpenAI to subsidize is you if no one reads /reports on your benchmark . They are only going to fund a few that are currently popular .
Community acceptance doesn’t automatically mean the best , it is combination of some level of technical quality and the ability of the promoter to socially influence or get support of influencers .
Wake me when it's going to spontaneously fix my leaky faucet because if it doesn't do it nobody else will. Until it has that capability I don't really care.
It’s not that different than a lot of real world economies. Often paying for someone or something with better quality can reduce total costs. You have less failures, less mistakes, so on, so while the expertise or quality of the product is higher than cheaper solutions, they can be more reliable and over time ultimately cheaper.
The question I have is how far back that curve can go without relying on economies of scale to just drag all the points back to the left. And without overfitting a specific metric that I don’t need (like this test).
Yep. In particular, ARC-AGI-3 is a series of games where if you fail, you keep trying again (until eventually hitting a timeout). So the sooner you succeed, the sooner you stop spending tokens retrying. If it was a benchmark where everyone got one attempt with no retries, you wouldn't see it bend backward.
From https://epoch.ai/latest/announcing-frontiermath-erdos
> Only GPT-6 Astra solved anything: 2 of the 68 problems. It disproved problem 74 by finding a counterexample, at a cost of $218 and 15 hours of working time, and it proved problem 126, at a cost of $247 and 16 hours
> Across all attempts, GPT-6 Astra solved 5 of the 68 problems at least once: the two above, plus problem 1, which it disproved, and problem 548 and problem 571, which it proved. Most of the remaining problems were attempted between two and five times in total (172 attempts), and none was solved. Reaching these five solutions took over $220,000 of compute across all attempts, compared with roughly $20,000 for the benchmark run itself.
Which implies a genuine improvement in capability, but there's a still a very long tail ahead that models will continue to need to improve to capture.
reply