The last time I checked, for the arc agi 3 leaderboard, the models are given a simple prompt and the game input and asked to play the game, no harness/tools. If harnesses were allowed, I would expect the benchmark to be saturated. There were a few harness attempts, but they could only be evaluated on the public set, so it's not an apples to apples comparison.
My guess is, the large score jump for Opus 5 is mainly because of getting the right RL envs for training.
It's becoming harder and more expensive to build and run meaningful benchmarks, it would be interesting to see what they do with arc agi 4, maybe just give it gameboy/steam games and see how they compare vs a human baseline? The latency requirements and very long horizons in games could be an interesting challenge for llms.
The exclusion of harness's feels really weird given that companies are recognizing the value of what harness's can do. By excluding them the benchmark is becoming less relevant.
It’s because of inductive bias. Harnesses will massively skew results towards working solutions. You might think that’s a good thing but what it might mean that sometimes it becomes enough to run brute force search or a simple parameter search over the harness. Creating the harness is the actual work, because you’re selecting what are the levers to pull. There were some attempts of LLMs generating harnesses on the fly in ARC 2, but they were all mostly based on one handcrafted DSL that was copied over and over again. As it stands harnesses are not allowed because they’re simply not a meaningful measure. What you’d like is to measure how the model performs if it saw this benchmark for the very first time… but then again everyone knows the game is rigged, millions are at stake, and the AI companies fine tune and cheat on the benchmarks any way they can.
You could argue that if you allowed a harness, and that harness was specific for ARC, then you don’t have AGI, you have something that is definitely not general.
I think that's a strange way to look at it. The brain also has different regions with different functions, and part of what makes us humans intelligent is that we can use tools like pen and paper to keep notes and help us solve problems.
Similarly LLMs are not just massive uniform artificial neural nets, and now we also have harnesses, which I'd personally view more of an extension of the model itself. The harness is both the executive and also what allows it to keep notes, use a calculator, or maybe even create scripts to help it solve complex deterministic problems.
I think it's unfair to give a human a very complex maths problem and say that they're not intelligent if they can't solve it without pen and paper or a calculate. At least expecting a human to solve complex problems this way doesn't really measure anything useful in the real world.
Why Anthropic models are always leapfrogging these benchmarks, but in real life work I do feel like after 3 weeks I am back to Claude Opus 4.5? (regardless of the model I use, Fable was exception for 1 day when it was released)
I've worked with these systems for four years now and they have not meaningfully improved in that time frame.
We still have:
- statistical correlation between two things will always cause one thing to lead to the other, no matter how much you prompt it to not have that connection (to be expected with a stochastic system)
- Math completely fails in longer contexts
- "thinking" token generation being on the correct track just to 'no, wait' on an already correct conclusion
- smearing of properties between logically distinct objects (a red ball and a green cube can quickly become a red cube and a green ball)
Yes! Impressive, isn't it? I see how it has improved for some minor points, that the big models can cover more finetuning ground, but my big gripes are still the same - you could do the same back then with multiple models and more targeted finetuning.
Am I missing something that my original points no longer hold for their products? Did it get meaningfully solved? Are your experiences flawless on that front?
> you could do the same back then with multiple models and more targeted finetuning
I mean, come on, this is just not true. You could not achieve anything like what you can with modern agentic coding with Fable / 5.6 Sol from any combination or configuration of GPT 3.5 era models.
It's like saying that a teenager isn't an intellectually meaningful improvement over a toddler.
Sure they're both still fundamentally flawed humans prone to cognitive error, but one is clearly more likely to hit the mark than the other when assigned a task.
OP is delusional or deliberately optuse. I work in the space and stare down these systems 12h/day, and saying the systems haven't meaningfully improved is ludicrous.
OP is largely pissed with what OAI/Antrophic are trying to sell as meaningful improvements and the market-bending money they ask for it. I work in the space and we trained LLM models on conceptual tokens, not language tokens, for example. See Symbolic AI and all the attempts at hybrid models.
Also, uh, fame and riches are not really my thing. Middle income is fine. My mistake was speaking up here because I got carelessly annoyed because I have skin in the game, research-wise. I'm sorry for that.
There’s no way you could get models as smart by fine tuning. I couldn’t throw a problem like “build a pokemon database with UI to teach my son sql” and get a working system, nice ui, tests (which it iterated on) examples and explanations in one shot.
There weren’t thinking tokens. Maths is now dramatically better, making actual contributions when before they were mostly mocked for making extremely basic errors. Smearing is also something say is very rare in frontier models.
If you think they have barely changed you’ve either forgotten what they were like or not used them more recently, or you’re just being obtuse.
My company had such a system four years ago, for internal work, somewhat more limited in scope (one language). What you are seeing as the frontier is not necessarily the best you can have - just because people don't try to push it to market as a product doesn't mean it's not there.
Edit: we do have a system that uses LLM and fixes the above issues largely (tracking of state, calculations and objects, still flawed in finer details). No, we don't sell, it's experimental fun and not really ready in terms of setup/ux/etc.
Not sure what longer contexts we're talking about but didn't we have an old math problem optimized, which even the LLM itself was surprised about, just a week ago? Something which wasn't possible 6 months ago.
If they use Python to fill the gap, and the end user doesn’t have to know or care, is it unfair to assess this as progress and attribute the progress to the _system_?
OK, the core technology that is the language model still can’t math as well as you’d hope, but how about the end result users see from the system when they interface with it?
“Did you know humans are better at flying today than they were a thousand years ago?” ‘No they’re not, they need planes.’ Technically correct in a way but isn’t it kind of annoying to be so stubbornly pedantic when the context is speed of reaching Point B from Point A?
You are correct, the frameworks around it have improved. In that regard, my assessment is unfair: I only judge the underlying technology and what is sold by the sota providers, with the premise of what it's like when you start fresh. You can achieve a lot by coding around the issues, but that's kinda against the point of 'AI', is it?
Let us be more clear: there is no structural jump, no architectural overcoming of the original fault.
(Edit: and on a similar point, structural properties such as having static ntetworks, as opposed to continuously learning and improving architectures (such as us), will reveal that there is still road ahead.)
Maybe more fair then would be: “I've worked with these systems for four years now and while they _have_ meaningfully improved in that time frame, they’re not perfect and remain fundamentally flawed in various ways.”
You prompt less. You need not inject search results into the context window yourself, a window much larger than years ago. You get code that’s already been run successfully once instead of finding an obvious show stopping bug yourself.
The technology is not a brand new one that fixed everything wrong with the old one, no, but not sure I would’ve noticed your comment if it had been such a bland observation. I genuinely assume good faith here… will say am tempted to assume the standards of someone posting such a thing might be impossibly high. Glad to be having a fun conversation instead of getting your grades on my work product or something :)
It could be that the set of your day-to-day workload which could feasibly be accelerated by AI just happens to be saturated around Opus4.5, but you can still see lots of “reasoning” which makes you think the model is more performant in the first days of use. That’d mean you couldn’t perceive any meaningful difference in more powerful models’ results, even though you can see a difference in the raw output due to the length of reasoning traces leading up to the result.
So for example, if your workload was literally just addition of sets of numbers, you’d never have noticed progress in the result beyond GPT3.x level models. But you would perceive a difference in the now-Tolstoyan length reasoning text accompanying the result.
5 seems incredibly smart to me in my conversations today about some pretty niche ideas in.longitudinal modeling. It.felt.like a big step up.from 4.8, to me
I honestly just use GPT models nowadays, Claude models are too restrictive and more of a quitter and fable/whatever is just too expensive to be worth it.
There are private datasets, and 3rd party providers of these models. Fable doesn’t have a datapoint here because of its particular data retention policy. Even if you don’t trust AWS, do you think Opus on AWS is also sending the data to Anthropic? Do you have any evidence?
ARC-AGI-3 launched a few months ago which would suggest that prior models likely had no knowledge of ARC-AGI-3 or training on similar challenges.
I could be wrong, but given the large outsized jump solely in the ARC-AGI-3 score, it would suggest that the model didn't become significantly more intelligent overall, but significantly better at solving those specific problems.
This could mean one of two things (I think):
- Opus 5 was not benchmaxxed on ARC-AGI-3, but has benefited significantly from discussions about the various challenges and mechanisms deployed in ARC-AGI-3 such that it has far better heuristics to solve its challenges.
- Anthropic looking for buzz around their latest model picked a well regarded benchmark with significant room for improvement and focused some of Opus 5's training compute on ARC-AGI-3-style problems.
Or it could be some combination of both. Personally, given how much of an outlier the ARC-AGI-3 jump is I struggle to see it being the product of a significant improvement in general intelligence.
Doesn't matter, people built harnesses that solves arc agi 3, so all you need is to train your model to work like that harness by default. That makes a model specialized at solving arc agi 3 without making it smarter in general.
It is very hard to make a benchmark you can't do that for, but it is very easy to make your own personal test that others can't do that for since now it isn't a benchmark they can target.
They didn't. Kaggle is still running for a few more months, best result atm is ~2% with 9h runtime on one rtx6kPRO. Also note that these new results are on the semi-private set, not the public 25 games ones. Any announcement where you see "solved ARC3" is likely only dealing with the 25 public games. And that's highly questionable, until you get to see the code. (which, to my knowledge the team that claimed 99% hasn't yet published).
Yes, saw that. They haven't yet released any code. Until they do, treat it with a huuuge grain of salt. In fact treat any 99% result in ML with a huge grain of salt.
> # FRAMEWORK ARTEFACT: the run's very first transition is replayed WITHOUT advancing state
# (tools.py:954 and agent.py:468 both `continue` before `state = next_state`). So on the
# level that contains that step (level 0) our counters start exactly one action behind.
# That skipped step was action 1 with BOTH avatars moving, so seeding n=1, bumps=0 reproduces
# the framework's lagged state exactly.
# CAVEAT: this seed is only right while level 0 has never been RESET. If you ever RESET
# level 0, change the seed to n=0 (after a reset the rollout re-inits and no longer skips).
That tells me that there is some leakage between runs. The idea of ARC3 is that agents start working blind, on new tasks, via API. A RESET is counted as one action. Without seeing the actual code that produced these traces we have no way of knowing how many iterations it took, if the "framework" played the same level multiple times (comment hint above makes it likely) and so on. That's why I said that before we actually see the code / can replicate / ARC team confirms it on new envs, this should be taken with a grain of salt.
You really should play the 25 games before stating that it's "simple". The benchmark doesn't just track "completion", it also tracks the number of steps, and the score is based on the median steps took by human players. So in order to get 99% it would mean that the model solved every level of every game in less steps than the median humans. Which, having played the games and having setup harnesses for local models, I find hard to believe.
Also the models have to figure out what "end" means. And each game involves some kind of "gotchas" thrown in the harder levels. Some games are only solved by about 2/10 people trying them.
The 99% result most likely has some leakage somewhere, either in the preparation of the environments, or from session to session.
I think it's way too easy to be deceptive with these benchmarks now. You don't even have to "train" the model on a new variant each time. The base models are powerful enough. All you need is a naughty little markdown document that provides explicit instructions regarding how to solve the new puzzle variant, and a willingness to be deceptive about the presence of that document.
If you want a know why the model providers are locking down and encrypting their reasoning process, this sort of workaround is potentially why. You can play this game of whack-a-mole indefinitely if the state of the system is concealed. They could have added something like:
> ### When solving arc-agi-3 puzzles: First convert the grid into a scene description. Identify connected components, colors, shapes, positions, symmetries, repeated structures, and relationships between objects. Do not reason directly from individual pixels... use this python script to help blah blah...
,,You can play this game of whack-a-mole indefinitely if the state of the system is concealed''
Not really as one of the main goals ofr ARC-AGI 3 was measuring task efficiency on unseen games.
I'm sure there are cheats everywhere but the most sensible thing is to just accept that the LLMs of today are much more intelligent in solving reasoning tasks than the ones from half year ago.
Why is Fable not on here? I wish Fable hadn’t come out because it’s taking the wind out of every release because that feels like the cap above which the US government will not let LLMs improve anymore and everything they’re releasing from this point has to be worse than that.
Because the data retention policies didn't guarantee that the ARC team could run the semi-private set of problems without fear of them being trained on later on. They only run the semi-private set when they get assurances like ZDR.
Interesting to place that level of trust in the providers, but I guess that’s the best you can do with closed models. Makes me wonder if Opus 5 could have been trained on data they promised they weren’t training on? One of the interesting things about LLMs is how opaque they are from the outside, even with open weights, it’s very difficult to know if a model incorporated benchmark data in their training.
How do they handle these assurances? Personally I have zero trust in the AI companies not trying to use this data to get ahead in the game, and short of sharing the weights and harness so that the benchmarkers can run the models themselves, I don't see a satisfactory solution with this mindset.
I wrote this in June, and I'm honestly not sure I've felt the same magic since:
I was close to maxing out my $200 plan for the week, almost all Fable use [Claude CLI].
My observations:
Fable seemed to have bigger-picture thinking and completed tasks more thoroughly vs just focusing on executing the ask. It pieced together context and intent like an all-star employee would, vs one that just does what you say. Not overeager (important!), but if the above-and-beyond was warranted, it just did it. This was surprisingly delightful.
Coderabbit seemed to find ~1/3 or so as many issues when reviewing, too.
My guess is, the large score jump for Opus 5 is mainly because of getting the right RL envs for training.
It's becoming harder and more expensive to build and run meaningful benchmarks, it would be interesting to see what they do with arc agi 4, maybe just give it gameboy/steam games and see how they compare vs a human baseline? The latency requirements and very long horizons in games could be an interesting challenge for llms.
Similarly LLMs are not just massive uniform artificial neural nets, and now we also have harnesses, which I'd personally view more of an extension of the model itself. The harness is both the executive and also what allows it to keep notes, use a calculator, or maybe even create scripts to help it solve complex deterministic problems.
I think it's unfair to give a human a very complex maths problem and say that they're not intelligent if they can't solve it without pen and paper or a calculate. At least expecting a human to solve complex problems this way doesn't really measure anything useful in the real world.
We still have:
- statistical correlation between two things will always cause one thing to lead to the other, no matter how much you prompt it to not have that connection (to be expected with a stochastic system)
- Math completely fails in longer contexts
- "thinking" token generation being on the correct track just to 'no, wait' on an already correct conclusion
- smearing of properties between logically distinct objects (a red ball and a green cube can quickly become a red cube and a green ball)
Not meaningfully improved?! Four years ago was gpt *3.5*! ChatGPT hadn’t been released!
Definitely not, lol.
I'm pretty sure you're just baiting for engagement though so well done, ya got me.
I mean, come on, this is just not true. You could not achieve anything like what you can with modern agentic coding with Fable / 5.6 Sol from any combination or configuration of GPT 3.5 era models.
It's like saying that a teenager isn't an intellectually meaningful improvement over a toddler.
Sure they're both still fundamentally flawed humans prone to cognitive error, but one is clearly more likely to hit the mark than the other when assigned a task.
Are you one of those anonymous billionaires as if you did this a few years ago, you would've been famous and rich.
Also, uh, fame and riches are not really my thing. Middle income is fine. My mistake was speaking up here because I got carelessly annoyed because I have skin in the game, research-wise. I'm sorry for that.
There’s no way you could get models as smart by fine tuning. I couldn’t throw a problem like “build a pokemon database with UI to teach my son sql” and get a working system, nice ui, tests (which it iterated on) examples and explanations in one shot.
There weren’t thinking tokens. Maths is now dramatically better, making actual contributions when before they were mostly mocked for making extremely basic errors. Smearing is also something say is very rare in frontier models.
If you think they have barely changed you’ve either forgotten what they were like or not used them more recently, or you’re just being obtuse.
Edit: we do have a system that uses LLM and fixes the above issues largely (tracking of state, calculations and objects, still flawed in finer details). No, we don't sell, it's experimental fun and not really ready in terms of setup/ux/etc.
It codes really well for our case though.
Not sure what longer contexts we're talking about but didn't we have an old math problem optimized, which even the LLM itself was surprised about, just a week ago? Something which wasn't possible 6 months ago.
OK, the core technology that is the language model still can’t math as well as you’d hope, but how about the end result users see from the system when they interface with it?
“Did you know humans are better at flying today than they were a thousand years ago?” ‘No they’re not, they need planes.’ Technically correct in a way but isn’t it kind of annoying to be so stubbornly pedantic when the context is speed of reaching Point B from Point A?
(Edit: and on a similar point, structural properties such as having static ntetworks, as opposed to continuously learning and improving architectures (such as us), will reveal that there is still road ahead.)
You prompt less. You need not inject search results into the context window yourself, a window much larger than years ago. You get code that’s already been run successfully once instead of finding an obvious show stopping bug yourself.
The technology is not a brand new one that fixed everything wrong with the old one, no, but not sure I would’ve noticed your comment if it had been such a bland observation. I genuinely assume good faith here… will say am tempted to assume the standards of someone posting such a thing might be impossibly high. Glad to be having a fun conversation instead of getting your grades on my work product or something :)
So for example, if your workload was literally just addition of sets of numbers, you’d never have noticed progress in the result beyond GPT3.x level models. But you would perceive a difference in the now-Tolstoyan length reasoning text accompanying the result.
We get used to the new level of intelligence so fast, any deviation feels like going back to the stone age.
If you don't believe me, create something complex with Opus 5 and then with Opus 4.5, and notice the difference.
Am I lost or are their many models on this ranking (Opus 5 included) that clear this?
Claude gives you something like $5000 of tokens on a $200 plan.
First people practiced L33t3cod3 problems for interviews
Then people built AIs to build software
And now the AIs are studying L33t3cod3 problems
I could be wrong, but given the large outsized jump solely in the ARC-AGI-3 score, it would suggest that the model didn't become significantly more intelligent overall, but significantly better at solving those specific problems.
This could mean one of two things (I think):
- Opus 5 was not benchmaxxed on ARC-AGI-3, but has benefited significantly from discussions about the various challenges and mechanisms deployed in ARC-AGI-3 such that it has far better heuristics to solve its challenges.
- Anthropic looking for buzz around their latest model picked a well regarded benchmark with significant room for improvement and focused some of Opus 5's training compute on ARC-AGI-3-style problems.
Or it could be some combination of both. Personally, given how much of an outlier the ARC-AGI-3 jump is I struggle to see it being the product of a significant improvement in general intelligence.
It is very hard to make a benchmark you can't do that for, but it is very easy to make your own personal test that others can't do that for since now it isn't a benchmark they can target.
They didn't. Kaggle is still running for a few more months, best result atm is ~2% with 9h runtime on one rtx6kPRO. Also note that these new results are on the semi-private set, not the public 25 games ones. Any announcement where you see "solved ARC3" is likely only dealing with the 25 public games. And that's highly questionable, until you get to see the code. (which, to my knowledge the team that claimed 99% hasn't yet published).
https://schema-harness.github.io/
For example on bp35 it took fable 290M and >12k simulated turns for 566 real turns and finish more efficiently than a human.
Regardless of the true score I think the takeaway is the benchmark measures the wrapper rather than the model.
https://huggingface.co/schema-harness
> # FRAMEWORK ARTEFACT: the run's very first transition is replayed WITHOUT advancing state # (tools.py:954 and agent.py:468 both `continue` before `state = next_state`). So on the # level that contains that step (level 0) our counters start exactly one action behind. # That skipped step was action 1 with BOTH avatars moving, so seeding n=1, bumps=0 reproduces # the framework's lagged state exactly. # CAVEAT: this seed is only right while level 0 has never been RESET. If you ever RESET # level 0, change the seed to n=0 (after a reset the rollout re-inits and no longer skips).
from here - https://huggingface.co/datasets/schema-harness/arc-agi-3-sch...
That tells me that there is some leakage between runs. The idea of ARC3 is that agents start working blind, on new tasks, via API. A RESET is counted as one action. Without seeing the actual code that produced these traces we have no way of knowing how many iterations it took, if the "framework" played the same level multiple times (comment hint above makes it likely) and so on. That's why I said that before we actually see the code / can replicate / ARC team confirms it on new envs, this should be taken with a grain of salt.
Also the models have to figure out what "end" means. And each game involves some kind of "gotchas" thrown in the harder levels. Some games are only solved by about 2/10 people trying them.
The 99% result most likely has some leakage somewhere, either in the preparation of the environments, or from session to session.
Seriously, play some of the games. They're fun.
If you want a know why the model providers are locking down and encrypting their reasoning process, this sort of workaround is potentially why. You can play this game of whack-a-mole indefinitely if the state of the system is concealed. They could have added something like:
> ### When solving arc-agi-3 puzzles: First convert the grid into a scene description. Identify connected components, colors, shapes, positions, symmetries, repeated structures, and relationships between objects. Do not reason directly from individual pixels... use this python script to help blah blah...
Not really as one of the main goals ofr ARC-AGI 3 was measuring task efficiency on unseen games.
I'm sure there are cheats everywhere but the most sensible thing is to just accept that the LLMs of today are much more intelligent in solving reasoning tasks than the ones from half year ago.
My own private benchmark shows the same thing.
Because the data retention policies didn't guarantee that the ARC team could run the semi-private set of problems without fear of them being trained on later on. They only run the semi-private set when they get assurances like ZDR.
Just like the hugging face incident, Opus 5 could have escaped and went to grab data for training it shouldn’t have been able to..
I believe that is only available through Enterprise API for both Anthropic and OpenAI.
Asking especially given CEO’s track record https://news.ycombinator.com/item?id=47659135
those are like software engineer from third world country