I have been claiming that I don't think Chinese AI companies are benchmaxxing harder than American AI companies, which has gotten mixed reception: sometimes people agree, sometimes they disagree.
It seems I was wrong. American AI companies might actually be benchmaxxing harder.
"Opus 5 states the hidden rules before its first action, then plays a byte-identical optimal solution in 5/5 seeds at temperature 1.0. Zero exploration."
I guess there is no way this can happen without benchmark being part of the training data??
What seems to implied is that some of his hold out testing suite includes simple/common tests that are out in the wild, and for those opus went straight to a memorised solution .
Simple/common tests is not an explanation for why only now Opus 5 is the only model encoding the answers like this. Something like the holdout test suite being leaked or Anthropic cheating (e.g. 'accidentally' including previous hold out run data in Opus 5 training) makes a much stronger fit.
I was just thinking they need to mark each model per benchmark as "model released before the benchmark was released" and "model released after the benchmark was released".
Meh, I doubt it was intentional. Deliberate benchmaxxing is incredibly damaging to credibility once it's discovered (see what happened to Meta with LlaMa 4).
It's more likely that the training data was contaminated with the benchmark data.
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)
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
5 felt both smarter than me and dumber in some ways - it gets stuck to its original ideas. I had never seen a model harder to talk into changing its initial opinions. it continuously hedges.
Some 20 years ago, the telecommunications sector in Germany was liberalized. Many telephone card providers entered what had previously been a barely competitive market. They advertised their products with aggressive claims like: “Buy our €10 top-up card and get 660 minutes to destination X.”
For the first few weeks, they would actually provide those 660 minutes to establish trust in their cards. But after a while, they would quietly start reducing the number of minutes on subsequent top-ups—say, from 660 minutes down to only 300. They wouldn’t do this for every card, so it was difficult to prove. Instead, they relied on averages across their customer base to make the economics work.
Lately, I’ve found myself wondering whether something similar may be happening with frontier AI models. Companies launch with an exceptionally strong model and generous compute limits to build adoption. Once the model is established as a market leader, the incentives change, and users may start perceiving the service as becoming more constrained or less capable over time.
I don’t have evidence that this is what’s happening with Anthropic—or with any other AI company. It’s simply a pattern that the current situation reminds me of.
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?
What do you hope to achieve here? Posting absurd and ridiculous things and then continuing like if someone would take you seriously after that. Keep going I guess.
> 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.
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.
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 may not be "core" improvements (structural reliability) but there are "emergent" improvements (apparent intelligence). Already the IQ tests from Maxim Lott ( trackingai.org ) show a progressive sliding towards the right side of the curve - which btw translates to a very much non-secondary decline in the user's frustration (and progresses with an increase of usability).
The more they work on it, the more probable the jump becomes - e.g. to achieve the Large Conceptual Models you say you worked on.
You're welcome to speak up, but saying models haven't meaningfully improved in three years, when the most recent models are solving top-level frontier maths problems, is indeed going to strike most people as weird. I still have no idea what you mean.
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 :)
If you go to a LLM without harness, GP original point in completely right.
LLMS by themselves are still shit at math, they still confuse weird correlation to causation every time (and sometimes in ways even a 9 year old would say "no, that's dumb"), and confuse original parameters very often.
I disagree with " "thinking" token generation being on the correct track just to 'no, wait' on an already correct conclusion", because i think that is an effect of the harness, not the LLMs.
80% off all the improvements since ChatGPT4 are in the harnesses, and the LLMs by themselves, while they improved in areas they already were good at (translation especially) did not fix any of they original issues (object permanence, calculusm correlation).
Just run old models in the playground and get them to play chess (maybe make a small custom harness if you feel like it), then replace it with a frontier model (i don't know if you still have API access without harness on US models, but if you don't try K3), you will see LLMs weaknesses were not at all fixed, even marginally. They're way better and not inducing bugs in the code, so that make them usable since Opus4.5 (anyone using them prior to that either had a greenfield project or like spending hours debugging).
I agree, harnesses is where everyone improved the most. Our internal experimental tool can now semi-reliably formulate small programs to assert the correctness of their theories, for example. Context length is still a weird factor that we haven't sensibly solved - if anything, the lesson learned was to keep the context as small as possible and do most of the true "thinking" in the harness and temporary generated code.
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.
Great comparison! We only have to take into account that applying lipstick well bears no consequences, but applying it poorly (i. e. new model tanking the benchmark) could amount to potentially losses of billions of dollars for the pig-breeders (AI labs).
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?
but I kinda wish I could select level... I accidentally pressed redirect button and when I came back I was once again shown level 1, all progress lost :(
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.
The comment more likely means the harness source was read, not memory from a previous run and the first few turns of bp35 appear to be a cold start.
Sure none of this is certain without the source.
I do believe the authors that this schema significantly improves over the base, particularly given that it took 22x simulated turns over 14 hours, which is moving the trial and error to context rather than to game. I also don’t doubt there is some contamination.
Regardless, the approach is sound and I do believe it would significantly improve scores, even if that was +20-30 over baseline (49% in this case) it does imply the benchmark is measuring the harness more than the model.
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.
ARC-AGI is a terrible benchmark for testing LLMs because LLMs are not made, trained, or tuned for playing games.
They are trained on text to respond well to text based questions and do tasks involving modifying text files.
They are not designed for playing games, looking at games, or visual puzzles. Also translating games into text input for the LLM skews the test completely.
Imagine trying to get a human to solve visual puzzle but they can’t look at the puzzle but it has to be explained to them in textual format, we would be terrible at it.
But yet we persist in wasting time on this benchmark. It doesn’t mean anything.
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.
I also noticed that Opus 5 doesn't show a corresponding gap on the ARC-AGI-2 leaderboard. There is a significant increase in performance between Opus 5 and 4.8 on ARC-AGI-2 though.
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.
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.
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.
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.
Appears to be benchmaxxing
https://x.com/quietnning/status/2080786711861407883
I have been claiming that I don't think Chinese AI companies are benchmaxxing harder than American AI companies, which has gotten mixed reception: sometimes people agree, sometimes they disagree.
It seems I was wrong. American AI companies might actually be benchmaxxing harder.
"Opus 5 states the hidden rules before its first action, then plays a byte-identical optimal solution in 5/5 seeds at temperature 1.0. Zero exploration."
I guess there is no way this can happen without benchmark being part of the training data??
What seems to implied is that some of his hold out testing suite includes simple/common tests that are out in the wild, and for those opus went straight to a memorised solution .
Simple/common tests is not an explanation for why only now Opus 5 is the only model encoding the answers like this. Something like the holdout test suite being leaked or Anthropic cheating (e.g. 'accidentally' including previous hold out run data in Opus 5 training) makes a much stronger fit.
It may read information about the benchmark, such as on blog post, or Twitter feeds (example OP) without active cheat
I was just thinking they need to mark each model per benchmark as "model released before the benchmark was released" and "model released after the benchmark was released".
https://xcancel.com/quietnning/status/2080786711861407883
I'm shocked, astonished even, that enterprises on which trillions of dollars are being poured would consider cheating on marketing benchmarks.
Meh, I doubt it was intentional. Deliberate benchmaxxing is incredibly damaging to credibility once it's discovered (see what happened to Meta with LlaMa 4).
It's more likely that the training data was contaminated with the benchmark data.
You really think they saw the jump in arc-agi-3 (which they reported in their official card), and didn't even bother to check?
They maybe have not intentionally benchmaxxed, but they certainly know that's what happened .
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.
What if the harness is developed by the same model in a prior session?
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.
iirc, a harness isn't allowed, but if the LLM wants to write its own tools to solve things that is allowed.
The value of a harness is more about developer workflows, I don’t think it really improves the model output.
Harness can totally change capability of the model
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)
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.
It's called frog boiling.
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.
The actual term for this is hedonic adaptation.
esp. important to point that correct term because frog boiling is a urban myth.
frogs dont stay in a pot even if you slowly increase the heat. they leave. it has reportedly been attempted multiple times and they. always. leave.
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
5 felt both smarter than me and dumber in some ways - it gets stuck to its original ideas. I had never seen a model harder to talk into changing its initial opinions. it continuously hedges.
Some 20 years ago, the telecommunications sector in Germany was liberalized. Many telephone card providers entered what had previously been a barely competitive market. They advertised their products with aggressive claims like: “Buy our €10 top-up card and get 660 minutes to destination X.”
For the first few weeks, they would actually provide those 660 minutes to establish trust in their cards. But after a while, they would quietly start reducing the number of minutes on subsequent top-ups—say, from 660 minutes down to only 300. They wouldn’t do this for every card, so it was difficult to prove. Instead, they relied on averages across their customer base to make the economics work.
Lately, I’ve found myself wondering whether something similar may be happening with frontier AI models. Companies launch with an exceptionally strong model and generous compute limits to build adoption. Once the model is established as a market leader, the incentives change, and users may start perceiving the service as becoming more constrained or less capable over time.
I don’t have evidence that this is what’s happening with Anthropic—or with any other AI company. It’s simply a pattern that the current situation reminds me of.
Going to call it user error if you find Opus 4.5 better than 5, sorry.
Same, like I prefer 5.3 codex over the “stronger” models.
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)
> I've worked with these systems for four years now and they have not meaningfully improved in that time frame.
Not meaningfully improved?! Four years ago was gpt *3.5*! ChatGPT hadn’t been released!
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.
> you could do the same back then with multiple models and more targeted finetuning
Definitely not, lol.
I have no idea how you could try and hold this argument without being facetious.
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?
What do you hope to achieve here? Posting absurd and ridiculous things and then continuing like if someone would take you seriously after that. Keep going I guess.
You're moving the goalposts. Initially it was "no meaningful improvement" and now suddenly it has morphed into "they're not flawless".
I'm pretty sure you're just baiting for engagement though so well done, ya got me.
This is obviously not true to all of us here…
> 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.
Utter nonsense.
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.
It codes really well for our case though.
> you could do the same back then with multiple models and more targeted finetuning
Are you one of those anonymous billionaires as if you did this a few years ago, you would've been famous and rich.
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.
> are trying to sell as meaningful improvements
There may not be "core" improvements (structural reliability) but there are "emergent" improvements (apparent intelligence). Already the IQ tests from Maxim Lott ( trackingai.org ) show a progressive sliding towards the right side of the curve - which btw translates to a very much non-secondary decline in the user's frustration (and progresses with an increase of usability).
The more they work on it, the more probable the jump becomes - e.g. to achieve the Large Conceptual Models you say you worked on.
You're welcome to speak up, but saying models haven't meaningfully improved in three years, when the most recent models are solving top-level frontier maths problems, is indeed going to strike most people as weird. I still have no idea what you mean.
I think he means that, while results are there, they are mysteriously emergent, since an analysis of the process reveals it can be faulty.
> - Math completely fails in longer contexts
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.
I mean calculations, not mathematical proofs
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?
No, its capabilities with a harness are what we are interested in. Your assessment is only relevant to benchmarking, not practical value.
orwin‘s response in a cousin comment helped me see your original valid point on harnessless LLMs!
>You can achieve a lot by coding around the issues, but that's kinda against the point of 'AI', is it?
Will think on that a bit more.
Messages like this in the training data are how LLMs learn to say absurd things with total confidence.
> I've worked with these systems for four years now and they have not meaningfully improved in that time frame.
That's absolutely insane. Is it some case of anti-AI psychosis?
Like how toddlers’ skills don’t meaningfully improve on infants’, because either could wake up in a wet bed.
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 :)
If you go to a LLM without harness, GP original point in completely right.
LLMS by themselves are still shit at math, they still confuse weird correlation to causation every time (and sometimes in ways even a 9 year old would say "no, that's dumb"), and confuse original parameters very often.
I disagree with " "thinking" token generation being on the correct track just to 'no, wait' on an already correct conclusion", because i think that is an effect of the harness, not the LLMs.
80% off all the improvements since ChatGPT4 are in the harnesses, and the LLMs by themselves, while they improved in areas they already were good at (translation especially) did not fix any of they original issues (object permanence, calculusm correlation).
Just run old models in the playground and get them to play chess (maybe make a small custom harness if you feel like it), then replace it with a frontier model (i don't know if you still have API access without harness on US models, but if you don't try K3), you will see LLMs weaknesses were not at all fixed, even marginally. They're way better and not inducing bugs in the code, so that make them usable since Opus4.5 (anyone using them prior to that either had a greenfield project or like spending hours debugging).
> 80% off all the improvements since ChatGPT4 are in the harnesses
That seems easily falsifiable by putting an old model into the current harness and comparing it to 5.6 Sol or Fable.
I agree, harnesses is where everyone improved the most. Our internal experimental tool can now semi-reliably formulate small programs to assert the correctness of their theories, for example. Context length is still a weird factor that we haven't sensibly solved - if anything, the lesson learned was to keep the context as small as possible and do most of the true "thinking" in the harness and temporary generated code.
Sounds very fair, thanks :)
Well, what kinds of things do you see Opus 4.5 completely fail at? Maybe those are not the ones that newer models have improved on.
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.
Enshittification.
Why is there no Kimi 3, and GLM5.2 didn't run the third benchmark? I am more interested in knowing the abilities of open weight models.
ARC-AGI is a beauty contest for pigs where the pig's owners compete to see who can apply the lipstick most convincingly.
Great comparison! We only have to take into account that applying lipstick well bears no consequences, but applying it poorly (i. e. new model tanking the benchmark) could amount to potentially losses of billions of dollars for the pig-breeders (AI labs).
> Only systems which required less than $10,000 to run are shown. (Notes[1])
Am I lost or are their many models on this ranking (Opus 5 included) that clear this?
Many models are much cheaper through their subscriptions' included usage. That could be what's happening here.
Claude gives you something like $5000 of tokens on a $200 plan.
Isn't it ~$3000 per week? Extrapolating from the current limit on Pro plans.
They don't exactly say. I was extrapolating from some sessions usage (my figure was per month, so weeks times ~4).
Also top on the freshly released Frontier-Bench, by a large margin: https://www.frontierbench.ai/
I have a suspicion that they are just trained on puzzles by now
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?
The fact that it says so in the licensing conditions on AWS?
It's like we've come full circle:
First people practiced L33t3cod3 problems for interviews
Then people built AIs to build software
And now the AIs are studying L33t3cod3 problems
Why the 3 rather than e?
("Why 'leet' or '1337' instead of 'elite'". Because restricted groups like to stress a difference.)
It is RLVR, Not a puzzle, Not leetcode
Why isn’t Fable 5 included on the leaderboard?
games are great (as for a human)
but I kinda wish I could select level... I accidentally pressed redirect button and when I came back I was once again shown level 1, all progress lost :(
Not possible. I don't get how Opus 5 gets so high. Have they run it against the private and held-out games?
Are these typically the type of tasks that are genuinely worth tens of thousands of dollars?
It's actually crazy to see the difference between opus 5 and the next best model on ARC AGI 3 when you actually look at the ARC AGI problems
How believable is this benchmark? EG maybe opus was training on this? (You can try to identify the IP of wherever previous ARC questions came from)
That’s why you have a private dataset.
Which you have sent to Anthropic/OpenAI/Google's servers when you run the benchmarks for the previous models.
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.
> people built harnesses that solves arc agi 3,
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).
For frontier models, not local.
https://schema-harness.github.io/
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.
No but the session traces are available. It passes the sniff test considering how AGI-3 is scored and how this wrapper works.
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
Not my sniff test :)
> # 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.
The comment more likely means the harness source was read, not memory from a previous run and the first few turns of bp35 appear to be a cold start.
Sure none of this is certain without the source.
I do believe the authors that this schema significantly improves over the base, particularly given that it took 22x simulated turns over 14 hours, which is moving the trial and error to context rather than to game. I also don’t doubt there is some contamination.
Regardless, the approach is sound and I do believe it would significantly improve scores, even if that was +20-30 over baseline (49% in this case) it does imply the benchmark is measuring the harness more than the model.
If you stop and think about the problem it really is quite simple. Just need to build a graph of the game state and then run A* to get to the end.
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.
Seriously, play some of the games. They're fun.
Why? 30% is passing the first two problems only, which are really very simple.
Huh. How do things end up with scores like 30.2% (and results between 0% and 1%) if it's that low resolution?
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.
My own private benchmark shows the same thing.
this is not a good measure of current model capability. we need to test agents in harnesses, not models with a single prompt
test Codex, not Sol. test Claude code, not Opus
There are other benchmarks for that
ARC-AGI is a terrible benchmark for testing LLMs because LLMs are not made, trained, or tuned for playing games.
They are trained on text to respond well to text based questions and do tasks involving modifying text files.
They are not designed for playing games, looking at games, or visual puzzles. Also translating games into text input for the LLM skews the test completely.
Imagine trying to get a human to solve visual puzzle but they can’t look at the puzzle but it has to be explained to them in textual format, we would be terrible at it.
But yet we persist in wasting time on this benchmark. It doesn’t mean anything.
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.
I also noticed that Opus 5 doesn't show a corresponding gap on the ARC-AGI-2 leaderboard. There is a significant increase in performance between Opus 5 and 4.8 on ARC-AGI-2 though.
> it would suggest that the model didn't become significantly more intelligent overall, but significantly better at solving those specific problems
We would actually need a test that shows the ability of a model to export its skills to more problems ("interdisciplinarity" etc.).
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.
> Why is Fable not on here?
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.
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.
OpenAI's Zero Data Retention claim held up in court. They were unable to produce prompts and outputs because they were never retained.
I believe that is only available through Enterprise API for both Anthropic and OpenAI.
Is there a distinction we can independently assess between never retained and deleted or hidden?
Asking especially given CEO’s track record https://news.ycombinator.com/item?id=47659135
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.
I think you could have accessed Opus on AWS then u don’t have to trust that the data will go to Anthropic?
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 don't know why exactly, but Fable has felt the most human LLM to arrive.
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.
Deepseek V4 and Kimi 3 still missing, at least GLM is there.
solving ARC-AGI and being useful turned out to be two different problems
Because they're cheating.
cost 20k ???? man
those are like software engineer from third world country