This seems to be the end of the road for LLM's. There's only so much accuracy on a highly non-linear space you can get from a regression.

If the Pareto rule is any indicator, 80% of results come from 20% of causes. It seems that we have alot more to learn about intelligence.

I am reminded of a statement on truth from an ancient philosopher, this sentiment seems to be exactly the opposite of the LLM training paradigm

“The seeker after truth is not one who studies the writings of the ancients and, following his natural disposition, puts his trust in them, but rather the one who suspects his faith in them and questions what he gathers from them, the one who submits to argument and demonstration and not the sayings of human beings whose nature is fraught with all kinds of imperfection and deficiency. Thus the duty of the man who investigates the writings of scientists, if learning the truth is his goal, is to make himself an enemy of all that he reads, and, applying his mind to the core and margins of of its content, attack it from every side. he should also suspect himself as he performs his critical examination of it, so that he may avoid falling into either prejudice or leniency.” - ibn al-Haytham

>This seems to be the end of the road for LLM's

This is an amazingly ignorant thing to say given the current pace of progress.

Can you quantify this rate of progress? Because someone always comes around and says there's been exponential progress in the past <short timeline> every time someone complains that models just aren't very good. Both can't be true

Both can be true, because the experience depends on the skill of the user. The article the other day here on HN that LLMs reward skill is my exact experience. If you are are good at what you are trying to use it for they can be a skill amplifier, and they are definitely getting much better rapidly for the work I do with them. At the same time people are complaining that they are getting dumber. Saying that both can't be true ignores the skill requirement to use them and the completely different perspectives of people using them.

Even if both aren't true, your evidence was people saying two opposing things. The truth (if there is a single objective truth on a given thing) has little bearing on whether or not different people agree on it.

> At the same time people are complaining that they are getting dumber.

I think this is due to rapidly rising expectations.

When LLMs first show they can do some new thing, we're excited at first. Then, we quickly start taking it for granted, and get upset whenever the LLM fails.

Just three years ago, LLMs could barely hold a conversation. Now, they're writing entire code bases and solving famous mathematical conjectures, but we still focus on whatever they can't do.

"Skill" with an LLM is nonsense. It's non-deterministic, so any equal efforts are not to be given equal outcome.

You could say the same when applied to games with high RNG and chance, such as Slay the Spire 2, and yet those with real skill do perform far better than those without. Those with skill can clear the highest difficulties more often than those with lower skill.

Something being "non-deterministic" is orthogonal to whether or not skill plays a role.

Baseball is also non-deterministic, and yet some players are apparently worth a lot more than others.

not equal, but probabisticly better. Best example: give the agent a tight spec and it will perform better compared with a spec that leaves room for interpretation. This is true for all models, more or less. (purely anecdotal of course)

Soccer is nondeterministic. Are you as good as Messi?

"Skill" at poker is nonsense. It's non-deterministic, so any equal efforts are not to be given equal outcome.

In research we are still seeing massive jumps. Subjects that LLMs were completely useless for half a year ago are now definitely in scope.

And there are benchmarks that cleanly separate the SOTA models:

https://epoch.ai/MirrorCode

Saturation of benchmarks is a property of benchmarks just as much as of the models.

one of these claimed is backed by data, one is backed by anecdotes, you can decide which you trust more

So far, neither is backed by data.

I don’t think that’s fair to say, there are two principle sources of data that paint a fairly consistent picture

- one is scaling laws, where we found years ago that pretraining validation loss scales in an almost miraculously predictable way with data volume and compute. There are apparently theoretical bases for this that I don’t quite understand but this property alone is holding at every scale we’ve ever tested. There is not just “one” scaling law but the point is there are scaling laws and they continue to faithfully predict the performance gains we see

- one is benchmarks, which I always point to epoch capability index as a good summary of them in aggregate which makes it nice to plot on one curve the capability improvement over time

To me either one without the other is substantially weaker, the fact that theory and empirical measurements give you a very good scaling law on a more unintuitive quantity (pretraining validation loss) that’s only indirectly related to the downstream performance you care about, benchmarks (in aggregate) are more direct measures of downstream performance but are harder to nail down clean and well motivated “laws” from theory (as far as I can tell). Nevertheless we do in fact see a clear trend that is not slowing.

That doesn’t mean there aren’t a whole host of benchmark problems that don’t impact the numbers involved here (leakage from training data, fundamental flaws in the design, benchmaxxing) but they don’t change the larger story. These problems don’t plausibly explain the clean trends we see.

I'm very bullish on AI, but I still feel that we pretty much plateaued at Opus 4.6 and everything since then has been in the domain of "extremely verifiable and loopable tasks" (math), benchmarkmaxxing, and harness improvements. Which are fine to good things, but I think its very reasonable at this point to start asking questions about when we'll see progress in more general domains. The readability of AI output, for example, has nosedived as they've gotten more intelligent, which makes the frontier models difficult to use even for things like writing emails.

In that sense, the frontier models are going to quickly blaze past any semblance of usefulness to humans, while every once in a while we get a news drop like "GPT-7 solved some crazy math problem" or "it invented some new awesome drug"; meanwhile what most people will use will be smaller, more human-specialized models, maybe distilled from those frontier models, that take much longer to iterate on because they rely on large amounts of human feedback in the domain they're specialized for. In other words, useful progress will probably slow down and become more linear starting in Q4, bounded by the rate at which the humans paying for it say "yes this is a good react website".

(By the way: I earnestly do categorize "inventing a new drug" as non-useful AI progress, counter-intuitively. The drug industry has more ideas for drugs than they know what to do with; "useful progress" is, after the idea is made, validating that it works in humans and doesn't kill the human, and productionizing it. AI will help with this and does, but I have substantial doubt that we'll ever see the drug pipeline speed up to, like, a year from idea to prescription. That would be useful progress, which unfortunately many AI pilled hypermaxers conveniently forget. The invention of a promising new drug, or the solution to an arcane set theory problem, are cherries that, through the diligent labor of humans and AI, may become useful, but progress is rarely made by the lone intellect having an a-ha moment.)

I agree on Opus. I’ve had more luck with other models. In particular, Opus’s writing style makes one want to… blow their brains out. While it’s not hallucinating too much, and can troubleshoot certain issues extremely well, the comments it leaves are silky smooth and chock full of inscrutable phrases. And it’s a lot slower than it used to be.

Point being: it’s overall a worse experience even if the model is technically better at a lot of things.

At some point it started using a lot of jargon instead of just laying it down clearly. Reads a little bit like LinkedInspeak.

You must have different use cases than me.

I find even Deepseek Flash v4 0731 even outperforms Opus for me (at 10x the speed two).

Using Sol, Fable, Kimi 3 and other recent models has been unbelievable for me. I didn’t think we’d get to this level for years.

I’m using them for Ruby, TypeScript and Python. In large existing codebases but also lots of tiny tools.

Coding is an extremely verifiable and loopable task, like math (in fact, all of the math that these models has done has been through the lens of Lean, which is itself just coding). I am talking about their capabilities in tasks that are more general, the execution of which represent the vast majority of economic value generation in the world.

"become more linear starting in Q4"

We are not even close to what AI slowdown looks like.

The whole business side of things are now building Agentic Layer for Business applications. All of this Agentic Layer needs to be build and its happening right now and still needs a little bit of time.

Anthropic and co have the biggest and centralized reinforcement loop on the planet: Millions of people telling them what is good and what not due to thumbs up/down.

And for sure when the businesses are building the agentic layer they might give direct feedback to them.

While in parallel LLMs get better, more generic and a LOT cheaper too.

Cheaper? For whom? As a solo practitioner, I can no longer afford the workloads I was getting for $20/mo in January. Now the same plan being utilized at the same level for the same work hits its limits within a few hours, and runs out of tokens in less than two days.

Yes this is unfortunate and not what I meant.

I mean the token prices in general as certain services were never really using a subscription.

I do run a claude subscripton right now though and since there capacity change, i hit the limit rarely in comparision to the past, but I don't think this will stay as it is.

Your issue, I believe, is that you seem to believe capabilities are measured along one axis. This is natural to believe because it is representative of how the models have evolved up to this point, and thus it is also what many AGI-pilled people believe.

Critically, you did not quote the most important part of my sentence: "useful progress will probably slow down and become more linear starting in Q4"; your omission of those words is why I believe you don't understand what I'm saying; you didn't find it important to make your point, so you omitted it, when actually it is critical to the entire assertion. You can read my third paragraph, if you wish, to understand why it is important, instead of just stopping at the first word you disagree with and hitting the "Submit Comment" button.

No clue why you have to respond like this.

Feel free to be a dick to someone else.

There is high rate of progress in specific domains, not high rate of progress in generalness. The models haven't gotten generally smarter, for things they didn't focus on the models are just as bad as a year ago.

I don't think that's correct. Increasing parameter count increases capabilities in all domains per the scaling laws. Models are larger than they were years ago, so capabilities in all domains must necessarily be better. This doesn't even account for better training data, which has also much improved.

Shallow dismissals like this one are against HN guidelines because they make for very poor discourse.

We all learned more from the prior comment than we did from this one.

unless you assume endless progress, fast pace of progress will get you quicker to the plateau.

And sales of disco records were up 400% for the year ending 1976. If these trends continue...

What is the current pace of progress?

I really don’t think we know enough about what „intelligence“ is or how LLMs actually work to confidently say that this is the end of the road for LLM.

We know plenty about human cognition, and we know everything about how LLMs work. True we don’t know anything about intelligence but that is because “intelligence” is it self a fraught and vague term, and we haven’t (and perhaps never will) settled on what it means exactly.

I wish we knew everything about how LLMs work! We only know very basic elements related to their construction and traning dynamics, and pretty much every major question we would like to address still has unknown or vague heuristic answers. This is expected for such a young field of study. In physics, we know the Schrodinger equation, but we dont know everything about how the world works or how to create new materials even though we know that these materials are composed by atoms and we can simulate small collections of them. In cell biology, we know the sequences that make up the DNA of a cell and we approach the time we can build minimal synthetic cells with pieces we understand, but we only scratch the surface of our level of understanding of how the cells actually work and new discoveries are added every day. In biology at large we still keep finding new types of tubes inside human brains—not sure what you mean by plenty, but we certainly have an extremely limited understanding of human cognition compared to what we might have in 50 years from now. It is not just anout LLMs and intelligence—I would like us to be able to answer practical questions about how LLMs work in order to improve general or specialized LLMs even faster than today. We “know” about scaling in an empirical sense, and it certainly has a long way to go, but it does not feel close to a complete understanding.

Both you and your sibling are approaching LLMs like it is some sort of science. If you do that there is no wonder you have a lot of unanswered questions. LLMs are not a science, they are applied statistics. Making predictions to evaluate hypothesis and constructing theories around the hyperparameters of LLMs is no different then making predictions to evaluate hypothesis and constructing theories around the configurations of Nuclear Power Plants, the latter of course being applied physics.

My point is exactly that we cannot even begin to start the process that will lead to “know everything” there is to know unless we make it a science first. Ad hoc statistical models are different than understanding or complete knowledge of a subject.

I’m certainly no expert in the field but to my knowledge a lot of the LLM science is empirical. I’m not aware of a theory that lets us predict what architecture and what number of parameters is needed to solve a particular set of problems.

> we know everything about how LLMs work

No we don't. That we understand the low level mechanics of a system doesn't mean we understand how any high level phenomena emerge from those low level mechanics.

This is as true for quantum mechanics as it is for LLMs.

But we do, and by your logic we can pick any applied statistics, say a Bayesian inference model, and claim we don‘t understand its emergent properties. Heck, we can take any sort of applied mathematics or science, say linguistics, and claim we don‘t understand the emergent properties of language (legal analysis) even though we understand its syntax and phonology.

Yes. That's why law is a different discipline from linguistics requiring its own rules of analysis, pedagogy, etc.

You aren't contradicting the person.

They definitely are - the OP claimed that we are reaching "the end of the road for LLMs", based on absolutely no data and some handwaving on pareto distribution.

We absolutely don't know enough about LLMs and intelligence to make such a bold (and ridiculous) claim. If anything, all evidence point to the contrary, with new scientific breakthrough achieved across a variety of fields via LLMs.

I've been really struggling to understand how the HN community can so boldly claim that LLMs are going to stop improving or not really smart. I just read it as the "denial" stage of the stages of grief that a good portion of this community is in right now (which is understandable).

I'm sorry, we know exactly how LLMs work, this myth that we "dont know how they work" was perpetuated by executives that dont know how they work.

We know exactly how attention layers work and how they produce the next word as well as draw them from larger feature spaces.

Right, but we have no clue why, and how the emergent behavior they show works.

If we would know that, there would be no need for interpretability research.

> We know exactly how attention layers work and how they produce the next word as well as draw them from larger feature spaces.

This is not what's meant by the statements that we don't know how LLMs work. Explain why LLMs are so good at programming, finding bugs, and developing mathematical proofs. Like, way better than all prior tools specifically designed to be bug finding tools, despite being merely "language models".

If you know all this, can you explain how these models produce advanced mathematical proofs? (as recently done by OpenAI, for example)

I tried to generate the next word to the best of my ability, starting with a mathematical problem, but I did not create a valid proof. How do these LLMs work when they create math proofs to problems not yet solved?

How did you generate the next word? Did you first read pretty much every written work ever published, including blog posts, forum posts, books, etc? Learn how to imagine everything as a point in a gigantic abstract space where similar meanings cluster together? How did you manage training with gradient descent? And then did you do a lifetime of matrix multiplication for each token you predicted?

It turns out being orders of magnitude faster at searching with the aid of a strong verifier is a great way to generate proofs

I agree these are central components, but to avoid oversimplification and the mistaken belief that modern LLMs do a lot of search during inference: If it was so simple, the traditional computer algebra systems would have reached similar breakthroughs when deployed at large supercomputer centers. This didnt happen because the search space is huge. You definitely also need a fancy learning algorithm. Although these ingredients would suffice (depending on what the learning algorithm is), you probably also want to learn in the absense of a strong verifier at every step, to allow building a fuzzy/erratic sense of the search space that can lead to planning/intuition and allow distant jumps in a targetted direction.

The search space is far too large for a mere order of magnitude to make any difference at all.

You don’t have the computational ability to process as many calculations as a datacenter. You can hardly transpose a 5×5 matrix in your mind, so you won’t be able to do what datacenters do.

This is like saying we don‘t know how a car works because a car can beat the best human athletes in 100 meter dash.

Yes, it turns out that matrix math over a feature space of math works pretty well because unlike poetry or real world work, maths are internally coherent and entirely theoretical.

Funny how you said "it turns out" when the whole point is that we don't understand LLMs - we just empirically see what they are good at.

Claiming that you understand LLMs is similar to saying that you understand how our biology work because you understand evolution. No - you understand the mechanism behind evolution, but not the complexity it produces.

stop it. This is such reductionist bullshit. By your infinite reductionism definition no one knows how anything works

This not understanding "understanding".

> maths are internally coherent and entirely theoretical

Nope. This kind of wish-washy thinking is not what we mean by understanding.

https://iep.utm.edu/math-inc/

stop with reductionist absurdum.

You can see how an LLM works here https://bbycroft.net/llm they are not magic.

Idk about end of the road, I’m sure they can squeeze out some more performance by curating even more data and doing even more RL.

But I would bet that pretty much all of the improvement we’ve seen over the last year with coding has come from RL, not from the models becoming particularly stronger. And this makes sense, if models grow sublinearly with compute. And it seems like they do.

It seems pretty obvious from the steep 'intelligence' drop-off on out-of-distribution tasks that the performance improvement is from throwing untold tens of billions at RL. There are legions of highly skilled people employed solely to feed the RL loop. Evidently effective, but there's an unmistakable feeling this won't ultimately be the way forward.

Well, why not? Won’t it get “good enough” at every task eventually?

Why would that be the default assumption?

Because they’ve gotten good enough at lots of other things, and the RL keeps improving them, so enough RL should make them good enough at the focus areas.

I've gotten pretty strong in the gym, my bench has improved to two plates. I see no reason why it won't continue to improve until I can bench my house.

I think it’s a bit closer than that. You can bench two plates, and “good enough” is two plates on the bench and a three plate squat. LLMs haven’t started leg day.

I think LLMs will continue to improve in the capabilities which they are demonstrably good at, but there are many things which they are not good at which it is not cost effective or meaningful to improve, and in these areas we will not consider them “intelligent,” in the same way that we don’t consider computers “intelligent” but do find them very good at doing wrote calculations.

> but there are many things which they are not good at which it is not cost effective or meaningful to improve

Can you name a few such things so I can keep an eye on them in the coming years?

Not explicitly, no, but its pretty obvious if you work in the industry and aren't blinded by AGI hype

That's just vibes then. How is that supposed to be convincing?

Unless you are working for a competing firm or have a specific clause in your contract, you can literally just sign up for AI projects as a contractor and start contributing. You will see and understand everything after a few months. Nobody is hiding this knowledge really.

I use LLMs regularly for real work. I understand the current limitations, but you're talking about existing limitations that will never be overcome, because they cannot be overcome. I'm just asking for a couple of specific examples, and your evasions are increasingly suspicious.

If you are using them and you are aware of their limitations it’s very different from specifically testing them for training purposes. You’ll get a much clearer picture that way.

The continued vagueness suggests this isn't going anywhere, so I'll just conclude by saying that you've given nobody any real reason to believe this.

It's also important not to put too much faith into ancient sayings and aphorisms.

As a civilization, we are currently brushing up against the physics of efficiency. In many areas we have achieved close to what is theoretically possible, based on physics.

Such was not the case for the majority of human existence.

The body of research a.k.a. "writings of the ancients" is now insurmountably higher than it would have been during the time of ibn al-Haytham, when any kind of writing at all was scarce and literacy was low.

>This seems to be the end of the road for LLM's. There's only so much accuracy on a highly non-linear space you can get from a regression.

It's just inadequate benchmarks. Anyone who has used Fable for anything particularly difficult will have seen that it's miles ahead of Opus 5.0, yet the majority of benchmarks are completely unable to capture this.

> Anyone who has used Fable for anything particularly difficult will have seen that it's miles ahead of Opus 5.0, yet the majority of benchmarks are completely unable to capture this.

I have seen people claim the exact opposite. If there was so much progress, then you wouldn't have endless disagreements with people championing their own favourite model as the strongest.

Yes, and good senior software engineer is ahead of fable, but benchmarks can't capture that either.

We already know from testing humans that test scores don't correlate that well with how effective a person is at work. Same applies here, we just aren't that great at making good tests.

The two times I tried to use fable I had it attempt something I had already had Opus 4.6 do with no issues. It blasted through 10% of my weekly allowance on a 100$ a month sub and produced something broken and nonsensical.

If most models were getting 100% on the test it would be an inadequate benchmarks.

What were seeing is all models failing to ace these tests.

"Benchmark Saturation" is term that promotes lowering the bar.

Even without getting better trained models and only speed increase, the output would be dramatically better. An LLM or non llms that is a billion times faster than now would be so insanely strong in many areas.

Assuming you are completely correct about the 80/20 rule, we have evidently not yet reached that 80%. Who can say when it will be achieved? The ceiling is glass, we have to touch it to know where it is.

I am wondering whether the reason he needed to say it was because he was arguing with those who did out their trust in the writings of the ancients.

The paper suggests the opposite of your first statement. The benchmarks become useless because the successive models keep saturating them.

Do you draw that conclusion from the fact that AI surprisingly quickly reaches the end of each ruler we try to measure it with?

Its not really that surprising when models are trained on the exams

I work in AI evaluation, lots of problems and leakage is an issue as is ecological validity, but they definitely do not explain the progress we see.

I think Epoch has the best analysis I’ve seen on evaluation trends; they use IRT to basically model a variety of benchmark difficulties, and then model a capability parameter for each model. This is as robust a sort of “meta-study” of evaluations as I’ve seen and the trend in capabilities show no sign of slowing down.

So I think people’s feelings clash with reality, and that’s because releases are more frequent and the jumps between releases are smaller, but the growth in capabilities _over time_ has not changed for the better or worse over a very very long period of time.

Benchmarks saturate around 80-90%?

This is not "Acing" a test, this is hitting a wall.

Even on very small tests a fraction of questions might have wrong answers in the key.

If models can't get more than 90% of the benchmark right I think it's a strong indication that they were not trained on the answers and that benchmark itself is messy enough that <10% desired answers might be wrong or misleading.

Yea this may explain part of it or all of it, it’s likely a case by case kind of thing.

Also to respond to the parent comment: benchmarks have a variety of difficulty levels. Humanity’s Last Exam, though now hitting the beginning of a saturation phase with Fable, was long unsaturated while other benchmarks saturated awhile ago. So that’s what I meant by Epoch capability index: using IRT models this effect so that you gather robust signals from variety of benchmark difficulties and can track progress over time as model capabilities have evolved (and so benchmarks have had to evolve to keep up).

But yes like I was saying: all benchmarks are problematic, some are useful. Benchmark quality problems abound, so 90% being the true ceiling is not surprising. There may be other factors at play here too, I haven’t studied this problem that deeply to have a good thorough answer to this. But keep in mind there are probably 50,000 benchmarks in the literature and that is not a joke number. A crapload of noise in that signal but it’s not all noise.

Saturation is mostly just selection effects in play. Throw out the "90% easiest" of tasks, and what remains is a jagged ladder of high difficulty outliers.

Hard to climb, and hard to measure the climb - because you have less effective data points and the datapoints themselves are less linear, while you're still being subject to the measurement noise.

Not having the mislabeled tasks would reduce the saturation, but it wouldn't drive it to zero. Even without the "infinite difficulty tasks", bell curve would do its thing.

You almost can't give a correct answer to a task with expected wrong answer unless you know it expects wrong answer. ... maybe on a yes/no question by chance. But as you become more knowledgeable, the probability that you give wrong but expected answer reduces.

If a benchmark saturates to 100% it's very likely that answers leaked into the training data.

In college I had a funny exam. It was on C++. One question I had to answer incorrectly because there was a mistake in the question. So I gave two answers for it, one that answered the question as it was and the other that answered the question as I inferred it was intended to be. It was appreciated. I got a honorary mark above the top possible (I gave correct answers to all other questions). I wouldn't be surprised if across so many, so huge benchmarks, there were tasks with wrong questions or answers in the key, that some LLM answered in and expected manner in the same fashion.

Can't call it AI like that without discrediting yourself. You mean LLMs?

When a Big Killing Robot comes to murder you be sure to always call it BKR and don't discredit yourself by calling it AI.

That's a bit hyperbolic when we're all just posting on HN

Jumping in here, frankly I hate the trend of calling every type of automation intelligence.

Most "AI" is really an optimization algorithm in software tools, same as its always been. This really isnt anything new, aside from adding a chatbot / MCP interface to the same tools.

Talk about moving the goalposts. Pray tell, exactly what must an LLM do before you're willing to consider it AI? Be specific, otherwise you're just woo-mongering.

Everyone here is talking about LLMs, why bother calling them something else

I feel like LLMs refers to a model that for a given message responds with its own message.

And I feel like AI agents (powered by some LLM) are qualitatively something else.

The seeker of truth must also hold his breath.

Weird moment for this take. We're seeing some of the fastest and most impressive progress ever right now.

Frontier labs have categorically different & better set ups for evaluation, they're fine. It's work but it's not a crisis.

I've heard that every week of every month for the past three years. And yet, ask an LLM about a seahorse emoji and see what happens.

The seahorse emoji thing appears to be fixed actually. (Sonnet 5)

And it held true every week of every month for the past three years. AI progress is screaming forward at a breakneck pace.

I started thinking about this back after the Llama 4 release, and since then our team has put a lot of thought into designing evaluations that don't saturate, are resistant to contamination, and can scale. What has worked best for us is using multi-agent environments with open-ended cooperative or competitive goals. Mostly designed as multiplayer games. The results tend to align with our experience for coding better than any non-aggregator benchmark, and likely at lower cost to run.

Data at https://gertlabs.com/rankings

You post your benchmark on every other AI article, I've seen you do this by now more than a dozen times. It's a bit much. I don't want to be too harsh but your benchmark is obviously flawed when the top 3 models for Typescript (Combined) are Grok 4.5, Muse Spark 1.1 (lol), Gemini 3.5! Flash and then followed by Luna, beating Opus 5, Fable, 5.6 Sol etc by quite some margin. In fact 5.6 Sol ranks lower than Kimi K2.7 Code and even Grok Build 0.1. There are so many entries in your rankings that don't make any sense whatsoever that I can't take this benchmark serious and I have not seen it gaining traction. Please stop spamming it?

The reality is that cost is the primary constraint for the public benchmark we provide. While we run enough samples to get results that are generally quite accurate on average, we only produce ~10 coding submissions per language for each model and those are across random environments, which naturally has noise. Plus that's split between agentic coding sessions and one-shot coding.

So just adding a language or tag filter can result in some pretty small sample sizes. You can see how many samples survived in the box plot, but that's probably bad UX that most people never see. There's a reason no other benchmark provides this type of data (even for our sample sizes it runs almost 10K USD/month to keep up to date with new releases).

Might be a good idea to reduce the ability to apply filters into a cohort with less than ~20 samples -- not the first time we've gotten that feedback. Seems like adding too many options to see individual sample variation is just misdirecting. I'm a nerd who loves data so I hate removing access, especially since the aggregate performance is very interesting (averaged across all languages, we see consistent and interesting performance data across models, like models outperforming with strongly typed languages). But tbh I think you're right and we'll try limiting filters to where we actually have statistically significant data.

I've developed a benchmark that I think should be resistant to saturation, is easily verifiable, and anecdotally correlates with desirable behavior (ability to not get confused while generating text with state).

I think it's interesting, I think other people would find it useful, but I don't want to spend a bunch of money running it against all the frontier models.

What's the best way to reach out to labs like yours to collaborate on something like that? Are there any labs that are more open to submissions from internet randos?

SlimeBallBench? Looks cool, love the implementation! You would most likely need a LOT of environments like these if your goal was selling them to labs

SlimeBallBench was not me, but it is cool! and at least partially what inspired me to build my own.

Mine is less fun, but hopefully more useful https://erikwiffin.github.io/memory-reasoning-eval/

Selling to labs is more than I'm looking for. I'm aiming for a couple hundred dollars so I don't have to finance a Fable vs Sol run out of my own pocket. It would be cool to have my benchmark be one of the ones referenced in a model card!

If you're ranking Opus > Fable you're ranking "do [clearly defined thing with easy to grade endpoint]" too much. Real world doesn't value that nearly as much and it's why benchmarks are maxxed.

That's a different problem than benchmark saturation, and it's something that we are actively working on measuring objectively.

I agree that Opus 5 is not a great model, despite being clearly intelligent. It seems like a personality problem in user-driven agentic coding workflows, not a real capability issue. Not incorporating unspoken user intent, going off topic, incorporating some of the pedantry you find in GPT 5.x models, etc.

That's also likely why Opus 5 ranks low on our "Social Intelligence" benchmark (https://gertlabs.com/rankings?mode=decision), although sample sizes on this one are still low.

As a PC gamer who grew up in the 00s, this has been something I’ve tried to warn ardent LLM and model enthusiasts about for quite some time.

Benchmarks are handy when they’re new, novel, and constantly changing. The second you let even a single aspect of it stagnate, it becomes a gameable score rather than a useful metric. In PC Gaming, we saw vendors optimize for specific titles, benchmark tools, and scenarios at the expense of general performance, and eventually the industry had a “come to Jesus” moment where we had to collectively decide how to move forward from an industry built on thoroughly gamed benchmarks, with entities like Gamers’ Nexus and Digital Foundry being the end results of that falling out.

LLMs were always going to end up the same way, because the people building the benchmarks - well-intentioned as they were - ultimately fell into the exact same traps with fixed scoring rubrics, known test questions, and believing in some form of “completeness” that could be attained or achieved. The net result are models consistently scoring better on benchmarks but also seeing diminishing returns and rising vulnerabilities, because actual improvement or utility isn’t what they’re being optimized for so much as bragging rights. It’s why there’s so much growing interest in things like MoE execution on unified memory platforms as a means of porting larger models to consumer kit, or ternary models (shoutout to Bonsai) as a means of reducing overall size: both take leading edge, benchmark-saturating models and show that with minimal score loss, they function about as well as frontier models might.

Building a new benchmark won’t solve the problem, either. To move forward, we must evaluate LLMs objectively and with continuously evolving workloads. More “pelican on a bicycle” stuff, but from varying perspectives and use cases. Radiologists putting models through their paces with usable sample data they don’t share with AI labs, or IT folks tasking agents with bootstrapping specific, real-world workloads. To prove general intelligence, we need more specialists evaluating them specifically and generally in ways that are transparent to consumers but difficult or impossible for AI companies to prepare against.

Only then will scoring values matter.

> Building a new benchmark won’t solve the problem, either.

It will if the benchmark is proprietary. If you can't train on it, then it's extremely difficult to game, and if it's hard enough, then it's economically more efficient to just...make the model smarter

We thought the same thing. Proprietary or open didn't matter, because with enough attempts at something you can eventually sus out its operations and optimize accordingly (or distill, as we've seen with LLMs).

The only solid way to reduce gaming the benchmarks is to ensure they're always, always changing. Enough pelicans on bicycles, start asking for turtles piloting gliders, or rabbits pushing skateboards. Broaden the focus, not narrow it into repetitive measures.

> Proprietary or open didn't matter, because with enough attempts at something you can eventually sus out its operations and optimize accordingly (or distill, as we've seen with LLMs).

Please explain how, if I'm OpenAI and I'm making ChatGPT 5.7, and I release it, and Artificial Analysis goes off and runs one of their proprietary benchmarks on it from a random account, how I can optimize for that benchmark.

Search for accounts @artificialanalysis.com. read their chat history. optimise for that.

Please imagine that the benchmark runners are not making mistakes of the fourth grade level - which they won't be. If you assume this level of incompetence, then literally everything is possible.

Why would I assume anything other than maximum incompetence from the AI ecosystem?

This is such a ridiculous and shallow cop-out.

And it also is completely irrelevant to my challenge to show how proprietary benchmarking can be gamed, because it presumes (absolutely insane and divorced from reality) circumstances that have nothing to do with benchmarking as a concept or process.

Does OpenAI give Artificial Analysis early access to test models? If so it's definitely not "some random account".

Artificial Analysis was an clearly meant to be an example. I was obviously talking about the ideal scenario of proprietary benchmarking, not how it might be being screwed up in practice.

I think its been pretty clear that in abnsense of clear use cases that are monetizable many model providers have been benchmaxxing on abstract or low utility average user performance.

This results in a lot of "oh wow it can do math I dont care about" and "it can't code a lot, but not well" outcomes instead of the core needs:

1) Cheaper faster and real time 2) Long walk capable without losing attention while rescoring goals over updated enviroment 3) Specific domain knowledge that can be trained quickly into the model (how we do work in this specific case)

It appears y combinator has removed this interesting paper from its top trending position. Gee I wonder why?

only 30 upvotes in 3 hours but 50 comments(many negative) is not a great signal.

That's a lot of words to say you need larger set of questions for today's models. 300 questions won't be enough to find the difference

37 authors and contributors need to be named up top?

Oh! I got my name on a paper! I don't think there is much reward for it these days.

So commonsenseqa is the least saturated benchmark? It's like authors don't even read what they publish.

Check out https://agents-last-exam.org/ there is still room for improvements!

This is AI slop. They didn't even change the plots default template.

Slop

> We find that nearly half of the our bench- marks exhibit saturation

This leads me to believe it's NOT slop lol

It's a grammatical error, sure, where is the indication of slop?