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
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.
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?
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.
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?
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.
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 giving thing) has little bearing on whether or not different people agree on it.
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.
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'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.
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.
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.
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.
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.
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?
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.
> 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
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)
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
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.
You aren't contradicting the person.
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'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.
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?
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.
You can see how an LLM works here https://bbycroft.net/llm they are not magic.
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?
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.
>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 giving thing) has little bearing on whether or not different people agree on it.
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
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.
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...
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.
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'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.
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.
The seeker of truth must also hold his breath.
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.
Can't call it AI like that without discrediting yourself. You mean LLMs?
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.
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
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
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.
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
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?
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
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.
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.
Check out https://agents-last-exam.org/ there is still room for improvements!
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)
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?