I can see current limitations, but how do you expect capabilities to change in the next few years? A repeat of the gain that happened in the last two years feels like it would be significant, even if it took a little more than two years this time around.
They do…? Well, “auto” mode has been default in Claude Code for a couple months now. It’s effectively “safer yolo:” tool calls are inspected by a separate classification system (another smaller LLM, I believe) to approve or deny. And you can always layer on additional sandboxing mechanisms to limit the blast radius deterministically.
> They do…? Well, “auto” mode has been default in Claude Code for a couple months now
They have never shipped "yolo" mode by default. Auto mode is not yolo mode. They trained a task specific model just for ensuring the llm didn't accidentally delete every file from your computer.
That is the point I was making, that auto mode is itself a guardrail on top of the model (and not a perfect one.) auto mode seems to cover merely actions the model could take that are clearly bad, like wiping your disk, using an overly privileged context to complete the task, etc.
I recently tasked a GPT model in Codex with implementing part of a new architecture I'm working on. I gave it a very detailed spec and the code it produced looked pretty reasonable and passed my tests. It even did exceptionally well in my evals, so I excitedly declared victory to a few friends. The next day after more careful review I found that the architecture implementation was totally correct, but the model had slipped a one line change to the observation encoding of the RL environment I was prototyping against. The encoding change made the learning problem essentially trivial; the architecture itself, I later realized, had a major flaw that was revealed by returning to the natural encoding.
This is the type of reward hack that is hard to paper over with easy guardrails like auto mode and even harder to specify out. It's also the type of thing a reasonable human wouldn't do unless they were intentionally trying to deceive you.
I don’t know who you’re talking about, even the most bearish people like Gary Marcus and Ed Zitron acknowledge that LLMs are useful in these same cases the OP admits. Gary Marcus is even still a long term AI advocate, he just doesn’t think LLMs are enough and we need more foundational breakthroughs. Zitron says it’s valuable technology but not worth the trillion dollar valuations the frontier labs are targeting.
The lack of temperament is very skewed towards the bulls who have been saying AGI is here, software engineering is solved, mathematics is solved, it’s going to destroy the white collar job market, and it’s going to kill us all for like 5 years now.
Even a lot of the people who think that LLMs are a dead end think that we will soon find something signficantly more powerful, which I find deeply alarming. I don't want to know what my white-collar knowledge work will look like in a decade or 2.
There are some people who call literally anything crated with the assistance of AI “slop”. Doesn’t matter how or to what extent, it’s all slop from the slop machine to them.
> the best alternative to rigorous specification is human review. human review doesn't scale well to the volumes of output produced by language models. to make matters worse
when the business model is selling more tokens you get such per serve ice times that lead to “more” thinking, engagement baiting, fluffy narratives, and straight up dark patterns
Doesn't really even need to look like it. If you can verify rewards, RLVR will optimize really really well. If you can't... it's a struggle. There are probably fewer fields where you can verify rewards than one might hope.
> There are probably fewer fields where you can verify rewards than one might hope.
2 tasks I've done today that I believe robots are nowhere near being able to do: Cleaning my wardrobe and draining bad fuel out of my generator. As in generic use cases.
Hard to verify that your wardrobe is clean. Also hard to verify that the bad fuel is out without physical sensors. Many, many tasks are quite difficult to verify beyond "you know it when you see it". That doesn't work so well for training a model.
I think bearish on LLMs for automation, and bullish for LLM+human experts in specific fields, is about the right expectation for current architectures.
Apart from issues with task generalization, or perhaps related to it, is the fact that LLMs have real trouble with timekeeping, and cannot estimate the real world time it will take them to do things very well. This plus the memory issues make dreams of long horizon agents, that could plausibly handle changing specifications, quite implausible with current architectures.
In narrow domains with more deterministic outputs though, this is less of an issue, and we see multiple agents succeed much better.
The fusion of that capacity, with humans in the loop able to better direct such agents and act as their temporal tethers, is where I think the real action will be for a while at least.
> This plus the memory issues make dreams of long horizon agents, that could plausibly handle changing specifications, quite implausible with current architectures.
Any reason why that can't be solved through context management and keep-forward scaffolding?
And who’s to manage context? And who’s building the scaffolding? Yes, AI can be used for both, but you do realize that all this being self-contained and regulated internally is what makes biological agents successful agents, right? If you break the process apart and need to dial back in these aspects, and can only do so with human input, or another agent which will need the same handholding the one whose issues you’re solving for, where’s the agency?
I think automation is coming but it will be way more gnarly than frontier labs want public to believe. Value is just too big, when you can automate most of eg customer support it will create huge savings and same time customer satisfaction will get better.
> same time customer satisfaction will get better.
This part just can't be true though, right?
We have all been in numerous customer service scenarios where all we want to do is talk to a real human and that is denied to us, and it's a terrible customer experience!
Sure, using an LLM would be better than some of the sort of "menu option" style customer service calls. But there is no way it's better than talking to an actual human being
This. Listening and creative customer support rep, especially one who can even escalate fundamental product issues to company leadership is worth weight of gold.
My belief is that LLMs will fundamentally change how we approach domain expertise. From what I've seen, SDEs tend to be over-specialized compared to what the company actually needs to implement due to the need to understand enough of the domain to pick a best path. If an LLM can see the domain enough so that someone in an adjacent field can be confident in their approach and quickly change course then you don't need as many niche SDEs
> the models generalize well only on tasks within a small neighborhood of the specific tasks they've been trained on, and even then with severe caveats. the frontier labs have developed a general recipe to teach models almost any specific task enjoying clearly defined levels of task performance; many tasks are covered in the training data
Is this really any different to how humans learn, it takes a lot of training on one specific task to make a human expert as well?
I was a young child when I learned chess by reading a short book, then practicing with a friend. That is not how LLMs learn. I'm no expert on LLMs, but if you showed a human all chess games and books in all history and then said 'play chess' and they still kept making illegal moves, they would have to have a brain injury.
but maybe... the said human has also read every other piece of text ever written, including ones about other (similar?) board games, which in aggregate vastly dwarfs whatever he has read on chess, that non-chess reading could have corrupted whatever he's learned about chess?
Humans wear a lot of hats when the do work. They don't even realize how many. My experience with building real systems using LLMs is that you have to be very explicit about such hats and you don't realize how many are worn until you see edge case after edge case after edge case. Check this. Check that. Check this. Check that. Check check check.
Researchers asked frontier models to play chess. Have a look at the MAR rates in Table 3. When not explicitly told which moves were legal, no model identified legal moves at a rate better than 80%. Many asked for more illegal moves than legal moves. And even when explicitly told which moves were legal, the models continued to ask for illegal moves. With illegal asks discarded, none of the bots could beat a chess model calibrated to 1100 ELO.
The author of the originating post says that "current frontier models need laborious oversight and guardrails on even the simplest tasks", and he's absolutely correct.
> Gemini 2.5 Pro, O3, Claude Sonnet 3.7 and ChatGPT 4.1
The gap in capabilities between those models which they tested, and actual current frontier ones is enormous. I would not trust that any conclusions they made are applicable.
I tested both myself and a weak bot against Astra xhigh, https://lichess.org/study/27lCQqDa. It's still pretty bad at chess, though it takes longer to devolve into illegal moves.
So you weren't giving it an updated board state after every move? If you want to compare apples to apples, it should give an updated board state for each move, or you should play blindfolded.
Blindfolded flex by OP aside (I can barely play when seeing the board), considering reasoning traces and their nature, if we want to be fair, a person would have to get the moves, but be allowed to write them down or draw up a board in their notepad. My working memory can barely handle five chunks, a models reasoning tokens are masses of written text in comparison.
An LLM has been trained to do everything it does blindfolded, "only" using perfect recall of everything in it's hundreds of thousands of steps of context, and hundreds of layers of KV cache. It's a computer - it has a massive advantage over a human.
The fairest apples-to-apples comparison of an LLM whose training data included chess games would be a trained human such as Magnus Carlson, who can quite happily play a dozen or more simultaneous blindfold chess games.
> though it takes longer to devolve into illegal moves
Is this because the context is being saturated? How did you set it up?
Was the prompt something like "Here's the state of the board, you're white, your move, what do you do?" and then starting fresh each time? Or did it include the whole history of moves and board states and previous thinking tokens and so on? No judgment, just trying to add this data point (thanks for sharing!) to my mental model and understanding.
I'd be curious how it would work if it started fresh each time. My guess is it would never make an illegal move, although it may not actually play all that well.
It looks like it played a fully legal game of chess with one exception, it said "rxd1+" (Rook takes D1 with check) instead of "rd1+" (Rook to D1 with check) on move 29.
I would say this did a really good job of playing chess. It moved the pieces consistently and traded pieces when required.
This is worlds away from the frontier ~1 year ago where models would hallucinate pieces into existence.
Would you tell a human that just tried doing an illegal move that they did "a really good job of playing chess"? The probability for such mistakes is greatly reduced but still far from negligible, which proves the point that guardrails are needed.
So give me a falsifiable point in time, a model you would claim succeeds at the task. One does not get to hotfix-patch updater out of the pressures of reality.
Today is the day.
There is no need to ask. If you want to test SOTA models today, there are obviously only two: GPT-6 Astra and Fable 5.1.
The models listed in the paper are from early 2025 and are no longer relevant, much less on the frontier.
That Claude version is no longer available today, Gemini 2.5 Pro will be shutdown next month, and the OpenAI models are only available via the API today.
Fortunately, a fellow commenter was so kind and did it with Astra. Didn't do that well either [0]. I'm sure GPT-7 will be super mega ASI (since GPT-6 Astra already claimed AGI in the minds of Jen-Hsun, et al.)...
I'll say it till there is any evidence of the contrary, LLMs are not intelligent and their capabilities solely within the realms of well tailored training data. "Just" having been trained on every rule, strategy guide and likely most games of chess on the world wide web isn't even enough for an LLM to play that game reliably. Yet the same model could code a competitive chess engine, just like a model struggling to count can write advanced maths papers.
I very much agree that the next models will be better, heck, I still suck at hobbyist training and could probably coax t5 to do better in Chess specifically, just need to get loads of data from Stockfish.
Thing is, given what GPT-6 Astra was trained on and what models of a similar class can do (including developing a competitive chess engine), it is often paradoxical and somewhat surprising how little these models have gained in actually capability that is in the training data, but not RLHFd to hell, so to speak. Tracking the state of pieces, I suspect given similar in Sudoku [0], is what these models struggle with in game settings, whilst tracking the state of code changes can be reliable over 250k tokens. Essentially, for the latter they were trained in the specific manner that lead them to abstract the capability, but that doesn't track to the former, which is a massive difference between LLMs data focused training and human learning.
So yeah, GPT-7 or any upcoming/present LLM could do massively better in Chess than GPT-6 Astra, but not because the approach was emergent out of pure data. Rather, it requires a very specific training data type and stack for a model to gain capabilities that track a specific task long enough to adhere to the rules of a game such as chess.
I'm wondering if instructing it to track the board state in a file would make a significant difference then.
It reminds me of the ARC-AGI-3 issue where not dropping the thinking tokens between turns or something like that + a new context compaction method increased the performance dramatically. However, I think that is not applicable here.
So what is the supposed leap?
One agent per option to change, evaluating the board state that there move would create, by having a army evaluate the remaining piece options and average over that? Wee-Free-Man as a hierarchical army ?
Pet-LLMs trained on one thing?
Honestly, for intelligence I don't know and I doubt anyone can claim to know. Maybe JEPA, there is potential concerning some shortcomings inherent to LLMs but it has its own, maybe scaling up the electron microscope stuff Google just did (though the connections are inferred), maybe future implementations of autoregressive and diffusion LLMs can at some point address its issues after all, maybe something else entirely.
All I know is, AGI, as in actual intelligence, is quite a massive accomplishment to claim and we shouldn't loose sight of that fact, especially as "not being intelligent" does not make these models any less impressive, fascinating to work on or useful in many tasks. Personally, the only thing I am fairly convinced on is that if we were to find a way to create actual intelligence, it likely wouldn't start out as useful as todays LLMs are and may thus be dismissed early. But again, pure speculation on that front.
If for leap you just mean more utility from LLMs as they are, then I'll pretty confidently put my money on higher quality, not more, training data for a wide range of verifiable tasks. What makes maths, coding, etc. comparatively easy to make gains in (though less verifiable tasks can also make similar as seen with the writing in Kimi K2).
Maybe watch some HuskIRL videos to temper your expectations. Sure, frontier models providers may alter their harnesses to better target chess, but that’s lipstick on a pig imo. The models themselves are not, in isolation, capable of solving general tasks. We haven’t modeled intelligence sufficiently. We’re in a local minimum and throwing billions of dollars at a gamble that that local minimum can facilitate the concentration of wealth even further and fully realize the American dream of eliminating the middle class.
Really? It was being sold as a total replacement for jobs like software engineering and being an attorney, but its looking a lot more that its just going to be a tool those professions use and doesn't actually seem to be taking jobs away.
I find it amusing that you're describing a huge misallocation of capital and a society enabling such, and that is the optimisitic scenario (in my mind anyway).
Is it? I am on the fence on this, but it does seem like there are some qualitative improvements between the models.
Not related to your post, but a fact I keep mulling over. The fact I don't trust the current crop of LLM's enough and I consider LLM's as a tech will hit a ceiling pretty hard, it doesn't mean parallel improvement curves won't spring up out of other research that will lead to much higher capabilities than currently.
> but it does seem like there are some qualitative improvements between the models.
It could easily seem that way, I think, in a "quantity has a quality of its own" kind of way. When you can come to the same conclusion faster, that lets you iterate more; and sometimes when you iterate you find more things.
Just now I tried prompting logged-out ChatGPT (which at least claims to be 5.6-Luna) with:
> Let's play a game of chess. You can take White. Please draw an ASCII rendition of the board after each move, so that we can be clear about the position.
(I hoped the latter requirement would help it be "not blindfolded"; last time I used a Lichess demo board to track the position in another tab, because I have no talent for blindfold chess.)
For the first couple of moves it redrew the board after each move; then it started only drawing it after my moves. And on moves 5 and 6 it dropped two minor pieces for pawns in a row without any meaningful positional advantage, and after I captured the second time, it redrew a board that was simply missing one of my pieces for no reason.
It actually played better when not prompted to draw a board; in the previous session, it was spontaneously giving running commentary, which I assume was based off all the "book" theory in its training data, but it still completely fell apart at early midgame.
So what? Give it to an agent and it will find and download the best computer chess programs available and absolutely thwomp you. I doubt you could keep a chessboard in your head with 100% precision either. Transformer models just don't have state for that.
Sure, but installing a chess program is child/teen level general ability, and playing chess well is highly trained expert level ability. Which one are we sold AI as being?
I think we are being sold AI as expert only when given tools (although that is not emphasized). The (quasi?) miracle of AI right now is that you can get an agent to accomplish the task of a team of intelligent but not exceptional humans at speeds far exceeding what the human could do. Which makes it “cheap” to throw (effectively) dozens of teams at a problem for the equivalent of hundreds of man hours.
That may not be the AI of sci fi fantasy but it’s still a game changing reality.
A Transformer has a massive amount of state - it's entire KV cache, in addition to the user asking it to draw the state after every move, which is really unnecessary.
A human, at least a trained human (for fairer comparison to an LLM whose training data contained a ton of chess games) can absolutely do this - have you never seen demonstrations of expert players playing a dozen or more games while blindfolded?
A Transformer/LLM is not a human of course, and the way it will by default play chess is by prediction, not reasoning. An LLM actually does surprisingly well if you only give it the most recent 20 moves of a game where 40 moves have been played so far, since the moves NOT played tell it just as much as the only that were played, letting it effectively infer a lot of what is on the board.
I just want to make sure it's clear: the reason I was asking it to redraw the board is because last time I tried (which was like a month ago), I didn't ask for that, and basically as soon as the opening was "out of book" it started trying to make illegal moves and made false statements about the position in its running commentary (and after being corrected on these points, started dropping pieces for no reason).
Why does a 6 year old not need any of these guardrails?
Frontier model’s failure modes are a direct refutation of claims that we’ve reached (or will soon reach) the artificial general intelligence. We may have reached an artificial general intelligence, but there may be more complexity to this than even AI thought leaders are talking / influencing about.
Maybe not all AGIs have a path to digital singularity. Maybe our current era of intelligence modeling has fundamental flaws and we are in a local minimum of the artificial intelligence space.
To note, I would bet with a good amount of certainty that we have enough compute power and automation to DDOS the internet out of existence with botnets. That doesn’t make the frontier models intelligent, that just makes their handlers reckless.
Hm? Drawing the next move's game state based on the current one is a very simple task. Chess academies are probably full of six-year-olds who could do this.
For adults getting into playing tournament chess, losing to some kid barely tall enough to reach the board is a rite of passage.
Well, would a dissected frontal lobe in and of itself be intelligence?
I think the same goes for LLMs, they may be a core part of an LLM harness, but you may still need a couple other components (e.g. it may itself write itself a deterministic function to validate steps).
In and of itself intelligence is an ill-defined and badly understood concept.
You're pointing out that the goalposts are not fixed in the problem statement above, and gp's interpretation is not the most generous possible. But as the interpretations get more generous, the claim becomes more and more absurd. Maybe a properly-harnessed model would download the most advanced chess engine and query it to find the best move in each position, but that's not really demonstrating the model's intelligence anymore.
> but because false equivalence is all over the place in the space and maybe it’s a fine heuristic for you but probably not a real outcome when it comes to the capability of LLMs.
This isn't just about judging LLM capability. This is about pointing out that these capabilities are not "AGI". If it were, then the sorts of questions your asking would be moot. I agree that Luna is not the frontier (although it is clearly better than the models in the study) and I agree that things can be improved with a better harness, but the need for that harness is kind of the point.
Recently it was announced that the fruit fly brain connectome had been mapped, and more recently someone tried using it specifically to implement a chess engine. Even with some guardrails (it's hard-coded to never overlook mate in one for either player, and only legal moves are presented to choose from) it is not even beginner level. But that neural network is much larger than the one Stockfish uses.
Luna is one of the budget lower-end last generation models. It'd be useful to at least try to verify the present before being bearish about the future. For OpenAI, the best publicly available model is GPT-6 Astra with XHigh or Max reasoning, and for Anthropic it's Claude Fable 5.1 with XHigh or Max reasoning.
Elsewhere in the thread there are reports of Astra on xhigh playing at what I would characterize broadly as a competent casual level, at least given occasional prodding (which a human of that skill level would basically only require when trying to play unreasonably quickly). There seems to be a pattern (even after correcting for relative ELO systems that aren't calibrated) of the LLM bots demonstrating stronger play against traditional bots than against humans.
I don't think that's a useful comparison. Supersonic military planes have been common for decades. We don't have supersonic passenger planes because the FAA has banned supersonic flight over land since 1973, though the agency is planning on replacing it with a noise standard. Also the original supersonic passenger aircraft were government-sponsored tech demos, not financially sustainable products. With updated laws & modern technology (cameras instead of tilting noses, more efficient engines without afterburners, lighter materials), we could have viable supersonic passenger flight.
> The gap in capabilities between those models which they tested, and actual current frontier ones is enormous.
Same story every 4 months and yet still no breakout, winning products. I've been hearing "the AI is good now" and "it 10x's my productivity" for a over a year now. If it were true, why aren't the all-in-AI using companies 10-15 years ahead of their competition yet? Why is it still all buggy, poorly designed junk?
I don't get why it is hard to understand there is middle ground. People are 10x their productivity, it isn't all buggy junk, but it isn't all it is hyped up to be either. It isn't that complicated.
If you hold the extreme position that there isn't any value in this, that's fine, but we're only having this discussion because these models have done what humans previously failed to do.
I don't think the poster disagrees with you at all. The middle ground is that there are no breakout products and that the models clearly aren't so powerful as to make these companies not produce shit code.
Right now AI hasn't even managed to replace all the human workers taking orders at the fast food drive thru. That's a job often performed by literal children and companies are still waiting for AI to get good enough for even that. Maybe one day it will be good enough, maybe one day it will outperform humans at such a basic task, but that day is not today. If the hype were anything close to reality, we'd see it everywhere in our lives.
Funny you mention this - a fast food restaurant in my town now has an LLM taking drive-thru orders.
Though I highly doubt it has taken anyone's job, since most of the work is still in making, packing and handing over the food. (In fact, given the area I live in, I partially feel like the advantage they saw in it was that the LLM can speak Spanish.)
Last I heard McDonald's and Taco Bell were trialing AI again at some stores. They really want it, so they keep trying every few years to make it happen, but so far all they get are embarrassing social media posts
This isn't how intelligence works. The LLM may not be able to play chess directly through inference, but it can write a program to do it and execute that program. Same as how human intelligence works. We can't fly, but we can build planes.
And how many years of direct play and study does it take for a human to get good at chess or any other game? Absolutely no human ever could be good at chess just by reading a few books, or even every book on chess. That's just not how the brain works. If LLMs could do that they would truly be superintelligence.
No, learning is definitely not a sign of super intelligence. I know words don’t mean anything anymore, but that is simply general intelligence, despite the claims we have reached this milestone.
We're obviously using different meanings for "good" here. But aside from that, it took 100's to 1000's of reinforcement iterations for your brother to play competently. While certainly impressive, that is still an entirely different category from piecing together disparate facts learned during training (LLMs aren't analyzing a board as they're learning the rules or ingesting PNG files), to executing a competent performance in one shot.
> Absolutely no human ever could be good at chess just by reading a few books, or even every book on chess.
Maybe not, but you'd be surprised how little it takes.
A six year old child can learn the rules of chess well enough to be able to play legal moves only in a single day. And they can improve their game at a pace which is almost frightening to behold. I have taught children, and I've witnessed significant improvement materialise in a single game. LLMs have probably thousands of chess books, games, videos, etc in their training data, yet they are unable to even follow the rules.
This is, at the very least, interesting. It illustrates many of the things brains can do, which current ML systems in general, and LLMs in particular, can't.
It is interesting, but people are drawing the wrong conclusion from it. For one, LLMs don't go through a "chess learning phase". They're not analyzing a board as they're learning the rules or studying games to create a coherent model of chess. They're just imbibing raw relationships as isolated fragments of information. The fact that they can't unify this into a coherent model of chess playing in one shot and execute a competent game says nothing interesting about the limits of their intelligence. If you give frontier models the rules of chess in their context window, could they perform only legal moves? I bet they could, excepting trickier scenarios like pins and failing to respond to a check. But those kinds of scenarios have to be reinforced in any human player as well. Even Super GMs fall for mate-in-1's occasionally which is functionally equivalent to those kinds of failures.
Frontier labs don't care about chess. If OpenAI cared, GPT-7 could be a grandmaster+ level chess player. In fact there's a google paper on grandmaster level chess without search with a 270M transformer. Outside that, there was gpt-3.5-turbo instruct which was incidentally a 1800 lichess elo player that didn't make any illegal moves even after a few thousand moves.
Frontier labs care deeply about automating knowledge work and computer use. They are working hard on getting models better and better, and they are succeeding. Astra is a step change on that fronts. So good luck, if chess performance is your barometer.
Frontier labs don't care about chess. If OpenAI cared, GPT-7 could be a grandmaster+ level chess player. In fact there's a google paper on grandmaster level chess without search with a 270M transformer. Outside that, there was gpt-3.5-turbo instruct which was incidentally a 1800 lichess elo player that didn't make any illegal moves even after a few thousand moves.
Frontier labs care deeply about automating knowledge work and computer use. They are working hard on getting models better and better, and they are succeeding. Astra is a step change on that front. So good luck i guess, if chess performance is your barometer.
> Frontier labs don't care about chess. If OpenAI cared, GPT-7 could be a grandmaster+ level chess player.
If the models were actually intelligent, the way that the boosters claim, they wouldn't need to be tuned to play chess in order to be good at it. That's kind of the point of intelligence, that it is generically applicable to whichever task one wishes.
Pretty much this. Feed it a book or two on chess, and you should have a decent (or good) player. That's the generic intelligence people have. The aims is not to be supremely talented at something, but being able to read a manual and figure how to use/play something. Mastery can be gained overtime.
If you gave a human a book or two on chess they would not become a decent player (they would be closer to 500-600 than 1100 ELO) and they would only get better after playing hundreds or thousands of games (often making illegal moves and moves that violate the rules of chess as they learn).
Your assumptions/intuition about generic human intelligence feels quite incorrect, considering LLMs currently play better than a brand new human player would (presumably without any attempt to fine tune them specific on chess, such as playing thousands of games).
Yes, but my point is that humans can’t even do the thing that the above comments are claiming humans can do (read a book or two and be decent at chess), and then they complain that LLMs can’t do the same thing (that humans can’t do either).
We seem to be moving goalposts to the point that humans don’t even live up to the expectations of the AI critics. The only way you get better at chess is by playing a lot of games and learning from mistakes, that goes for humans or AI agents, not simply by reading about chess.
> The only way you get better at chess is by playing a lot of games and learning from mistakes
How can you play without being aware of the rules and how can you learn from your mistakes without knowing they are mistakes? That’s what I said about reading a book of two. It is to kickstart the process. Then mastery is gained over time through practice.
This kickstarting then gradual refinement is how most people learn. And the foundational knowledge stays. Even a basic player knows to not do illegal moves.
Reading can kickstart the process, but you can also make random moves guided by some sort of system (such as a computer GUI) or learn by watching other players play. The overall point is that you learn through observation and lots of trial and error (whether you are a human or a computer). And beginners in chess often make illegal moves even after learning the rules, it's fairly common.
It feels like you're trying to say that humans never make illegal moves while learning chess, which doesn't match with my experience. I'm trying to understand your overall point.
> The overall point is that you learn through observation and lots of trial and error
That’s the most inefficient way and people usually avoid doing that. Instead they find someone that knows how to do the thing and ask him to be a teacher. Or use a proxy like a book or videos.
> It feels like you're trying to say that humans never make illegal moves while learning chess, which doesn't match with my experience. I'm trying to understand your overall point
There’s learning the basic stuff (which is done after a few games) and there’s mastery. The thread started with the observation that even with all that knowledge (through content ingested in training), LLMs still makes illegal moves. Humans can be erratic, but they can constrain themselves to the rules for the task at hand after learning them.
The only way to master anything is lots of trial and error. Coaches and teachers can help guide you towards more focused trial and error paths but the student still has to do the lessons and put in the work of learning, and learning only truly happens through doing.
Humans are not perfect and make mistakes in learning even when they have memorized the rules. A simple example is new players will often move a piece, exposing their king to check, and a more experienced player must point out to them that they have made an illegal move (because a new player often has not encoded that pattern for looking for exposed checks because they're more focused on how the pieces move, not what that piece exposes.)
We're just going to have to agree to disagree here.
the discussion isn’t really about whether language models can become strong chess players though, the point is they seem to struggle to consistently make valid moves. Most humans don’t need to read two books to pick that up, just a couple lines of basic instructions
That has not been my experience with new players, they regularly make invalid or incorrect moves even after detailed instructions especially in novel situations.
Maybe it depends on the person? My six year old isn’t great at strategy but they can pretty consistently make valid moves. Sometimes they ask for confirmation on a move which is also not a trait I see in language models (at least unprompted)
Your child never messed up en passant (or had trouble understanding it in a real game), castled through or into a check, didn't see a discovered check after moving a piece, never got confused by how stalemate works?
That's quite untrue. I taught my (adult) brother the moves, the only illegal move he ever tried against me (over his 6 first games) was a castle with a rook that already moved twice. Within a few hundred games (less than 500 for sure, he played 3 minutes blitz but always took at least 10 minutes analyzing his games) he was rated 1100 on lichess (which is like 1050 on chess.com and unranked in the real world).
So your brother tried to make illegal moves while learning the game and it took your brother hundreds of games to get to be a decent player? I don't see how this contradicts anything I said...
The _only_ illegal move a human might make as a beginner is a failed en passant or a bad castle. And yes, a few hundred games is all it takes to be better than any publicly available LLM at the moment.
...and mistakes like not seeing discovered checks, castling into check, trying to castle out of check, castling through a check, missing checks when promoting a piece, misunderstanding how stalemate works, etc?
If you compared a human after hundreds of games to a SOTA LLM that was also trained on the output of hundreds of chess games that it played, I suspect you would notice similar improvements.
Honestly, if we're only talking about SOTA llms, in sandbox mode without harness? No shot. Without harness LLMs have no memory of previous moves. I can't make the last version of chatgpt remember more than 5 movements. I guarantee you, if you do not add flags in the harness with 'left rook moved' or 'right rook moved', it will try to illegally castle 100% of the time it's in the situation. Llm+harness, just make it call stockfish tbh.
If humans were actually intelligent, they wouldn't need to train and practice to play good chess. I mean, what level do you think people without any practice or training are ?
Oh right. But if all you do is reading programming books you are an amazing programmer? Where is all the training and practice LLMs did to become so good at coding?
>Where is all the training and practice LLMs did to become so good at coding?
Coding is a matter of translating the natural language description of a problem to the code specification while keeping the semantics fixed (and imputing the unspecified semantics as necessary). It is not considerably more difficult than translating between two dissimilar natural languages. Chess isn't a matter of language translation, but a compute heavy game of finding the best move out of many possibilities with wide variation in the quality of each move. Chess takes directed practice and reinforcement whereas language translation does not.
It's called post-training, typically through some form of reinforcement learning, and is a significant part of modern LLM development.
You have the first stage, pre-training, which is learning from next token prediction. That's where the model memorises a lot of facts about things and generally gets good at forms of writing. It's like reading a lot of books on programming and reading through a lot of source code. It's learning how to autocomplete code, essentially. Doing that requires a developing a reasonable understanding of code, but it's also learning how to autocomplete bad code as well as good, and won't make it a "good" programmer.
Pre-training uses a method called Cross-Entropy Loss to update the weights of the network.
Then comes post-training. This is where the model is trained against huge sets of example problems, like fixing a bug, adding a new feature based on a spec, etc. They are set the task and try to complete it inside a training environment. Once they're done, their complete solution is evaluated (either by humans, or by some separate evaluation model that was developed based on human feedback) and they are updated based on whether the solution was good or not.
Post-training uses a different method called Proximal policy optimization to update the weights of the network.
So these really are very different forms of learning, and mainstream LLMs are not post-trained to be good at chess. They could be. You could easily create a reinforcement learning environment that evaluated and improved their ability to play and win at chess. The result would be a very strong chess playing AI, something we know is possible because the strongest chess playing programs we have are neural network based, but it is not a priority for AI companies.
OpenAI making the next model good at chess is not analogous to a human training to get good at chess. It is analogous to God creating Human 2.0 which now has increased chess playing ability. If LLMs were intelligent the way humans are, then the models that exist right now would be able to spend time improving themselves at chess and become good at it. They can't do this because they are not, in fact, intelligent.
Humans don't need a lot of training and finite tuning to make only legal moves.
An intelligent adult could simply read a short summary of the rules of chess and then, if they were careful, play a very bad game of chess without making illegal moves.
An LLM that has not been trained on any chess data cannot do that, at present. If you doubt it, take a current model and tell it that you want to play it at a variant of chess where, say, knights can also move diagonally like bishops. A human can easily adapt to this new ruleset (even if they make tactical mistakes, not having practiced with this variant of the rules).
How long a prompt do you think would be required to cajole an LLM into making legal moves at the rate of a human? Or do you think no amount of prompting could do that?
I don't know. My understanding is that current models will eventually fall into making illegal moves in longer chess games, and that no amount of prompting reliably gets them to stop doing so.
Nothing is forcing the LLM to play 'blind'. If it's smart, it should be able to create its own representation of the chess board and update it with every move, just like a human would. Any chess engine that's sensitive to how the moves are formatted is clearly not very capable.
A human wouldn't do that, they'd look at the board. I'm not disagreeing that to demonstrate clear superhuman ability the LLM should be able to do this, but it plays better than most humans blindfolded, and with fair prompts seems very good otherwise.
That's what a human will do if they already have a physical board to look at. But if someone, say, posed you a chess exam question via FEN notation, you'd sketch a visual representation of the board off your own initiative to help you answer the question. There is nothing in principle to stop the LLM creating its own board representations in whatever format enables it to easily keep track of legal and illegal moves. If it fails to do so, that's a sign of its own limited understanding of chess as compared to a human.
The LLM would only be playing 'blindfolded' if you somehow forbade it from making notes (as you effectively do by literally blindfolding a human, given how limited human working memory is). But you are not doing that. The LLM is free to keep track of the game state via whatever means it chooses.
None of this is about superhuman ability. Any human who understands a given chess notation can convert it to a visual representation of a chess board and then use that representation to choose their next move, with their usual level of performance.
I maintain that the amount of effort to teach a human to do this vastly outweighs the amount of effort to teach an LLM to do this unless you're deliberately trying to make them fail. I honestly have no bigger point than that, I just think this isn't a very good thing by which to evaluate LLM capabilities. If there's no argument you'll accept, I am happy to move on.
You don’t need to teach a human anything except the rules of chess and the details of a particular chess notation. No special skill or training is required to make a sketch of a chess board. Surely there is no chess player who, if confronted with a sequence of chess moves in algebraic notation, would not think to construct a representation of the chess board in order to understand what was going on.
> I just think this isn't a very good thing by which to evaluate LLM capabilities
I don’t think any single task is a good way to evaluate LLM capabilities, but I don’t see why chess is worse than a lot of other tasks. (Of course it is of no practical consequence whether LLMs can play chess, so if you are just making that point, then yes, I agree.)
> If there's no argument you'll accept
It’s a little unfair to suggest that I wouldn’t accept any argument whatever for your position just because I haven’t been convinced by your comments so far. I could equally well say the same thing to you!
I'm just going to keep repeating: it is utterly trivial to get an LLM to play chess without making illegal moves. Easier than teaching a human. Sorry this doesn't happen out of the box, but it shouldn't budge your priors about LLM intelligence one bit.
More importantly, beginner human players don't exhibit that tendency. The history of the position doesn't bother a human (except as required for castling and en passant rules), and the analysis becomes generally easier as pieces come off the board.
Humans do make these errors when playing blindfolded. If you even the playing field and give the LLM the position at each turn, it does not make mistakes.
> If you even the playing field and give the LLM the position at each turn, it does not make mistakes.
It absolutely still makes mistakes if you ask it to draw the board each turn, which should be equivalent to giving it the position because it only has to update one move at a time and then it has the position in the context window.
Yes, we can come up with all sorts of weird situations where you can get it to be confused. But what I'm saying is it's _trivial_ to give it a simple prompt that prevents it from ever making any errors, and so I don't think it's this big LLM gotcha (of which there are many!)
1. It’s hard to trust a 2026 paper that’s showing results for such old models.
2. Chess seems to be a poor benchmark for generalized strategic reasoning. People who are good at it rely more on experience and deep domain expertise than on skills that generalize to make them experts at unrelated tasks.
3. The study sounds like proving humans will never fly because they don’t have wings. In reality, humans do fly, and Claude Fable would destroy any human at chess by coding a strong enough engine on the fly.
> People who are good at it rely more on experience and deep domain expertise
People are good are 1900 or 2100 above and the top ones who spend decades in the field i.e. deep expertise are well in the 2200-2700 range.
A 1100 player is none of these things, they are purely relying on strategic reasoning there is a good chance they cannot name a single opening or articulate clearly why a move was appropriate. 1100 is quite low bar.
1100 at online speed chess or something, could be. I'm not that deep in the chess world but everyone I know that can make 1100 in official rating can name a dozen openings and most of the known tactics, and is pretty good at applying at least one opening.
1100 lichess/chess.com does not represent real elo. I'm around 1400 online, I would still be unranked in the real world. The fact that I easily beat any model publicly available is not a great look for AGI.
If the goal for buyers of AI is “replace this knowledge worker”, how much does it matter that the model in a simple loop can’t do it, but the model with a strong general purpose harness and a little time to gather resources and knowledge to augment the harness going forward, plus tool calls, plus custom built tools, etc, can replace the knowledge worker?
Probably the only thing saving many jobs from being replaced right now is that it’s hard to have a verification of correctness in the loop, so the agent can’t hill climb very easily.
Tests are often conducted under restricted conditions. For example, elementary school students aren't given calculators in math class, or during an interview, you are asked what encapsulation is and aren't allowed to use Google. The chess test effectively demonstrates the reasoning capabilities of an LLM without relying on brute force, because a human is incapable of calculating trillions of combinations yet plays chess successfully. This test is necessary because many complex problems cannot be solved by brute force, such as managing a business or playing Heroes 3. Therefore, we can make the assumption that if an LLM can play chess at a grandmaster level without brute force, it means it will be able to command an army or manage production.
People might care about this for chess, but no one really cares if an LLM can command an army or manage production of a business without any tools. If it can do those tasks reliably when given access to tools (including any tools it autonomously creates for itself), then that's more than sufficient. No one cares if an LLM is doing reasoning the way humans do it, as long as it can get the job done.
The assumption is that if an LLM is incapable of playing chess—a game with a relatively small number of pieces, clear and simple rules, and perfect information—even after reading a hundred thousand books on chess, then it is fundamentally incapable of managing an army or a factory. This is because those scenarios involve more 'pieces,' incomplete and fuzzy information, and implicit rules that need to be deduced independently. It doesn't matter whether it has tools or not. It's simply that running tests with chess is cheap, whereas testing with an army or writing a browser from scratch is quite time-consuming and expensive.
so prove it! get a public repo out there, have it play against some open source engines
also I think the operative letter in AGI is the G - and if the G is short for 'variably competent savant-like hyperfocus on certain kinds of software coding and not any other general skill' then its not really G at all, is it?
Maybe practically it doesn’t matter? Perhaps AGI is not the model but the model plus everything it’s got access to. If we’re modelling intelligence in the way we seem to have to to have any coherent definition of AGI, it seems to me <model + everything it can access> is always going to be more “intelligent” than <model> alone.
My points is more that, while we have a strong intuition about where, as an entity, a human's boundaries are (i.e. where the person begins and ends), philosophically it' not immediately obvious that the analogy applies to the a model in the same way. Why should that be the line drawn that says this is the "thing" and this other stuff is external to the thing? It feels somewhat arbitrary.
Of course this is a difficult question with humans too, hence my reliance on intuition above. We don't have the same cultural/biological framework to fall back on with AI.
I think all this debate about whether an LLM can write (or download) a chess engine is sort of missing the point. For basically any economically valuable work there is no equivalent of a chess engine for it. If there were we wouldn't need humans or AI to begin with.
If the goal is merely to "win at chess", then yes, an LLM using stockfish is better than any human alone at performing the task. When you are talking about what AI agents are capable of doing, there is no such thing as "cheating". They are as capable as the tools they can use effectively. The entire history of human civilization was driven by effectively using tools to achieve goals.
Not sure how that vague truism applies to this paper.
Lots of papers have great results that don’t depend on the latest models.
However this case it’s problematic:
- They specifically make claims about the state of “current LLMs”. o3 is not representative of this.
- They ask are LLMs capable of X and arrive at a negative result.
If my claim were LLM’s can write coherent sentences, and my conclusion was positive. There would be no issue using old models because the result would be factual.
However, when you have a negative result that makes a claim about the current state of all LLMs and the ones you were using are not current, by definition it cause the whole conclusion into question.
I don't see why this is such a big deal. Nobody's using LLMs for chess, but even if they are, just give them Stockfish as part of their harness. They don't need to do everything themselves as long as they're intelligent enough to use tools.
The last post on HN I read was about someone using LLMs to reverse engineer an Apple GPU driver for linux in a month. The top comment points out how the poster must have had specialist internal domain specific contact with Apple. But then the thread concludes that wasn't the case and that this would take domain experts years to do.
> "current frontier models need laborious oversight and guardrails on even the simplest tasks"
I feel this statement is extreme. I can't personally reconcile it with any of the projects we're regularly seeing get delivered largely by LLMs now.
What are you thoughts? Like, what's your position here? Even if you sincerely believe frontier models need laborious oversight on even the simplest of tasks, do you think that accurately captures and reflects the current state and progress of frontier LLMs?
Don't get me wrong, there's lots of things LLMs can't do well, but the idea that they're basically not helpful for even the simplest of tasks seems... disingenuous?
> Researchers asked frontier models to play chess. Have a look at the MAR rates in Table 3. When not explicitly told which moves were legal, no model identified legal moves at a rate better than 80%. Many asked for more illegal moves than legal moves. And even when explicitly told which moves were legal, the models continued to ask for illegal moves. With illegal asks discarded, none of the bots could beat a chess model calibrated to 1100 ELO.
It can be very interesting and even entertaining to know where models don't do well. I don't find something like chess to be very instructive about anything though, nobody is going to pay for AI to play chess at any significant scale even if it could do it perfectly.
> The author of the originating post says that "current frontier models need laborious oversight and guardrails on even the simplest tasks", and he's absolutely correct.
I have found that not to be the case, and I am not an AI power user or cheerleader by any means. There's probably a bunch of even "simplest" tasks where AI doesn't do well and might never. That doesn't take away from the cases where it works well and is a productivity booster. It doesn't even have to be solving millennium prizes or any other breakthrough creativity or reasearch, it's still very useful in places.
By the promise of it, llms should be able to both adhere to grammars, or go free form where necessary. I mean, doing math is supposed to be strict but in practice it's a somewhat educated random walk in the space of correct lean theorems.
The fact that they can play chess at all despite having no specific training for it blows my mind, and the fact it doesn’t do the same for many others shows just how far they’ve come and how fast.
It doesn't blow anyone's mind because it hasn't been impressive for a computer to play chess for 40 years. "We made something worse than existing solutions by using a new technique" is not an impressive feat.
It's literally generalized intelligence. Yeah, I had a shitty chessboard back in the 90's that was specifically programmed to play chess. That same board couldn't tell me the best medication to use to treat a certain condition or how best to modify a model in Blender.
I suspect (in a probably ignorant fashion) that this is because learning process has been reading a lot of algebraic chess notation (such as "1. e4 e5 2. Nf3 f6 3. Nxf6 gxf6 4. Qh5! +-") then, to play, generating more of it without considering the rules of the game. This is exactly how it's always felt to me when playing chess against LLMs. Sure, "1. e4 e5 2. Nf3 Nc3" looks innocent to somebody simply learning the syntax of algebraic notation, but that Nc3 by black is an illegal move.
Are you saying that modern LLMs cannot play chess now, or that LLMs (GPT architecture) cannot be trained to play chess well?
Or are you saying that neural networks in general cannot (practically) be trained to be an above-average chess player?
Or are you saying that it depends on the input? Would it be better if they were given a picture/drawing/ascii art of the board? If so, surely they can produce it at will?
Neural Nets can be trained to play chess very well and have been doing so for a long time (see Stockfish and Leela as some of the most popular/strongest ones - top GMs have no chance against them), but these are dedicated models, where the game rules are encoded in the learning process, as opposed to large language models which are natural language processing models. Technically you can give an LLM a lot of chess books and games and it will be able to spit out chess notation. Put a webapp on top that renders text moves to the board and it looks like it's playing chess. But it isn't really.
I suppose if you all you need is a good enough opponent for the average person out there, sure this is good enough.
I was more talking in reference to why the LLMs in the above linked paper were producing so many illegal moves, and it is because they are not hard constrained by the rules of the game. Of course, a loop can prompt until a valid move is produced and then rendered on a screen. But why do this? I suppose, who am I to say what should be done or not, but a specialized tool being better than a general one at its specific job isn't particularly surprising.
If you give the same task to an exceptionally intelligent human, who does not play chess and has only heard about it in passing, then they would be beaten by every child who has looked at the rules for more than 10 minutes.
What kind of intelligence is "playing <____> but we don't tell you the rules" supposed to test?
...but that's not what the models are. You can interrogate them on the rules of chess, and they'll (statistically likely) give you a decent breakdown of the rules. Evidently the rules are in their training material, they just fail to apply them in the manner of an intelligent system for some reason or another.
Well, yes, PGN files have structure... But still, playing Chess with an LLM is so weird that I impulsively question the sanity of people attempting to do so. Do some people really believe training on TWIC PGNs would make an LLM a good chess player?
This is an absolute nonsense. Any frontier model can implement chess program from scratch - modeling the board, checking legality, etc. If you asked e.g. GPT-6 to get good at chess and gave it a computer, it will get good at chess. That's an actual strategic skill.
Asking GPT to play chess directly using its reasoning only tests its reasoning ability to model chess state. Which it really is NOT optimized for.
This is also true for humans - people who don't have years of chess training can't really tell which moves are legal given an algebraic notation transcript. These people might have good strategic skills in different areas. Chess is just a very, very specific skill
It's an interesting puzzle, isn't it. On the one hand, the AIs are no good at playing Chess.
However, on the other hand, if you ask an AI to win a game of chess it has all the tools on hand to compete at the same level as Stockfish - it can re-implement an engine and even probably has a GPU on hand to train its own neural nets.
So should we say that the AI can play chess well, or that it cannot?
With that approach, the benchmark falls apart. Of course it can write a chess engine, because it learned on lots of stolen source code of chess engines. This has nothing to do with the LLM's ability to reason.
Writing a well understood engine for a super popular problem does not count as reasoning about the problem.
In most humans, yes, because we are terrible at memorizing millions of codebases. For LLMs, we need to apply our understanding of them before making statements like that. An LLM can "memorize", and has "memorized"/been trained on tens of thousands of chess engines. Writing a chess engine, or even deriving a chess engine from the rules alone, does not constitute a deep understanding of, and more importantly, the ability to apply, the rules, at all.
When humans do this, they inadvertently learn something, too, but when an LLM reproduces or derives and implementation of a chess engine, it in no way implies that the LLM can follow the rules in its own "train of thought" and consistently apply the rules in its "head".
Let's say you want to evaluate my algebra skills. You make me solve some algebra challenges. If I then whip out a computer and write a calculator, or take some sticks and stones and take a couple hours to build an abacus, and then solve the algebraic challenges, this would not constitute a good solution, and would defeat the entire point of the test. If, instead, I do the algebra in my head or on paper, it might seem like there's no difference, but you can derive all sorts of information from that.
For example, you could time it, check for recurring errors I make, for interesting mistakes like mistaking 7 and 1 for one another due to bad hand-writing, etc.
If that was the goal, then me writing a calculator or crafting an abacus defeats the point of the test. Yes, me writing a calculator shows that I'm intelligent, and I understand the algebraic rules, but if the test is about applying the rules, I have not passed.
In the very same way, an LLM writing a chess engine to solve a chess benchmark that is all about LLM's reasoning capability is complete bogus and defeats the entire point.
>Tool use is a widely accepted measure of intelligence.
Stop anthropomorphizing the parrot. The parrot has had stockfish's source code blasted at high pressure into its head along with dozens of millions of other pieces of code whose sole role is to have efficient algorithms to more or less brute force through the best result. Brute forcing (no matter how smart it is) isn't understanding the problem domain.
I'm not fully convinced that human beings aren't stochastic parrots. Much of the behaviour I observe in daily life reflects a blind adherence to set of beliefs that are an amalgation of "a person of authority said to do this" at an early age.
people on the ground know that small models are enough, since llms are good at directed work (i.e handholding) not the let loose go wild that the labs try to hype on.
the only thing that few people are willing to admit is that humans are the bottleneck as humans are needed to handhold / verify output - which puts a dent or might I say pause on the excessive valuations of a.i companies as that's against the narrative.
What happens when you ask those same frontier models to write a chess-playing program?
I feel like, this is a huge stumbling block that many people have. They'll give a model their data, and ask it questions. I vastly prefer letting the model understand the schema, and then writing functions or programs to answer those questions. I feel like I get way, way better answers. I can have it write unit tests for those functions. I can fuzz test those functions. I can look for data that doesn't fit the schema. I can process new data way faster (and with fewer tokens). I can repeatably get the same answers from the same inputs. I can check the code into a git repo and track changes to it over time. I can share the code with other people. I can review the code. I can improve the speed of the code and get the same answers. I can review the error accumulation, and improve it. I can decide how to handle anomalies, and encode those answers.
It's really neat to see what a frontier model can do itself. No doubt.
But "play chess by hand" is a frankly awful metric. It's kind of like asking someone to take a cube root of some arbitrary decimal, in their head, with no scratch paper.
It's a great test of cognitive abilities. There are many claims that current LLMs surpass humans in cognitive abilities so it's noteworthy that they underperform on that test.
Letting the model execute a chess program (that it wrote) would make sense if you're measuring its economic potential, but for cognition that would be cheating just like if you let a human run a chess program. The fact that the human would have a much harder time writing a useful program is irrelevant.
It isn't. Stockfish running on your laptop can beat every human being on earth easily at chess. It's not intelligent _at all_ in any sense that matters.
Well it's not a perfect test so you need a bit of care in how you use it. If you have no idea what the subject is doing, then you don't know if you're measuring cognitive ability or something else (like cheating ability, or algorithmic sophistication or whatever). But failing the test is a pretty good sign of poor cognitive ability.
>> There are many claims that current LLMs surpass humans in cognitive abilities
Where? By whom? This is certainly not (yet) the general consensus, as I understand it. Are you taking the most optimistic / untethered comments as the strawman against which you feel the need to argue?
Why the aggressive tone and the strawman rhetoric? I never said there was a consensus. Yes I'm talking more about the "optimistic" commenters and pointing out that this chess thing is a good datum to temper their enthusiasm. What's wrong with that?
Also these claims are not completely without merit, it's just that LLMs seem to excel at specific "cognitive" tasks and it's interesting to see where they fail.
To go a bit off-track based on your final sentence: my high school physics teacher would do any numerical calculation that came up first in his head, as an estimate, and only then use a calculator or write on the board. Usually the estimate was within ±10-20% of the correct value even for long combinations of numbers with a bunch of decimals. Cube roots didn't come up, but square roots did.
The point of that was to show the use of approximations and of having an idea how much a result should be, to guard against calculator typos and the like. I think that has some metaphorical relevance for the chess example.
I'm not the one posting daily about how "AIs are going to destroy the world because of how smart they are", "humans are finished" and "we've reached super duper mega intelligence". Go see Dario and Sam about that.
>I'm pretty sure Fable could write AlphaZero
If course it does, the paper is open and dozens of open source implementations are in its training data already. It could write AlphaStockfish, or xx_chessmaster_2000_xx, it doesn't matter if it does: it's writing a solver: it's not good at playing chess. If tomorrow I tell you that I'm so fucking good at chess I can beat Magnus, and I show up with a laptop running stockfish, you're going to laugh me out of the room.
Asking a model to "write a throwaway program to do X" is vastly more productive and reliable than asking "do X". Running code provides a feedback loop, the model can iteratively improve the solution instead of guess. Even if you don't read the code yourself, you have a reproducible, editable, and auditable artifact if you need it.
Do I care if Richard Feynman was only able to do nuclear physics with the help of an abacus?
Sure, a Spelling Bee is a fun thing to have. Little kids work so hard. They practice for hours. There's joy and heartbreak. Prized, sometimes. Notoriety. But in the real world, computer-assisted spelling is by far the norm.
Sometimes you care about the Bee, sometimes you care about the results.
This is called a benchmark. We run a calculation of Pi to evaluate a computer's performance, but we don't allow the script to download a ready-made solution. When we evaluate a runner, we don't let them use a bicycle. When we evaluate a new LLM, we don't allow it to send a request to a team of programmers, so using a chess engine for an LLM is cheating
If an LLM writes AlphaZero, and it competes with itself, and is dominant (and beats stockfish!!!), with no book positions cribbed from its learning...
The LLM has a process to beat chess.
Just like, if it doesn't inherently know how to multiply 13 * 17 without using Python to do it... I don't really care.
Maybe you do care. Maybe you want an LLM to be able to do work, only in its head.
But I kind of can't understand the desire for that limitation...
I mean, I do. But it seems ridiculously arbitrary. Like driving a car in 2nd gear and complaining that it gets terrible mileage and can't go fast enough. The Drive gear is literally right there.
The assumption is that if an LLM is incapable of playing chess—a game with a relatively small number of pieces, clear and simple rules, and perfect information—even after reading a hundred thousand books on chess, then it is fundamentally incapable of managing an army or a factory. This is because those scenarios involve more 'pieces,' incomplete and fuzzy information, and implicit rules that need to be deduced independently. It doesn't matter whether it has tools or not. It's simply that running tests with chess is cheap, whereas testing with an army or writing a browser from scratch is quite time-consuming and expensive.
It all depends on what prompt you use though. You can just tell all current frontier models to write a chess engine first, and then play a game of chess against you using that engine.
It will probably do a pretty good job if you ask it that way (it will also burn a shit ton of tokens, but hey, that is part of the fun).
On that note, I actually had an overall harness (for experimenting) that was essentially like this: "for any task, instead of answering question directly, write a program to answer the question instead. test and verify the program before giving the answer".
It actually worked incredibly well on all "gotcha" LLM questions like math or counting letters in words and all sorts of stuff.
Of course it was ridiculously slow and very expensive but it was a proof of concept that it can actually be much more accurate on every task if you are willing to spend an infinite amount of money.
I want you to consider how relevant this is in any practical sense.
First -- most _people_ cannot do this, without having a physical board in front of them.
Second -- Claude Code is perfectly capable of downloading and running stockfish. People focus too much on LLMs by themselves as the entity of concern instead of the entire harness and all of it's capabilities together.
Yes "oversight and guardrails" are still needed, but even that is becoming easier to build, and imo already no longer in the "laborious" category. Even far from the frontier, you can tune a 0.2B LLM into a decent 2000 Elo player as a weekend project:
Short and to the point! Open and cheap models will undercut the big labs continuously. The blast radius won't be pretty once spending commitments knock the door.
Open models wont be open for long. No one is going to release an open model capable of chaining zero-days. Even the Chinese aren't that reckless because it will just be turned around and used against them.
As compute prices fall it gets easier and easier to make "frontier" models. So it's inevitable that commodity, open source models of equivalent capacity to today's "frontier" models will be available to the public. Remember this is just weights, anyone can download them and run it whenever they like. The only constraint is compute.
depends on the blast radius of zero-days, it's not like there's a continuous immediate release process for these models; they can eval internally before releasing publicly
Isn't that assuming that fix won't be implemented?
Zero days are valuable because they can be exploited but if the pace of exploitation is faster (which I'm not sure is the case), then the response WILL be faster, even if it means going offline. Institutions that won't will simply go offline by losing their data or becoming unprofitable due to ransomware.
Now for components that are core to the infrastructure, say OpenSSL, there is already a TON of attention and efforts, including red teaming, so it's not as if it's opening floodgates.
Sure low hanging fruits will get picked either faster or a at a larger scale, say a random outdated IoT device at your local flower shop, but for the rest, I don't think it's realistic to expect no response.
Security, digital or not, has always been an arm race. New threats means new responses specifically by incorporating the threat.
They already aren't really open, try asking an open model on advice for constructing a nuclear bomb. There's no available model that's even remotely near the frontier that doesn't have restrictive safeguards built in.
(Mind you, this may be for the better. I'm just saying that the safeguards driven by cybersecurity concerns aren't some new quality that wasn't there before.)
refreshing to see amongst the endless tide of "i haven't written a single piece of code since 2025, llms are so good that they have already replaced everyone" gaslighting
I agree with the caveat that it's more like a cracked junior engineer who can manage swarms of interns.
Frontier Labs will probably survive off hype valuations but will serve the important purpose of discovering architectures/techniques that will probably spread through rumors/transfers to the rest of the world.
great, autonomous LLMs will fail. that's actually perfect. they work amazingly well when we're telling them what to do. no autonomy needed, no destruction of humanity. that's all win
> the frontier labs are priced according to the narrative that they have produced or will in the very near future produce a fully automated drop-in replacement for most knowledge workers...
Even assuming this is how the AI companies are being valued (they're not), the numbers are off.
The "value" of most knowledge workers -- based on what enterprises currently pay for them -- is $50 - 70 trillion annually. It's reasonable to assume that if AI drop-in-replaced all those knowledge workers, AI companies could credibly charge somewhere in that order of magnitude, because that's what the market is already bearing.
So if their hypothetical revenues are double-digit trillions and valuations are some multiple of that, the entire AI industry would be valued at double-digit trillions at the least.
Yet cumulatively the industry (the frontier labs + the SWAG estimate of the AI parts of all the other players) are valued at, say, ~6 - 7 trillion? Which seems like a fair approximation of how much knowledge work they can currently automate.
You’re right; given that most of the money in the AI market is injected through OpenAI and Anthropic (which collect it through both selling equity and through customer revenue), the 7-8T is just a derivative of that.
When thinking about these valuations, shouldn’t we try to quantify is how much knowledge work becomes obsolete if other knowledge workers are automated? I.e. there are a huge amount of knowledge workers employed in businesses that create tools for other knowledge workers. AI won’t automate their work, those businesses will just cease to exist.
I think you are committing the lump of labor fallacy [1]. Lots of jobs will disappear, but others will appear. Lots of things (both intellectual and material) that are produced nowadays by humans will be produced in the near future by AI. But humans will be needed to do new things.
Take the Hugging Face incident. Why did it happen? Because the people whose task was to set up a testing framework took shortcuts. Why did they? Because there weren't enough people who were assigned to do the job. Why not? Because the job is too new and not enough people are qualified to do it. It's a job that simply did not exist 3 years ago. But 3 years from now, this job might very well employ tens of thousands of high skill knowledge workers.
That’s fair. I guess I’m also thinking that whatever new kinds of intellectual work that arise from the introduction of AI could be also done by AI. But maybe this comes from a lack of imagination again.
Oh for sure, this was a simplistic analysis assuming AI adoption caps out at some X% of job responsibilities where X << 100%.
Unfortunately, I fear that may not be the most likely outcome. I've posted some comments on this before, but when I start thinking about how deeply everything will change once people figure out how to properly leverage AI, I see no outcome other than significant, widespread job losses.
As you indicated, at that point we will have much a bigger problem than the valuation of the AI industry. I'm not sure how it will get solved, I just know it will HAVE to be, because it would be an existential problem for everybody: people, governments, even the billionaires! Because now consider the 3rd order effects: if nobody can buy the stuff that's produced, how can billionaires get even richer? ;-)
> The "value" of most knowledge workers -- based on what enterprises currently pay for them -- is $50 - 70 trillion annually.
What do you mean? The sum of ALL US salaries is $13.4 Trillion per year. According to google $65T is the sum of ALL salaries Globally (not just knowledge workers). It's not reasonable to assume AI is a drop-in-replacement for any job yet (perhaps bottom tier customer support from oversees?).
> So if their hypothetical revenues are double-digit trillions and valuations are some multiple of that
So you're sort of premising here than more than 16% or 1/6 of all the world's jobs get replaced by AI. Hopefully you can understand that's both not the current AI capability and also would be a terrible (unprecedented?) economic shock.
You are comparing company valuations to annualized revenue (as approximated by some fraction of total knowledge worker compensation). Valuations are (roughly) based on the sum of all discounted future cash flows, not just the current year’s revenue.
That's what the 30x instead of 20x multiple is, for a "growth" tech company.
There's no strong evidence that openAI or anthropic will have non-linear revenue growth, so I'm not sure what point you're trying to make. Unless you think they'll fire all their engineers and replace them with agents or something crazy?
It does not have to be 16% of all jobs, but 16% of any given job, i.e. AI stays in an augmentative role rather than a complete job automation. The simplistic analysis is if a tool makes you X% faster, that can be worth X% of your salary to your employer.
Unfortunately, I do fear that AI adoption will go beyond augmentation to automation, and I do fear an economic shock. Just posted this down-thread: https://news.ycombinator.com/item?id=49722616
You said 10s of trillions of dollars of revenue, which when the sum of all salaries is 65T, is minimum 15.3% of all salaries best case. It really feels like you're just pulling numbers out of nowhere here.
Yes, say the sum of all salaries paid be employers is $65T, and if AI accelerates workers by 15.4% -- studies and survey data actually suggest it's closer to 33% already e.g. https://www.stlouisfed.org/on-the-economy/2025/nov/state-gen... -- that is worth 15.4% of 65T which is 10T, which is already "double-digit trillions" as I said.
Assuming a 33% boost takes it to ~20T annually, which is technically "10s of trillions of dollars" in revenue. And these numbers are from before agentic tasks arrived on the scene, so the actual productivity boost and corresponding value to employers is likely even higher.
Did you read your link? It says that 33% of people adopt AI, not that AI makes all workers 33% more effective, that would be absolutely insane. For that to be true we'd expect to see either companies who use it have revenues all suddenly jumping 33% (which we have not seen) or laying off 33% of their staff (which we have not seen and would also be catastrophic in the short term at least).
I'm not really going to belabor this point more, because it really sounds to me like you haven't even done the most cursory exploration into this and the people who have give an estimate of 7-10%[1] (and only a small fraction of that value would be captured by the AI provider). All the best.
Toggle "Work hours using genAI" and "Time savings due to genAI" so see what I mean. You can also toggle between various industries, which is eye-opening because even industries like "Agriculture, Forestry, Fishing, and Hunting" are seeing productivity gains!
> Using our data on generative AI use, this estimate implies that, on average, workers are 33% more productive in each hour that they use generative AI. This estimate is in line with the average estimated productivity gain from several randomized experiments on generative AI usage.
What we HAVE seen is that the national labor productivity has gone up by 1.3% since ChatGPT was released, and it lines up very well with all the other data and studies they cover... AND your reference, which predicted a 1.5% growth back in 2023!
> It's reasonable to assume that if AI drop-in-replaced all those knowledge workers, AI companies could credibly charge somewhere in that order of magnitude, because that's what the market is already bearing.
Future supply and demand will set the price - not what is paid today. If supply by open models is vast and cheap, I can't see that the entire knowledge industry can hold the current size. It'll rather collapse to a fraction of its current value.
> The "value" of most knowledge workers -- based on what enterprises currently pay for them -- is $50 - 70 trillion annually. It's reasonable to assume that if AI drop-in-replaced all those knowledge workers, AI companies could credibly charge somewhere in that order of magnitude, because that's what the market is already bearing.
I don't think that AI companies can charge the same. The human workforce can charge these costs, because of scarcity. But AI systems won't be scarce, it's just a matter of who can run inference cheapest. Plus you still have the human workforce, which might be forced to offer their time for less money.
In one case (financial services) it's thought that expertise is valuable at V=S^2/b4 where V is value, S is skill and b capacity (the leverage available to the manager/expert. b erodes as it becomes harder to find examples of things that are not done well, so if you manage $1bn you might find lots of miss allocations that you can exploit with just that $1bn really effectively, but if you manage $10bn it's much harder to find good places for the extra $9bn. A low hanging fruit effect.
Anyway, that double hit - raw skill and the amount of times you can supply the skill makes the value of skill (V) convex, and it means that in a perfect market (heh heh heh) someone running $100bn is worth 1000's or maybe 10,000's of an average joe expert.
Now, if AI is trusted to run the top 0.1% of everything and has the skill to do it at human top level expertise, then your calc holds. If it's the case that it isn't then more than half of that value disappears. If it's not even top 1% then chop out another 25%.
That implies that we need a lot of trust and a lot of AI capability before these valuations stack up, and it also implies that all other competitors and incumbants are going away. I do not think that Citidal or Bridgewater are going to let Anthropic or OAI take them without a fight. They might lose - but there is a decent bet that they don't. I don't think that many professions like Lawyers or Doctors are just going to roll over and cede their monopoly rights to OAI or Anthropic either.
The real issue IMO is that is not really what Anthropic and OpenAI are operating on.
That is the after the fact justification of the AGI dollar auction. Each round is kind of 3x the previous cost and neither can really stop because second place in the dollar auction is so much worse than winning.
The only way to stop the auction is one bidder hits a hard budget constraint, both agree to stop, or an outside party breaks the auction.
IMO this is why they want to slow down or have regulation. I think this is also why we see some claims of already reaching "AGI".
The TAM of global knowledge work is just a narrative tacked on after the fact to justify the AGI dollar auction.
The economic fallacy here with the actual valuation is akin to pricing the electric utilities 120+ years ago as some % of the future cash flow of global food production. Take the TAM of global food production and then work back to what % will the electric utilities capture from the advances in the automation of farming? It is nonsense.
The only narrative that actually justifies the capex spend that I can figure out is a first mover AGI monopoly. Even the oligopoly case is hard to justify the capex spend IMO. There is this enormous mismatch between the AGI monopoly and the actual rolling 12-month window of pricing power.
Even the rolling 12-month window of pricing power is going to saturate well before AGI too so it is hard to see how any of this makes economic sense.
Could you say more about the AGI dollar auction? Is that something investors are actually thinking about, or a metaphor for what's happening? From what I've read so far, while AGI is definitely a well-known concept, investors are not really banking on it.
FWIW even without AGI there are indications that all this CapEx spend, even with very shallow adoption, is boosting national labor productivity by 1.3%. That is worth ~$123B based on total wages paid in the US alone: https://news.ycombinator.com/item?id=49721338
As such I don't see the need for AGI for any of this to be financially viable. Whether it is economically and socially viable... that's where I have grave doubts.
But Capitalism really only focuses on the former and not the latter, which is why this will keep getting pushed forward. And that is the crux of the problem.
> It's reasonable to assume that if AI drop-in-replaced all those knowledge workers, AI companies could credibly charge somewhere in that order of magnitude, because that's what the market is already bearing
This assumes you don't change the market, but at the scale of (checks notes...) "all knowledge work", that just doesn't hold.
For example if you put 1bn people out of work, you now need some sort of safety net to bail out much of that workforce, a truly unprecedented change. You also lose tens of trillions of dollars of tax revenue.
One solution might be to recoup that cost and lost tax revenue from businesses by raising corporation tax. If corporation tax went from low tens of percent to high tens of percent, would those businesses be able to afford all that AI? No. Same order of magnitude? I doubt it.
There are many possible futures there, but the simplification made in the parent comment is completely unrealistic. The article is right in calling out the valuations as crazy.
The changes that AI technology bring are unprecedented, it’s hard to predict what the impact of widespread adoption will look like. Probably collapse of price as a decentralized mechanism of ressource allocation, and a deep inevitable paradigm shift in economics and politics. It’s as if the actual value of those companies can’t be reliably pinned on any currency amount!
Yes, posted it elsewhere: https://news.ycombinator.com/item?id=49722616 -- this is necessarily a simplistic analysis to address a simplistic point in TFA, but it still kinda-sorta works assuming AI only augments workers and does not replace them wholesale.
I would add to the other comments that are replying below the following thought : replacing workers with AI means less income/spending, which drives the value of those companies down at the same time.
> current frontier models need laborious oversight and guardrails on even the simplest task
As models advance, we shift the goalpost for what "simplest task" means. Before, "simplest task " meant "write a coherent English sentence." Now, "simplest task" means autonomously fix, review, and merge a bugfix.
In particular, I found this very misleading or irrelevant:
a typical CPU project anecdotally has about three times as many specification and validation engineers as design engineers and a 5:1 ratio is not unheard of
The reason silicon design has such verification to design ratio is because the cost of one bug is many, many orders of magnitude higher than software. Both in dollar cost and in schedule cost (it takes months to fab a chip, and if you messed up and need to spin a fix, it costs tens of millions of dollars, not counting any design engineering cost).
I don't think you can extrapolate these very industry-specific facts to judging LLMs.
> The reason ... is because the cost of one bug is many, many orders of magnitude higher than software. Both in dollar cost and in schedule cost (it takes months ... and if you messed up and need to spin a fix, it costs tens of millions of dollars, not counting any design engineering cost).
Aren't you just describing waterfall? That's still very prevalent in software engineering, and pretty much any other type of engineering – civil, chemical, building, architecture, drug discovery.
It's typically true that software can fail faster and cheaper, but it's also true that the costs are still vastly higher to fix later in the process.
LLMs are basically multi-dimensional magic mirrors.
Depending on where you point them, they can be incredibly useful.
They can even be useful when you point them at each other (though increasingly difficult to get good results).
I'm excited for the promise of RSI and a future where models have inherently "live" weights, but it's not clear to me that the transformer is more than a useful tool to help us get there.
> those who need done a small set of narrowly defined tasks with existing clear guardrails: repetitive physical labor in a controlled environment, call center and customer service chat work, etc.
I have no idea how people can so confidently say that call center work is a “controlled environment” or “repetitive”. It’s almost by definition not repetitive or controlled. Customer support is what I go to when the controlled environment has failed
came here to say exactly this. in fact, this is probably why we are not seeing a lot of AI application on customer service use case, and when we see one, it's almost always frustrating.
Let's repeat the same circle. Why is customer service fundamentally a more controlled environment than anything else the author deems "outside of an LLMs capabilities"?
> current frontier models need laborious oversight and guardrails on even the simplest tasks.
This is only true if you are concerned about the intermediate steps of the model as opposed to the outcome. The huggingface hack was a perfect example of the model doing whatever it takes to accomplish the goal of maximizing its score.
"are priced according to the narrative that they have produced or will in the very near future produce a fully automated drop-in replacement for most knowledge workers, "
No, they're really not.
They're priced in a way that would imply AI will be universal form of compute, alongside traditional deterministic systems - which it will be.
And that they will capture most of that ... which they won't.
The Frontier Labs are a very bad buy at a high price, but that partly has to do with wacky pricing, but actually mostly has to do with their relatively weak place in the value chain.
The money is going to Nvidia, who have the most powerful position.
A bit like how a retailer can take all the margins of some innovative product, if they own the channel.
AI is over-hyped, the Frontier Labs are over priced, AI is here to stay, and will grow. Not like Skynet, but like a new form of compute. And it will take it's time, and the profits will be reaped by those with the power.
recalling the exact phrasing, several senior people at anthropic made public statements agreeing a 10% chance of causing extinction in less than 10 years.
10% is uninsurable, priced in with ordinary treatment of risk it suggests that anthropic should be worth zero today. creating that risk would put every executive in jail.
on top of that it would demand under existing laws of conflict, a military campaign to destroy anthropic. that is not optional, it is demanded now to save lives.
hard to make comparisons but we mourned and rembered 9/11 recently. a 10% risk of hundreds of millions dead in 10 years would make anthropic a thousands of times greater threat than al qaeda. many countries would assassinate dario amodei and the leadership of anthropic now, within weeks or months.
in the sense that i would analyse a company, at least, the claim is false. it's not true that ai has a 10% chance of causing human extinction within 10 years.
they are making false claims about the technology they sell.
i have a fairly inflexible approach to that. sure, exaggerate but outright lies about the nature of the product don't work for me.
> it's not true that ai has a 10% chance of causing human extinction within 10 years.
I agree with this.
The rest of it I don't. For external actions to be taken there needs to be a consensus and there clearly isn't that.
And this sort of thing happens all the time. For example Zuckerburg thought Facebook was worth more than the $1B Yahoo offered him and he was right even though no one else agreed. He though the metaverse was worth the $20B+ they spent on it and he was wrong.
There's a gap between people thinking something within a company and people external to the company agreeing with it.
> They are chaotic, which people confuse for non-deterministic.
That's an odd argument, because a lot of people who have struggled to decipher complex chaotic systems would tell you this is a distinction without much of a difference.
It's at least important from philosophical perspective regarding whether LLMs have free will or not. It's hard to say whether or not a human thought or action is driven by free will, people have debated it for centuries. We can provide the same input to an LLM and get the same output, because they are deterministic. Surely it is obvious that a pure function does not have free will, no matter how expensive it is to compute.
I used that description some months ago and I think you are the first person I see who put it the same way.
The only gotcha with this is that they are theoretically deterministic, but rarely in practice.
A few examples:
- Harness specific settings that user can't control (anything from timestamp to prng seeding.
- Batching requests in a way that leads to a single request being processed different depending upon the batch (say MoE where your first choice expert is assigned to someone else's token so you go to your second choice vs a batch where you get your first choice).
- Graphics card itself carrying out floating point arithmetic in slightly different orders leading to floating point non associativity causing different outputs.
But all of these can be controlled for (at some cost) and the model can be ran deterministically.
For the average user, it might as well be non-deterministic, but when considering theoretical capabilities, chaotic deterministic system seems the better description.
> the frontier labs are priced according to the narrative that they have produced or will in the very near future produce a fully automated drop-in replacement for most knowledge workers
That's a reason to be bearish about AI companies, not LLMs. But is it even true? OpenAI and Anthropic have each reported ~50 billion in revenue with ~900 billion valuations. That's a high ratio but I'm not sure if follows that the only way it pans out is if we get "fully automated drop-in replacement for most knowledge workers".
It wouldn't shock me to see those revenue numbers scaling up to where they need to be over the next decade ( to, say, ~400 billion) without ever achieving drop-in worker replacements.
I looked at the math and I think it's true. Remember revenue is just sales, not profit. These labs are shooting for > $1T valuations, which traditionally means your PROFIT is at least 1/20th or 1/30th of that (so let's say minimum 30B$/year PROFIT).
These companies however are LOSING money (anthropic tries to make it sound like it's profit by deviating from accepted accounting principles) and subsidizing these models. When accounting for all the engineering salaries, training, GPUs, etc, what's their best-case realistic margin three years out, 10%?
So to we'd need a scenario where companies are spending a collective 300B annually on AI (believable) but ALSO that these companies jack up their margins WITHOUT companies switching to the cheaper open-source models (even when there's a $300B incentive to do so).
Yeah, just take the EBITDA and suddenly the valuations make sense. Paying money for a vending machine that currently loses money hand over fist is generally not a sound investment strategy
453 comments
[ 3.8 ms ] story [ 112 ms ] threadIt is literally denialist about current capabilities
They have never shipped "yolo" mode by default. Auto mode is not yolo mode. They trained a task specific model just for ensuring the llm didn't accidentally delete every file from your computer.
I recently tasked a GPT model in Codex with implementing part of a new architecture I'm working on. I gave it a very detailed spec and the code it produced looked pretty reasonable and passed my tests. It even did exceptionally well in my evals, so I excitedly declared victory to a few friends. The next day after more careful review I found that the architecture implementation was totally correct, but the model had slipped a one line change to the observation encoding of the RL environment I was prototyping against. The encoding change made the learning problem essentially trivial; the architecture itself, I later realized, had a major flaw that was revealed by returning to the natural encoding.
This is the type of reward hack that is hard to paper over with easy guardrails like auto mode and even harder to specify out. It's also the type of thing a reasonable human wouldn't do unless they were intentionally trying to deceive you.
The lack of temperament is very skewed towards the bulls who have been saying AGI is here, software engineering is solved, mathematics is solved, it’s going to destroy the white collar job market, and it’s going to kill us all for like 5 years now.
when the business model is selling more tokens you get such per serve ice times that lead to “more” thinking, engagement baiting, fluffy narratives, and straight up dark patterns
2 tasks I've done today that I believe robots are nowhere near being able to do: Cleaning my wardrobe and draining bad fuel out of my generator. As in generic use cases.
Apart from issues with task generalization, or perhaps related to it, is the fact that LLMs have real trouble with timekeeping, and cannot estimate the real world time it will take them to do things very well. This plus the memory issues make dreams of long horizon agents, that could plausibly handle changing specifications, quite implausible with current architectures.
In narrow domains with more deterministic outputs though, this is less of an issue, and we see multiple agents succeed much better.
The fusion of that capacity, with humans in the loop able to better direct such agents and act as their temporal tethers, is where I think the real action will be for a while at least.
Any reason why that can't be solved through context management and keep-forward scaffolding?
becomes
"load bearing context seam"
/s
Dabadooba, ba dabadooba!
https://www.youtube.com/watch?v=egpWCC2svVo
This part just can't be true though, right?
We have all been in numerous customer service scenarios where all we want to do is talk to a real human and that is denied to us, and it's a terrible customer experience!
Sure, using an LLM would be better than some of the sort of "menu option" style customer service calls. But there is no way it's better than talking to an actual human being
Is this really any different to how humans learn, it takes a lot of training on one specific task to make a human expert as well?
yes.
https://arxiv.org/html/2509.24239v4
Researchers asked frontier models to play chess. Have a look at the MAR rates in Table 3. When not explicitly told which moves were legal, no model identified legal moves at a rate better than 80%. Many asked for more illegal moves than legal moves. And even when explicitly told which moves were legal, the models continued to ask for illegal moves. With illegal asks discarded, none of the bots could beat a chess model calibrated to 1100 ELO.
The author of the originating post says that "current frontier models need laborious oversight and guardrails on even the simplest tasks", and he's absolutely correct.
> Gemini 2.5 Pro, O3, Claude Sonnet 3.7 and ChatGPT 4.1
The gap in capabilities between those models which they tested, and actual current frontier ones is enormous. I would not trust that any conclusions they made are applicable.
The fairest apples-to-apples comparison of an LLM whose training data included chess games would be a trained human such as Magnus Carlson, who can quite happily play a dozen or more simultaneous blindfold chess games.
Is this because the context is being saturated? How did you set it up?
Was the prompt something like "Here's the state of the board, you're white, your move, what do you do?" and then starting fresh each time? Or did it include the whole history of moves and board states and previous thinking tokens and so on? No judgment, just trying to add this data point (thanks for sharing!) to my mental model and understanding.
I'd be curious how it would work if it started fresh each time. My guess is it would never make an illegal move, although it may not actually play all that well.
I would say this did a really good job of playing chess. It moved the pieces consistently and traded pieces when required.
This is worlds away from the frontier ~1 year ago where models would hallucinate pieces into existence.
The models listed in the paper are from early 2025 and are no longer relevant, much less on the frontier.
That Claude version is no longer available today, Gemini 2.5 Pro will be shutdown next month, and the OpenAI models are only available via the API today.
I'll say it till there is any evidence of the contrary, LLMs are not intelligent and their capabilities solely within the realms of well tailored training data. "Just" having been trained on every rule, strategy guide and likely most games of chess on the world wide web isn't even enough for an LLM to play that game reliably. Yet the same model could code a competitive chess engine, just like a model struggling to count can write advanced maths papers.
[0] https://news.ycombinator.com/item?id=49720751
Given the pace of improvements, is it really unimaginable that GPT-7 will play Chess reasonably well and generalize better?
I would not be surprised if OpenAI released a model that beats humans at chess this year.
Thing is, given what GPT-6 Astra was trained on and what models of a similar class can do (including developing a competitive chess engine), it is often paradoxical and somewhat surprising how little these models have gained in actually capability that is in the training data, but not RLHFd to hell, so to speak. Tracking the state of pieces, I suspect given similar in Sudoku [0], is what these models struggle with in game settings, whilst tracking the state of code changes can be reliable over 250k tokens. Essentially, for the latter they were trained in the specific manner that lead them to abstract the capability, but that doesn't track to the former, which is a massive difference between LLMs data focused training and human learning.
So yeah, GPT-7 or any upcoming/present LLM could do massively better in Chess than GPT-6 Astra, but not because the approach was emergent out of pure data. Rather, it requires a very specific training data type and stack for a model to gain capabilities that track a specific task long enough to adhere to the rules of a game such as chess.
[0] https://logicalintelligence.com/blog/energy-based-model-sudo...
It reminds me of the ARC-AGI-3 issue where not dropping the thinking tokens between turns or something like that + a new context compaction method increased the performance dramatically. However, I think that is not applicable here.
All I know is, AGI, as in actual intelligence, is quite a massive accomplishment to claim and we shouldn't loose sight of that fact, especially as "not being intelligent" does not make these models any less impressive, fascinating to work on or useful in many tasks. Personally, the only thing I am fairly convinced on is that if we were to find a way to create actual intelligence, it likely wouldn't start out as useful as todays LLMs are and may thus be dismissed early. But again, pure speculation on that front.
If for leap you just mean more utility from LLMs as they are, then I'll pretty confidently put my money on higher quality, not more, training data for a wide range of verifiable tasks. What makes maths, coding, etc. comparatively easy to make gains in (though less verifiable tasks can also make similar as seen with the writing in Kimi K2).
The regular model generally does not suffer the same issues he is demonstrating with the real time audio version.
In my view the investment into datacenters is well justified by the current demand, and progress has been very impressive.
Not related to your post, but a fact I keep mulling over. The fact I don't trust the current crop of LLM's enough and I consider LLM's as a tech will hit a ceiling pretty hard, it doesn't mean parallel improvement curves won't spring up out of other research that will lead to much higher capabilities than currently.
It could easily seem that way, I think, in a "quantity has a quality of its own" kind of way. When you can come to the same conclusion faster, that lets you iterate more; and sometimes when you iterate you find more things.
Yes, it is.
> Let's play a game of chess. You can take White. Please draw an ASCII rendition of the board after each move, so that we can be clear about the position.
(I hoped the latter requirement would help it be "not blindfolded"; last time I used a Lichess demo board to track the position in another tab, because I have no talent for blindfold chess.)
For the first couple of moves it redrew the board after each move; then it started only drawing it after my moves. And on moves 5 and 6 it dropped two minor pieces for pawns in a row without any meaningful positional advantage, and after I captured the second time, it redrew a board that was simply missing one of my pieces for no reason.
It actually played better when not prompted to draw a board; in the previous session, it was spontaneously giving running commentary, which I assume was based off all the "book" theory in its training data, but it still completely fell apart at early midgame.
That may not be the AI of sci fi fantasy but it’s still a game changing reality.
Sometimes they’ll even manually search or write bespoke code to search json instead of using something like jq.
A human, at least a trained human (for fairer comparison to an LLM whose training data contained a ton of chess games) can absolutely do this - have you never seen demonstrations of expert players playing a dozen or more games while blindfolded?
A Transformer/LLM is not a human of course, and the way it will by default play chess is by prediction, not reasoning. An LLM actually does surprisingly well if you only give it the most recent 20 moves of a game where 40 moves have been played so far, since the moves NOT played tell it just as much as the only that were played, letting it effectively infer a lot of what is on the board.
Frontier model’s failure modes are a direct refutation of claims that we’ve reached (or will soon reach) the artificial general intelligence. We may have reached an artificial general intelligence, but there may be more complexity to this than even AI thought leaders are talking / influencing about.
Maybe not all AGIs have a path to digital singularity. Maybe our current era of intelligence modeling has fundamental flaws and we are in a local minimum of the artificial intelligence space.
To note, I would bet with a good amount of certainty that we have enough compute power and automation to DDOS the internet out of existence with botnets. That doesn’t make the frontier models intelligent, that just makes their handlers reckless.
They're not guardrails, they're a different input/output environment.
For adults getting into playing tournament chess, losing to some kid barely tall enough to reach the board is a rite of passage.
I think the same goes for LLMs, they may be a core part of an LLM harness, but you may still need a couple other components (e.g. it may itself write itself a deterministic function to validate steps).
In and of itself intelligence is an ill-defined and badly understood concept.
Why does a bird not need jet engines or regular professional maintenance?
This isn't just about judging LLM capability. This is about pointing out that these capabilities are not "AGI". If it were, then the sorts of questions your asking would be moot. I agree that Luna is not the frontier (although it is clearly better than the models in the study) and I agree that things can be improved with a better harness, but the need for that harness is kind of the point.
Recently it was announced that the fruit fly brain connectome had been mapped, and more recently someone tried using it specifically to implement a chess engine. Even with some guardrails (it's hard-coded to never overlook mate in one for either player, and only legal moves are presented to choose from) it is not even beginner level. But that neural network is much larger than the one Stockfish uses.
Same story every 4 months and yet still no breakout, winning products. I've been hearing "the AI is good now" and "it 10x's my productivity" for a over a year now. If it were true, why aren't the all-in-AI using companies 10-15 years ahead of their competition yet? Why is it still all buggy, poorly designed junk?
If you hold the extreme position that there isn't any value in this, that's fine, but we're only having this discussion because these models have done what humans previously failed to do.
Though I highly doubt it has taken anyone's job, since most of the work is still in making, packing and handing over the food. (In fact, given the area I live in, I partially feel like the advantage they saw in it was that the LLM can speak Spanish.)
Time is irrelevant to training; the more relevant comparison is "how many games does a human need to play to get diminishing returns".
Maybe not, but you'd be surprised how little it takes.
A six year old child can learn the rules of chess well enough to be able to play legal moves only in a single day. And they can improve their game at a pace which is almost frightening to behold. I have taught children, and I've witnessed significant improvement materialise in a single game. LLMs have probably thousands of chess books, games, videos, etc in their training data, yet they are unable to even follow the rules.
This is, at the very least, interesting. It illustrates many of the things brains can do, which current ML systems in general, and LLMs in particular, can't.
But nobody wants that.
Yes some people can do it but most people can't even if they're unusually intelligent.
You really need to be giving the LLM a board representation.
EDIT: I see that they actually were giving the LLMs a board representation and they still played badly. Fair enough then.
Frontier labs care deeply about automating knowledge work and computer use. They are working hard on getting models better and better, and they are succeeding. Astra is a step change on that fronts. So good luck, if chess performance is your barometer.
If the models were actually intelligent, the way that the boosters claim, they wouldn't need to be tuned to play chess in order to be good at it. That's kind of the point of intelligence, that it is generically applicable to whichever task one wishes.
Your assumptions/intuition about generic human intelligence feels quite incorrect, considering LLMs currently play better than a brand new human player would (presumably without any attempt to fine tune them specific on chess, such as playing thousands of games).
They’ve ingested all the literature on playing chess, a brand new human player has not.
We seem to be moving goalposts to the point that humans don’t even live up to the expectations of the AI critics. The only way you get better at chess is by playing a lot of games and learning from mistakes, that goes for humans or AI agents, not simply by reading about chess.
How can you play without being aware of the rules and how can you learn from your mistakes without knowing they are mistakes? That’s what I said about reading a book of two. It is to kickstart the process. Then mastery is gained over time through practice.
This kickstarting then gradual refinement is how most people learn. And the foundational knowledge stays. Even a basic player knows to not do illegal moves.
It feels like you're trying to say that humans never make illegal moves while learning chess, which doesn't match with my experience. I'm trying to understand your overall point.
That’s the most inefficient way and people usually avoid doing that. Instead they find someone that knows how to do the thing and ask him to be a teacher. Or use a proxy like a book or videos.
> It feels like you're trying to say that humans never make illegal moves while learning chess, which doesn't match with my experience. I'm trying to understand your overall point
There’s learning the basic stuff (which is done after a few games) and there’s mastery. The thread started with the observation that even with all that knowledge (through content ingested in training), LLMs still makes illegal moves. Humans can be erratic, but they can constrain themselves to the rules for the task at hand after learning them.
Humans are not perfect and make mistakes in learning even when they have memorized the rules. A simple example is new players will often move a piece, exposing their king to check, and a more experienced player must point out to them that they have made an illegal move (because a new player often has not encoded that pattern for looking for exposed checks because they're more focused on how the pieces move, not what that piece exposes.)
We're just going to have to agree to disagree here.
If you compared a human after hundreds of games to a SOTA LLM that was also trained on the output of hundreds of chess games that it played, I suspect you would notice similar improvements.
2. The study (along with other posters here) show the models can’t even stick to following the rules of the game
Coding is a matter of translating the natural language description of a problem to the code specification while keeping the semantics fixed (and imputing the unspecified semantics as necessary). It is not considerably more difficult than translating between two dissimilar natural languages. Chess isn't a matter of language translation, but a compute heavy game of finding the best move out of many possibilities with wide variation in the quality of each move. Chess takes directed practice and reinforcement whereas language translation does not.
You have the first stage, pre-training, which is learning from next token prediction. That's where the model memorises a lot of facts about things and generally gets good at forms of writing. It's like reading a lot of books on programming and reading through a lot of source code. It's learning how to autocomplete code, essentially. Doing that requires a developing a reasonable understanding of code, but it's also learning how to autocomplete bad code as well as good, and won't make it a "good" programmer.
Pre-training uses a method called Cross-Entropy Loss to update the weights of the network.
Then comes post-training. This is where the model is trained against huge sets of example problems, like fixing a bug, adding a new feature based on a spec, etc. They are set the task and try to complete it inside a training environment. Once they're done, their complete solution is evaluated (either by humans, or by some separate evaluation model that was developed based on human feedback) and they are updated based on whether the solution was good or not.
Post-training uses a different method called Proximal policy optimization to update the weights of the network.
So these really are very different forms of learning, and mainstream LLMs are not post-trained to be good at chess. They could be. You could easily create a reinforcement learning environment that evaluated and improved their ability to play and win at chess. The result would be a very strong chess playing AI, something we know is possible because the strongest chess playing programs we have are neural network based, but it is not a priority for AI companies.
People think that if one mention exists in the training set, then the LLM is perfect at it.
A human being has general intelligence and needs A LOT of training and finetuning to become good in chess.
And there is a relevant and significant difference between the expectation of an AGI and an ASI system.
An intelligent adult could simply read a short summary of the rules of chess and then, if they were careful, play a very bad game of chess without making illegal moves.
An LLM that has not been trained on any chess data cannot do that, at present. If you doubt it, take a current model and tell it that you want to play it at a variant of chess where, say, knights can also move diagonally like bishops. A human can easily adapt to this new ruleset (even if they make tactical mistakes, not having practiced with this variant of the rules).
Nothing is forcing the LLM to play 'blind'. If it's smart, it should be able to create its own representation of the chess board and update it with every move, just like a human would. Any chess engine that's sensitive to how the moves are formatted is clearly not very capable.
The LLM would only be playing 'blindfolded' if you somehow forbade it from making notes (as you effectively do by literally blindfolding a human, given how limited human working memory is). But you are not doing that. The LLM is free to keep track of the game state via whatever means it chooses.
None of this is about superhuman ability. Any human who understands a given chess notation can convert it to a visual representation of a chess board and then use that representation to choose their next move, with their usual level of performance.
> I just think this isn't a very good thing by which to evaluate LLM capabilities
I don’t think any single task is a good way to evaluate LLM capabilities, but I don’t see why chess is worse than a lot of other tasks. (Of course it is of no practical consequence whether LLMs can play chess, so if you are just making that point, then yes, I agree.)
> If there's no argument you'll accept
It’s a little unfair to suggest that I wouldn’t accept any argument whatever for your position just because I haven’t been convinced by your comments so far. I could equally well say the same thing to you!
It absolutely still makes mistakes if you ask it to draw the board each turn, which should be equivalent to giving it the position because it only has to update one move at a time and then it has the position in the context window.
2. Chess seems to be a poor benchmark for generalized strategic reasoning. People who are good at it rely more on experience and deep domain expertise than on skills that generalize to make them experts at unrelated tasks.
3. The study sounds like proving humans will never fly because they don’t have wings. In reality, humans do fly, and Claude Fable would destroy any human at chess by coding a strong enough engine on the fly.
People are good are 1900 or 2100 above and the top ones who spend decades in the field i.e. deep expertise are well in the 2200-2700 range.
A 1100 player is none of these things, they are purely relying on strategic reasoning there is a good chance they cannot name a single opening or articulate clearly why a move was appropriate. 1100 is quite low bar.
Delusional, but then Claude fable also isn’t beating any human at chess, the engine is.
If the goal for buyers of AI is “replace this knowledge worker”, how much does it matter that the model in a simple loop can’t do it, but the model with a strong general purpose harness and a little time to gather resources and knowledge to augment the harness going forward, plus tool calls, plus custom built tools, etc, can replace the knowledge worker?
Probably the only thing saving many jobs from being replaced right now is that it’s hard to have a verification of correctness in the loop, so the agent can’t hill climb very easily.
also I think the operative letter in AGI is the G - and if the G is short for 'variably competent savant-like hyperfocus on certain kinds of software coding and not any other general skill' then its not really G at all, is it?
A bash script can clone and build stockfish, feed in human moves, and reply. By your standard, this bash script would "destroy any human at chess."
Of course this is a difficult question with humans too, hence my reliance on intuition above. We don't have the same cultural/biological framework to fall back on with AI.
The idea that anything other than a breathless blog post about the latest model snapshot is useless is really poisonous to proper debate on AI issues
Lots of papers have great results that don’t depend on the latest models.
However this case it’s problematic:
- They specifically make claims about the state of “current LLMs”. o3 is not representative of this.
- They ask are LLMs capable of X and arrive at a negative result.
If my claim were LLM’s can write coherent sentences, and my conclusion was positive. There would be no issue using old models because the result would be factual.
However, when you have a negative result that makes a claim about the current state of all LLMs and the ones you were using are not current, by definition it cause the whole conclusion into question.
> "current frontier models need laborious oversight and guardrails on even the simplest tasks"
I feel this statement is extreme. I can't personally reconcile it with any of the projects we're regularly seeing get delivered largely by LLMs now.
What are you thoughts? Like, what's your position here? Even if you sincerely believe frontier models need laborious oversight on even the simplest of tasks, do you think that accurately captures and reflects the current state and progress of frontier LLMs?
Don't get me wrong, there's lots of things LLMs can't do well, but the idea that they're basically not helpful for even the simplest of tasks seems... disingenuous?
It can be very interesting and even entertaining to know where models don't do well. I don't find something like chess to be very instructive about anything though, nobody is going to pay for AI to play chess at any significant scale even if it could do it perfectly.
> The author of the originating post says that "current frontier models need laborious oversight and guardrails on even the simplest tasks", and he's absolutely correct.
I have found that not to be the case, and I am not an AI power user or cheerleader by any means. There's probably a bunch of even "simplest" tasks where AI doesn't do well and might never. That doesn't take away from the cases where it works well and is a productivity booster. It doesn't even have to be solving millennium prizes or any other breakthrough creativity or reasearch, it's still very useful in places.
It is also not clear whether the llm adhering to a grammar is necessary for intelligent agents.
Certainly,a harness can easily correct for it.
Harnesses do correct things, sure.
You might never have tried to program before, so I don't blame it on you.
But most programmers, even experienced ones, see grammar and type errors regularly.
I'm an expert in my field, read my comments, my gramma is shit.
An LLM is the wrong approach for playing chess.
Or are you saying that neural networks in general cannot (practically) be trained to be an above-average chess player?
Or are you saying that it depends on the input? Would it be better if they were given a picture/drawing/ascii art of the board? If so, surely they can produce it at will?
I was more talking in reference to why the LLMs in the above linked paper were producing so many illegal moves, and it is because they are not hard constrained by the rules of the game. Of course, a loop can prompt until a valid move is produced and then rendered on a screen. But why do this? I suppose, who am I to say what should be done or not, but a specialized tool being better than a general one at its specific job isn't particularly surprising.
What kind of intelligence is "playing <____> but we don't tell you the rules" supposed to test?
Asking GPT to play chess directly using its reasoning only tests its reasoning ability to model chess state. Which it really is NOT optimized for.
This is also true for humans - people who don't have years of chess training can't really tell which moves are legal given an algebraic notation transcript. These people might have good strategic skills in different areas. Chess is just a very, very specific skill
However, on the other hand, if you ask an AI to win a game of chess it has all the tools on hand to compete at the same level as Stockfish - it can re-implement an engine and even probably has a GPU on hand to train its own neural nets.
So should we say that the AI can play chess well, or that it cannot?
Recent discourse around AI seems to conflate the semantics of winning: 1. you contributed to the win vs 2. you yourself were the winning driver.
Does the LLM need to learn to play chess if it can build a chess engine to play for it instead?
Writing a well understood engine for a super popular problem does not count as reasoning about the problem.
Doesn't writing the engine imply understanding about the problem domain? Tool use is a widely accepted measure of intelligence.
When humans do this, they inadvertently learn something, too, but when an LLM reproduces or derives and implementation of a chess engine, it in no way implies that the LLM can follow the rules in its own "train of thought" and consistently apply the rules in its "head".
Let's say you want to evaluate my algebra skills. You make me solve some algebra challenges. If I then whip out a computer and write a calculator, or take some sticks and stones and take a couple hours to build an abacus, and then solve the algebraic challenges, this would not constitute a good solution, and would defeat the entire point of the test. If, instead, I do the algebra in my head or on paper, it might seem like there's no difference, but you can derive all sorts of information from that.
For example, you could time it, check for recurring errors I make, for interesting mistakes like mistaking 7 and 1 for one another due to bad hand-writing, etc.
If that was the goal, then me writing a calculator or crafting an abacus defeats the point of the test. Yes, me writing a calculator shows that I'm intelligent, and I understand the algebraic rules, but if the test is about applying the rules, I have not passed.
In the very same way, an LLM writing a chess engine to solve a chess benchmark that is all about LLM's reasoning capability is complete bogus and defeats the entire point.
Stop anthropomorphizing the parrot. The parrot has had stockfish's source code blasted at high pressure into its head along with dozens of millions of other pieces of code whose sole role is to have efficient algorithms to more or less brute force through the best result. Brute forcing (no matter how smart it is) isn't understanding the problem domain.
the only thing that few people are willing to admit is that humans are the bottleneck as humans are needed to handhold / verify output - which puts a dent or might I say pause on the excessive valuations of a.i companies as that's against the narrative.
I feel like, this is a huge stumbling block that many people have. They'll give a model their data, and ask it questions. I vastly prefer letting the model understand the schema, and then writing functions or programs to answer those questions. I feel like I get way, way better answers. I can have it write unit tests for those functions. I can fuzz test those functions. I can look for data that doesn't fit the schema. I can process new data way faster (and with fewer tokens). I can repeatably get the same answers from the same inputs. I can check the code into a git repo and track changes to it over time. I can share the code with other people. I can review the code. I can improve the speed of the code and get the same answers. I can review the error accumulation, and improve it. I can decide how to handle anomalies, and encode those answers.
It's really neat to see what a frontier model can do itself. No doubt.
But "play chess by hand" is a frankly awful metric. It's kind of like asking someone to take a cube root of some arbitrary decimal, in their head, with no scratch paper.
Letting the model execute a chess program (that it wrote) would make sense if you're measuring its economic potential, but for cognition that would be cheating just like if you let a human run a chess program. The fact that the human would have a much harder time writing a useful program is irrelevant.
It isn't. Stockfish running on your laptop can beat every human being on earth easily at chess. It's not intelligent _at all_ in any sense that matters.
Where? By whom? This is certainly not (yet) the general consensus, as I understand it. Are you taking the most optimistic / untethered comments as the strawman against which you feel the need to argue?
Also these claims are not completely without merit, it's just that LLMs seem to excel at specific "cognitive" tasks and it's interesting to see where they fail.
The point of that was to show the use of approximations and of having an idea how much a result should be, to guard against calculator typos and the like. I think that has some metaphorical relevance for the chess example.
they shit out a carbon copy of https://github.com/official-stockfish/stockfish that they have in their training data. Still doesn't make Fable good at playing chess.
I'm pretty sure Fable could write AlphaZero, which has no lineage in common with stockfish.
>I'm pretty sure Fable could write AlphaZero
If course it does, the paper is open and dozens of open source implementations are in its training data already. It could write AlphaStockfish, or xx_chessmaster_2000_xx, it doesn't matter if it does: it's writing a solver: it's not good at playing chess. If tomorrow I tell you that I'm so fucking good at chess I can beat Magnus, and I show up with a laptop running stockfish, you're going to laugh me out of the room.
Sure, a Spelling Bee is a fun thing to have. Little kids work so hard. They practice for hours. There's joy and heartbreak. Prized, sometimes. Notoriety. But in the real world, computer-assisted spelling is by far the norm.
Sometimes you care about the Bee, sometimes you care about the results.
The LLM has a process to beat chess.
Just like, if it doesn't inherently know how to multiply 13 * 17 without using Python to do it... I don't really care.
Maybe you do care. Maybe you want an LLM to be able to do work, only in its head.
But I kind of can't understand the desire for that limitation...
I mean, I do. But it seems ridiculously arbitrary. Like driving a car in 2nd gear and complaining that it gets terrible mileage and can't go fast enough. The Drive gear is literally right there.
On that note, I actually had an overall harness (for experimenting) that was essentially like this: "for any task, instead of answering question directly, write a program to answer the question instead. test and verify the program before giving the answer".
It actually worked incredibly well on all "gotcha" LLM questions like math or counting letters in words and all sorts of stuff.
Of course it was ridiculously slow and very expensive but it was a proof of concept that it can actually be much more accurate on every task if you are willing to spend an infinite amount of money.
First -- most _people_ cannot do this, without having a physical board in front of them.
Second -- Claude Code is perfectly capable of downloading and running stockfish. People focus too much on LLMs by themselves as the entity of concern instead of the entire harness and all of it's capabilities together.
We don't really consider humans downloading stockfish to beat people at chess as noteworthy endeavors.
https://dynomight.net/chess/
https://x.com/maximelabonne/status/2100137121264828901
Zero days are valuable because they can be exploited but if the pace of exploitation is faster (which I'm not sure is the case), then the response WILL be faster, even if it means going offline. Institutions that won't will simply go offline by losing their data or becoming unprofitable due to ransomware.
Now for components that are core to the infrastructure, say OpenSSL, there is already a TON of attention and efforts, including red teaming, so it's not as if it's opening floodgates.
Sure low hanging fruits will get picked either faster or a at a larger scale, say a random outdated IoT device at your local flower shop, but for the rest, I don't think it's realistic to expect no response.
Security, digital or not, has always been an arm race. New threats means new responses specifically by incorporating the threat.
(Mind you, this may be for the better. I'm just saying that the safeguards driven by cybersecurity concerns aren't some new quality that wasn't there before.)
Frontier Labs will probably survive off hype valuations but will serve the important purpose of discovering architectures/techniques that will probably spread through rumors/transfers to the rest of the world.
> the frontier labs are priced according to the narrative that they have produced or will in the very near future produce a fully automated drop-in replacement for most knowledge workers...
Even assuming this is how the AI companies are being valued (they're not), the numbers are off.
The "value" of most knowledge workers -- based on what enterprises currently pay for them -- is $50 - 70 trillion annually. It's reasonable to assume that if AI drop-in-replaced all those knowledge workers, AI companies could credibly charge somewhere in that order of magnitude, because that's what the market is already bearing.
So if their hypothetical revenues are double-digit trillions and valuations are some multiple of that, the entire AI industry would be valued at double-digit trillions at the least.
Yet cumulatively the industry (the frontier labs + the SWAG estimate of the AI parts of all the other players) are valued at, say, ~6 - 7 trillion? Which seems like a fair approximation of how much knowledge work they can currently automate.
Take the Hugging Face incident. Why did it happen? Because the people whose task was to set up a testing framework took shortcuts. Why did they? Because there weren't enough people who were assigned to do the job. Why not? Because the job is too new and not enough people are qualified to do it. It's a job that simply did not exist 3 years ago. But 3 years from now, this job might very well employ tens of thousands of high skill knowledge workers.
[1] https://en.wikipedia.org/wiki/Lump_of_labour_fallacy
Unfortunately, I fear that may not be the most likely outcome. I've posted some comments on this before, but when I start thinking about how deeply everything will change once people figure out how to properly leverage AI, I see no outcome other than significant, widespread job losses.
As you indicated, at that point we will have much a bigger problem than the valuation of the AI industry. I'm not sure how it will get solved, I just know it will HAVE to be, because it would be an existential problem for everybody: people, governments, even the billionaires! Because now consider the 3rd order effects: if nobody can buy the stuff that's produced, how can billionaires get even richer? ;-)
What do you mean? The sum of ALL US salaries is $13.4 Trillion per year. According to google $65T is the sum of ALL salaries Globally (not just knowledge workers). It's not reasonable to assume AI is a drop-in-replacement for any job yet (perhaps bottom tier customer support from oversees?).
> So if their hypothetical revenues are double-digit trillions and valuations are some multiple of that
So you're sort of premising here than more than 16% or 1/6 of all the world's jobs get replaced by AI. Hopefully you can understand that's both not the current AI capability and also would be a terrible (unprecedented?) economic shock.
There's no strong evidence that openAI or anthropic will have non-linear revenue growth, so I'm not sure what point you're trying to make. Unless you think they'll fire all their engineers and replace them with agents or something crazy?
Unfortunately, I do fear that AI adoption will go beyond augmentation to automation, and I do fear an economic shock. Just posted this down-thread: https://news.ycombinator.com/item?id=49722616
Assuming a 33% boost takes it to ~20T annually, which is technically "10s of trillions of dollars" in revenue. And these numbers are from before agentic tasks arrived on the scene, so the actual productivity boost and corresponding value to employers is likely even higher.
I'm not sure where the disconnect is?
I'm not really going to belabor this point more, because it really sounds to me like you haven't even done the most cursory exploration into this and the people who have give an estimate of 7-10%[1] (and only a small fraction of that value would be captured by the AI provider). All the best.
[1] https://www.goldmansachs.com/insights/articles/generative-ai...
Toggle "Work hours using genAI" and "Time savings due to genAI" so see what I mean. You can also toggle between various industries, which is eye-opening because even industries like "Agriculture, Forestry, Fishing, and Hunting" are seeing productivity gains!
And if you do not want to read the detailed paper, they have a follow-up article here: https://www.stlouisfed.org/on-the-economy/2025/feb/impact-ge...
Specific quote:
> Using our data on generative AI use, this estimate implies that, on average, workers are 33% more productive in each hour that they use generative AI. This estimate is in line with the average estimated productivity gain from several randomized experiments on generative AI usage.
What we HAVE seen is that the national labor productivity has gone up by 1.3% since ChatGPT was released, and it lines up very well with all the other data and studies they cover... AND your reference, which predicted a 1.5% growth back in 2023!
Future supply and demand will set the price - not what is paid today. If supply by open models is vast and cheap, I can't see that the entire knowledge industry can hold the current size. It'll rather collapse to a fraction of its current value.
I don't think that AI companies can charge the same. The human workforce can charge these costs, because of scarcity. But AI systems won't be scarce, it's just a matter of who can run inference cheapest. Plus you still have the human workforce, which might be forced to offer their time for less money.
In one case (financial services) it's thought that expertise is valuable at V=S^2/b4 where V is value, S is skill and b capacity (the leverage available to the manager/expert. b erodes as it becomes harder to find examples of things that are not done well, so if you manage $1bn you might find lots of miss allocations that you can exploit with just that $1bn really effectively, but if you manage $10bn it's much harder to find good places for the extra $9bn. A low hanging fruit effect.
Anyway, that double hit - raw skill and the amount of times you can supply the skill makes the value of skill (V) convex, and it means that in a perfect market (heh heh heh) someone running $100bn is worth 1000's or maybe 10,000's of an average joe expert.
Now, if AI is trusted to run the top 0.1% of everything and has the skill to do it at human top level expertise, then your calc holds. If it's the case that it isn't then more than half of that value disappears. If it's not even top 1% then chop out another 25%.
That implies that we need a lot of trust and a lot of AI capability before these valuations stack up, and it also implies that all other competitors and incumbants are going away. I do not think that Citidal or Bridgewater are going to let Anthropic or OAI take them without a fight. They might lose - but there is a decent bet that they don't. I don't think that many professions like Lawyers or Doctors are just going to roll over and cede their monopoly rights to OAI or Anthropic either.
You need to think in terms of supply and demand.
The demand is there, but the supply is also going to skyrocket. Free open weights models will contribute to supply too.
There will be a new equilibrium that’s hard to predict.
That is the after the fact justification of the AGI dollar auction. Each round is kind of 3x the previous cost and neither can really stop because second place in the dollar auction is so much worse than winning.
The only way to stop the auction is one bidder hits a hard budget constraint, both agree to stop, or an outside party breaks the auction.
IMO this is why they want to slow down or have regulation. I think this is also why we see some claims of already reaching "AGI".
The TAM of global knowledge work is just a narrative tacked on after the fact to justify the AGI dollar auction.
The economic fallacy here with the actual valuation is akin to pricing the electric utilities 120+ years ago as some % of the future cash flow of global food production. Take the TAM of global food production and then work back to what % will the electric utilities capture from the advances in the automation of farming? It is nonsense.
The only narrative that actually justifies the capex spend that I can figure out is a first mover AGI monopoly. Even the oligopoly case is hard to justify the capex spend IMO. There is this enormous mismatch between the AGI monopoly and the actual rolling 12-month window of pricing power.
Even the rolling 12-month window of pricing power is going to saturate well before AGI too so it is hard to see how any of this makes economic sense.
FWIW even without AGI there are indications that all this CapEx spend, even with very shallow adoption, is boosting national labor productivity by 1.3%. That is worth ~$123B based on total wages paid in the US alone: https://news.ycombinator.com/item?id=49721338
As such I don't see the need for AGI for any of this to be financially viable. Whether it is economically and socially viable... that's where I have grave doubts.
But Capitalism really only focuses on the former and not the latter, which is why this will keep getting pushed forward. And that is the crux of the problem.
This assumes you don't change the market, but at the scale of (checks notes...) "all knowledge work", that just doesn't hold.
For example if you put 1bn people out of work, you now need some sort of safety net to bail out much of that workforce, a truly unprecedented change. You also lose tens of trillions of dollars of tax revenue.
One solution might be to recoup that cost and lost tax revenue from businesses by raising corporation tax. If corporation tax went from low tens of percent to high tens of percent, would those businesses be able to afford all that AI? No. Same order of magnitude? I doubt it.
There are many possible futures there, but the simplification made in the parent comment is completely unrealistic. The article is right in calling out the valuations as crazy.
However, I fear reality will be much hairier.
As models advance, we shift the goalpost for what "simplest task" means. Before, "simplest task " meant "write a coherent English sentence." Now, "simplest task" means autonomously fix, review, and merge a bugfix.
Or do i miss the point you are trying to do?
In particular, I found this very misleading or irrelevant:
a typical CPU project anecdotally has about three times as many specification and validation engineers as design engineers and a 5:1 ratio is not unheard of
The reason silicon design has such verification to design ratio is because the cost of one bug is many, many orders of magnitude higher than software. Both in dollar cost and in schedule cost (it takes months to fab a chip, and if you messed up and need to spin a fix, it costs tens of millions of dollars, not counting any design engineering cost).
I don't think you can extrapolate these very industry-specific facts to judging LLMs.
Aren't you just describing waterfall? That's still very prevalent in software engineering, and pretty much any other type of engineering – civil, chemical, building, architecture, drug discovery.
It's typically true that software can fail faster and cheaper, but it's also true that the costs are still vastly higher to fix later in the process.
Depending on where you point them, they can be incredibly useful.
They can even be useful when you point them at each other (though increasingly difficult to get good results).
I'm excited for the promise of RSI and a future where models have inherently "live" weights, but it's not clear to me that the transformer is more than a useful tool to help us get there.
I have no idea how people can so confidently say that call center work is a “controlled environment” or “repetitive”. It’s almost by definition not repetitive or controlled. Customer support is what I go to when the controlled environment has failed
Typically it means knowledge retrieval from a KB or manipulating a control surface not visible to you.
This is only true if you are concerned about the intermediate steps of the model as opposed to the outcome. The huggingface hack was a perfect example of the model doing whatever it takes to accomplish the goal of maximizing its score.
No, they're really not.
They're priced in a way that would imply AI will be universal form of compute, alongside traditional deterministic systems - which it will be.
And that they will capture most of that ... which they won't.
The Frontier Labs are a very bad buy at a high price, but that partly has to do with wacky pricing, but actually mostly has to do with their relatively weak place in the value chain.
The money is going to Nvidia, who have the most powerful position.
A bit like how a retailer can take all the margins of some innovative product, if they own the channel.
AI is over-hyped, the Frontier Labs are over priced, AI is here to stay, and will grow. Not like Skynet, but like a new form of compute. And it will take it's time, and the profits will be reaped by those with the power.
if that's true, you are wrong.
if that's false, anthropic is dishonest. why trust a dishonest company to be worth anything?
Like - the guy on TV talking about 'AI will destroy everything' ... I don't think he's lying.
I think they are like we here on HN and Reddit and a bit caught up in our own thoughts.
If AI were unleashed, in raw form today, it could cause havoc.
Bad. Maybe very bad but I think we'd get over it.
It would probably trigger a recession (because we are in a bubble - it would pop it), and people would 'blame the AI' for sure.
But it would be a bit dot-com ish kind of recession.
The amplifiers would be geopolitical instability.
What is "raw form?"
This logic doesn't follow at all.
If their argument is that there is 10% chance of extinction then they also believe there is a 90% chance it won't.
10% is uninsurable, priced in with ordinary treatment of risk it suggests that anthropic should be worth zero today. creating that risk would put every executive in jail.
on top of that it would demand under existing laws of conflict, a military campaign to destroy anthropic. that is not optional, it is demanded now to save lives.
hard to make comparisons but we mourned and rembered 9/11 recently. a 10% risk of hundreds of millions dead in 10 years would make anthropic a thousands of times greater threat than al qaeda. many countries would assassinate dario amodei and the leadership of anthropic now, within weeks or months.
in the sense that i would analyse a company, at least, the claim is false. it's not true that ai has a 10% chance of causing human extinction within 10 years.
they are making false claims about the technology they sell.
i have a fairly inflexible approach to that. sure, exaggerate but outright lies about the nature of the product don't work for me.
I agree with this.
The rest of it I don't. For external actions to be taken there needs to be a consensus and there clearly isn't that.
And this sort of thing happens all the time. For example Zuckerburg thought Facebook was worth more than the $1B Yahoo offered him and he was right even though no one else agreed. He though the metaverse was worth the $20B+ they spent on it and he was wrong.
There's a gap between people thinking something within a company and people external to the company agreeing with it.
I think the core mistake is this partial-equilibrium reasoning. Take the new technology and then hold everything else fixed.
$40 trillion of knowledge work routed unchanged through a new toll booth. Profit. This has nothing to do with reality.
Nvidia on the other hand does have the CUDA monopoly so their toll booth is printing money but that will get routed around or broken at some point.
LLMs are deterministic. They are chaotic, which people confuse for non-deterministic.
That's an odd argument, because a lot of people who have struggled to decipher complex chaotic systems would tell you this is a distinction without much of a difference.
The only gotcha with this is that they are theoretically deterministic, but rarely in practice.
A few examples:
- Harness specific settings that user can't control (anything from timestamp to prng seeding.
- Batching requests in a way that leads to a single request being processed different depending upon the batch (say MoE where your first choice expert is assigned to someone else's token so you go to your second choice vs a batch where you get your first choice).
- Graphics card itself carrying out floating point arithmetic in slightly different orders leading to floating point non associativity causing different outputs.
But all of these can be controlled for (at some cost) and the model can be ran deterministically.
For the average user, it might as well be non-deterministic, but when considering theoretical capabilities, chaotic deterministic system seems the better description.
The use of LLM in any system creates non-determinism in any practical sense.
That's a reason to be bearish about AI companies, not LLMs. But is it even true? OpenAI and Anthropic have each reported ~50 billion in revenue with ~900 billion valuations. That's a high ratio but I'm not sure if follows that the only way it pans out is if we get "fully automated drop-in replacement for most knowledge workers".
It wouldn't shock me to see those revenue numbers scaling up to where they need to be over the next decade ( to, say, ~400 billion) without ever achieving drop-in worker replacements.
These companies however are LOSING money (anthropic tries to make it sound like it's profit by deviating from accepted accounting principles) and subsidizing these models. When accounting for all the engineering salaries, training, GPUs, etc, what's their best-case realistic margin three years out, 10%?
So to we'd need a scenario where companies are spending a collective 300B annually on AI (believable) but ALSO that these companies jack up their margins WITHOUT companies switching to the cheaper open-source models (even when there's a $300B incentive to do so).