Another big thing is govt aligned sponsorships are not for software but for generating jobs/work within their borders.

> About their ELO ratings from their own website:

> A field-relative rating calculated within ChessBench. It compares performance among the tested models and is not a direct equivalent of a human chess rating.

I am around 1600 elo in over the board I can mop up Astra Fable etc even if I give them literal infinite time and all the subagents and internet access..

Please folks at least use your AIs to read stuff before making claims.

AI is not GM level, it's not even 1600, I am 1600 by using memorized openings people frequently fall for with very basic intuitions.

A GM is 2600 they can beat me in under 20 moves...

Why do I even scroll through this website. For a moment I truly felt fooled, but then I read like a human should.

Maybe I should stop doing that will be a happier life, don't think just believe in the AGI.

Chess is a deterministic game won by a combination of known movesets and constrained multi-level forward search.

LLMs do neither of these things. They don't reproduce training data exactly, their next response is more 'inspired by' prompts and its own memory than produced deterministically, and they don't have the capability to do general forward search on their own.

So when you ask an LLM to play chess you're getting the equivalent of a very compressed and lossy JPEG of chess rules and strategies with added per-turn random noise.

They also don't have the ability to design their own chess engine, although it would be interesting to see what happens if you ask for one.

Monte Carlo Tree Search is stochastic.

I, as a human AGI, would jever just forget and remove a piece from the board from one turn to the next.

I mean we've done all this before in a task-specific fashion. It's useful to know that LLMs haven't managed to do that in the process of learning to represent the entire text on the web. On the other hand they have gotten say very good at machine translation without being trained exclusively (and I select the preceding word carefully) on machine translation.

Edit: I'm saying this because there is this idea expressed by e.g. Ilya Sutskever, that in order to predict the next token accurately an LLM has to learn something about all of underlying reality. See for example this interview with Dwarkesh:

https://x.com/biobootloader/status/1640512444958396416

Where Sutskever claims that "Predicting the next token well means you understand the underlying reality that led to the creation of that token".

If that were true, we should have seen LLMs play good chess by now. There is a huge amount of data on playing chess floating around on the web in the form of algebraic chess notation and if LLMs were capable of learning the "underlying reality" of chess, they would already have. They haven't. Because they can't. What Sutskever is saying flies in the face of literally hundreds of years of statistical modelling, which is to say, building predictive models that, very explicitly, do not have to understand any "underlying reality" and only have to be good at modelling a dataset.

Not at all. LLMs learn by imbibing a mass of relationships as isolated fragments of information. There is a certain amount of sorting and indexing that happens during the training phase. There is also a certain amount of compute executed on these relationships during inference. LLMs can model processes that fit within the compute budget. Language translation works well because language is lookup-heavy while being light on compute.

Chess is a compute heavy game of finding the best move out of many possibilities with wide variation in the quality of each move. Humans cut through the compute requirements by reinforcement and learning intuition. LLMs don't get reinforcement on chess so they must compute during inference a unified model of chess. Developing a strong model of chess from raw fragments of information is simply not in their compute budget.

You gotta be careful how you use the word "relation" here because there's an informal meaning (I'm related to my cousin) and a more strict, formal meaning, that is used in computer science e.g. in the "Relational Calculus" etc. In the formal sense, the one relation that LLMs learn during training is the co-occurrence of tokens in a corpus of text, what's called more technically a "collocation" relation. Nothing says that this is enough to play chess, so I'm indeed doubtful that they can.

That's a separate question from whether an LLM (unaided by a C complier) can learn to play chess as well as human (also unaided by a C compiler). Certainly humans can't become grandmasters only by reading chess transcripts on the web, and certainly humans require many "thinking tokens" during a game to play effectively. Do you know for sure that a transformer can't reach grandmaster level if it is allowed to learn by playing games (as humans do) and is given a sufficient number of thinking tokens during the game? It seems near certain that they could, if someone wanted to spend the money (and I don't see why anyone would.)

Yes, I do mean that the LLM's weights are set so that it will execute minimax or MCTS when it needs to. That has nothing to do with whether humans can do the same or not.

I don't disagree that a Transformer could learn to play chess if it was explicitly trained to do that. My argument is that LLMs, trained to predict the next token, have not learned to play chess. That's LLMs, not Transformers.

Just to make sure this is not taken as splitting hairs, the point is that there's all sorts of claims made about what LLMs learn when they train on text. For example, there was a claim by Sundar Pichai that one of their models had learned to translate Bengali without explicitly being trained to do so. It later emerged that Bengali was indeed included in the model's training set [1]. It's not clear whether that included parallel texts, e.g. between Begnali and English or another intermediary language, in any case Sundar Pichai's claim was that the ability to translate Bengali was "emergent".

So I'm interested in understanding the extent to which these "emergent" abilities are real or not. With chess, given the amount of textual data tracing games that floats about on the open internet, I would totally except some ability to play chess to "emerge". Maybe the reported 700-800 ELO level is even that sort of ability. Maybe we should only expect LLMs to learn to play at the level of an untrained, casual player. Maybe not. I have no idea.

On the other hand, the fact they keep making elementary mistakes like illegal moves must be taken to mean that, so far, LLMs haven't learned to play chess.

__________________

[1] https://www.buzzfeednews.com/article/pranavdixit/google-60-m...

Sarcasm aside, I think this is an easy cognitive trap to fall into. It does sometimes feel like the LLM must have some world model because it converses somewhat coherently. Examples like this failure to understand chess, or to count the number of Rs in "strawberry", seem difficult to explain if the models are intelligent. But that doesn't stop people believing they are anyway. I think there must be something about the conversational interface that fools us easily. I wonder if people trained in interrogation techniques are also fooled?

This is unfair to HN readers all of whom but one did not post the comment you replied to. You can't just tar everyone with the same brush. There are thousands (hundreds of thousands?) of users on this site.

Not everything I have the time and energy to reply to. This chess one is just ridiculous claims on top of ridiculous claims all the way and 0 push back in the comments except mine.

I don't even know if there is critical thought or we believe what we read/shared/etc

And there's plenty of pushback on here about the chess thing besides your very valid points.

EDIT: anyway if I can offer a bit of unsolicited advice, it won't do you or anyone any good to accuse everyone who doesn't agree with you of laziness, even if you can see e.g. they haven't really read an article. Just say the thing you wan to say and let them figure it out. Most people will appreciate that much better and you will feel better about yourself for acting like a mature adult.

It's even in the site guidelines:

Please don't comment on whether someone read an article. "Did you even read the article? It mentions that" can be shortened to "The article mentions that".

But despite that you aren't wrong and the only reason I even visit this website is because people sometimes did/do take time to reflect on things based on their experience and knowledge.

And in hindsight pointing out that hn has issues wasn't even the point but I feel frustrated when everyone is readily agreeing to things on here without reading. When that in this moment feels like the one thing that separates humans from machines that we get to think and learn.

I possibly should just drop reading this place until we have most noisy people go away. I have for one tried to always only comment on things where I could be a value add, this one does feel like I could I have done better.

In the moment I probably thought if they are GM level and I can beat them, is this some interesting find, my disappointment honestly led me to making a rather incorrect call on this one.

Either way I still do think HN as a whole has devolved into mindless herd follower mindset, I can point to more than a few posts that just say adopt the hacker mindset aka move fast don't care about the consequences.

And I for one find this laughable even though that's the reality of my job/work as well.

Not to disappoint you but I don't think they ever will. HN is free to join and use so people will join and use it and say whatever they want to say whether it makes sense or not. Filtering out noise is a useful skill to have especially since one can't block users or mute conversations and so on.

>> In the moment I probably thought if they are GM level and I can beat them, is this some interesting find, my disappointment honestly led me to making a rather incorrect call on this one.

Sorry, I didn't get this? What was the incorrect call you made?

Talking about people's inability to read rather than just pointing out that the article pointed at something else.

I have been proved wrong I should have quit while I was ahead. /s

But given how easily I can crush them and how often they want to make illegal moves (btw above bench seems to use a harness that pokea the model until it gives valid moves).

I would rate them around 500-800 big range but at that level it's all about if the model can recall an opening or not. If it plays good first 4-8 moves the person on the end will fumble for certain and they win.

I can play good/best moves till 14-15 moves if I remember the lines and find someone who falls for it.

If you could give them the lines as prompts like the best 20-30 openings then they will be around 700-800.

700 is around the rating for a human who doesn't know the tricks but can do bare minimum calculations and understands the rules thoroughly.

Anyone who casually plays on a regular basis can beat them more often than they lose. As you said if you just know the core openings (and end games, both of which you can get a handle on with modest effort) you will generally win.

Edit: reminder we had computers beating the best players in the world literally decades ago. LLM’s are remarkable tools but the current promises and expectations are ridiculous

They played 4...c6, followed by 5...Nc6, somehow forgetting about the pawn the just put on c6. (My move in between was 5. Nc3, and apparently they were trying to mirror me.)

In some ways this is reflective of the AI experience at large, sometimes shockingly competent but then also sometimes ludicrously incompetent.

Because other HN bring in their own experience telling us what is real and what is BS. Maybe next time it will be someone else with experience in something else that will call out BS and you will see it. I didn’t really think LLM:s are any near good in chess but I don’t play chess so don’t know what 1600 means. So you helped me by calling BS.

Don't use the word infinite in any CS claims. They can recreate or approximate monte Carlo tree search and it technically is still a correct solution in your framing of the problem so long they defeat you.

I don't believe this.

You refer to "subagents", so this is not just an LLM but an LLM with some kind of agentic harness. Any reasonable harness and prompt, given internet access and appropriately prompted to succeed on this task, is more than capable of firing up Lichess or chess.com and relaying moves back to you. The free levels will be enough to beat you.

A frontier model can also likely one shot a chess engine that plays at your level, again if given an environment in which it can do that.

I completely believe the LLM on its own can't play a full game of chess at your level. Though I'd bet that with enough reinforcement learning it is possible to train a pure transformer architecture to do that. We just don't do it because there are other approaches that play chess much better.

This argument is fundamentally incompatible with all the breathless rhetoric about "AGI" coming from the providers' general direction.

So many commenters here see it as their ... duty? to argue against the most optimistic/unhinged (take your pick) arguments from "the other side" and then treat everybody who disagrees as a shill or an idiot.

Why is "being good at chess" a proxy for whatever AGI strawmen you want to argue against?

Maybe step back from your black-and-white ledge and think about discussing what's actually under discussion? For example, why or why not would an LLM be good at chess? Will they be good at chess? What technical limitations might preclude that?

The labs frequently apply their raw models to problems that do not make economic sense for their customers but that demonstrate the power and capability of their systems. These experiments can cost millions of dollars. That's not customer-shaped.

They're not going to give you access to that. It's not a product. The government might have an interest in this, but that's not something you'd be privileged to know about.

And when these labs do develop "AGI", they more than likely won't be selling it to end users. They've pretty much already said this.

>If they cared to have it perform well in chess games, you'd see a different shape and behavior.

So the things they claim are on the verge of AGI actually aren’t? They need to be trained for specific tasks?

It can’t even count the R’s in strawberry

It can’t even add numbers

It can’t even solve a millennium puzzle

It’s not even a chess GM

It’s not even beyond human capability in Go

It can’t even drive a car

It can’t even self replicate

It can’t even build weapons

It doesn’t even have feelings

So how could someone conceivably convince everyone that some system is AGI when there are still tasks that some human or group of humans can do that the system cannot?

This will only happen, in my opinion, when the model/system can self-improve at a rate that scares people.

One, you're not addressing what I wrote above and two, yes, that's absolutely correct. Doing X doesn't qualify something as AGI. If you can't X you can't be AGI. The inverse doesn't hold though.

Notably, if you have to retrain the model in order to X then it can't possibly be AGI since if it were _general_ it would be capable of figuring X out on its own having never seen it before.

Even if they solve 99% of whatever problems LLMs have, the 1% will remain the goal post, forever.

Until you get RFC-whatever from some standards body that defines what an AGI system is, it’s pointless to argue about whether something fits your own personal definition or not.

And for what it’s worth I just watched GitHub Copilot figure something out. So your definition is once again lacking.

However there are plenty of disqualifiers that are more or less universally accepted (ie the negative). In the above case it is literally by definition. Something cannot be termed general if it is incapable of generalizing.

Appealing to a standards body won't do you any good here. Those are composed of people. They exist to facilitate wide scale coordination. Their documents aren't always widely accepted. They aren't the arbiters of truth.

Which is exactly the point I’ve made repeatedly, there will always be something that they cannot do, and thus there will never be AGI. There will always be a long tail of capabilities that whatever system is created doesn’t have, and a long line of social media commenters eager to list them.

An AI controlled robot will be standing over the cooling corpse of the last human who will die certain that it wasn’t done by AGI.

Delusion runs deep in HN circles.

I say that as someone heavily invested in AI startups and projects and as someone working in the field.

I think most people on HN should touch grass and find real human contact. Lmao

Incredible reasoning all around here.

And no an AGI system doesn't need to play chess on a certain level to be disruptive to you and me and whole industries. It only needs to be as good as a person and cheaper.

Just because you define AGI as something it doesn't has to be,doesn't mean i need to touch grass.

This chess comparision is one of the most ignorant and stupid arguments i have heard after the parrot thing

That doesn't mean it has to be extraordinary at these things. But to be AGI, it has to have some level of competency when used on problems outside its training set. In particular, it the LLMs were to install a known chess engine and run that to get the moves when asked to play chess, that would qualify for more AGI-like behavior. But really, chess is such a simplistic game that they should be able to do decently well at it even without even needing that. At the very least, they should be able to consistently play without making illegal moves - something that many 7-year olds manage quite well.

I’m sure these models know the rules and can explain them when prompted, but that doesn’t seem to be the way they actually complete this task. Will they get there? Maybe

Also AI bros: LLM can’t beat an avg chess player. But that doesn’t mean anything. It doesn’t count

So we have a situation where very powerful and influential people are saying we will have AGI in 6 months (if we don’t already), yet the facts on the ground are so clearly pointing in the opposite direction.

>GPT-6 Astra xHigh: 0.06% rejected moves

I could see myself messing up something at some point if the board is complicated enough and trying an illegal move, perhaps if a piece somewhere would attack my king if I moved another piece. Even through I do know the rules of chess, and I have played a few games once every so often.

And the stuff i'm using LLMs daily is just fake?

I see i see. I will see myself out of this weird discussion while I let an LLM continue doing a lot of interesting things.

No, because we can, in fact, generally read the rules of a game and then follow them. It's actually a hobby for many of us.

> And the stuff i'm using LLMs daily is just fake?

This misses the point completely.

How many times do you think chess.com prevents illegal moves from being executed? Even Super GM's fall for mate-in-1's occasionally, which is functionally equivalent to missing a pin or a check. This idea that LLMs failing to only ever make legal moves undermines their intelligence doesn't pass the smell test.

It only matters if you are claiming it to be general purpose.

If you admit that it's just a collection of narrow capabilities - whose strength is mostly confined to the 1000 or so RL environments it was post-trained in, then there is of course no expectation of it being general purpose.

The AI companies seem to heavily want you to believe it is some some near human level general intelligence, so therefore pointing out all the things it can't do is very relevant.

Because we want to use this as a replacement for humans, and the average human can learn the rules of chess without needing to see the rules explained hundreds of thousands of times in millions of games.

So, yeah, it matters if a model has millions of examples of something in its training set and still cannot follow the rules.

Okay, lets go with that: it's the "shown the rules" bit that we are arguing about.

The argument is that a human may play maybe a dozen games after learning the rules, after which they won't be inadvertently attempting illegal moves. What we are observing with SOTA models is that, even after seeing millions of chess rules, rulebooks, actual games, etc, they still attempt illegal moves.

This does not point to generalisable and adaptable intelligence, such as we see in the average human.

See my comment here for more: https://news.ycombinator.com/item?id=49725306

HOW you do it makes a big difference in how you should assess the capability of the thing doing it. Stockfish will trounce any LLM, and any human, at chess, so should we say that Stockfish is smarter than both?

The claim here is not about intelligence, it is about generality. There's no doubt for me the LLMs are intelligent.

Sure they could, but that's irrelevant.

A chess position is just a matter of remembering what piece number is on each square - just a list of 64 numbers. A trained model may store a trillion numbers (weights). It could store a TON of chess positions if it needed to.

However, that's not how LLMs work. They don't memorize inputs - they predict them, based on discovering predictive patterns, and those predictive patterns are not input patterns (e.g. board positions). They are deep patterns (maybe 100 layers of abstraction removed from the input), representing partial inputs, generalized across many training samples.

> Don't you know the legend about rice grains on a chess board?

Sure, but this has nothing to do with chess, and nothing to do with how many games were in the LLM's training data.

> The claim here is not about intelligence, it is about generality. There's no doubt for me the LLMs are intelligent.

Intelligent humans created the training data, and the LLM attempts to predict (copy) the training data, so of course it looks intelligent. If I say "E=mc^2", does that make you think I am Einstein?

> don't memorize inputs - they predict them

I feel some tension here.

> rice grains on a chess board? Sure, but this has nothing to do with chess, and nothing to do with how many games were in the LLM's training data.

> just a list of 64 numbers

> remember even a few positions? Sure they could, but that's irrelevant.

I don't think you do. Or rather you do know the legend but for some funny reason seem to be unable to apply its lesson here, because you are talking about enormous terabytes of training data.

> Intelligent humans created the training data, and the LLM attempts to predict (copy) the training data, so of course it looks intelligent.

If for you it is about intelligence, I am out of this discussion.

If not, then what are you talking about ?

If yes, then what is the relevance to an LLM playing chess ?

> remember even a few positions? Sure they could

A rough estimate of number of positions across all X move games is X^10. For 15 moves it is hopeless to remember even a relatively small part of them. Typical game has 40 turns, 1 move per player, so 80 moves.

2) An LLM is not going to memorize vs generalize when there is no training pressure to do so. You might expect it to memorize book openings that occur over and over in the training data, but not some random non-celebrity game that occurs once in the Lichess dataset and is never again referred to.

> They can't possibly remember even a few positions. Don't you know the legend about rice grains on a chess board?

If the wise man was a bit wiser, he'd have asked for his rice on a snakes & ladders board (100 squares, not 64) and would have had 2^36 more rice, which is equally irrelevant.

We all know AI can code, but the question it all stemmed from what if it's AGI or GM level in chess on it's own.

You can't just back pedal from the statement that apparently being able to code a chess engine is the same as being good at chess.

I can write a chess engine that beats Magnus Carlson without AI that alone neither makes me GM level or AGI or any of the other claims the above comments seem to be making?

We know no such thing. LLMs are quite bad at generating code, worse than any capable human.

He definitely needs to touch grass.

Is code omnipotent, I have been in software all my life and I would hard agree here.

Sure stuff LLMs can do with being good at parts of code reproduction is incredible. And honestly it's the new way to do a lot of things but I have not see an iota of proof that it can scale across the board.

For instance Maths is just code with different symbols and slightly less universally legible concepts.

AI is the best invention at figuring out or walking the search space and directionally doing logically computation over general software adjacent stuff.

But that's it, I am certain a bunch of companies will make a lot of money despite no AGI.

I think people either don't understand AGI or don't understand how real world works.

Until an LLM can bow it's head take responsibility for mistakes made and ensure they aren't repeated again with 100% confidence to the leadership it's inarguably a tool a rather questionable one at that.

So.. like chess?

Anyway, do you have any prediction on what LLM's can or can't do in a few years?

> The system is built and delivered as it is because that's how the providers make the most money. If they cared to have it perform well in chess games, you'd see a different shape and behavior.

These comments indicate a complete failure to understand the technology.

I won't respond again.

I see no reason meta quest headsets couldn't do it next gen.

My personal ambition is to see VR/AR be cheap enough.

Both side get to feel like they got more out of the initial investment.

1:3 2:3

Everyone wins inside their heads and its nice to live life where everyone wins...

Also easier to justify charity when you think you are winning..

I feel like it's pretty good for what it is.

But otherwise you are right, this is done because it makes everyone donating feel like a winner other than few cynics like me. :D But that said I do not mind these drives.

They might in the future but future is in the future ofc

Edit: to be clear I think it's ridiculous they don't but from a company's stand point it doesn't make much sense

Unfortunately, the growth potential that software industry generally provides has generally prevented governments for creating and enforcing similar laws for software products and such.

I personally think countries should start sanctioning and international warrants for all large investors and senior executives in a company if something like this happens.

But it won't happen unless US changes their stance on companies being liability shiled.

Sure it should require very exceptional cases but I know companies whose negligence has cost lives of hundreds of people, the reason they and all their investors weren't jailed is the real scam of the democraciez we have gifted ourselves.

For anyone curious the big thing with jj is you can undo. Basically if you were running a rebase and bungled a commit sha or branch name, no worries undo your way out, if you forgot to push but abandoned a commit not issues undo it. Same with all jj commands.

Delta db takes this a step further and everything you edit is a change so your actual edits are also all saved.

I find non destructive-ness for as long as possible approach very nice, since there aren't as many constraints on us not being able to do it nowadays, we had limited storage and stuff in 2000s perhaps but in 2026 even with the killer SSD prices it's fairly cheap to buy more SSD, we can definitely waste 100MB per project storing redundant history we may not ever need.

Same way in photoshop non destructive editing took over, especially useful for humans, but works for AI and LLMs too, they can undo their mistakes as well.

You worked on multiple things, you sync with the upstream, then boom now you have conflicts in one or more of your local branches.

Instead of immediately presenting you a conflict resolution view (usually 3 pane vimdiff) and you'll have to resolve at the sync time, the commits are simply marked as "have conflict", and you can continue working on other things until you have time to deal with it.

And for a chain of commits that all have conflicts, chances are fixing the first conflict automatically resolves all other conflicts in the chain, thanks to the auto-rebasing feature.

What happens if one is unlucky and the conflicts are not automatically resolved from changes to the first commit? Does that potentially imply more work overall than if one was using a VCS that forced conflict resolution earlier, like git?

The main advantages are that you can do the changes piecemeal, in any order, at any time, and by jumping around in the commit history however you need or want to.

Sometimes the best way to solve a conflict is by changing something before the conflict, or by moving pieces from one commit to another. Git doesn’t allow this, you can only do: fully fix commit 1, fully fix commit 2, fully fix commit 3, and so on.

Why is git's "forced conflict resolution" better, assuming the same workflow (local change+sync upstream) would produce the same # of conflicted commits?

2. Yeah submodules aren't supported but since you can use jj in a git repo you can commit git stuff via git like submodules and then enjoy the world of jj. Since submodule interactions should be rare imho. I use it for a bunch of projects with submodules.

For instance, if you are interactively using `git bisect` and you mark commits as good or bad, and you accidentally mark a commit incorrectly, you have to do something like:

    git bisect log > /path/to/file.txt
    git bisect reset
    # modify file.txt to "look correct"
    git bisect replay /path/to/file.txt

Another example is when you do something like screw up an interactive rebase. Let's say you rebase 20 commits and then you get a conflict on commit 8. You fix the file conflicts, and continue. You accidentally solve the merge incorrectly, continue and get another conflict -- but only realize your mistake after you start solving it. The reflog can't save you here. You have to completely abandon the rebase and start over. (This specific example might be handled better these days).

I think the biggest thing about `jj undo` is that it works everywhere. You can undo rebases, merges, conflict resolutions, copies, deletions, whatever. The secret behind it all is that internally, jj is architected in a way where implementing a feature looks like you are working with a transactional database. You actually have `begin_transaction()` and `commit()` methods in the codebase that will make changes to the commit graph visible in an atomic way. When a command like `jj rebase` happens, all of the changes it makes are inside a transaction and committed at once. Every operation in the repo is a transaction, and it all goes into a log, which records the effects of a transaction -- very much like a database system!

So "undo" just means "undo the effects recorded in a transaction" and that is about all. And so it works for everything! And this design is very easy to intuitively understand and program against, as a maintainer, along with our other high level internal APIs. Any developer can easily write code that Just Does The Right Thing and the user can undo it and it's no big deal. When I develop and work on Jujutsu myself -- like I'm actively developing new features or prototyping ideas -- I almost always _use my own jj repository_ as a test repo while testing my builds.

In contrast, Git does not have one unified "transactional" layer for things like this. But not all is lost, there has been work on 'git undo' and it was implemented by... Someone who is now a Jujutsu maintainer[2]. :)

[1] Technically we do not yet have "step by step" bisect with good/bad yet (only "automatic" bisect that is one-shot), so that is something Git can do we can't do at all right now, but bear with me. :')

[2] https://blog.waleedkhan.name/git-undo/

Since jj snapshots the commit you're editing a lot, you can often even jj undo an accidental edit to a just-created file that in git would have still been uncommitted.

¹ https://drive.google.com/file/d/1dVzug1lHoOxdbFu8gcCJCu-G_uV...

Points of no return.

I think this misses the actual limits here.

The problem isn't demand it's, "how much people are willing to spend on it".

Cheap AI has to be served on cheap compute, and if inference gets cheap enough to unlock massive usage numbers, by definition it also doesn't require anywhere near as much infrastructure per unit of demand.

Take DeepSeek serving ~100T tokens/day, depending on workload and utilization, you're potentially talking about only a few thousand last-gen GPUs. With current-gen GPUs maybe closer to ~1,000, and with Rubin even fewer I will be damned if I could get my hands on one.

That's the part I think people are missing when they extrapolate token demand into enormous infrastructure or AI revenue.

Yes usage will explode. But if the cost per unit collapses, the revenue doesn't necessarily go up with it.

You can't simultaneously argue that intelligence becomes so cheap that everyone uses enormous amounts of it, while also assuming customers will somehow spend trillions of dollars a year consuming it.

There is no obvious $1T customer-facing AI revenue number at the end of this rainbow in the short/medium term.

The average person isn't going to spend anything remotely comparable to what they spend on a car every year for an AI service. Even businesses have budgets now, huge demand doesn't matter if the willingness to pay isn't there.

The only path I can see to numbers like that is AI consuming existing business domains, even then it's very thin.

Say SaaS + legal + consulting + BPO + various other service industries collectively represent something like $10-20T globally.

Even if AI eventually replaces an enormous portion of that, it's probably not doing so at the same price. Why would customers switch otherwise?

Either the AI product has to be dramatically better, which is difficult for mature workflows, or dramatically cheaper which is much more plausible.

If it replaces $10-20T of existing services at roughly 1/10th or 1/100th (more likely) the cost, then you're looking at maybe a ~$1T AI revenue opportunity after replacing an absurdly large fraction of the existing service economy.

Who are now unemployed and can't pay for shit.

And that's before competition.

I think it's crazy to assume AI companies won't compete aggressively on price. As capabilities diffuse, smaller models catch up, inference hits pareto frontier the open-source alternatives have already improved and caught up, margins on routine intelligence should compress "hard" (emphasis on "hard").

We've already seen how difficult adoption can be even when the technology looks impressive on paper. Cheap here means 100x cheaper for 10x more demand that's a net 10x loss before any software or hardware optimizations.

So yes, I completely agree that cheap intelligence can bring an enormous amount of new usage.

"I just don't think usage means revenue." (you can plaster it on a wall if you want to, "usage doesn't mean revenue", if you want to find that out I have foss software bridge to sell)

The PC analogy actually reinforces this if you really think about it. Compute became "vastly more useful" while the cost per unit of compute collapsed. Society captured enormous value, but all computer companies are literal failing giants without the AI hype. Value got caught by people who provided productionization.

Now if people expect AI to self productize itself I am happy to tell your try it. We all saw how OpenAI fell behind Anthropic because they thought that would work...

Google couldn't productize the search, instead they sold the eye balls and web-real-estate. Maybe that's the AI business model, but that's not $1T worth given you need to unglue people from other stuff.

Unless we get something approaching genuine ASI producing so much additional economic value that entirely new trillions, I don't see a path to $1-2T in direct AI revenue from customers.

The market simply can't absorb that level of spending.

Demand can be effectively infinite at the right price. But I think people are delusional on HN and SF if they think that number is in Trillions like the investments seem to suggest.

I am not saying Nvidia will fall tomorrow but someone will have to pull the breaks before this car goes to hell.

Lol its not even that - its what can I do with it? Which eventually has to show up somehow in the financials - from a macroeconomic stand point. Software production is microeconomic.

This is the right way to look at it, but a few of your estimates are a bit off. AI is being sold as an accelerator (or, if you're in a dystopian mood, total replacement) of knowledge workers. Currently knowledge worker salaries are $50 - 70 trillion a year globally, $10 - 11T in the US alone: https://gist.github.com/danielmiessler/2dc039762a202b083753b...

> Even if AI eventually replaces an enormous portion of that, it's probably not doing so at the same price. Why would customers switch otherwise?

AI is wayyyyyyy easier to wrangle than humans; no sick leaves, health insurance, perks, HR issues... heck they don't even sleep! If companies could replace us with robots, they would do so in a heartbeat. Capitalism!

So in a "what the market will bear" sense, we have an upper bound on the TAM. Indeed, I expect this is where Anthropic's ridiculous "$30 trillion" number is coming from... except now we see how they came to it.

If AI makes workers even 1% more efficient, that's a $500 - 700 billion value annually. In reality AI makes workers way more efficient (studies from the ancient era of 2024 showed about a 30% boost) so AI companies could realistically charge that much more. But then all the other factors you mentioned -- smaller models, competition, self-hosting, etc -- come into play, which put a downward pressure on revenues.

It's impossible to predict how these dynamics will play out, but the numbers involved are astronomical. This is why everyone from the frontier labs to Big Tech to VCs to nation states are scrambling to get in on it.

And which company profits the most from exponential growth in token generation?

Here we go, I just drew a "circle" for these naysayers.