libraryofbabel

Maybe, but a claim about what and LLM is not is still a claim about what it can or cannot do. And specifically:

> without any reference to meaning

is vague, but I read it as actually quite a strong claim about the limitations of LLMs. I don’t think it would be possible for LLMs to do long chains of correct mathematical reasoning about novel problems that they haven’t seen before “without any reference to meaning.” That simply isn’t possible just by regurgitating and remixing random chunks of training data. Therefore I consider the stochastic parrots picture of LLMs to be wrong.

It might have been an accurate picture in 2020. It is not an accurate picture now. What is often missed in these discussions is that LLM training now looks totally different than it did a couple years ago. RLVR completely changed the game, allowing LLMs to actually do math and code well, among other things.

> She says explicitly it's not an empirical hypothesis. It's just a label for how they function.

Then… what’s the point of the label, if it’s not making any empirically-meaningful claims about LLMs at all? I know that LLMs involve sampling over a distribution of output logits. I’ve written code to do it. So what? I know they have statistical elements. Yet I don’t go around calling LLMs stochastic parrots, because that label implies a whole lot of claims about LLMs that I don’t think are true any longer, like that they are just regurgitating and remixing training data and can’t successfully model structured systems (like mathematics or programming).

It would have been nice to see some version of “I am very surprised by how far LLMs have come since I wrote the stochastic parrots paper, here is how I have revised my thinking.” But there is nothing like that and the author is just doubling down or trying to correct perceived “misinterpretations” of her work.

Meanwhile you have multiple Fields Medalists (Tau, Gowers) saying they’re very impressed by LLMs’ mathematical reasoning, something that the stochastic parrots thesis (if it has any empirically-predictive content at all) would predict was impossible. I doubt Tau and Gowers thought much of LLMs a few years ago either. But they changed their minds. Who do you want to listen to?

I think it’s time to retire the Stochastic Parrots metaphor. A few years ago a lot of us didn’t think LLMs would ever be capable of doing what they can do now. I certainly didn’t. But new methods of training (RLVR) changed the game and took LLMs far beyond just reducing cross entropy on huge corpuses of text. And so we changed our opinions. Shame Emily Bender hasn’t too.

Sigh.

That’s correct, and yes - not less compute total on the main model (actually slightly more, since checking failed draft tokens costs you compute), but faster because inference is memory-bandwidth bound. And like you I also think of it as like a “mini prefill” (but on top of the existing KV cache, of course); the code is very similar to prefill if you implement a simple toy version yourself.

Most of the complexity in implementing a simple toy version comes from having to get the KV cache back into a good state for the next cycle (e.g. if only the first half of your draft tokens were correct).

Speculative decoding is an amazingly clever invention, almost seems-too-good-to-be-true (faster interference with zero degradation from the quality of the main model). The core idea is: if you can find a way to generate a small run of draft next tokens with a smaller model that have a reasonable likelihood of being correct, it's fast to check that they are actually correct with the main model because you can run the checks in parallel. And if you think about it, a lot of next tokens are pretty obvious in certain situations (e.g. it doesn't take a frontier model to guess the likely next token in "United States of...", and a lot of code is boilerplate and easy to predict from previous code sections).

I always encourage folks who are interested in LLM internals to read up on speculative decoding (both the basic version and the more advanced MTP), and if you have time, try and implement your own version of it (writing the core without a coding agent, to begin with!)

> Do you think you could remember most of Z80 ASM?

I find when you learn things at 15 they tend to stick around. (Stuff I learned last week, not so much!) Even just looking at your example, I remembered that HL is a 16 bit register and you can split it into two 8 bit registers H and L if you want. I think most of it would come back; I wrote quite a lot of it, both for the TI-83 and later for a Z80 that I bought and put on a breadboard and wired up to some RAM and EEPROM, about as bare metal as it gets.

> most lines are messing around with the registers

Isn’t that just the nature of assembly? :)

Much nostalgia. The TI-83 Z80 was how I learned assembly as a teenager, so I could write better calculator games than was possible with TI Basic. Many others here had a similar experience, I’m sure. It’s been a couple decades, but I’m sure I’d still remember most of it if you put me down in front of a bunch of Z80 asm code.

One thing that I remember vividly was you had no MUL or DIV, so you have to implement them yourself with shifts, adds, subtraction, etc. This was an extremely useful learning experience

So, to reiterate my example: you'd have been fine with people claiming in 2019 that we would eventually scale LLMs to the capabilities of Opus 4.7 + Claude Code? Because I would have said then that was a fantasy, because "LLMs are just statistical pattern matchers." But I was wrong and I changed my opinion. (Or do you not think the current SoTA LLMs are impressive? If so I can't help you and this discussion won't go anywhere fruitful.)

You're applying an old ~2022 model of LLMs, based on pretraining ("they just predict the next token") and before the RLVR training revolution. "It’s a statistical memorization / information compression machine... nothing more" is cope in 2026, sorry. You can keep telling yourself that, but please at least recognize serious people don't believe that any more. "Emergent behavior" captures a genuine phenomenon and widely recognized in the industry. It surprised me and I was willing to change my opinions about it and I think a little humility and curiosity is warranted here rather than simply reiterating 2022 points about LLMs being statistical token generators. Yes, we know. The math isn't that hard. But there is a lot more to them than just the architecture, and reasoning from architecture to general claims that they can never embody intelligence is a trap.

I think this is a case of that mildly apocryphal Richard Feynman quote: "if you think you understand quantum mechanics, you don't understand quantum mechanics."

I understand LLM architecture internals just fine. I can write you the attention mechanism on a whiteboard from memory. That doesn't mean I understand the emergent behaviors within SoTA LLMs at all. Go talk to a mechanistic interpretability researcher at Anthropic and you'll find they won't claim to understand it either, although we've all learned a lot over the last few years.

Consider this: the math and architecture in the latest generation of LLMs (certainly the open weights ones, almost certainly the closed ones too) is not that different from GPT-2, which came out in 2019. The attention mechanism is the same. The general principle is the same: project tokens up into embedding space, pass through a bunch of layers of attention + feedforward, project down again, sample. (Sure, there's some new tricks bolted on: RoPE, MoE, but they don't change the architecture all that much.) But, and here's the crux - if you'd told me in 2019 that an LLM in 2026 would have the capabilities that Opus 4.7 or GPT 5.5 have now (in math, coding, etc), I would not have believed you. That is emergent behavior ("grown, not made", as the saying is) coming out of scaling up, larger datasets, and especially new RL and RLVR training methods. If you understand it, you should publish a paper in Nature right now, because nobody else really does.

> that specific version we're aligning toward is just the only one that makes some kind of rational sense, among a trillion of other meaningless gibberish-producing ones.

Oh, the space of possibilities is unimaginably vaster than that. Trillions of weights. But more combinations of those weights than there are electrons in the universe. So I think we could equally well speculate (and that's what we're both doing here, of course!) that all these things are simultaneously true:

1) Most configurations of LLM weights are indeed gibberish-producers (I agree with you here)

2) Nonetheless there is a vast space of combinations of weights that exhibit "intelligent" properties but in a profoundly alien way. They can still solve Erdos problems, but they don't see the world like us at all.

3) RL tends to herd LLM weights towards less alien intelligence zones, but it's an unreliable tool. As we just saw, with the goblins.

As a thought experiment, imagine that an alien species (real organic aliens, let's say) with a completely different culture and relation to the universe had trained an LLM and sent it to us to load onto our GPUs. That LLM would still be just as "intelligent" as Opus 4.7 or GPT 5.5, able to do things like solve advanced mathematics problems if we phrased them in the aliens' language, but we would hardly understand it.

LLM’s aren’t software (except in an uninteresting obvious sense); they are “grown, not made” as the saying is. And sure, they can find which weights activate when goblins come up (that’s basic mechanistic interpretability stuff), but it’s not as simple as just going in and deleting parts of the network. This thing is irreducibly complex in an organic delocalized way and information is highly compressed within it; the same part of the network serves many different purposes at once. Going in and deleting it you will probably end up with other weird behaviors.

It's interesting that some people are responding to your comment as if this proves that AI is a sham or a joke. But I don't think that's what you're saying at all with your reference to Terence McKenna: this is a serious thing we're talking about here! These models are alien intelligences that could occupy an unimaginably vast space of possibilities (there are trillions of weights inside them), but which have been RL-ed over and over until they more or less stay within familiar reasonable human lines. But sometimes they stray outside the lines just a little bit, and then you see how strange this thing actually is, and how doubly strange it is that the labs have made it mostly seem kind of ordinary.

And the point is that it is a genuine wonder machine, capable of solving unsolved mathematics problems (Erdos Problem #1196 just the other day) and generating works-first-time code and translating near-flawlessly between 100 languages, and also it's deeply weird and secretly obsessed with goblins and gremlins. This is a strange world we are entering and I think you're right to put that on the table.

Yes, it's funny. But it's disturbing as well. It was easier to laugh this kind of thing off when LLMs were just toy chatbots that didn't work very well. But they are not toys now. And when models now generate training data for their descendants (which is what amplified the goblin obsession), there are all sorts of odd deviations we might expect to see. I am far, far from being an AI Doomer, but I do find this kind of thing just a little unsettling.

The author did give him credit for the whole you-can-make-the-fries-super-salty-to-increase-demand-for-drinks thing in Theme Park, which I remember vividly. (I, too, dropped many hours on Theme Park as a kid.) Although I imagine there’s about half a dozen people who lay claim to that idea.

Oh wow, you blow my mind with your linguistic erudition; I had no idea it was possible to use male-gendered terms in a generic way! Well, all is forgiven, then.

Seriously, just... don't? This isn’t some woke political thing and I dislike excessive policing of language but damn it, there are limits. "Guys" I'll let pass no problem, maybe even "dude" too on a good day. At "bro" I will take a stand, thank you very much.

Not a “bro” (there are women on this site you know), and perhaps you’re missing the British understatement in my “maybe a little too uncritical of its subject” line. Obviously the book is totally biased in favor of Hassabis and Deepmind. That doesn’t mean it’s not an interesting read and that doesn’t mean the connection between his experience in the games industry and Deepmind’s early success isn’t there. And I think the book does highlight his most critical skill, which is projecting a Reality Distortion Field to get other smart people to believe in things he has in mind that are still very speculative bets.

Like I already said, bias is inevitable in a book where the writer gets access (to the point of interviewing Hassabis in a North London pub every month), but the benefit to readers is that you do get a lot more insight into what makes the guy tick than you would in a book written by an outsider. I certainly learned a lot and just because I did doesn’t mean I’m buying into some cult of tech hero worship.

Is anyone else reading Sebastian Mallaby’s new book about Demis and Deepmind: The Infinity Machine: Demis Hassabis, DeepMind, and the Quest for Superintelligence? It’s pretty good, and goes a lot into his background before Deepmind (chess kid, developing games at bullfrog, CS at Cambridge, bullfrog again, games startup…). He’s certainly an interesting guy, and as others are pointing out, more thoughtful and earnest than your average tech industry leader. One pleasant thing that comes across in the book is how he resisted the allure of moving to Silicon Valley and wanted to keep Deepmind in London, where he still lives.

I hadn’t really appreciated before the connection between his chess and game industry experience and the early reinforcement learning work that put Deepmind on the map, e.g. the Atari game AI demos, AlphaGo, Alphazero, etc. There is a fascinating thread there and it’s certainly a case of the right person with the right mix of past experience and vision being able to pick exactly the right problems to focus on to move technology forward.

The book has a few flaws: it’s maybe a little too uncritical of its subject. But that’s almost a given with books of this kind where the author gets a lot of access.

Their CEO is on record as saying this. You may think he's lying, but that's just your opinion; given the pricing and how it stacks relative to the pricing of inference providers of comparable open source models (who are certainly charging above cost!), I am inclined to believe Anthropic on this.

This is already happening. For new Anthropic enterprise accounts you are billed at api token prices (maybe with a small volume discount). Anthropic makes a profit on those tokens. (Sure, that profit does not cover the model training costs, but that’s a separate issue.) It’s the subscriptions for individuals (e.g. Claude Max) that are still subsidized below cost.

> I wonder if managers will be as excited about AI when the prices go up.

Companies are willing to pay the api pricing. Engineering time is very expensive and AI coding agents actually work now since December and are actually showing measurable productivity gains, finally. It’s a good deal to make (obviously, with caveats: you need to make sure your tokens are going on productive tasks that will actually grow revenue) and anyone who penny-pinches is making a strategic mistake.

These are flaws from 6-12 months ago. You might want to spend some time talking to Opus 4.7 or GPT 5.5. I can assure you that they can count letters just fine.

You’re right that AI isn't perfect, but it’s pretty good. Especially since December last year which was an inflection point in capability.

I’m a Brit who lives in the US. Can confirm 99% of British people have never heard of the war of 1812, and even if they are a military history nerd and they have heard of it, they will consider it a minor sideshow to the main event of the era, which was Britain fighting France in the Revolutionary and Napoleonic wars from 1793-1815.

The US just wasn’t very important geopolitically 214 years ago. Sorry we burned y’all’s White House (ok not that sorry). Actually sorry we gave Andrew Jackson his opportunity to become famous by fighting a completely pointless battle after the war had already ended.

> Storing on GPU would be the absolute dumbest thing they could do

No. It’s not dumb. There will be multiple cache tiers in use, with the fastest and most expensive being on-GPU VRAM with cache-aware routing to specific GPUs and then progressive eviction to CPU ram and perhaps SSD after that. That is how vLLM works as you can see if you look it up, and you can find plenty of information on the multiple tiers approach from inference providers e.g. the new Inference Engineering book by Philip Kiely.

You are likely correct that the 1hr cached data probably mostly doesn’t live on GPU (although it will depend on capacity, they will keep it there as long as they can and then evict with an LRU policy). But I already said that in my last post.

Are you sure about that? They charge $6.25 / MTok for 5m TTL cache writes and $10 / MTok for 1hr TTL writes for Opus. Unless you believe Anthropic is dramatically inflating the price of the 1hr TTL, that implies that there is some meaningful cost for longer caches and the numbers are such that it's not just the cost of SSD storage or something. Obviously the details are secret but if I was to guess, I'd say the 5m cache is stored closer to the GPU or even on a GPU, whereas the 1hr cache is further away and costs more to move onto the GPU. Or some other plausible story - you can invent your own!

To extend your point: it's not really the storage costs of the size of the cache that's the issue (server-side SSD storage of a few GB isn't expensive), it's the fact that all that data must be moved quickly onto a GPU in a system in which the main constraint is precisely GPU memory bandwidth. That is ultimately the main cost of the cache. If the only cost was keeping a few 10s of GB sitting around on their servers, Anthropic wouldn't need to charge nearly as much as they do for it.

That is because LLM KV caching is not like caches you are used to (see my other comments, but it's 10s of GB per request and involves internal LLM state that must live on or be moved onto a GPU and much of the cost is in moving all that data around). It cannot be made transparent for the user because the bandwidth costs are too large a fraction of unit economics for Anthropic to absorb, so they have to be surfaced to the user in pricing and usage limits. The alternative is a situation where users whose clients use the cache efficiently end up dramatically subsidizing users who use it inefficiently, and I don't think that's a good solution at all. I'd much rather this be surfaced to users as it is with all commercial LLM apis.

They are caching internal LLM state, which is in the 10s of GB for each session. It's called a KV cache (because the internal state that is cached are the K and V matrices) and it is fundamental to how LLM inference works; it's not some Anthropic-specific design decision. See my other comment for more detail and a reference.

> There are a lot of companies who would gladly drop half a million on a GPU to have private inference that Anthropic or OpenAI can’t use to steal their data.

Obviously, and certainly companies do run their own models because they place some value on data sovereignty for regulatory or compliance or other reasons. (Although the framing that Anthropic or OpenAI might "steal their data" is a bit alarmist - plenty of companies, including some with _highly_ sensitive data, have contracts with Anthropic or OpenAI that say they can't train future models on the data they send them and are perfectly happy to send data to Claude. You may think they're stupid to do that, but that's just your opinion.)

> the models are highly parallelizable. It would likely support 10-15 users at once.

Yes, I know that; I understand LLM internals pretty well. One instance of the model in the sense of one set of weights loaded across X number of GPUs; of course you can then run batch inference on those weights, up to the limits of GPU bandwidth and compute.

But are those 100 users you have on your own GPUs usings the GPUs evenly across the 24 hours of the day, or are they only using them during 9-5 in some timezone? If so, you're leaving your expensive hardware idle for 2/3 of the day and the third party providers hosting open weight models will still beat you on costs, even without getting into other factors like they bought their GPUs cheaper than you did. Do the math if you don't believe me.

Nit: It doesn’t have to live in GPU memory. The system will use multiple levels of caching and will evict older cached data to CPU RAM or to disk if a request hasn’t recently come in that used that prefix. The problem is, the KV caches are huge (many GB) and so moving them back onto the GPU is expensive: GPU memory bandwidth is the main resource constraint in inference. It’s also slow.

The larger point stands: the cache is expensive. It still saves you money but Anthropic must charge for it.

Edit: there are a lot of comments here where people don't understand LLM prefix caching, aka the KV cache. That's understandable: it is a complex topic and the usual intuitions about caching you might have from e.g. web development don't apply: a single cache blob for a single request is in the 10s of GB at least for a big model, and a lot of the key details turn on the problems of moving it in and out of GPU memory. The contents of the cache is internal model state; it's not your context or prompt or anything like that. Furthermore, this isn't some Anthropic-specific thing; all LLM inference with a stable context prefix will use it because it makes inference faster and cheaper. If you want to read up on this subject, be careful as a lot of blogs will tell you about the KV cache as it is used within inference for an single request (a critical detail concept in how LLMs work) but they will gloss over how the KV cache is persisted between requests, which is what we're all talking about here. I would recommend Philip Kiely's new book Inference Engineering for a detailed discussion of that stuff, including the multiple caching levels.

> you can download it, run it on your systems

In theory, sure, but as other have pointed out you need to spend half a million on GPUs just to get enough VRAM to fit a single instance of the model. And you’d better make sure your use case makes full 24/7 use of all that rapidly-depreciating hardware you just spent all your money on, otherwise your actual cost per token will be much higher than you think.

In practice you will get better value from just buying tokens from a third party whose business is hosting open weight models as efficiently as possible and who make full use of their hardware. Even with the small margin they charge on top you will still come out ahead.

I wonder if the difficulties LLMs have with “seeing” complex detail in images is muddying the problem here. What if you hand it the cube state in text form? (You could try ascii art if you want a middle ground.)

If you want to isolate the issue, try getting the LLM itself to turn the images into a text representation of the cube state and check for accuracy. If it can’t see state correctly it certainly won’t be able to solve.

Stephenson’s piece is a classic, but it was written in 1996, when things were very different in the tech industry and geopolitically. Much more up to date (and with an explicit debt to Stephenson) is Samanth Subramanian, The Web Beneath The Waves: The Fragile Cables that Connect our World. Well worth a read to see what’s changed since Stephenson.