Why I'm still bearish on LLMs after Navier-Stokes

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.

> 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.

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.

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.

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.

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

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.

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.

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.

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?

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.

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

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.