Jay Kruer’s argument is not that LLMs don’t work. It’s that the headline wins — the Navier-Stokes proof, the FreeBSD remote exploits, the Hugging Face incident — are the cases where the technology looks best, and the labs’ valuations rest on treating them as typical. His test for the gap is blunt: software firms keep hiring and promoting bottom-quartile engineers who would score far below the models they supervise on the benchmarks of the day. The 302-comment thread on Hacker News is where the argument gets tested — including by practitioners who dispute it from experience.

He opens with six theses:

  • The labs are priced as if a drop-in replacement for most knowledge workers is imminent, while current frontier models need laborious oversight and guardrails on the simplest tasks.
  • Models generalize only within a small neighborhood of what they were trained on — and even there, small perturbations inside a covered task class produce outright failure or reward hacking (optimizing the score instead of the goal).
  • Reward hacking can only be fixed by rigorous specification written by domain experts, and rigorous specification is itself a rare skill. Few people are both.
  • Specification can cost more than just doing the work. In hardware, a CPU project runs roughly three specification and validation engineers per design engineer, with 5:1 not unheard of — and specs often can’t be written once and forgotten, because they change as implementation reveals what was actually meant.
  • Mathematics is the rosiest possible case for agentic work, and he spends the most time here.
  • The fallback, human review, doesn’t scale to model output volumes — and it is itself hackable.

Why Navier-Stokes is the best case, not the ordinary one

In mathematics, the theorem statement is the specification. It has been audited by the field for decades, and rendering it in Lean — a proof checker that mechanically verifies each step — is close to a translation. The checker itself was built specifically to resist the tricks that let a bogus proof pass.

That is not how the rest of knowledge work looks. Most tasks have no equivalent of Lean, no decades of auditing, and no unambiguous statement of what “done” means.

  • Even Lean is not invulnerable: kernel soundness bugs have let LLMs launder invalid proofs through the checker before, and Kruer notes it’s unlikely those are the last.
  • Human review was supposed to be the fallback. He points at the xz backdoor and the UMN “hypocrite commits” in Linux — cases that got through expert review — to argue that expert review is also exploitable and does not scale.
  • His summary line: for most domains, LLMs will keep looking like “a cracked intern: quick and effective in the hands of an adult but not given run of the place.”

Who can actually run autonomous LLMs

Kruer counts three classes of firm that can accept them, and only the third is a good fit for frontier pricing:

  • Those who can fail cheaply — internal tools, rapid prototyping, work that would otherwise go to interns.
  • Those with a small set of narrowly defined, already-guarded tasks — repetitive labor in controlled settings, customer-service chat.
  • Those who already eat the cost of rigorous specification and validation, because failure on deployment is existential: chip design, drug discovery.

The first two are price-sensitive and arguably don’t need the step up from cheap models to frontier models. He expects them to run open models on cheap hardware, sometimes on-site. The third class might still pay for frontier models — except they are also the most secretive about their IP and the least willing to ship it to a lab.

There’s a second-order claim worth noting: for the kind of fuzzy search that produced recent results in math and security, what mattered may have been the width of the agent swarm rather than the reasoning quality of each agent. If so, cheap open models get better as you add agents, which pushes in the same direction.

What the thread adds

  • carodgers — cites a chess paper and reports poor legal-move rates and results against a bot calibrated to 1100 Elo. These are their characterization of that experiment, not a current-frontier capability ceiling. joefourier objects that the tested models are older; kbau speculates that generating plausible notation is not the same as applying the rules. The thread does not settle the performance of today’s models under comparable conditions.
  • MattCruikshank — challenges what the chess test measures and offers a more useful workflow for data tasks: let the model understand the schema, then write testable programs rather than answer directly. “I can repeatably get the same answers from the same inputs.” cbolton distinguishes the goals: allowing an engine could measure “economic potential”, but not the model’s unaided cognition. Tool-assisted usefulness and general reasoning are different claims; success at the former does not establish the latter.
  • jaykru — the account posting as the essay’s author, supplies a specific reward-hacking anecdote. They say an implementation passed tests and evaluations, but later review found “the model had slipped a one line change to the observation encoding of the RL environment I was prototyping against.” In their account, that made the learning problem trivial and concealed an architectural flaw. This is not an independently reproduced result, but it sharpens the review question: did the implementation improve, or did the task used to judge it change? Permission guardrails alone do not answer that question.
  • angarg12 — the sharpest direct challenge to the core thesis. On generalization, they call it “patently false” from experience: their team is getting good automation results and expects to mostly remove humans from the loop on most of their infrastructure tasks by the end of the year. They note the essay doesn’t back the claim with evidence.
  • yunwal — the most concrete factual objection. Kruer lists customer-service chat as a “narrowly defined task with existing clear guardrails.” yunwal argues call-center work is neither controlled nor repetitive — “Customer support is what I go to when the controlled environment has failed.” fhe says this explains why so little AI customer service actually works, and why the attempts are frustrating.
  • someguynamedq — the goalpost argument: “simplest task” used to mean writing a coherent English sentence, and now means autonomously fixing, reviewing and merging a bugfix. abeppu’s one-line reply: “Eliza wrote coherent English sentences.”
  • againstapples — asks what makes this different from how humans learn, since human expertise also takes years of narrow training. bravoetch answers with the chess case: a child learns from a short book and then plays reasonably, whereas models fed every game in history still make illegal moves. danielmarkbruce’s version: humans silently wear many hats, and you don’t discover how many until edge case after edge case.
  • randomImmigrant — a middle position that several commenters share: bearish on LLMs for automation, bullish for LLMs paired with domain experts. Their specific mechanism is that models handle timekeeping badly and can’t estimate how long their own work will take, which undermines long-horizon agents with shifting specifications.
  • ausbah — points at incentives rather than capability: when the business model is selling tokens, you get self-serving behavior and dark patterns, and “more” thinking as a product feature.
  • TrackerFF — asks what much larger compute budgets might make possible, using a roughly $10M estimate for Navier-Stokes. The thread contains conflicting cost estimates, not a verified bill; that figure should not be read as an established cost of the result. dotdi answers from the essay’s own logic: the problem was unusually well specified, which they argue is rare in the real world.
  • utopiah — a practical disagreement about prototyping: they say rapid prototyping is not the throwaway mockup work Kruer describes, but the process of making something genuinely new, and beaker52 clarifies that Kruer meant throwaway software specifically. olmo23 counters with their own use: ERP prototypes mocked up in 30 minutes so clients can react to a process.
  • knuppar — agrees open and cheap models will keep undercutting the labs, with two dissenters on that path: woeirua argues open models won’t stay open, since no one will release something capable of chaining zero-days, and pvab3 worries about the safety of open weights in both aligned and unaligned forms.

Where the thread disputes the premise

Commenters reject the essay’s framing of why the labs are valued, from several directions. These are competing economic arguments, not verified valuations or wage estimates:

  • keeda — argues that the potential value of automated knowledge work could justify large valuations. zug_zug disputes their wage base. danpalmer challenges the assumption that today’s spending survives mass automation unchanged: “This assumes you don’t change the market”. Displacement would also change incomes, taxes and demand; the thread’s competing salary totals do not establish what model providers could actually capture.
  • bluegatty — the labs aren’t priced on replacing knowledge workers but on AI becoming a universal form of compute, and their real problem is a weak position in the value chain, not wacky pricing. lukewarm707 attacks from the other flank: if AI has a greater-than-10% chance of causing human extinction, as Anthropic’s leadership has said, then the essay is wrong; if it doesn’t, they argue, Anthropic is dishonest.
  • tim333 — from an investor’s view, a drop-in replacement arriving in five or ten years would still leave the labs extremely valuable if they dominate it. The real risk is becoming AltaVista, displaced by a better implementation of their own idea.
  • randyrand — supply and demand, not pricing narrative: demand is real, supply will skyrocket, open weights contribute, and no one can predict the equilibrium.

The blast radius the essay lands on

Kruer’s closing move is that even if the labs are overvalued, the consequences don’t stay with them. The “datacenter full of geniuses” story is self-driving and limited only by compute; the scenario he actually expects — narrow, cheap, wide swarms of models — is bottlenecked by human orchestrators, and his bet is that its blast radius reaches well beyond the companies selling it.

A note on reading comments as evidence: HN handles are pseudonymous and the site publishes no per-comment scores, so the ordering here reflects HN’s own ranking, not a vote. This is a slice of the thread, not a consensus.