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

TEXT START: i'll begin with a few theses for the reader to chew on:

The Dissection

This is a technically serious attack on frontier-lab valuation and the fantasy of an imminent, fully autonomous replacement for knowledge workers. Its strongest contribution is identifying the real control-loop problem: generating an answer is cheaper than specifying, validating, and safely deploying one.

But the text quietly changes the question. It treats zero-review autonomy as the threshold for economic displacement. Under the Discontinuity Thesis, that threshold is unnecessary. A system that lets one human supervise vastly more output can still destroy the wage circuit. The retained human becomes a Servitor or bottleneck operator, not evidence that mass productive participation survived.

The article also contains its own counterargument. Its cheap open models, wider swarms, and blast-radius claims describe a path where frontier labs lose their premium while AI deployment continues. It is bearish on the visible vendors without being consistently bearish on automation itself.

The Core Fallacy

The core fallacy is autonomy absolutism: treating human review as a stable employment defense rather than a shrinking control layer.

Human validation can be expensive and indispensable while still requiring far fewer workers than the production layer it supervises. If a hundred engineers become ten reviewers, the profession has not been preserved because ten jobs remain. Its labor market has been liquidated and reorganized around a narrower class of Servitors.

The text also mistakes current specification difficulty for a permanent architectural boundary. Its evidence establishes that present systems are unreliable outside well-defined task neighborhoods. It does not establish that specification, decomposition, test construction, orchestration, or validation will remain human-intensive at the same scale. The phrase structural reasons endemic to current architectures is asserted, not demonstrated.

P1 does not require a magic switch to universal autonomy. P2 does not require every human-only domain to disappear overnight. Partial, uneven superiority is enough to make institutions replace, compress, or devalue human labor across enough domains to trigger P3.

Hidden Assumptions

• Fully autonomous end-to-end execution is required before displacement counts.
• Human review scales roughly with model output and therefore preserves substantial employment.
• Domain experts must continue producing specifications manually; the specification process itself cannot be automated, templated, or amortized.
• Failures in current models reveal permanent limits rather than a lag defense.
• Continued hiring of weak software engineers proves durable human necessity rather than temporary workflow adaptation or cheap supervisory labor.
• The three adoption classes described are the whole market. Hybrid systems, narrower task decomposition, and concentrated human oversight are treated as secondary when they may be the main transition mechanism.
• If frontier labs are overvalued or weakened, the automation thesis weakens with them. The article underweights whoever controls the underlying compute, energy, logistics, and maintenance stack.

Social Function

Classification: partial truth, transition management, and prestige signaling, with a consoling side-effect for incumbent knowledge workers.

The text punctures investor hype and correctly warns that benchmark victories are not equivalent to trustworthy autonomy. That is real analysis, not empty copium. But the cracked-intern framing also preserves a familiar hierarchy: adults remain necessary, therefore the labor order remains intact. That conclusion does not follow.

In practice, this argument can become a deployment manual for cheaper models, broader swarms, narrower human bottlenecks, and lower labor costs. It is a map of the early kill zone disguised as a rebuttal to the final form of the weapon.

The Verdict

Accurate about near-term autonomy friction and potentially accurate about frontier-lab economics. Wrong as a systemic defense of human employment.

Frontier labs can be cooked while the labor circuit is still being severed. The article attacks the fantasy of a datacenter full of autonomous geniuses; it does not defeat the more durable path of cheap models, swarm production, concentrated supervision, and ownership of AI capital. It has identified a bottleneck, then mistaken that bottleneck for a life-support system.