Fine-tuning an LLM to write docs like it's 1995

The Dissection

This is an engineer's toy project that accidentally documented the automation of professional cognitive labor—and then immediately buried the evidence under reassuring platitudes. The experiment: fine-tune a 7B parameter model on 37 million words of 1990s Microsoft technical documentation, spending roughly $58 total, to produce period-accurate corporate technical writing. The results were striking enough that the author felt compelled to end with "a human tech writer can't be replaced."

The Core Fallacy

The final paragraph is a ritual propitiation—performative human-centering designed to manage cognitive dissonance. The author literally wrote that a 7B model trained on period documentation "produced a convincing chapter opening that could be mistaken for genuine period material" and that "Qwen 192k was the strongest." Then: "reassuring takeaway is that such a model can never replace a human tech writer, only augment them."

This is logically incoherent. If a model produces output indistinguishable from a human professional, and does so at trivial cost ($58, consumer hardware, a day of compute), the argument that it's "only augmentation" collapses. The author is performing hedge, not reasoning.

Hidden Assumptions

1. "Replacement" requires the model to be the complete worker. It doesn't. A $58 fine-tune that handles 80% of first-draft technical documentation work eliminates 80% of the economic value of that labor category. The remaining 20% doesn't justify full human salaries—it justifies part-time oversight.
2. Professional judgment is doing the work. The author claims fine-tuned models "have the same lack of judgment as their non-tuned siblings." But most technical writing is not judgment work. It's structured knowledge transcription. The model demonstrated this by generalizing style to novel concepts (REST API in 1990s voice) and hallucinating convincingly for fictitious functions. That's not lack of judgment—that's production-grade behavior for the task domain.
3. High-quality training data is a meaningful barrier. The author spent effort finding Bitsavers corpus, cleaning 192k examples. This is a one-time cost. The resulting adapter (a small file) is infinitely replicable. The "difficulty" is a startup cost, not a structural constraint.

The Verdict

This article is a carbon date on the profession. Technical writing—structured, domain-specific, style-governed documentation work—is now automatable for $58 and a day of compute on a 7B model. The author documented the mechanism and then flinched from the conclusion.

The Discontinuity Thesis says the mass employment circuit severs when AI achieves cost and performance superiority for economically necessary labor. At $58 and consumer hardware for professional-grade domain output, this domain has already crossed that threshold. The labor doesn't vanish immediately—it attenuates across years as tooling matures and adapters get cheaper. But the trajectory is terminal.

The author built a machine that produces Microsoft technical writer output at zero marginal cost and called it "augmentation." That sentence will age like milk.