hawkes.aiKey ideas
- The one idea — Supervised fine-tuning: a comparatively tiny set of worked instruction→response demonstrations, run through the same next-word ritual, snaps the completer into assistant shape.
- How it is shown — The demonstration binder; the same furnace at whisper scale; the before/after — quiz-continuer to answerer.
- The trap to avoid — SFT teaches the SHAPE of answering, not new knowledge — the knowledge was already in the base; demonstrations cast it, and flaws in the binder become flaws in the character.
- What it sets up — Demonstrations can't cover taste — what when answers are matters of degree?
Trillions of tokens built the raw model. A binder of a few thousand worked examples made it an assistant — and whoever writes that binder designs the character.
The one idea
Supervised fine-tuning: a comparatively tiny set of worked instruction→response demonstrations, run through the same next-word ritual, snaps the completer into assistant shape.
How do you turn a text-completer into an assistant? Not with rules — with examples. A binder of worked demonstrations: instruction, exemplary answer, thousands of pages. Watch what a small book does to a giant machine. The mechanism is nothing new. Supervised fine-tuning is the same ritual from episode sixty: predict the next word, compare, nudge. The only change is the diet: curated scripts in chat format, where the text after the assistant-marker is always an exemplary answer. Days, not months; thousands of pages, not trillions of tokens. Same furnace, finishing heat. Why does a binder steer what a mountain built? Because SFT adds almost nothing — it aims. Knowledge, languages, reasoning: all already in the weights from pretraining. The demonstrations teach posture: when text looks like an instruction, be the figure who answers it.
How it works — the demo
The demonstration binder; the same furnace at whisper scale; the before/after — quiz-continuer to answerer.
A rudder, not an engine. And the transformation lands. Same question that spawned quizzes now gets an answer — direct, shaped, finished with the end-of-turn token it learned from these very pages. Episode seventy-four's mask, installed. Episode fifty-eight's goodbye, taught. The weirdo holds a shape, because thousands of pages showed it one figure to be. Which makes the binder destiny. Whoever writes those demonstrations designs the character: its tone, its hedges, its refusal style, its formatting tics. Verbose demonstrations breed a verbose assistant. Evasive pages breed evasion. That quirk you'd swear was intentional? Often it's page four hundred's author. The trap — and teams burn real budgets on it: fine-tuning to teach facts.
The trap to avoid
SFT teaches the SHAPE of answering, not new knowledge — the knowledge was already in the base; demonstrations cast it, and flaws in the binder become flaws in the character.
Why it matters — and what’s next
Demonstrations can't cover taste — what when answers are matters of degree?
A few thousand pages can't press new knowledge into vaults that took trillions of tokens to fill; SFT shapes behavior, not memory. New facts belong on the desk — retrieval, Act Eight. Shape with the binder. Inform with the desk. Mixing those up is the most expensive confusion in applied AI. But demonstrations have a ceiling. Two answers, both correct — one warmer, one sharper. Which should it prefer? You can't write a page for every shade of better. Teaching taste needs a different mechanism entirely: humans comparing, ranking, judging — and a model that learns the judge. Next: RLHF, the strangest and most consequential step in the whole finishing school.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 2:30 of narration
How do you turn a text-completer into an assistant? Not with rules — with examples. A binder of worked demonstrations: instruction, exemplary answer, thousands of pages. Watch what a small book does to a giant machine.
The mechanism is nothing new. Supervised fine-tuning is the same ritual from episode sixty: predict the next word, compare, nudge. The only change is the diet: curated scripts in chat format, where the text after the assistant-marker is always an exemplary answer. Days, not months; thousands of pages, not trillions of tokens. Same furnace, finishing heat.
Why does a binder steer what a mountain built? Because SFT adds almost nothing — it aims. Knowledge, languages, reasoning: all already in the weights from pretraining. The demonstrations teach posture: when text looks like an instruction, be the figure who answers it. A rudder, not an engine.
And the transformation lands. Same question that spawned quizzes now gets an answer — direct, shaped, finished with the end-of-turn token it learned from these very pages. Episode seventy-four's mask, installed. Episode fifty-eight's goodbye, taught. The weirdo holds a shape, because thousands of pages showed it one figure to be.
Which makes the binder destiny. Whoever writes those demonstrations designs the character: its tone, its hedges, its refusal style, its formatting tics. Verbose demonstrations breed a verbose assistant. Evasive pages breed evasion. That quirk you'd swear was intentional? Often it's page four hundred's author.
The trap — and teams burn real budgets on it: fine-tuning to teach facts. A few thousand pages can't press new knowledge into vaults that took trillions of tokens to fill; SFT shapes behavior, not memory. New facts belong on the desk — retrieval, Act Eight. Shape with the binder. Inform with the desk. Mixing those up is the most expensive confusion in applied AI.
But demonstrations have a ceiling. Two answers, both correct — one warmer, one sharper. Which should it prefer? You can't write a page for every shade of better. Teaching taste needs a different mechanism entirely: humans comparing, ranking, judging — and a model that learns the judge. Next: RLHF, the strangest and most consequential step in the whole finishing school.