Training at a glance

Training happens in stages, each with a different job. Confusing them is the source of half the wrong claims about models — so here is the map.

Stage 1 — Pre-training: learn the world's text patterns

The model reads trillions of tokens (web, books, code) with one task: predict the next token. No labels, no teachers — the text supervises itself. Out of this emerge grammar, facts, reasoning patterns, and also every bias and falsehood frequent enough in the data.

This stage is absurdly expensive (thousands of GPUs for months) and is why only a handful of organizations train frontier base models. What it buys: broad capability. What it doesn't buy: obedience — a base model completes prompts, it doesn't follow instructions.

Stage 2 — Instruction tuning: learn to be helpful

Fine-tune on tens of thousands of instruction → response pairs (written or curated by humans). The model learns formats: answer the question, refuse politely, show steps, use tools. Capability barely moves; behavior transforms. This is cheap relative to pre-training — the reason small labs can ship useful assistants on top of open base models.

Stage 3 — Preference tuning: learn taste and guardrails

Humans (or stronger models) rank pairs of responses; training (RLHF, or simpler DPO-style objectives) pushes the model toward preferred ones. This shapes style, honesty-habits, and refusal behavior. It does not reliably implant new knowledge — mostly it changes which of the model's existing behaviors surface.

Alignment tuning is largely skin-deep: capabilities come from pre-training, manners from tuning. Pressure-test accordingly.

What training can't fix

• Knowledge cutoff. The model froze when training data did — it cannot know last week unless given tools or fresh context.
• Hallucination. Next-token prediction rewards plausible continuations, not verified ones. Tuning reduces it; the incentive structure that causes it remains.
• Reasoning ceilings. More training scales pattern mastery, but genuinely novel multi-step reasoning stays brittle — verify outputs that matter.