L12 · Post-Training: From Predictor to Assistant

L12 · Phase 4 · 20 min

Post-Training: From Predictor to Assistant

“If the model only learned to predict the next likely word, why doesn't it ramble forever, argue with itself, or answer a different question than the one I asked?”

Raw Completion vs. Assistant Response

Same underlying model, same prompt. Toggle to see what changed.

YouThe solar tank is

AI Tool (raw completion)the solar tank is mounted on the roof of the building and connects via a series of copper pipes to the collectors below, which absorb sunlight and transfer heat through a working fluid that circulates continuously during daylight hours, and the tank itself is usually insulated to reduce heat loss overnight, and modern systems often include a backup heating element for days with insufficient sunlight, and…

The raw model just continues the sentence plausibly — it has no notion that this fragment was actually a question waiting for an answer.

Depth ladder

A raw, freshly pre-trained model only knows how to continue text plausibly — asked 'The solar tank is', it might just keep rambling rather than actually answering your question. A second training stage teaches it to behave like a helpful assistant instead: read the question, answer the question, stop.

Knowledge check

Given a raw completion and an assistant-style response to the same prompt, identify which is which and why.