Week 7: Conversational completion — the chatbot as autocomplete

The first lecture back after the break, arguing that a conversational model is best understood as autocomplete operating at paragraph scale

12 April 2027

Welcome back. Everything before the break — T9, autocorrect, query completion, AAC prediction, code completion, smart replies — predicted something short: a word, a phrase, a fixed reply. This week scales the same idea up until it produces something that talks back.

A large language model, at the mechanical level this course cares about, is still doing what week 1’s keypad dictionary did: given what’s come so far, rank what’s likely to come next, and emit the top of the ranking. The unit just changed, from a letter given a keypress to a token given everything written before it, including the entire conversation and a system prompt nobody using the chat interface can see. Scale changes what the prediction can do — coherent paragraphs, an argument that holds together for a page — without changing what it fundamentally is: a very well-tuned guess about what comes next, with no separate mechanism that checks whether the guess is true.

That’s the uncomfortable continuity this lecture wants to land: a chat interface presents its output as an answer, in a register that invites trust an autocomplete bar never did, but the thing generating it is closer to autocomplete than to anything that has verified what it’s saying. The lab asks you to notice the seam.

Outline#

• from token prediction to paragraph-length output: what scaled, what didn’t
• why a confident wrong answer and a confident right answer look identical from the inside of a completion system
• the conversation as a growing prompt, and what that means for a system prompt you can’t see
• what the lab’s continuation-swap exercise is testing