From Completion to Conversation
A language model only knows how to complete text —
so what if I fabricate a chat log and ask it to finish it?
You can only predict the next token, right?
So I'll craft a format: "Here's an unfinished chat log — fill in the assistant's reply."
Step 1
Assemble a "fake chat log":
write the user's message, then deliberately leave the assistant's line blank
write the user's message, then deliberately leave the assistant's line blank
Round 1: What does the model see? What does it complete?
User: Who is Zixia the Fairy? (紫霞仙子, a character from the classic Chinese film A Chinese Odyssey)
Assistant:
Assistant:
↓ Model starts completing from "Assistant:" onward
Assistant completes:
Step 2
Feed the entire text to the Base model —
it will do exactly one thing: complete from the end
it will do exactly one thing: complete from the end
Step 3
When the user sends another message, concatenate "the previous reply + the new question"
into a longer pending-completion text, then send it to the model again
into a longer pending-completion text, then send it to the model again
Round 2: The user speaks again — the first-round answer is appended and sent
Assistant completes:
Well, well — the model actually nailed it!!!
Hold on — did it reason its way there?
Read round two again. The previous line only says Zhizun Bao loved Zixia most deeply — it never says who Zixia loved. Strictly from the text, "who did she love?" is unanswerable.
The model fills it in anyway because it has seen this story countless times in training. That is not reasoning; it is continuing the words that most often go together. Sounding right and having worked it out are two different things — and that gap is exactly where the chapter on hallucination picks up.