The model is frozen and stateless

What this lesson shows

Weights are fixed after training; no memory between calls; it confabulates.

The question it opens with

Ask the model today's news or your private doc: it can't know.

The walkthrough, in the lesson's own words

• It got the date wrong, confidently. Tap the next word you think it scored highest.
• It can only do one thing: predict the next word. So it does.
• Your turn. Ask it anything it couldn't possibly know.
• It doesn't lie on purpose, it just always finishes the sentence.
• The weights froze at training: no clock, no memory of you, a hard cutoff.
• Nothing in next-word scoring flags low confidence as “I don’t know.” The model computes scores live from your question using frozen weights, then states the top guess, even when every option is a guess. Whatever happened since training is sealed out.
• It does, and that's the point. The app pastes today's date (and often live search results) into the model's context before it predicts a single word. The bare model still can't know; the product hands it the answer. The rest of Act 5 is about that trick.
• The model remembers you and learns a little from each chat.
• The weights are frozen between chats. Any 'memory' is the system re-feeding past notes into the context window; the model itself never updates from talking to you.
• An assistant says it remembers your last conversation and just learned your preference. What is really happening, given the model is frozen?
• The weights don't change between chats: the model is frozen and stateless. Any memory is the system re-feeding past notes into the context each turn. It didn't learn; something stored your preference and pasted it back in. That's why memory is a system feature, not the model getting smarter.
• Exactly. It only ever predicts the next word, “I don’t know” is just one more option it almost never ranks first.
• Notice “I don’t know” sits near the bottom. The model only does next-word prediction, and that option almost never scores highest, so it states the confident guess instead.
• Close, but a hedge like that still isn’t the top score. The model only does next-word prediction; the confident first word scored highest, so that’s what it said.
• “I don’t know” is on the list, it just scores near the bottom. The model computes these scores live from your question, using weights frozen back at training, but nothing in that scoring flags low confidence as “I don’t know,” so it just states the top guess.

Key takeaway

Motivates RAG and tools.

What you can do after this lesson

You can explain why a frozen model hallucinates and has a knowledge cutoff.

Where it leads: What if we hand it the right documents?

This is the written summary. The lesson itself is interactive: you predict, drag and operate the mechanism above, and the reveal answers you.