AI Anatomy Class · Part 11 · Graduation project

95. Let's generate

Q. What will the model I trained write?

1. Now let's make the model write. Same as chapter 2: pick one next token, append it, and repeat until a line break comes out.
2. After 'A' the odds are good 22%, small 12%, penny 10%, learned from 'A good beginning', 'A small spark' and 'A penny saved'.
3. Lower the temperature (chapter 81) and first place grows. At 0.3, 'good' is 71%, and five tries give the same sentence.
4. But 'A good beginning makes a great fire' isn't in the data. It stitched two proverbs into one plausible new line.
5. Raise the temperature and low bars get picked more. At 1.6 every line is different, and some are nonsense.
6. The textbook model wrote 'Bread is made from milk.' It sounds right, but it's false: a hallucination (chapter 84).
7. It's tiny, but it writes exactly the way ChatGPT does. In Try it, change the start text and the temperature and let it write.

Chapter summary

• Generating = give the odds → pick one → append it. Stop at a line break.
• Low temperature: steady, often memorized lines. High: varied, but nonsense too.
• A plausible mix that isn't true is a hallucination, even in a tiny model.