Loading slide

Scale and Emergence

  1. 01Scale and Emergence
  2. 02Memory card
  3. 03The loop that was being scaled
  4. 04What happens when you just make it bigger
  5. 05The improvement had a shape
  6. 06Hold this contrast
  7. 07Abilities that appear from nowhere
  8. 08Take a breath
  9. 09The GPT-3 moment
  10. 10A raw model is not yet helpful
  11. 11Teaching it to answer
  12. 12Letting humans steer it
  13. 13The side effect: confident and wrong
  14. 14Reinforce your understanding
  15. 15Question: What emergence means
  16. 16Question: Making the model helpful
  17. 17Question: Confident and wrong
  18. 18Quiz: answer
  19. 19Bigger changed what was possible
  20. 20Want to go deeper?
11 / 19
BackNext

Teaching it to answer

The first step in shaping a base model into an assistant is to show it what a good answer looks like.

Humans write thousands of example conversations by hand. A question, then an ideal reply. Another question, another ideal reply. Each one written carefully to be helpful, honest, and clear. Together these examples become a new training set, far smaller than the internet the model first learned from, but pointed at a specific target.

Then the model trains on them the exact same way it always has. Look at the input, guess the output, measure how wrong it was, nudge the weights. The same guess-and-adjust loop, the same machine. The only thing that has changed is the diet. Now the inputs are questions and the targets are good answers.

After enough passes, the model has absorbed a new habit. Not "continue whatever text came before me" but "respond to this person helpfully." This is fine-tuning, and we have met it before: it is the same trick from the transformer module, where one pretrained model was cheaply re-pointed at a specific job. Here the specific job is simply being a good assistant.

Fine-tuning on its own gets you most of the way. The assistants people actually use usually add one more step, and that step changes the tone of the whole system.

Citations