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?
10 / 19
BackNext

A raw model is not yet helpful

There is a gap between what GPT-3 could do and the assistant you actually talk to, and it is worth being clear about.

Straight out of training, a language model has exactly one instinct: continue the text. That is all the loop ever rewarded. So if you typed a question into the raw model, it would not necessarily answer it. It might continue your question with three more questions, because that is what a list of questions looks like in its training data. Ask it for advice and it might write what looks like a forum thread arguing with itself. It is not being difficult. It is doing the only thing it was trained to do: produce the most likely continuation.

This is the base model, the same idea we met back in the learning module: a model trained on a huge general pile of data, capable but unaimed. Enormously knowledgeable, and pointed at nothing in particular.

Turning that into something that answers you, helpfully and on purpose, takes a second, much smaller phase of training laid on top. That phase is where a raw text-continuer becomes an assistant, and it is the next two slides. It is also, as we will see, where one of the model's most talked-about flaws gets built into its behaviour.

Citations
Citations