AI Anatomy Class · Part 7 · Teaching GPT: pre-training
64. Studying without an answer key
Q. Text on the internet has no answer key. So what did GPT check its answers against?
- In chapter 4, people had to label each email 'spam' or 'not spam'. But no one can label every text on the internet.
- The secret: the text itself is the answer key. Let's take one sentence: 'I walk to school.'
- Show the beginning, hide the next piece, and it's a question. What comes after 'I'? The answer is right there: 'walk'.
- Add one piece at a time and you get more questions. After 'I walk' comes 'to', and after 'I walk to' comes 'school'.
- A 5-piece sentence gave us 4 questions. The text, not a person, provides the answers. That's self-supervised learning.
- The model solves all 4 questions at once. Thanks to chapter 51's mask, each position can only see what comes before it.
- All of GPT-3's roughly 300 billion tokens are questions like these. About 300 billion questions, without a single answer key.
Chapter summary
- Show only the beginning of a sentence and ask for the next piece: that's a question.
- The answer is the very next piece in the text. No person needs to label anything = self-supervised learning.
- A sentence of N pieces → N − 1 questions. The whole internet becomes a workbook.