How large language models actually work, in 24 minutes

Key takeaways

1. 01Models read text as tokens: chunks of words, not letters.
2. 02Attention lets each token look at every other one to work out meaning from context.
3. 03Training is predicting the next token billions of times, then tuning with human feedback.
4. 04“Hallucinations” come from fluent guessing, not from looking facts up.