Let's Build GPT - From Scratch, in Code, Spelled Out

Let's Build GPT - From Scratch, in Code, Spelled Out

Andrej Karpathy via YouTube Direct link

conclusions

30 of 32

30 of 32

conclusions

Class Central Classrooms beta

YouTube videos curated by Class Central.

Classroom Contents

Let's Build GPT - From Scratch, in Code, Spelled Out

Automatically move to the next video in the Classroom when playback concludes

  1. 1 intro: ChatGPT, Transformers, nanoGPT, Shakespeare
  2. 2 reading and exploring the data
  3. 3 tokenization, train/val split
  4. 4 data loader: batches of chunks of data
  5. 5 simplest baseline: bigram language model, loss, generation
  6. 6 training the bigram model
  7. 7 port our code to a script
  8. 8 version 1: averaging past context with for loops, the weakest form of aggregation
  9. 9 the trick in self-attention: matrix multiply as weighted aggregation
  10. 10 version 2: using matrix multiply
  11. 11 version 3: adding softmax
  12. 12 minor code cleanup
  13. 13 positional encoding
  14. 14 THE CRUX OF THE VIDEO: version 4: self-attention
  15. 15 note 1: attention as communication
  16. 16 note 2: attention has no notion of space, operates over sets
  17. 17 note 3: there is no communication across batch dimension
  18. 18 note 4: encoder blocks vs. decoder blocks
  19. 19 note 5: attention vs. self-attention vs. cross-attention
  20. 20 note 6: "scaled" self-attention. why divide by sqrthead_size
  21. 21 inserting a single self-attention block to our network
  22. 22 multi-headed self-attention
  23. 23 feedforward layers of transformer block
  24. 24 residual connections
  25. 25 layernorm and its relationship to our previous batchnorm
  26. 26 scaling up the model! creating a few variables. adding dropout
  27. 27 encoder vs. decoder vs. both ? Transformers
  28. 28 super quick walkthrough of nanoGPT, batched multi-headed self-attention
  29. 29 back to ChatGPT, GPT-3, pretraining vs. finetuning, RLHF
  30. 30 conclusions
  31. 31 Oops "tokens from the _future_ cannot communicate", not "past". Sorry! :
  32. 32 Oops I should be using the head_size for the normalization, not C

Never Stop Learning.

Get personalized course recommendations, track subjects and courses with reminders, and more.

Someone learning on their laptop while sitting on the floor.