RLHF & Alignment
How a raw next-token predictor becomes a helpful assistant — the three-stage RLHF pipeline behind InstructGPT and ChatGPT, and the simpler DPO alternative that followed.
— Access required
This piece is private.
Some series on Black Strat — notes on reinsurance, the data foundations, transformer architectures, hardware & compute — are restricted to readers with the access password.
If GK has shared the password with you, enter it once and your browser will stay unlocked for 30 days.
Unlock with password →Don’t have the password? Write to GK.