Ever wonder why ChatGPT gives helpful answers instead of random text?
Reinforcement Learning from Human Feedback (RLHF) — that's the secret.
After training on the internet, the model still doesn't know "good" from "bad." So we:
- Generate multiple answers
- Humans rank them (this one's better, that one's worse)
- Train a reward model to score answers like humans would
- Tune the AI to maximize that score
Result: answers feel helpful, not just statistically probable.
Newer method: DPO (Direct Preference Optimization) — skips the reward model, trains directly on human preferences. Meta used it for Llama 3.
AI alignment = teaching AI our values, one ranked answer at a time.
Hope this helps someone.
About me : I am a staff product analyst having interest in Product, Analytics, ML, GenAI, DE, Physics, Art, Literature. You can take mini AI course or read my mini blogs here : https://aisimplified.live
Top comments (0)