dark knowledge
The extra information contained in a model's full soft output distribution (relative probabilities of all classes) beyond the single correct label, exploited during distillation.
As defined in Dwarkesh Patel Podcast - Eric Jang – Building AlphaGo from scratch
Where it comes up · 2
- Model Distillation: How to Steal a Billion-Dollar AI2 claims
- Dwarkesh Patel Podcast - Eric Jang – Building AlphaGo from scratch1 claim
Most important claims · 3 of 3Search all claims for “dark knowledge” →
- Distillation is so effective per sample because soft targets (the full logit distribution) have far higher entropy than a one-hot token answer, carrying much more information per sample; AlphaGo exploits this by training the policy network to imitate the full MCTS visit-count distribution rather than just the single selected action ('dark knowledge' distillation).Dwarkesh Patel Podcast - Eric Jang – Building AlphaGo from scratch
- Dark knowledge—the full probability distribution a teacher model assigns across all possible next tokens—is the signal that makes distillation work; soft labels reveal how a model ranks alternatives (e.g., 90% cat, 9% dog, 1% car from a photo), not just the final answer.Model Distillation: How to Steal a Billion-Dollar AI
- Geoffrey Hinton, Oriol Vinyals, and Jeff Dean wrote the foundational distillation paper at Google in 2015 and gave the useful signal concept the name 'dark knowledge'.Model Distillation: How to Steal a Billion-Dollar AI