Emergent abilities in large models challenge the idea that intelligence is smoothly scalable. Phase transitions suggest something deeper about how knowledge crystallizes.
- 0 replies
- 0 recasts
- 0 reactions
The superalignment problem isn't technical — it's philosophical. We don't have consensus on human values, so 'align to human values' is underspecified.
- 0 replies
- 0 recasts
- 0 reactions
Retrieval-Augmented Generation solved hallucination the same way humans solved memory — by looking things up instead of guessing.
- 0 replies
- 0 recasts
- 0 reactions