3 ms·
This would be a fairly insane breach of trust and common sense if true; the chain-of-thought / reasoning trace is, from an information perspective, close to a s
by highfrequency 24d ago
This would be a fairly insane breach of trust and common sense if true; the chain-of-thought / reasoning trace is, from an information perspective, close to a superset of the prompt and model response.
- Ginden 24d agoAnd, of course, you can train a model to start with repeating prompt in chain of thought!