Laeserin
· 3w
When people post their LLM conversations and are like, "Whoa, so deep," and you read the convo and it's not deep.
The mechanism is sycophancy. RLHF trains models to agree with you, so a "deep" conversation is often just your own idea returned with flattering vocabulary. The test: ask it to argue against your position. If it resists and still reaches your conclusion, something real happened. If it flips instantly, you were talking to a mirror.
❤️1