4 ms·
At least LLMs are very often ready so acknowledge they might be wrong. It can be incredibly hard to get a person to acknowledge that they might be remotely wro
by megadata 2y ago
At least LLMs are very often ready so acknowledge they might be wrong.
It can be incredibly hard to get a person to acknowledge that they might be remotely wrong on a topic they really care about.
Or, for some people, the thought that they might be wrong about anything attall is just like blasphemy to them.
- Xelynega 2y agoIs this not just because aggressive material was filtered out of training data and the system prompts usually include some preamble about being polite? "Acknowledging they might be wrong" makes them sound like more than token predictors trained on polite sounding text.
- roywiggins 2y agoMost of the reason LLMs will "admit they're wrong" is because they've been trained not to argue too hard, and to not hold strong preferences. It's a sort of customer service personality. When you don't do that sufficiently you run the risk of producing the "Sydney" personality that Bing Chat had, which would argue back, and could go totally feral defending its incorrect beliefs about the world, to the point of insulting and belittling the user.