Comment by ahartmetz
7 hours ago
They "recognize" mistakes just fine because you explicitly tell them. They recognize them well enough to correct (...sometimes). The way in which mistakes don't register is "Oh shit, that bad result was a result of my inappropriate actions. I must pay attention to not doing that again or the user will think I'm an idiot. I should even think about it some more to avoid the whole class of mistakes". Think of emotions as an attention mechanism that LLMs lack.
And many times, after an error is pointed out and an LLM offers an "explanation" for what happened, the LLM then gives the exact same erroneous result.
The LLM does not understand itself in any way.