3 ms·
They don't model knowledge, they model statistical matches to desired text. An LLM can't "know" that it is fabricating or conveying real-world fact. It can tak
by slowmovintarget 3y ago
They don't model knowledge, they model statistical matches to desired text.
An LLM can't "know" that it is fabricating or conveying real-world fact. It can take as input: "that statement is false." Depending on the sophistication of the model, it will then retain that correlation for that session. But start another session, and it can generate the same or similar text that is inconsistent with real-world fact, or even with input.
- dragonwriter 3y ago> They don’t model knowledge, they model statistical matches to desired text. I am not convinced that this is not a method of modelling knowledge. > But start another session, and it can generate the same or similar text that is inconsistent with real-world fact, or even with input. I’m not sure why this is meaningful, aside from the misleading metaphor of a named LLM as an individual; named LLMs are more analogous to genetic templates that code for some instinctive capabilities; if you want an analog to an individual, a single session (for a chat-oriented LLM-centered system) is the most appropriate unit of analysis.