3 ms·
Yes, the short term behavior/output of the LLM could reflect an implicit goal, but I doubt it'd maintain any such goal for an extended period of time (long-term
by HarHarVeryFunny 2y ago
Yes, the short term behavior/output of the LLM could reflect an implicit goal, but I doubt it'd maintain any such goal for an extended period of time (long-term coherence of behavior is a known shortcoming), since there is random sampling being done, and no internal memory from word to word - it seems that any implicit goal will likely rapidly drift.
- Nevermark 2y agoAgreed. They can’t accurately model our communication for very long. So any implicit motives are limited to that. But their capabilities are improving rapidly in both kind and measure.