3 ms·
We really don't know why LLMs have the most of their emergent capabilities. The entire GPT-3 paper was about being surprised about it.
by spion 4y ago
We really don't know why LLMs have the most of their emergent capabilities. The entire GPT-3 paper was about being surprised about it.
- throwaway290 4y agoMaybe we don't disagree then. We know how LLM work. Parameters. Training data. Random number generator. That stuff. We don't know why it outputs what it outputs because rngs are notoriously unpredictable and we know it. So we are surprised but that in itself is unsurprising.