Re:Telling GPT-4 you're scared or under pressure improves pe...
by seydor on Saturday November 04, 2023 @07:24AM
(#38140036)
> enough training data reinforcing the close correlation of those tokens.
The basic component is 'attention' which is a map correlating entity A to B,C etc, which creates a vast network of correlations, and this is repeated on multiple modalities. Some LLM researchers (who are trying to make sense of why they work) call those modalities 'skills'. It's simplistic to call it token correlation in the training data. It's likely that emotional words 'trigger' some skills more than others and this enables better performance.
The basic component is 'attention' which is a map correlating entity A to B,C etc, which creates a vast network of correlations, and this is repeated on multiple modalities. Some LLM researchers (who are trying to make sense of why they work) call those modalities 'skills'. It's simplistic to call it token correlation in the training data. It's likely that emotional words 'trigger' some skills more than others and this enables better performance.