2 ms·
From what I understand, these models generate the response one word at a time. Every time you see a new word appear at the end, the model is taking into conside
by trekkie1024 4y ago
From what I understand, these models generate the response one word at a time. Every time you see a new word appear at the end, the model is taking into consideration the entire chat history + its own answer so far to generate that next token.
- ehnto 4y agoThanks for the comment, that's so fascinating since it seems to put limitations on thinking in general. A human for example can imagine future possibilities concurrently while speaking and correct themselves as they go. It doesn't seem to map well tk how I put together a thought either, but admittedly I wouldn't really know how the mechanics of my brain do it, maybe it's not so different just with some auxiliary modules bolted on ha.