3 ms·
> In fact, I think in near future it will be the norm for MLLMs to "think" and "reason" without outputting a single "word". It will be outputting something, as
by limaoscarjuliet 1y ago
> In fact, I think in near future it will be the norm for MLLMs to "think" and "reason" without outputting a single "word".
It will be outputting something, as this is the only way it can get more compute - output a token, then all context + the next token is fed through the LLM again. It might not be presented to the user, but that's a different story.
- LudwigNagasena 1y agoThat’s the only effective way to get more compute in current production LLMs, but the field is evolving.