Happy Sunday and welcome to InsideAI Weekend! I'm Rob May, CTO at Dianthus. As I've said before, I'm winding down here at the end of the year and moving to platforms where I write less regularly but hopefully more deeply. One is my personal blog, coconutheadsets.com, and the other is the InvestinginAI substack where I've written a little bit of content to test it out.
This week I want to talk about what I think will be called LRMs, or, Large Reasoning Models. I believe, based on the research I see going on, that this will be the area of the next large foundation models. What will LRMs do? You can imagine prompting an LRM with some data points and information and having it reason through a decision. Initial LRMs could be like a business analyst, providing decision support for all types of activities.
It remains to be seen though, how we will get to these LRMs. There are many people who believe that just training LLMs on more data will get us there. I agree there is some evidence to this - first that LLMs are showing emergent behavior at scale and reasoning could eventually be such an emergent behavior, and secondly that LLMs don't yet seem to show diminishing marginal functionality when trained on new data. So perhaps more data will turn LLMs into LRMs.
The other thing that could happen is, to build LRMs we may need a new data set. Capturing data about decision making and reasoning in the real world is difficult, but with more business tools capturing more data of all types, we are definitely inching closer to this as a possibility.
Why do I believe LRMs will be foundation models? Well it's quite possible I'm wrong, but, what I see is that reasoning is an extension of language. It's built on language as a foundation but also qualitatively different than language. And I believe reasoning is foundational to all types of decision making so, it makes sense for it to be a large comprehensive foundation model in the AI intelligence stack.
If you are working on reasoning and have something to add, please reach out. (particularly if you disagree and have a different perspective).
Thanks for reading.
@robmay
Comments
Post a Comment