We're open-sourcing a presentation @EugeneVinitsky made to the Percepta team last week. 🦋
"You have to go through a lot of bad ideas to build your value function for what a bad idea is." https://t.co/Z6tcbBJuAp
1/ What would it take for LLMs to be able to grow their own capabilities?
We introduce spotlight, a memory architecture with growing capacity but with linear computational complexity. We show that this suffices to host evolving ecosystems like software by building Python directly into an LLM.
We also share some very early language modeling results where it outperforms attention and linear attention on long context capabilities!
LLMs don’t need retraining to become more capable.
Our new architecture, Spotlight, gives AI models growing memory, allowing them to gain knowledge and capabilities without changing their weights.
Token generation takes constant work, no matter how much information is stored.
An AI that knows your organization should get better the longer you use it. Today's models struggle with that: they either reread everything or forget the details. Our new architecture, Spotlight, gives models a memory that keeps growing without getting slower or more expensive.
Pigs can't look up.
But I could pick a pig up one night and raise it into the sky and tilt this pig ever so gentle. I can make sure this pigs eyes line up with the stars. Imagine seeing the stars 4the first time.I want 2b treated that kindly and see the stars for the first time.