When I started working on “memory” i started out building “general person embeddings.” What has actually taken shape first is adding state to LLMs. But memory can be much more than this context. It is about personalization, next-gen recommendation systems, dynamic UI, etc.

In this house, Jerome Powell is a hero x.com/federalreserve…

The best comp for modern AI is the tower of babel

I should have clarified in blog (but am now afk and can’t edit articles from phone, plz fix Elon): We don’t use an actual filesystem. We use Postgres but have a wrapper on top of it expose it to the LLM as a filesystem

I remember at the beginning of this first AI wave everyone seemed so focused on such esoteric algorithmic improvements in foundation models and scaling. Either that or AGI. Then Karpathy left OpenAI to work on agents and over time the focus has shifted towards more narrow,

New generations of men arrive and inevitably must face consequences of decisions of generations before them. I'm optimistic about my generation but unsure of where we will take our stand. This Dostoevsky quote rings in my ears often though: “And there is still more: how many

a 32M parameter multi vector model outperforms 600M parameter models and comes close to a 8B model. multi vector is the only way forward to make retrieval better. Rikiya Takehi cooked big time