Yuya's Blog

The default is model-invoked.

The smartest AI agents fail without good UI for reviewing their output.

A deep dive into how modern AI applications mirror cloud infrastructure, breaking down each component from LLMs as compute engines to memory systems as databases, and exploring why a Heroku-like platform for LLMs could be transformative.

Models are good enough.

A look at MIRIX, a modular memory architecture that helps AI assistants get past the "day one" problem

The memory AI needs depends on what you are using it for.

My notes on Tuomas Artman's talk about Linear's sync engine: the object graph, the transaction queue, and how the read and write paths stay in sync.

Local-first software is a new approach to building applications that prioritize the user experience and data ownership.

My note from "CRDTs for Mortals" by James Long at dotJS 2019

How Apollo Client uses "optimism" to improve the performance of the cache