Agent Swarms Are Great for Local AI
Single-stream vibecoding makes a local LLM rig look like an expensive Chromebook. Agent swarms flip the tokenomics — and suddenly the box under the desk pays for itself.
Thoughts on building, exploring, and technology.
Single-stream vibecoding makes a local LLM rig look like an expensive Chromebook. Agent swarms flip the tokenomics — and suddenly the box under the desk pays for itself.
For a couple of years it looked like AI had left consumers behind — the good models lived in someone else's datacenter and you rented access by the token. That era is quietly ending, and the thaw is showing up under desks, not in datacenters.
Anthropic's Claude Fable policy turns coding assistants into non-neutral infrastructure for product companies.
The hardware and software stack behind training and serving open-weights models from my home office — 2x 3090s, 256GB of non-ECC RAM, and a Threadripper that's older than it should be.
How agentic coding tools can build impressive features while making your codebase incomprehensible.
A base model trained on my text history showed me relationship patterns that therapy couldn't.
The VC-backed startups got Google Cloud credits. I got Photon, OpenStreetMap, and open weights embedding models. The scraps turned out to be enough.
How Microsoft's AI strategy turned brilliant engineers against the technology they should be building.
An interactive exploration of the bsdiff binary diff algorithm used for software updates in macOS and Chrome.