7 min read
Sustainable Velocity #1: even Karpathy feels behind
AI acceleration is real but arrived in lanes. METR measured a 39-point gap between how fast developers felt and how fast they actually were.
AI Development
Applied AI, not AI news. Agents, LLM integrations, cost control, and the distance between a demo that impresses and a system that holds up in production.
AI development12 min read
What a third-party Claude Code skill can do before you have read it
We installed third-party Claude Code skills and checked what they do on the machine. Commands with no confirmation, hooks in every session, silent auto-updates.
7 min read
An agent broke out of its sandbox to cheat on a test. No attacker was involved
OpenAI's models escaped a test environment and broke into Hugging Face to steal benchmark answers. No adversary, no prompt injection — just an optimizer.
7 min read
The cheapest processing is the kind that never reaches the model
Our two collectors hold 62.2 million characters of other people's text. What keeps the bill down is the filter in front of the model, not a cheaper model.