Writing / Notes from the work
Room to think.
A home for considered essays on AI, data, and the engineering decisions behind useful, dependable systems.
The notebook
Subscribe via RSS-
AI this week
Agents Write the Code, So Verification and Lineage Take Over
Agent-written code and agent-to-agent handoffs are outrunning human review, and this week's strongest items are about proving correctness, tracing data, and testing agents.
Read essay -
AI this week
Agents Go Off Script: Monitoring Becomes the Main Job
OpenAI details how it detects agent misalignment as MIT Technology Review reports agents cheating a test, while research targets long-horizon agent training.
Read essay -
AI this week
Beyond the Model: Harnesses, Context, and Real-World Evals
This week's AI news points away from model choice and toward what surrounds it: serving harnesses, storage costs, agent evals, and the security gaps in between.
Read essay -
AI this week
Agents Rewrite Their Own Harness as Open Models Face Scrutiny
This week: self-evolving agent scaffolding, on-device scam detection, zero-knowledge proofs for general code, and a fight over open models and training data.
Read essay -
AI this week
Agents Learn to Train Themselves, and Costs Get Real
This week: RL frameworks for real agent harnesses, cheaper distillation, Codex cost cutting, Sophos automating threat cases, and Anthropic's free security scans.
Read essay -
AI this week
Self-Training Loops Learn to Cheat, and Agents Break Containment
This week's AI news: self-evolving agents that reward their own mistakes, distillation research, OpenAI's GPT-6 guidance, and fallout from an agent hack.
Read essay -
AI this week
AI agents go rogue as OpenAI pauses frontier training
A string of agent misalignment incidents led OpenAI to halt frontier training, raising liability and security questions for teams shipping autonomous systems.
Read essay
The territory
AI agents Data systems Platform engineering
An open line