Applied AI reviews, build logs and essays on systems, validation, agent workflows, decision-making and how ideas survive contact with reality.
Why I write Working Stories Kitchen Notes
An independent researcher used Claude Code, Codex, and NASA’s archived telescope data to identify a possible exoplanet 116 light-years away. Now the discovery faces a rather inconvenient test: reality. Somewhere, approxi
Read essay ->
OpenAI released hundreds of AI-generated mathematical results. Anthropic helped overturn decades-old complexity conjectures. The bigger question is what happens when machines produce discoveries faster than humans can un
Read ->
Giving every developer an AI agent won’t fix a broken software development lifecycle. You need to redesign the entire SDLC, from tickets and documentation to CI/CD, QA, and team rules. Imagine you own a restaurant. You’v
Read ->
A new robotics paper shows how tactile feedback can train a vision-language-action model that no longer needs touch sensors during deployment. The result: 86.7% success handling jars, waffles and eggs. One of the biggest
Read ->
Claude used Lean and TLA+ to find bugs in the Agent SDK and generated 16 PRs. The harder problem is not proving the model. It is proving…
Read ->
A new Trinity College Dublin paper turns budget cuts, channel launches and campaign pauses into quasi-experiments, then estimates the sales that would have happened anyway. There is a very common marketing measurement pr
Read ->
OpenClaw proved persistent agents were useful. Grok Bot, Instinct and Muse are turning the same idea into something normal people can…
Read ->Asking an LLM “is this a complaint?” works surprisingly well.
The usual argument says any U.S. restraint on frontier AI simply hands the race to Beijing. Matt Sheehan’s conversation with Ezra Klein, and China’s own 2026 AI safety framework, suggest a more complicated reality. Yeste
Google DeepMind has mapped the predicted effects of 9 billion DNA mutations. Elsewhere, researchers compile Verilog into living cells and simulate an entire bacterium on a computer. Biology is becoming strangely programm
Models are getting better at understanding us. That still doesn’t make “do something with this” a good instruction. People have been declaring prompt engineering dead for years. I’ve never particularly liked the term any
A new NBER experiment suggests that consent interfaces don’t need to remove choices to change behavior. They just need to make one choice slightly more annoying. I love research papers about extremely mundane things that
