Notes from the work.
Guides for people with an app built on AI tools, and weekly notes on software development and AI.
- 2026-10-26Fix it or rebuild it? How to decide what to do with an app your AI tool wrote
When every change breaks something else, starting over is tempting. It is usually the wrong call, and the signs that it isn't are easy to spot once you know where to look.
- 2026-10-26Is your AI-built app safe to launch? Five things to check before real users arrive
The tools that build an app in a weekend are good at making it work and much less careful about who else can make it work. Here is what to check, and how to check it without reading every file.
- 2026-10-23A popped LLM gateway steals your cloud keys, a nation-state actor runs AI pentests at scale, and Anthropic gives away the standard it built
LiteLLM lost a release to a typosquatted lookalike repo that harvested credentials from everything on the machine, Microsoft caught a Chinese state actor running autonomous AI pentests against 2,250 tenants through a compromised ConnectWise app, and Anthropic handed its Agent Skills standard to a neutral foundation already running inside ChatGPT, Cursor and VS Code.
- 2026-10-16Claude solves a 25-year cryptography mystery and helps catch an actively exploited Windows bug, while OpenAI pays mathematicians to check its proofs
Anthropic traced NIST's unexplained 1990s curve seeds to a forgotten random number generator and helped flag a zero-day under active attack, while OpenAI puts $5 million toward mathematicians who can tell it whether its AI-generated proofs actually hold up.
- 2026-10-09Python 3.15 slips to the wire, Mistral names a trillion-parameter model after a meme, and Copilot keeps retiring models under you
A three-year-old Python feature forced two last-minute release candidates, Mistral's new trillion-parameter model still loses to a rival a third its size, and GitHub's Copilot model list keeps changing whether your workflow is ready or not.
- 2026-10-02OpenAI bets DevDay on agents that outlast you, and Next.js ships a patch you can't skip
OpenAI's biggest event of the year sells persistent agents and a $500-a-month tier to match, while Next.js quietly patched a critical bug on schedule.
- 2026-09-25OpenAI and Anthropic cut their workhorse model prices the same day, and VS Code quietly drops a disclaimer
GPT-6 Sol and Luna and Claude Opus 5.5 landed hours apart with near-identical pitches, and a small, unexplained edit to VS Code's release notes says more than the changelog itself.
- 2026-09-18JDK 27 ships, VS Code puts its agents in a box, and two AI CEOs disagree about what Claude is
A boring, load-bearing Java release, an editor update that finally treats coding agents as something to contain rather than trust, and a public argument about whether Claude's "wellbeing" language is a design choice or a mistake.
- 2026-09-14Three products shipped this week that assume your AI agent won't need you around
OpenAI, Cursor, and Microsoft each launched infrastructure this week for coding agents that run for hours or months without a human watching, and that's a bigger shift than any single feature in it.
- 2026-09-06App Store review gets stricter, and the model release cadence isn't slowing down
A new response requirement on App Store submissions, and why a week with releases from five different labs says more about the industry than any one of them does.