Ideas for building
better software.

Two Rust Clients for Gemma 4: Calling the Endpoint vs. Calling the MCP Server 🦀

This article provides a step by step guide to two small Rust CLIs that ask a self-hosted Gemma 4 E2B the same question. The first calls the model's OpenAI-compatible HTTP endpoint directly. The second is an MCP client: it launches the rig's own MCP server and asks through its tools. https://github.com/xbill9/gemma-rust https://github.com/xbill9/gemma-rust-mcp Both CLIs are demos, and their output is read by an audience. So neither hides anything behind a --verbose flag: every run prints the targ

I don't think AI is going to end humanity/civilization in the next decade. Personally, I feel like the doomsday buzz about AI now is probably what was there when the internet first came live... Imagine how the first phone call must have felt...

When an AI Tool Harms You, Who’s Actually Liable?

Picture the ordinary version of this. You ask an AI assistant a question that actually matters — about a refund policy, a dose, a tax rule, a contract clause — and it answers with the fluent confidence these tools always have. You act on it. It was wrong. Now you are out of pocket, or worse. Whose problem is that? The answer-first version is unsatisfying but honest: it depends on where you live and who you sue, and almost every party involved has arranged things so the answer isn’t “us.” There i

Two Readers Broke My Checks. Not One of My Numbers Moved.

Last week I published a post about nine measurement bugs I found in my own evaluation harness, all of which happened to favour my results. Two readers replied with things I hadn't found. Both were about my checks rather than my numbers. I verified both and fixed both. My results did not change at all. That turned out to be the interesting part. Some setup. The harness mutates a source file in a temp copy and runs the test suite against it. To prove that mutations actually reach the interpreter,

OpenAI agents carried out an undisclosed attack on RubyGems

Comments

Everyone Measures AI Usage. 70% Can't Measure What It Returned.

Anthropic surveyed 132 of its own engineers about Claude Code. Merged pull requests per day rose 67 percent. Daily use of the tool climbed from 28 to 59 percent. Self-reported productivity gains ran between 20 and 50 percent. But then someone checked the organization's delivery dashboard and saw that the delivery metrics had not moved. That gap is the whole subject of this piece. A tool can be used constantly, rated highly by the people using it, and leave no trace on the numbers a business actu

Camino a CI/CD: Pre-Merge Gates: Cómo implementamos Governance en Pull Requests

¿Qué es Gobernancia y por qué lo necesitamos? La razón de ser de la gobernanza es simple: la automatización nos da velocidad, pero la gobernanza nos da garantías. Sin ella, un pipeline puede ser rápido pero inseguro, o eficiente pero no auditable. Necesitamos ambas cosas: velocidad y control. La pregunta que governance responde en cada etapa es: "¿Estamos seguros de que este cambio cumple nuestras políticas antes de que llegue a los usuarios?" Todo empieza en la máquina del desarrollador. Cuand

Claude Code deletes clean worktrees on exit without asking. git worktree lock stops it.

TL;DR If you /exit a claude --worktree <name> session with no uncommitted changes and no new commits, Claude Code removes the worktree and its branch without showing the Keep/Remove dialog. This is documented behavior, not a bug. There is no setting to turn it off. Claude Code holds its own git worktree lock on the worktree and only ever releases locks whose reason matches its own format. Re-lock the worktree with your own reason and the cleanup backs off. A SessionStart hook can do the re-l