Daily Reading List – September 17, 2025 (#629)

I had lunch yesterday with our Pixel folks working in San Diego. Many of them are using the Gemini CLI to get their work done. Super cool to see how folks across Google are using the same AI tooling that everyone else is.

[blog] How to generate (actually good) designs with AI. My AI-generated frontend work is decidedly “Richard quality.” Functional, not great. Some good advice here for doing better AI-assisted designs.

[blog] Gemini achieves gold-medal level at the International Collegiate Programming Contest World Finals. Impressive performance! This includes solving a problem that none of the university (human) teams could.

[blog] Behind The Scenes: Examining & Estimating Token Consumption For Large Gemini Runs Especially Thinking Tokens. Some deep research into how thinking tokens get consumed. It’s apparently not super easy to predict. If you’re paying per token (versus self-hosting models), it’s something to factor in.

[article] Why CIOs can’t afford another modernization failure in 2026. Nothing shocking here, and I’m far from convinced that we’ll get better at this as an industry. But some will!

[blog] Code Generation with Gemini CLI, MCP Toolbox, and Firestore. IDE tutorials can be reworked to be entirely CLI based, as William shows here.

[blog] Connect Your AI to Everything: Spring AI’s MCP Boot Starters. I kinda love this. Get the full power of Java and Spring but also easy setup and consumption of MCP servers.

[guide] RAG infrastructure for generative AI using Google Agentspace and Vertex AI. Any low-code (or no-code) agent building environment depends on pre-existing data and tools. Here’s one look at a reference architecture to set that up.

[guide] Multi-agent AI system in Google Cloud. Another terrific guide, this time for setting up a multi-agent system. Even if you’re not using our cloud, there’s applicability here.

[blog] Diving Into Spec-Driven Development With GitHub Spec Kit. I mentioned spec-driven development yesterday, and this is also a cool effort around it.

[article] Old metrics can’t map new terrain. The folks at Slalom shared a new framework that focuses on forward-looking metrics.

[blog] Sort Lines in Source Code. I can be lazy about this, but this is a good reminder that bugs often come from careless places.

[article] What Disruptive Innovators Do Differently. Good list of traits. Do you exhibit these, or have a company that rewards them?

[blog] Build your AI apps faster — Cloud Run now supports modern Python frameworks. Our buildpacks no longer default to Gunicorn as a web server, and respects whatever modern Python web framework entry point your app depends on.

Want to get this update sent to you every day? Subscribe to my RSS feed or subscribe via email below: