#ai

This site has a light and a dark theme, and the cover image cross-fades between them — so every illustrated cover needs a genuine night twin. Faking one by dimming the day image is easy and looks wrong. Here's how the covers get generated, and the three gates — two of them running on a GPU — that prove a dark cover is a real re-light and not a graded fake.

Two weeks ago I said the DGX Spark's bandwidth wall was physics, not a driver update — fast or big, pick one. Then I ran a 122B model on the same box at six times the speed of a 70B. The law was right. The number I kept feeding it wasn't.

I bought Andromeda, a DGX Spark, to fine-tune models — but first I owed my last post a rematch. I re-ran the same local/free/paid LLM inference benchmark on 120 GB of Grace-Blackwell instead of a CPU NUC. It fixes the one thing the NUC got fatally wrong, hits a bandwidth ceiling I didn't see coming, and makes clear why the real reason I bought it isn't inference at all.

Google's Android Bench scores LLMs on real Android engineering tasks. A look at what the June 2026 leaderboard actually says — why the top is a statistical tie, where the real cost-efficiency winners are, and the trap hiding in the cost and latency columns.

Google shipped ADK for Android in May. I built a Compose app — clean architecture, MVI, Hilt, persistent chat history — where a Gemini agent calls real on-device functions, and hit every sharp edge the docs don't mention.

A hands-on walkthrough of Android's experimental AppFunctions API. I build a small but complete notes app, expose four functions to on-device agents with @AppFunction, and trace exactly what the KSP compiler generates and how the system discovers and invokes them.

A deep technical look at the multi-agent workflow that scraped the entire awesome-selfhosted dataset — nine runs, every architectural failure, exact agent counts and token numbers per run, and how it ended with Python bypassing agents entirely.

I had Portainer running but it made figuring things out harder, not easier. And there was no single place to just click a link and open a service. I described the problem to Claude Code and it built me exactly what I needed.

I bought a NUC to run local LLMs and escape API costs. Then I tested local, free cloud, and paid cloud against the same prompts. The result wasn't 'pay for inference' or 'run it yourself' — it was that 'fast and free' isn't a quadrant that exists right now, and that has implications for anyone planning a personal AI setup.

I've spent months building Android apps — todo, drawing, live wallpaper, launcher, Kotlin Multiplatform, full-stack with Ktor — entirely with Claude Code, without writing a line myself. This is an honest account of what worked, what didn't, what it cost, and what I think it means for how software gets built next.