Two colleagues who owned a critical subsystem left within a week of each other, and a leftover task in their area — one I barely knew — landed on me. The migration plan for it was freshly AI-generated, and what I didn’t fully trust was my own grasp of it. So I asked my AI to quiz me — not to grade me, just to see what I’d actually absorbed. ...
The Lie We Tell About the Code
Every codebase has two states: the code, and the lie we tell about the code. Most documentation systems make the lie comfortable. One I came across recently doesn’t. Documentation has always been the thing engineering teams promise to do and quietly skip. Sprint deadlines win, the wiki rots, the README becomes archaeology. We all know the cycle. Recently I came across a project (by Franco Dominguez Noriega) that does it differently — and the difference is not just “we tried harder this time.” It’s a structural change that only makes sense in a world where AI agents are part of the development loop. After spending some time with it, I’m convinced the pattern is worth sharing, because it solves the documentation-decay problem in a way I haven’t seen before. ...
The Lattice Hypothesis
An engineer’s conjecture on distributed biological intelligence, dreams, and the uncomfortable trajectory of AI In computing, thick clients carry their own processing power, storage, and local decision-making — but they become dramatically more capable when networked. A thought lodged in my head recently and wouldn’t leave: A human brain looks remarkably like a thick client. Andy Clark and David Chalmers argued in their “Extended Mind” thesis (1998) that cognition doesn’t stop at the skull — it extends into notebooks, tools, and cultural artifacts. They didn’t use the client-server vocabulary, but the implication is the same: the brain is a node in a larger system. ...
The Most Dangerous Hallucination Is the One That Sounds Right
Your AI coding assistant quotes a specific line from a specific file. The file is real. The section name is close. The quote sounds exactly right. But the line doesn’t exist — and that’s precisely what makes it dangerous. A Quote That Never Was I’ve been working with Claude Code on a large healthcare integration project — syncing an on-premise practice management system (I’ll call it “SimplePractice”) with a cloud FHIR platform. Over several weeks, we’ve built push/pull sync, environment tooling, documentation, and dozens of shell scripts. It’s been genuinely productive. ...
Gemma 4 Structured-Task Performance: Field Report from a Local-First App
Gemma 4 Structured-Task Performance: Field Report from a Local-First App Benchmark data and prompt-format findings from deploying Gemma 4 E4B in a real application. Intended for LLM teams (Gemma, Ollama) and developers building structured-output pipelines on local models. Context We build Gary, a privacy-first personal assistant CLI for macOS. It runs entirely locally — an encrypted database, a daemon process, and a local LLM via ollama. The LLM handles three structured tasks: ...
The Vibe Coding Trap
Let’s be real. When the first AI coding agents dropped, we all nodded solemnly and said, “Of course, a human will always review every single change. Safety first.” We lied to ourselves. Or, more accurately, we succumbed to the seductive illusion of frictionless productivity—a 10x illusion where we feel like we’re coding faster, but we’re actually just accumulating debt we can’t afford to pay. The recent Stack Overflow post on AI as a second brain identifies the core issue: we are offloading our judgment. This isn’t a future sci-fi risk; cognitive offloading is happening now, and it’s reshaping both our codebases and our minds. ...
10x Illusion
The 10x Illusion: If AI Codes 10x Faster, How Much Faster Do Projects Actually Ship? AI coding tools are getting shockingly good. So it’s natural to ask: if the coding part gets 10x faster, shouldn’t the whole project get 10x faster too? The answer is surprisingly counterintuitive — and backed by a growing body of data. The Speed Is Real. The Extrapolation Is Not AI coding tools deliver genuine speed on implementation tasks. GitHub Copilot studies show developers completing isolated coding tasks 55% faster. AI agents can generate entire modules in minutes. The speed is not the illusion. ...
I Benchmarked Apple’s Secret On-Device LLM. It’s Not Ready. macOS Tahoe ships a local language model you’ve never heard of. I tested it against the same benchmark I use for ollama models. Here’s what happened. The Discovery Apple quietly ships a language model in macOS Tahoe (26.x). It powers Apple Intelligence features — Siri, Writing Tools, notification summaries — but it’s locked behind system frameworks with no public API. Then I found apfel, a tool that wraps Apple’s FoundationModels framework as an OpenAI-compatible API server. Suddenly, Apple’s on-device model is accessible to any app — including mine. ...