2026.05.22 The Growth Bet #ai Every technology leap offers two paths: cost savings, which runs out, and growth, which does not. The AI jobs question turns on which one we pick. 2026.05.17 The Reaction Is the Character #product Reactions, not writing or voice acting, are the load-bearing structure of emotional connection in games. The half-second between a player acting and a character responding is where the work happens. 2026.05.08 The Field Got Wider #career The guilt engine raised the ceiling on what you can build. The orchestrator role widens the field of what you are managing right now. Notes on agent management overhead and the cost of running parallel. 2026.05.03 The Wild West of AI Tooling #ai A new AI tool goes viral on X every week. Some of it is influencer pump. The data underneath shows something more interesting, and the convergence is already happening. 2026.04.29 The Tip and the Wave #ai ChatGPT just hit 900 million weekly users. That's about 10% of the world. The exhaustion of keeping up with AI is real, but it's mislocated. 2026.04.25 AI Writes the Code. Who's Testing It? #ai AI coding tools shipped first and shipped loud. The agentic testing wave arriving now is the more interesting story, and the benchmarks are starting to lean its way. 2026.04.19 A Weekend with Claude #ai A weekend of using Claude Design, Claude Code, and Claude Cowork. Notes on what Anthropic's stack looks like from the inside, and why the handoffs are what changed. 2026.04.17 The Guilt Engine #career Why the more you build with AI, the further behind you feel. Notes on the ambient pressure to keep shipping, and how to tell useful guilt from the kind that just burns fuel. 2026.04.14 The Llama That Wouldn't Fit #ai I tried to ship a 3B parameter Llama inside an iOS app. iOS had other ideas. Here's what I learned about on-device AI, jetsam, and why Apple Intelligence won the round. 2026.04.10 The Models That Work Too Well #ai Anthropic won't ship Claude Mythos. OpenAI is staggering Spud. Two frontier labs, one week, the same flinch. Default-ship is quietly becoming default-gate. 2026.04.09 Everyone Hates AI in Games. Everyone's Already Using It. #ai 85% of players reject AI in games. 90% of developers already use it. The gap between those numbers is where the future of game design actually lives. 2026.04.08 AI Gave Robots a Brain. The Hard Part Just Started. #ai Jensen Huang called 2026 the ChatGPT moment for physical AI. The real story is more interesting: the robot brains are here, but deployment is where the race begins. 2026.04.06 The Three-Year Forecast for Knowledge Work #ai AI hasn't wiped out knowledge jobs yet. But the next three years will reshape them beyond recognition. Here's what the data says is coming. 2026.04.05 The Spark Came Back #career AI coding tools didn't just make me faster. They brought back the part of building I thought I'd lost for good. 2026.04.04 Work Work: How Claude Code Joined WoW's Addon Arms Race #ai When Blizzard killed WeakAuras, RWF guilds turned to Claude Code to build custom boss addons in real-time. AI coding tools just became part of competitive raiding. 2026.04.03 OpenClaw: The AI Agent That's Selling Out Mac Minis #ai OpenClaw turned the Mac Mini into a personal AI appliance. Here's what it is, why people are buying dedicated hardware for it, and what can go wrong. 2026.04.02 What's the Difference Between LLMs and World Models? #ai Yann LeCun left Meta to bet on world models over LLMs. What's the difference, and why does it matter for the future of AI? 2026.04.01 Strawberries, Sisters, and Carwashes: What Viral LLM Failures Actually Reveal #ai How many r's in strawberry? How many sisters does Alice's brother have? Should you drive or walk to the carwash? Each breaks AI at a different layer. 2026.03.31 The Software Developer of Today Writes No Code #career Since December 2025, I've handed off basically all code writing to Claude Code. The shift already happened. Here's what the job actually is now. 2026.03.31 Botwash: A Post-Mortem for a Product That Got Out-Evolved #product Botwash was a car wash for AI text. It never found product-market fit because the problem it solved disappeared before launch. Here's what I learned. 2026.03.30 Slipstream: Building a 51M Parameter F1 Expert LLM in a Weekend #ai I pair-programmed with Claude and built a 51-million parameter transformer trained on Formula 1 Wikipedia articles. Here's what I learned about how LLMs actually work.