My bet for 2026: companies chasing margins and hitting diminishing returns on frontier models will shift to fine-tuning small, domain-specific models. Fine-tuning is getting cheaper, open-weights models are catching up to frontier ones, and differentiation through your own data beats competing on prompt engineering alone.

AI won't kill programming jobs or leave them untouched. It's a new abstraction layer, like every one before it, that will create vastly more developers building vastly more software. Existing programmers keep their jobs; a long tail of cheaper new roles unlocks demand we couldn't previously afford to meet.

I scaled back my ambitions from "how to write AI apps" to what I've actually learned. Key lessons: LLMs excel at condensing text, not creating it. Give them the data they need rather than relying on training. Let them self-correct. Use regular code wherever possible. And stop trying to replace humans, especially doctors and lawyers.

After leaving Netlify, I've been exploring AI/ML/LLM opportunities. Here's my explainer: "AI" is marketing, ML learns from examples, and LLMs are essentially massive word-completion engines. But at scale, something remarkable emerges that looks like understanding, and that changes everything about what software can do.

I've been saying for years that there are no fundamentals of web development, and people keep pushing back. My argument is simple: "fundamental" means *necessary*, and no single skill, not HTML, not CSS, not JavaScript, is necessary to get a web page into the world. Gatekeeping helps no one.

I push back on Alex Russell's "Market for Lemons" post blaming JS frameworks and their advocates for poor mobile web performance. Developers aren't stupid or deceived: React dominates because of rational economics. Frameworks save developer time, enable hiring, and compound via ecosystems. Better frameworks, not fewer frameworks, is the answer.

I dig into what's genuinely impressive about crypto (distributed trust, abuse prevention, true cloud computing) and what's deeply problematic: boundary interactions, governance concentration, transaction cost tradeoffs, and unclear incentives. My conclusion: crypto may only be useful for financial engineering, and that's probably fine, just smaller than the hype suggests.

After 25 years in web dev and 10 doing data work, I finally feel like I know what I'm doing. Key lessons: hire data engineers not scientists, buy your warehouse don't build it, use dbt and Airflow like you'd use Rails, and keep domain experts close. Still figuring out dashboards and discovery.

Twenty years of blogging, from sandwich diaries to AWS takedowns, coming out (twice) to becoming American. The blog has outlasted Slashdot, nearly outlasted Twitter, and documented almost my entire adult life. I was right about who I am and what I do. Here's to whatever comes next.