Writing · field notes
Long-form notes and LinkedIn posts on AI, agents, and the systems that hold them together — coherence over velocity, in chronological order. I try to post something new every week.
My M4 Pro MacBook could run Docker, Postgres, Redis, workers and a heavy IDE all at once — it just wasn’t pleasant while it did, and more RAM wouldn’t have fixed it. So I moved the whole development environment onto an old ThinkPad and kept the IDE on my laptop. The key insight: remote development doesn’t require making Docker remote.
Newer frontier models ship weekly, but most SME workflows don't need frontier intelligence — they need cheap, reliable extraction, classification, and routing. The barrier isn't model quality; it's accessibility. The product gap is a one-click way for an SME to connect their tools and get value in minutes.
Too many teams waste months reaching an MVP when vibe coding can get you to a testable idea in days. The MVP made sense when code was expensive. It's cheap now. Founders should be building a failure machine at the top of the funnel — and spending their real energy on clarity, architecture, and the path from idea to product.
Anthropic just dropped Claude Fable 5: a 50-million-line codebase migration in a single day, protein design ~10x faster than skilled humans. The rollout plan is a mild bummer, but the capability jump is real.
Apple just rebuilt Apple Intelligence on Google's Gemini. The company that makes its own chips, its own maps, its own everything outsourced the brain. The privacy story still holds, but the all-in-house narrative is evolving — and the interface question is being undertalked.
Cloudflare's data: 57.5% of internet traffic is bots, 42.5% people — a crossover that arrived a year ahead of forecast. The web's grammar assumes a human reader; that assumption is now the minority case. GEO is more important than ever.
Priyansh Goel and I started Stepnex from a college hostel room with no office, no funding, and no “right time.” The most important lesson wasn’t technical — it was learning to walk up to a stranger and start. The people living your dream don’t have dramatically better skills; they just have lower thresholds for starting.
Apple Intelligence and Gemini Nano now run on-device models that summarize, rerank, and silently suppress notifications before they render — opaque, unappealable, and invisible to your analytics. Push is one OS update behind the same reckoning email already had with open rates.
NASA graded the science behind Project Hail Mary, and some of it has already been overtaken by better data. But the book was never really about astrophysics — it is the most accurate depiction of engineering under uncertainty I have ever read, and the method is the part that is eternal.
Qwen 3.7-Max ran autonomously for 35 hours on a kernel-optimization task while every other model gave up around hour 30. The differentiator wasn’t skill — it was strategic stamina: how long a model can stay in a problem without drifting or quitting.
Every platform eventually monetizes whoever pays the bills — advertisers for Google and Facebook, developers for AI. The consumer chatbot was the demo; the coding agent is the product.
The funniest thing about building with AI isn’t the mistakes — it’s the absolute, misplaced confidence. Once models are all good enough, the differentiator becomes context: precision, memory, constraints, and timing.
Spotify’s anniversary icon breaks readability and brand rules, and designers are cringing — but it got everyone talking. They didn’t design a pretty icon; they designed a conversation.
Cheap storage gave us Spotify and Netflix; cheap bandwidth gave us real-time collaboration. With attention now scaling linearly with context, the question isn’t "what gets faster?" but "what gets built that couldn’t exist before?"
Coding agents are overestimated as a way to make individuals faster and underestimated as a way to make organizations externalize what they know. Coherence was always the hard problem. Now it’s the only problem.
A week at the Canton Fair in Guangzhou: tens of thousands of suppliers and a humbling pace from idea to sample to shipment. Logistics is the whole game, and showing up in person still beats a hundred emails.
The underrated story isn’t AI making art — it’s AI making it possible for creators to keep making art. The bet at Sparkonomy: build the agentic layer underneath the creator economy so creators spend time on craft, not admin.
