A thoughtful final batch, part cautionary, part analytical, part fun.
Haris Khan’s We Are Building Data Breach Machines and Nobody Cares is the most important security read in this batch and probably this month. His argument: AI agents are just loops making API calls, they have no internal morality, they run at machine speed, and the industry is moving too fast to have standardized security primitives. LLM APIs are inconsistent (Anthropic, OpenAI, and Google handle tool-calling schemas completely differently), agent behavior is non-deterministic (you can’t reproduce a bug reliably), and the industry’s response, “AI-enabled defense mechanisms”, is science fiction that would introduce new non-determinism into your security layer. His actual recommendation: use the boring tools. Anomaly detection models, circuit breakers, JIT credential vending with short-lived tokens. Govern the ball, not the moving goalposts.
Niki Tonsky’s Claude Is an Electron App Because We’ve Lost Native pushes back on the “just rewrite it in SwiftUI” crowd. His argument: native has nothing to offer anymore. OS vendors actively make native development painful. Native apps no longer look better than web apps, Apple’s own UI guidelines are inconsistent and change every year. Performance advantages are real but never realized because nobody cares enough. The real problem isn’t the runtime, it’s a lack of care. You can build slop in any stack.
Martin Alderson’s No, It Doesn’t Cost Anthropic $5k Per Claude Code User is the most useful piece of financial reasoning I’ve read about AI inference economics. The Forbes article confusing retail API prices with actual compute costs is a meaningful error: open-weight models of comparable architecture on OpenRouter cost roughly 10% of what Anthropic charges in API fees. At that ratio, even the heaviest Claude Code Max users cost Anthropic ~$500 in actual compute, not $4,800. The average user, per Anthropic’s own /cost data, costs around $18/month to serve. The “inference is a money pit” narrative benefits the frontier labs by making their moat look deeper than it is.
LukeW’s Durable Patterns in AI Product Design draws from four years of building AI-native products. His framing: every major platform shift (mainframes to PCs, desktop to web, web to mobile) changes almost everything about what an “application” is, and we always spend the first years trying to cram the old paradigm into the new one. The patterns he identifies as durable: suggested questions that help users understand what the system can do, context management that scales gracefully, and progressive disclosure that doesn’t overwhelm users with the system’s internal complexity. Worth reading alongside the Böckeler context engineering piece.
Chris Kirk-Nielsen’s CheffMPEG is a delightful palette cleanser, a writeup about building a little browser-based video conversion tool using FFmpeg compiled to WebAssembly. The technical journey is fun (getting FFmpeg’s WASM build to work in a browser, handling SharedArrayBuffer requirements, dealing with codec licensing) and the project is charming. Sometimes bookmarks are just “someone built a fun thing and explained how.”