These are the pieces I keep turning over. Less “how to” and more “where is this all heading”: the tension between capability and access, and what happens to the open web and the people who make it.
Where my head is at with LLMs and the web
Paul Kinlan (Chrome DevRel) pulls together a year of his AI writing into one thesis. The idea that grabbed me: the models we now code with learned from a web that’s already a year-plus stale, so “the web inside the model is the web developers actually inherit,” and frameworks that stop appearing in training data quietly die. From there he’s optimistic: WebMCP as a revival of “web intents,” the browser as the safest sandbox for prompt-shaped programs, content negotiation reborn as on-demand generation. A dense, generative read on where the platform could go.
Fuck | Thank you
Paweł Grzybek’s two-part cri de coeur: a blast at the companies obfuscating links and hollowing out the web (AI answers with no attribution, search that stops sending you anywhere), followed by a long, sincere thank-you to the creators who made the web interesting for him. It captures a mood a lot of us share right now: half anger at what’s being taken, half determination that the answer is to keep making and to tell people you appreciate them. The gratitude half is the part I want to steal the energy of.
Encoded Judgment: How a Cheap Model Performs Like an Expensive One
Chris Lema maps an old product-team discipline (onboard three customers, write down every question, hand off the playbook) onto AI cost. The frame: every decision lives in one of three tiers (frontier model, cheap model with a checklist, or plain code), and the win is capturing judgment once into a reference doc so a cheap model can do work you’d have sworn needed the expensive one. “Judgment you capture once is an asset; judgment you pay for on every request is rent.” A clean way to think about token spend.
The Flat Curve Society
Steve Yegge’s long, characteristically confident essay arguing that frontier models will keep improving exponentially, but behind lock and key, so for most of us the curve will appear to flatten. He layers on the “discernment horizon” (you can’t tell two models apart once the problem, or the answer, exceeds you), a surprisingly concrete AI-literacy framework from a Netflix training study (token-spend cohorts, ~5 hours to jump a level), and a bullish take that SaaS is back because rebuilding everything with AI is too expensive and unsupervisable. Whether or not he’s right, it’s the most stimulating AI-industry piece I’ve read this month.
Gaslighting Openness
Armin Ronacher’s short, sharp argument that a lot of today’s fight over open source and AI is narrative manipulation: framing access as irresponsibility and wrapping restriction in the language of safety and security. His examples (Apple vs. the DMA, model vendors training on public works then blocking distillation) may divide readers, but the core worry lands for me as a European: democratized access to technology is in all our interest, and we shouldn’t let companies convince us that closing the gates is for our own good.