Ben Thompson argues the AI wave has hit the same wall every platform shift does: subsidies end, aggregators form, and the middle gets squeezed. The question is which layer becomes the aggregator this time — the model, the application, or the inference runtime.
Local inference with a 200B-class model on consumer silicon was a 2027 prediction six months ago. It just happened. Worth reading for anyone who thought OpenAI and Anthropic had a durable API moat.
Mollick makes a case that LLM-default prose is a style, not a capability ceiling. The fix is not a better prompt — it is being specific about whose writing you want the model to sound like, and doing the work to show it.
Ships with pricing, a tuned budget parameter, and a surprise: the public leaderboard numbers for the same budget went up after the beta run. Good news if you were holding off until the contract was stable.
Nathan Lambert on why "open" is now a gradient, not a binary. Weights without training data, without preference data, without eval sets is a reproducibility crisis in motion.
Swyx argues documentation is getting read by agents before humans, and the best docs are now structured for the agent reader first. Implications for anyone shipping an API in 2026.
Jack Clark walks through the internal spreadsheet math of a frontier lab — why 70% of compute is inference, why that ratio flips annually, and what it means for who funds what.
Another step down the inference cost curve. The parts of your product that you argued were too expensive to AI-augment three months ago are probably not anymore.