Google replaced its Flash model in under three weeks, and Vercel already exposes it on AI Gateway at 50% off through December 31. The point isn't the model: it's that the model identifier stopped being a constant in your code.
If you're building AI for a regulated domain, the real work is no longer picking a model — it's ingestion, versioning and traceability of documentation that isn't on the internet. With everything that implies.
The ChatGPT/Codex app ships a complete copy of LibreOffice inside it. This isn't a fun fact: it shows how agent-based desktop clients are being built, and what you should audit before shipping yours.
Netflix is giving a Spanish film the widest theatrical distribution in its history. Underneath the cultural gesture there's a product decision anyone building platforms should read closely.
Anthropic shipped Fable 5.1 with safety classifiers enabled, and some routine coding requests can get refused. That turns the AI Gateway's fallback array into an architecture decision, not a cost optimization.
Google cut off Aurora Store's access and GrapheneOS users were left without a clean channel for installing apps. The practical takeaway: on mobile, the distribution channel is part of your attack surface.
Memory stops being elastic and becomes a quota: the background process is no longer a safe place to keep state. What to measure, what to persist, and why rearchitecting anything is still premature.
Going from a personal LLM to a corporate one changes who decides, who audits and who is legally on the hook. A practical guide to evaluating it before you sign anything.
Auto mode isn't a convenience checkbox: it moves human consent out of the loop. How to contain the blast radius when your agent reads content you don't control.
Tencent released a 770B MoE with a 1M-token window and Vercel already routes it. What changed isn't the model: it's that trying it costs one string in your config. And that makes deciding badly more expensive.