Field notes on shipping AI past the demo.
Original commentary on enterprise AI news, deployment surfaces, infrastructure trade-offs, and what actually changes for teams trying to put AI into production.
Mira Murati's lab shipped a giant open model — and told you to make it smaller
Thinking Machines' first model, Inkling, is a 975B open-weight release under Apache 2.0 — but the pitch is not the size. It is that you fine-tune it into something narrower and cheaper. The bet is that owned, specialised models beat one-size-fits-all frontier APIs on the work that actually pays.
Kimi K3 is the largest open model ever — and it splits "open" from "runnable"
Moonshot's 2.8-trillion-parameter K3 matches frontier coding models and ships its weights under a permissive license. Downloading them is free; standing up the eight-GPU node to serve them is not. That gap is the real story.
Murati's lab shipped a 975B open model under Apache 2.0 — the license is the news
Thinking Machines released Inkling, the largest open-weight model yet from a US lab, under a genuinely permissive license, with a per-request reasoning dial and day-one support across vLLM, SGLang, and llama.cpp. The weights are free. The data center is not.
OpenAI just turned the emailed deck into a shared, updatable surface
ChatGPT folded Codex, Work, and Chat into one app and shipped Sites — a way to build interactive web apps by prompt and share them with your team behind a workspace login. The unit of delivered work quietly moved from a file you attach to a link you keep current.
Copilot sells the tool. Autopilot sells the work — and services firms are rebuilding around it
Sequoia's "Service-as-Software" thesis says the next trillion-dollar company will look like a services firm, not a SaaS vendor. Legal, accounting, and compliance firms are already the proof — and the shift is structural, not a new toolbar.
A frontier model went offline for three weeks — then came back with a gate
US export controls pulled Anthropic's Fable 5 from global availability over a reported jailbreak, then lifted them. The model is back, but with a new safety classifier that quietly reroutes some coding requests. Two reminders that a hosted frontier model is a dependency, not a fixture.
MCP is going stateless — and that is the part worth reading
The next Model Context Protocol spec drops the session handshake so agent servers scale like ordinary web services. The capabilities — MCP Apps, Tasks — get the headlines; the boring transport change is what actually makes agents shippable.
Your agents just became identities you have to govern
Microsoft moved AI agents into the same identity system that governs employees — with their own IDs, human sponsors, and expiry dates. The product is Microsoft's; the shift is everyone's.
NVIDIA collapses physical AI into one open world model
Cosmos 3 folds scene understanding, world generation, and action prediction into a single open omni-model. The headline is robots; the quieter story is that synthetic data for vision systems just got a lot cheaper to make.
MiniMax M3 makes a million-token context cheap — and ships the weights
An open-weight model that claims frontier coding scores is the headline. The quieter, more useful story is the attention rewrite that drops the cost of running a 1M-token context — and the fact that you can run the whole thing yourself.
Anthropic moves agent execution onto your infrastructure
Self-hosted sandboxes and MCP tunnels keep the agent's hands behind your perimeter while the model loop stays in the cloud. It is a useful split — and a line worth reading carefully.