Writing about what it actually takes to adopt AI that lasts, from someone who spent eighteen years making automation survive the real world.
A 29 MB model switches real GPIO pins on a Raspberry Pi 4, fully offline, in about 1.4 seconds per command. It is not yet reliable enough to trust on its own, and that is the useful result.
ReadYou do not need new PLCs to monitor asset health. Most of the signal is already being measured and thrown away. Here is what a brownfield plant already has, where new sensing actually pays, and how to do it read-only.
ReadMost false positives in industrial anomaly detection are not model failures. The model correctly found a statistical deviation that an engineer would have recognised instantly as a normal change of state. Here is how to fix that.
ReadA safety system has to be provably correct, and a language model cannot be proven correct in the way IEC 61508 means. That does not rule AI out of the plant. It rules it out of one specific layer, and the boundary is drawable.
ReadMost plants move data by wiring every system to every other system. A Unified Namespace replaces that mesh with a single real-time hub. Here is what OPC UA and MQTT each actually do, where Sparkplug B fits, and how these builds fail.
ReadMaintenance records are the one plant dataset RAG was actually designed for. They are real text written by people. Here is why it still goes wrong, and the step-by-step build that works.
ReadA language model forgets everything when the session ends. Long-term memory is what turns a stateless model into an agent that knows you. Here is the architecture underneath it, and what breaks.
ReadHistorians compress your data on purpose, and standard RAG has no idea. Here is why exception and compression exist, why vector search cannot answer historian questions, and the query-first architecture that works instead.
ReadGeneric RAG works on documents. Plant data is not documents. Here is why retrieval-augmented generation breaks on the factory floor, and the architecture that makes it work.
ReadA coalition of the biggest names in American tech just told policymakers open weights are a national necessity. The real story is bigger than one country: open models are becoming the ground the US, China, and Europe are all trying to build their own AI sovereignty on.
ReadAll three store plant data, but they answer different questions for different people. A first-hand guide to which you actually need, what it costs to guess wrong, and the layered setup most plants land on.
ReadMoonshot's Kimi K3 is the first open 3-trillion-parameter model: fast, cheap, and huge on context. Here is how it actually compares to Claude Fable 5 and GPT-5.6 on the benchmarks, and where each one wins.
ReadThinking Machines has released Inkling, its first open-weights model: a natively multimodal model you can download from Hugging Face, try in the Playground, or fine-tune on your own data with Tinker. Here is what it is and how to use it.
ReadAnima is an open-source operating system for running a small company on AI agents. Five agents handle the operations; the human signs off in a few hours a week. Here is the idea, and why the human stays at the boundary.
ReadGPT-5.6 is cheaper, comes in three tiers, and leads OpenAI's coding benchmark. Fable 5 is generally available today and still edges it on some agentic work. Here is how to choose.
ReadFable 5 wins the benchmarks. The useful question is narrower: is it better enough to justify twice the price for the work you actually do?
ReadGetting operational data off the plant floor comes down to three moves: collect it with OPC UA, move it with MQTT, and land it in a store the cloud can query. Here is the architecture, and where each piece belongs.
ReadA chatbot is built to respond. An agent is built to act. The thing that turns one into the other is almost embarrassingly small: a loop.
ReadWe've stopped building chatbots and started building digital coworkers. Underneath nearly every capable agent are the same five architectural pillars.
ReadYour machines have been talking all shift. The question is whether anyone is listening.
ReadOperational data spent decades locked in the systems that produced it. The story of the OT/IT divide: why it existed, and what finally made plant data reachable.
ReadAutomation makes a process run on its own. Optimization makes it run better. The difference is most of the value — and most of the skill.
ReadEvery era of operations got better when it got a new instrument. AI is the latest one — not a break from the work I've always done, but a continuation of it.
Read