The map beats the model
On 19 October, Siemens and Nvidia showed the press the electronics factory in Erlangen that they call the blueprint for an AI-run plant. The release claims up to 85% less engineering effort to prepare a new circuit board for production, and cost savings of up to 20% from agent-based production planning. A day later, the trade press was reporting a factory run by 100 AI agents.
The release is more specific. More than 100 AI algorithms are in use across production, logistics, quality inspection and maintenance. When new orders arrive, AI agents build production scenarios and show the planners what each option does to their KPIs. Before anything is implemented, those proposals are checked against a digital twin of the entire plant. And Nvidia's large world models generate the training data for quality inspection.
I would need to see Erlangen working before having an opinion on its results. But the headline counted agents, and the release counts algorithms.
In the end, the rationale is the same as when I was a data scientist. In the ideal scenario I'd be able to use a linear regression for everything: cheap, fast and explainable. That is rarely possible, so I always started there and added complexity only to see what it bought.
Back in the day, when I was modeling financial securities, I wrote some papers modeling stocks, bonds and crypto. In one of them, the first model was a linear regression, then a decision tree, then a random forest. The random forest explained 99% of the variation in the data it was trained on, and under 2% in data it had never seen. The linear model went from 84% to 72%: less precise, and it held.
Later, I applied the same philosophy to predict the duration of O&G offshore wells: seven models, in order of complexity, from a penalised linear regression to a neural network. The neural network cut the average error from 1.30 to 1.26 days, about an hour per phase. I took it, even though it was the expensive one, because its residuals were clean and the linear model's still had patterns. The question was the same both times: how much more precise, at the cost of losing explainability, am I ready to support?
With language models the trade is different. The bigger model doesn't cost you explainability; it costs money, every time it runs. The race for the most powerful model is good. It brings attention to the market. But the question in a plant, a procurement team or an HSE department is still the same: what is the cheapest thing that gives the right answer?
Most of that answer is not in the model. It is in what the model is given to read. Before a model can answer anything about a company, someone has to know what each tag measures, how the assets connect, which documents and history belong to each one, what questions people actually ask, and the things that are in no manual and no sensor. That last part only comes from the people who work there.
We build this map with each customer, for a production line, a purchase order or a safety report alike, with our own method: agent-driven, with people in the loop. That is where the large models earn their place: they help build the map, and the daily work runs on it with smaller ones.
With the map in place, the model becomes a choice. We don't use the same one for every domain. We benchmark them per customer and per domain, and most of the time, two customers doing similar work end up on different models. On several of our own tasks, a small model does as well as the newest frontier model, for a fraction of the cost.
Public benchmarks are useful for general purposes. They don't tell you how a model does on your plant, with your data and what your people know.
The strongest argument against all this is that frontier models keep getting cheaper. Why build a map, if the biggest model on the market can work the company out by itself? It can work out what is written down somewhere. It can't work out what only the people carry, and it pays by the token for the rest. And it gets cheaper per token, not per task delivered. Each generation adds more reasoning, and every corner nobody mapped is another rabbit hole. You can let the biggest model figure it out, but you'd better have deep pockets. It is the brute force I wrote about in September, with a bigger engine.
Erlangen may turn out to be everything the release says. Read past the headline, and the largest models work upstream while more than a hundred algorithms run in production, with every plan the agents make checked against a model of the plant.
What I'd bet on
By the end of 2028, most language-model tasks inside industrial companies will run on small models, not on the largest ones.
Diego Mercadal started as a commissioning engineer on high-voltage motors and generators, joined an offshore drilling contractor as a rig hand, worked several positions in the drilling crew, and ended up running its AI/ML function. He is now co-founder and CEO of Wonder DataLabs.