The Hands-On Experience Gap

Alfio Quarteroni is professor emeritus at the Politecnico di Milano and the EPFL in Lausanne, and one of Italy's most prominent applied mathematicians. He built, with his team, the first mathematical model of the human heart, and his recent book, L'Intelligenza Creata. L'AI e il nostro futuro (Hoepli, 2025), maps where artificial intelligence is taking us. Across a set of recent conversations, he keeps returning to a distinction that is easy to lose in the noise: artificial intelligence is not one thing, and it is not magic. It is a human construction of mathematics, data, energy, infrastructure, and responsibility. That sober framing leads him to a warning worth hearing.

The warning

The tension is not between humans and machines. It is inside the way we introduce the machines. When we hand automated systems the tasks that used to be a beginner's apprenticeship, we quietly close the door those beginners walked through to become experts.

His point is not a rejection of technology. It is a condition attached to it. As he puts it, technology can strengthen the human being, but without concrete paths of training and experience, the central role of the human being risks weakening. For a field like biotechnology, where judgment is built at the bench and in the field over years, that condition is not a detail. It is the whole game.

What the machines themselves teach us

Here is the part that should give us pause. The systems we are handing the work to did not learn from a manual. They learned from experience. As Quarteroni describes modern machine learning, a computer is not programmed step by step to carry out a task. It learns from experience, analysing large quantities of data. Through the right algorithms the system is trained on vast training sets, and progressively learns to recognise patterns and hidden relationships in the data, turning them into a form of working knowledge. That is exactly how artificial neural networks operate.

Read that back against the warning and the paradox is hard to miss. We are building machines that become competent the way an apprentice does, through patient, accumulated experience with real cases. At the same time we are tempted to remove exactly that experience from the humans who will have to design those machines, choose their training data, and check their answers. You cannot judge whether a model was trained on good data if you have never generated data yourself at the bench. You cannot catch a plausible but wrong result if you have never watched the real process succeed and fail with your own hands.

Not all of it is a chatbot

Part of the confusion comes from collapsing all of artificial intelligence into the chatbot we talk to. Speaking to enterprises in Il Sole 24 Ore, Quarteroni is blunt that the biggest gains often come not from conversational systems but from digital twins, mathematical replicas of a machine, a plant, or a process that let you test a scenario before touching reality. Artificial intelligence, he adds, is not a general-purpose technology you simply plug in. You have to decide beforehand which specific problem you want to solve, then build the training data and choose the algorithm to match.

AI as accelerator, not replacement

None of this makes Quarteroni a skeptic of AI in the lab. He points to AlphaFold, which predicted protein structures that had resisted biology for decades, and to Halicin, the antibiotic a neural network surfaced in 2020 by screening tens of millions of compounds, effective even against strains resistant to existing drugs. His reading of these successes is precise: artificial intelligence does not replace scientists. It lets them explore enormous spaces of possibility, millions or billions of molecules, that would be impossible to analyse by traditional means. In that sense it becomes a powerful accelerator of scientific discovery.

But he adds the caveat that defines the whole argument. The real scientific discovery still requires something more: intuition, the capacity to interpret, and an understanding of what the results actually mean. Those capacities are not downloaded. They are grown, slowly, through direct contact with the work.

The last mile belongs to those who have done the work

Quarteroni describes a decisive stretch he calls the last mile. Artificial intelligence can analyse vast quantities of information, spot patterns, and suggest solutions. The final decision, though, always belongs to the human, because only the human knows the context, understands the purpose of the action, and can take responsibility for the consequences. The system, by contrast, has a structural limit: it has no awareness, it does not know why it produced a given answer, and it holds no ethical principles. The human being brings experience, critical sense, and responsibility.

The uncomfortable implication is simple. The last mile is only walkable by someone who has walked the earlier miles. Judgment under uncertainty, the sense for when a result is too clean to be true, the instinct for what to check next: these are earned hands-on. Remove the early years of doing, and the person who is supposed to own the last mile arrives without the legs to walk it.

Where the Foundation stands

The bioERGOtech Foundation develops Engineered Living Systems through synthetic biology, AI-driven automation, and human-centred design. That last phrase carries the weight. We adopt automation to remove drudgery and to widen what a small team can attempt, not to remove people from the bench. Our answer to the experience gap is integration, not substitution.

Quarteroni's own advice to young people entering the field reads almost like a mission statement for this approach: study mathematics, physics, and computer science well, but stay open to the problems of the real world. Innovation is often born precisely at the meeting of different disciplines, when a solid theoretical training confronts concrete problems to solve. Colleagues trace his own limits-aware, concrete view of technology partly to his rural upbringing, a culture accustomed to confronting limits, timing, and material reality. Concrete before abstract. Problem before promise.

It is telling that when he speaks to enterprises, he lands in the same place. Artificial intelligence, he insists, does not decide the destiny of companies. At most it calculates some of their trajectories. The rest still depends on the people who decide. For a foundation that invests in people and in a territory, that is not a caveat. It is the point.