Outgrowing the System
We tend to talk about AI as the disruption. Reading through this issue, I kept circling back to a different possibility.
What if AI itself is no longer the main bottleneck, and the systems around it are?
The same tension surfaces across wildly different fields. AI capabilities are moving fast, while organizations, factories, infrastructure, energy systems, healthcare, and law move at their own, slower pace. So the interesting question is no longer simply what AI can do. It is what happens when those capabilities land in environments that were never designed for them.
At first, the articles here seemed to pull in completely different directions. Business transformation. Industrial systems. Chips and energy. Healthcare. Explainability. Contracts. Agentic AI. And yet, somewhere between them, the same friction kept surfacing.
Somi Jeong starts with the business itself in AI Adoption Is Not Digital Transformation. Companies are buying tools, automating processes, launching pilots, and calling it transformation — but adoption and transformation are not the same thing. You can bolt AI onto an organization without changing much about how that organization actually works. The harder part comes afterward, when processes, roles, decisions, and sometimes the business model itself have to reshape themselves around the technology.
That gap between capability and reality is the one I explore in The Factory That Remembers: How Continual Learning Could Transform Industry 4.0. Smart factories already generate enormous amounts of information, but collecting experience is not the same as remembering it. A factory that learns over time needs more than a model capable of updating itself. The processes around it have to support validation, retention, rollback, human judgment, and change. Otherwise the intelligence evolves while the factory around it stands still.
In my second contribution, The AI Infrastructure Stack: Why the Next AI Race Will Be Won Below the Surface, I move further down the stack. We spend an extraordinary amount of time talking about models when so much of what determines AI’s future sits underneath them — compute, semiconductor manufacturing, cloud infrastructure, energy, geopolitical dependencies. The visible AI economy rests on a physical and industrial one, and some of its most consequential strategic choices are being made far from the chatbot window.
Lucy pulls the thread somewhere unexpected in Jensen Huang’s First X Post Reveals a Second Layer of AI Governance. What reads at first like the same tired open-versus-closed-model spat turns into a blunter question — who owns the market sitting under the models? Compute. Energy. Talent. Cloud access. The right to compete. Pile those into too few hands and an “open” ecosystem is open in name only. And that’s her real move: she drags governance away from how individual systems behave and toward the plumbing that decides who even gets to build, compete, and ship in the first place.
Cristina Teleki takes that physical constraint straight to energy in Nuclear Energy for AI – Two (Bad) Strategies. AI can scale in a hurry. Power systems cannot simply do the same on command. And once electricity, grids, cooling, location, and sustainability enter the equation, scaling AI stops looking like a purely technological decision. Some solutions remove one constraint only to spawn another. Scale, it turns out, is not a neutral goal.
Efstathios Iliopoulos and Olivia Zumbach find a similar mismatch in healthcare with Investment Is Not Impact, where investment in medical AI is enormous and the expectations around it are larger still. But impressive performance in controlled settings does not automatically translate into better care in hospitals and clinics. Between the model and the patient sit workflows, clinicians, procurement, regulation, evidence, trust, and institutions that keep their own pace. The problem isn’t progress. It’s translation.
Then the problem moves into law.
Marina Danielyan’s The Contract Nobody Signed: AI Agents and the Limits of Agency Law starts from an assumption so ordinary we barely notice it: contracts are made by people. Intent comes from somewhere. Consent belongs to someone. Responsibility can eventually be traced back to an actor. Autonomous systems make every one of those assumptions less comfortable. Once software begins negotiating, accepting, or acting with meaningful autonomy, legal concepts that looked settled are suddenly asked to do work they were never designed for.
Feyisayo Lari-Williams pushes that further in When AI Becomes Your Shopping Agent: Rethinking Agency. Systems are beginning to search, compare, decide, and transact on our behalf, and at that point AI is no longer simply supporting human judgment — parts of that judgment are being handed over. The trouble is that consumer and legal frameworks still largely assume a human being somewhere in the loop who understands the choice being made. What happens when that assumption stops holding?
That is where the eight articles in this issue finally meet.
The bottleneck keeps moving. It slides from the model to the organization, from the algorithm to the factory, from compute to the electricity bill, from investment to what actually happens at a patient’s bedside — and then from an AI-generated action to some court squinting at it, trying to work out what the thing even means. And now from “open” models to the concentrated infrastructure and markets humming underneath them. It never sits still.
AI doesn’t fail or succeed in a vacuum. It walks into systems that already have histories, constraints, incentives, and baked-in assumptions of their own. Grids hit hard limits. Organizations drag legacy processes behind them. Factories pile up operational mess. Healthcare has to turn a benchmark score into actual care. Markets live and die on access to scarce resources. And law? Law still expects a name, an intention, someone to hold responsible.
And AI is beginning to push against all of them.
Maybe that’s the more interesting story now. Not simply that AI is becoming more powerful, but that its progress is exposing, one system at a time, exactly where the world around it can no longer keep up.
My thanks to every contributor, reviewer, editor, and volunteer who made this issue what it is. Every month, I’m reminded that the most interesting thinking about AI rarely comes from inside a single discipline. It happens at the edges — where fields meet, assumptions collide, and the questions get harder.
And a special thank you to this month’s three Featured contributors. Anne Schwerck, in The Faithfulness Gap: Explainability in Medical AI Agents, takes us into one of medical AI’s hardest problems: whether an explanation that sounds convincing actually tells us how an AI system reached its conclusion. Elena Pasquali shifts the sovereignty debate toward a different question: as increasingly autonomous AI agents access and act on data, where is control over that data actually enforced? And Dr. Phoebe Koundouri, together with Angelos Alamanos and Georgios Feretzakis, looks at Europe’s environmental and water data from another angle: having vast quantities of open data means little for AI if that data is not sufficiently interoperable and machine-ready. Three different fields, but the same reminder. Sometimes the constraint is not the intelligence of the model. It is everything the model has to encounter once it leaves the lab.
