Two Glimpses Into the Same Machine
Written by Louize Clark
Over the course of just one hundred days, Anthropic unintentionally gave the world two rare glimpses into the future of artificial intelligence. Most people saw two unrelated news stories. I believe they were looking at the same machine from opposite directions. One revealed the operational infrastructure surrounding modern AI. The other offered an early view of the internal architecture emerging within it.
Neither was meant to happen. For business leaders, the significance isn't the technology itself. It's that these events exposed parts of an infrastructure that organisations are increasingly relying upon without ever consciously deciding to build. Every industrial revolution has created a new infrastructure that businesses eventually forgot they depended upon. AI is the first time we are watching that infrastructure form in real time — which means, for a short window, we get to understand it before it becomes invisible.
The Evidence
On 31 March 2026, a routine software update to Anthropic's coding tool, Claude Code, was published with a debugging file that should never have left the building. It pointed to the tool's complete source code — roughly half a million lines, covering how the system manages memory, handles permissions, and orchestrates the tools around it. Anthropic confirmed it was human error in a release process, not a breach; no customer data, credentials, or the model itself were exposed. What became visible instead was the scaffolding — the operational layer that turns a language model into something capable of acting inside a business.
Just over three months later, on 6 July 2026, Anthropic published research pointing somewhere else entirely: not the infrastructure around the model, but a pattern that appears to exist inside it. Their paper describes a small, distinct zone of internal activity, a fraction of the system's total processing that behaves, functionally, like a shared workspace, echoing an established theory of how human attention and awareness work. Anthropic are careful and consistent about what this isn't: it is not a claim of consciousness, and they say so themselves. What it is, though, is a tool for seeing what a model is holding "in mind" before it produces an answer.
Two disclosures, three months apart, neither intended. I don't read them as two stories about Anthropic. I read them as one story about all of us: organisations are becoming dependent on two infrastructures simultaneously one they can see and influence, and one they cannot.
An Operational Environment, Not a Software Tool
For years, the conversation around AI has centred on the model itself. Which one is smarter. Which one reasons better. Which one is faster or cheaper. Those questions still matter, but they are becoming less important than a different one: what sits around the model, and what, increasingly, sits inside it.
Modern AI, in any organisation using it seriously, has stopped being a piece of software you switch on and off. It has become an operational environment — something an organisation now works inside, the way it works inside its finance systems or its supply chain, whether or not anyone signed off on that shift. It is quietly embedded in the daily operations of organisations that never made a conscious decision to become dependent on it.
I'd go further than "infrastructure," actually. What I think we're really watching form is something closer to organisational cognition. Businesses have always thought through people — through judgement, memory, and conversation carried between colleagues. Increasingly, they think through models, search, retrieval, memory systems, automation, recommendation engines, agents, and knowledge bases instead. That isn't software any more. It's how the organisation thinks. And once thinking itself is distributed across tools an organisation doesn't fully understand, "AI adoption" stops being the right frame for what's actually happening.
That's also why intelligence, in this context, is misleading if you imagine it sitting neatly inside one model. It no longer sits in a single place. It is becoming distributed across models, software, workflows, people, and operational memory — which means understanding "the AI" tells you almost nothing about understanding the organisation that has come to depend on it.
Why This Is a Leadership Question, Not an Engineering One
Very few organisations are building foundation models. Almost every organisation, however, is becoming dependent on them — and that dependency rarely arrives through a board decision. It arrives through a software update. A CRM starts scoring customers automatically. An accounting platform begins forecasting cash flow. A recruitment system starts ranking applicants. Staff quietly adopt AI assistants because they make the working day easier. None of these, taken individually, look like strategic choices. Collectively, they have become one.
This is where years spent moving across industries, roles, and functions has shaped how I read moments like this. Organisations rarely change because of new technology. They change because of new behaviour — and behaviour shifts long before governance catches up with it. The March leak and the July research are both, in their own way, evidence of exactly that gap: two disclosures that reveal how far ahead the behaviour has already run, and how little most organisations have paused to notice.
Together, they point to a strategic reality that boards, policymakers, and business leaders can no longer afford to overlook. Organisations are building operational dependency on technology whose internal workings — both around it and within it — are still being explored, including by the companies that created it.
The Question Worth Asking
The question is no longer "should we use AI?" That question is already out of date. The more useful one is: how dependent have we become, and do we understand what we're dependent on?
That is not a technical question. It is a governance one. It's about knowing which parts of the business would be exposed if an underlying system changed, failed, or turned out to work differently than assumed. It's about recognising that interpretability research — however academic it sounds — has direct implications for how these systems can eventually be audited, monitored, and trusted at scale.
I suspect that within five years, boards will spend less time asking which AI model they use, and more time asking how much of their organisation's operational memory now exists outside the organisation itself.
Where This Leaves Us
History rarely tells us when we are living through an infrastructure shift. People didn't describe themselves as entering the electricity era, or the internet era. They simply adapted until the infrastructure became invisible.
I believe the same thing is happening with artificial intelligence. The difference, this time, is that we have the opportunity to watch the infrastructure forming while it is still visible.
The organisations that understand it now won't simply adopt AI more effectively. They'll understand the systems they are becoming dependent upon before that dependency disappears into everyday operations, indistinguishable from how the business simply works.
That, to me, is Infrastructure Intelligence™ — the discipline of understanding the invisible operational systems that increasingly shape how organisations think, decide, and operate, before those systems become too familiar to question.
Further information on the Invisible AI Infrastructure™ framework and Infrastructure Intelligence™ is available at aipolicies.uk.