Four Essays in Ten Days. The AI Industry Just Changed Its Mind.

In the last ten days, four of the most influential people in artificial intelligence each published a major essay. Satya Nadella at Microsoft. Mira Murati at Thinking Machines. Demis Hassabis at Google DeepMind. Amjad Masad at Replit.

Different companies. Different audiences. One wrote about enterprise economics, one wrote philosophy, one wrote civilizational governance, and one just published his own company's numbers.

Read them together and they are describing the same shift.

For three years the story of AI was the model. Whoever had the biggest, smartest frontier model won, and everyone else rented it. That story is quietly being retired — by the people who built the models. The new story is about what happens around the model: who owns the knowledge, who controls the learning loop, and whether humans stay in the driver's seat.

Here is each essay, and then what the convergence means for anyone actually running a business.

Nadella: you pay for intelligence twice

Nadella's piece, the Reverse Information Paradox, reframes a classic economics problem. The old paradox — Kenneth Arrow's — was the seller's: to sell information you have to reveal it, and once revealed, you have effectively given it away. AI flips the problem onto the buyer. To make a model useful on your business, you feed it your prompts, corrections, evaluations, and workflows — your proprietary know-how. So you pay once in cash, and again in the knowledge you had to hand over.

His warning: if that learning only ever flows one way, value pools with whoever owns the infrastructure, not the company that created the knowledge. His fix is a "trust boundary" — a hard line around your data, evals, memory, and adapted model weights that nothing crosses without consent. He lists five things enterprises should keep hold of: control, capability, choice, cost, and compounding. Own your evals. Keep your learning inside your own walls. Do not get locked into a single model.

It is a striking argument coming from the company that sells you the models. The core point stands on its own anyway.

Murati: the future worth building is human

Murati's essay is the philosophical one, and it lands in the same place from a different direction. Most AI today, she notes, is trained in a handful of places and then frozen — it never really learns from the people it serves. She argues that is a waste, because the knowledge that actually runs the world is tacit, local, and dispersed: the chef's feel for a recipe, the shopkeeper's read of their customers. You cannot scrape that into one central model.

So the answer is AI that is distributed and customizable — models organizations can fine-tune and own, that keep learning alongside their people. She makes the same point about values: a single place deciding how every AI thinks becomes a single point of power to capture. Better to have an ecosystem of diverse AIs, shaped in many places.

Her throughline: keeping humans in the loop is not a tax on capability. It is a technical problem worth solving, because it makes the whole system both smarter and safer.

Hassabis: the dawning of a new age

Hassabis goes the widest. He argues AGI — a system with the full range of human cognitive abilities — is probably only a few years out, and that its impact is less like the internet and more like electricity or fire: potentially ten times the Industrial Revolution at ten times the speed, opening an era of abundance in science, energy, and materials.

With that upside comes risk that is outrunning our understanding — cybersecurity now, bio and nuclear later, and eventually agentic systems that improve themselves. His proposal is the most centralizing of the four: a U.S.-led Frontier AI Standards Body, modeled loosely on FINRA, that would test frontier-class models against evolving benchmarks, review them before release (voluntarily at first, up to 30 days), and set a shared bar for safety. Smaller players exempt; open to becoming an international standard over time.

He closes on the part technologists cannot answer alone — what a post-scarcity world does to work, meaning, and the human condition — and says those questions belong to everyone.

Masad: the self-driving company

Masad's is the one with receipts. At Replit, he reports, agents are now woven through the whole company — not just writing code, but reviewing pull requests, investigating incidents, analyzing business data, triaging support, researching sales accounts, even improving the agent systems themselves.

The numbers: holding the engineering team constant, roughly 2.9x the code output, with review latency flat and quality steady — because an agent does a first pass on reviews and only pulls in a second human when the risk warrants it. He calls the result the "self-driving company": humans set direction, taste, and responsibility; agents run the execution loops and escalate when judgment is needed. Employees are not replaced so much as promoted — from doing the work to directing it. And his internal agents beat off-the-shelf tools, precisely because they are wired deep into how Replit actually runs.

Where they agree

Four very different essays, and they rhyme:

• The moat is not the model. All four move value away from raw model capability and toward what you own and do around it — your knowledge, your data, your loop.
• Distributed beats centralized. Nadella and Murati say it outright; Masad builds it internally; even Hassabis wants the safety layer coordinated rather than owned by any single lab.
• The loop is the point. Continuous, organization-specific learning — Nadella's compounding, Murati's ongoing collaboration, Masad's self-improving agents — beats a frozen model you rent.
• Humans stay in the driver's seat. Not out of nostalgia, but as the design goal. Direction, judgment, and responsibility stay human.

Two of them — Nadella and Murati — reach for the same economists, Hayek and Polanyi, to argue that the most valuable knowledge cannot be centralized. That is not a coincidence. It is the intellectual center of gravity this whole conversation is settling on.

Where they split

They are not the same essay. The differences matter:

• Scope. Masad is on-the-ground and empirical — here is what happened at my company. Nadella is enterprise economics. Murati is philosophy and interface design. Hassabis is civilization-scale.
• Risk. Hassabis is focused on frontier, catastrophic risk — bio, cyber, self-improvement — and thinks it needs outside oversight. The others worry more about the economic and power risk of centralization, and about losing human judgment.
• Centralize or decentralize? This is the real fault line. Hassabis wants a central body to hold the safety line. Murati explicitly warns against any single locus of alignment. Nadella wants boundaries against the providers. Masad just builds his own.
• When. Masad, Nadella, and Murati are largely talking about the next year or two. Hassabis is looking a few years out and asking what a post-scarcity society even means.

What this means if you run a business

Strip away the philosophy and the governance, and there is a directive hiding in all four essays, aimed straight at you:

Stop treating the model as the thing you are buying. The model is a rented commodity — it gets better every quarter whether you do anything or not. The durable advantage is the layer the frontier leaders are now all pointing at: your proprietary knowledge, your data, your evals, and a system that learns from your actual work and keeps that learning inside your walls.

This is the case we have been making since day one, so we will say it plainly. The companies that win the next few years will not be the ones with access to the best model — everyone has that. They will be the ones who turned their own knowledge into systems they own and operate, instead of pouring it into someone else's.

The most powerful people in AI just spent ten days telling you the same thing from four directions. The moat was never the model. It is what you build around it — and whether you own it.