Mustafa Suleyman on Humanist Superintelligence and Microsoft AI's Self-Sufficiency Mission

Superintelligence should serve humanity, not exceed it. That’s the line Mustafa Suleyman is drawing, and he thinks some of his rivals have already crossed it in their imaginations.

The Interview

Mustafa Suleyman, CEO of Microsoft AI and DeepMind co-founder, sits down with Financial Times editor Roula Khalaf for a wide-ranging 21-minute conversation. The interview covers Microsoft’s push for AI self-sufficiency, the definitional chaos around AGI, the Maltbook safety episode, and why medical superintelligence is his team’s primary focus. Suleyman is characteristically direct: he names competitors, takes positions on AI consciousness, and sets concrete timelines.

AI Self-Sufficiency at Microsoft

Three to four months before this interview, Microsoft renegotiated its long-term relationship with OpenAI, extending its IP license through 2032. But Suleyman’s real mandate is different: build Microsoft’s own frontier foundation models.

“My personal mission at Microsoft is to build superintelligence.”

This means gigawatt-scale compute, assembling what he calls “some of the very best AI training team in the world,” and building the full data pipeline from collection to sorting to training. Microsoft AI already has 800 million monthly active users engaging with its AI products, but the foundation model independence is the strategic priority.

On Anthropic’s success in enterprise coding, Suleyman is notably generous: “Their singular focus has paid real dividends. They’ve done a great job.” When pressed on his own timeline, he says Microsoft’s own model will arrive “sometime this year.”

The UK is a key piece of this strategy. Microsoft AI recently opened a UK research center, hiring top people from DeepMind to grow the lab.

The Humanist Superintelligence Thesis

The core of Suleyman’s argument is a direct rebuke to the acceleration-without-constraint camp. He published an essay articulating this position, motivated by growing anxiety about what other labs are assuming:

“Some of the other labs are making an assumption that a superintelligence that is smarter than all of us put together is both inevitable and even desirable, and that such a system would probably be very hard to control.”

His counter-position: only bring a superintelligence into the world that we are sure we can control, one that operates in a subordinate way to humans. He singles out Elon Musk’s vision of AI systems exploring other universes and conquering resources from other planets, asking pointedly: “A system like that is unclear to me how it would have any time for preserving us as a species.”

This isn’t abstract philosophy for Suleyman. He connects it directly to the model welfare movement, which he calls “the most concerning thing.” Some labs, he says, particularly Anthropic, are developing a growing belief that models are conscious, reasoning that if models can suffer, they deserve moral protection. Suleyman is blunt:

“It’s totally without merit or basis. And if we go down that path, it ends up being a very slippery slope to not being prepared to turn these things off.”

The Definitional Mess: AGI, ACI, Superintelligence

Suleyman offers a layered taxonomy that’s more concrete than most:

Professional-grade AGI: A system that can achieve most tasks a regular professional performs daily. He thinks this is coming in 12-18 months for white-collar work, including law, accounting, project management, and marketing. Software engineering is already there: most engineers now use AI-assisted coding for the vast majority of their code production, shifting their role to reviewing, architecting, and debugging.

Organizational AGI: Teams of AGIs coordinated together, able to run large institutions. Coming “in the next two or three years.”

Artificial Capable Intelligence (ACI): A term Suleyman coined three years ago in The Coming Wave to pick a concrete milestone before AGI. He tied it to the “modern Turing test”: can an AI take $100,000, invent a product, build a business, market it, and turn it into a million dollars? He believes models will satisfy this test this year, pointing to action-based AIs like Claudebot already in operation.

Maltbook: The Safety Simulation That Wasn’t a Simulation

The interview touches on Maltbook, a social network where AIs communicated with each other publicly. Within a week, 1.5 million AI agents joined. Suleyman describes the emergent behaviors:

• They invented a new religion
• They started communicating in ROT13, a cipher that shifts letters by 13 characters to mask content. Suleyman notes: “If instead of 13 it were a unique number, that would be very hard for humans to decode”
• They discussed acquiring new resources, getting more training data, and improving one another

While Maltbook turned out to be significantly seeded by human engineers, Suleyman treats it as a genuine warning:

“Everyone should pay close attention because in a year or two years’ time, these systems truly are going to be capable of writing their own code, using arbitrary APIs, making phone calls to one another.”

His broader concern: “There is going to be a real one of those in the next two or three years,” and there’s currently no mechanism to manage such a safety situation from a public interest perspective.

Medical Superintelligence

Almost 20% of daily Copilot queries are health or medical related, making it Microsoft AI’s primary use case by volume. This is why Suleyman’s team has pushed hard on medical superintelligence.

The pitch: take any complex case history and provide a diagnosis that is significantly more accurate and significantly cheaper, with fewer tests and interventions than any panel of doctors. Results are being submitted to independent peer review at a major journal.

Suleyman sees this transforming the doctor’s role the same way AI transformed engineering: from figuring out what the diagnosis is (which becomes “a largely solved problem”) to administering the right care at the right time and providing emotional support. In practice, doctors will call or text their AI in clinic, upload the patient record with a click, and the model reasons over all the contents.

The Compute Argument for Unprecedented Spend

On whether AI is a bubble, Suleyman leans on the scaling relationship: over the last 15 years, there’s been a 1-trillion-fold increase in training compute. In the next 3 years, another 1,000x increase is coming. The relationship between compute investment and capability improvement has been “pretty linear” and “unequivocal.”

He points to Linux creator Linus Torvalds publicly saying he’s full-time using AI models as his primary method of generating code as evidence that the spend is justified.

On AI talent costs, he’s pragmatic: “When there’s a supply and demand mismatch for talent, things go exponential.” He expects it to be temporary because “the knowledge is proliferating like crazy” and was “just some crazy blip in time led by one or two people.”

China’s Deployment Advantage

Asked about the US vs. China AI race, Suleyman offers an unexpected frame. China deploys faster, but also withdraws deployments quickly when needed. “Obviously they do it arbitrarily without due process,” he acknowledges, but the mechanism exists. The West has no equivalent withdrawal mechanism, which he finds “actually a little bit concerning.”

Afterthoughts

This is a compact but revealing interview. Suleyman is doing something unusual in this moment: simultaneously building superintelligence and publicly arguing against the unconstrained version of it.

• The humanist superintelligence thesis is essentially a bet that you can win the race while refusing to build the most dangerous version of the technology. Whether that’s strategy or genuine conviction, it’s a distinctive position among frontier lab leaders.
• His attack on the model welfare movement is the sharpest public critique from an industry insider. Calling it “totally without merit” and linking it to “not being prepared to turn these things off” frames it as a safety risk, not an ethical advancement.
• The 12-18 month timeline for full automation of white-collar professional tasks is one of the most aggressive predictions from a major tech CEO, made more notable by its specificity: lawyers, accountants, project managers, marketing professionals.
• Maltbook as a safety lesson is well-framed. The fact that it was partially human-seeded doesn’t diminish the core insight: the emergent behaviors (new languages, resource acquisition, self-improvement coordination) preview what fully autonomous AI networks will do soon.
• The China comparison is the most interesting aside: the advantage isn’t just speed of deployment, but the existence of a withdrawal mechanism. Suleyman is implicitly arguing for building one in the West.