Microsoft declares it's building "humanist superintelligence" to keep AI safe. Reality check: They're 2 years behind OpenAI, whose models they'll use until 2032. The safety pitch? Product differentiation for enterprise clients who fear runaway AI.
Three Stanford professors just raised $50M to prove OpenAI and Anthropic generate text wrong. Their diffusion models claim 10x speed by processing tokens in parallel, not sequentially. Microsoft and Nvidia are betting they're right.
Good Morning from San Francisco, OpenAI wants taxpayers to guarantee its trillion dollar dreams. Sarah Friar floated federal backstops Wednesday,
Microsoft declares it's building "humanist superintelligence" to keep AI safe. Reality check: They're 2 years behind OpenAI, whose models they'll use until 2032. The safety pitch? Product differentiation for enterprise clients who fear runaway AI.
Claims humanist approach while rivals accelerate. Suleyman admits frontier models still years away.
Microsoft launched its "Humanist Superintelligence" team Thursday, positioning AI safety as competitive advantage just as Washington signals it wants the brakes off.
The timing creates an immediate contradiction. AI chief Mustafa Suleyman insists Microsoft will build "superintelligence" that keeps humans "at the top of the food chain." Meanwhile, the incoming Trump administration's AI czar David Sacks champions unconstrained acceleration. Microsoft needs both messages to work.
That's the bind.
The Breakdown
• Microsoft launches "humanist superintelligence" team, promises AI that keeps humans in control while rivals chase raw performance
• Suleyman admits frontier models are 2 years away; Microsoft depends on OpenAI access through 2032
• Every major AI lab now claims superintelligence focus despite none achieving AGI yet
• Healthcare becomes Microsoft's wedge: diagnostic AI beats doctors 4x, leverages existing enterprise infrastructure
Every major AI player now claims to be building superintelligence. Meta renamed its entire AI division "Meta Superintelligence Labs" in June. OpenAI CEO Sam Altman says his company has already "figured out" AGI and moved on to superintelligence. Anthropic runs a dedicated superintelligence safety team. Even Ilya Sutskever's startup is literally called Safe Superintelligence.
None have achieved artificial general intelligence yet.
The rebrand serves three purposes. It signals ambition to investors burning through billions monthly. It differentiates positioning when everyone's using similar transformer architectures. And it lets companies claim progress without delivering AGI.
Microsoft's "humanist" variant adds a fourth: creating distance from OpenAI while still depending on their models through 2032.
Here's what Suleyman admitted to Fortune: "It's still going to be a good year or two before the superintelligence team is producing frontier models."
Microsoft just renegotiated its OpenAI deal last week, securing model access until 2032 and a 27% stake in OpenAI's new structure. The company needs that runway. Despite hiring Karén Simonyan as chief scientist and poaching researchers from DeepMind, Meta, and Anthropic, Microsoft starts from behind.
The numbers explain why. OpenAI burns about $5 billion annually on compute. Anthropic raised $13.8 billion total. Google's DeepMind has direct access to TPU clusters. Microsoft hasn't disclosed its new team's GPU allocation, but Suleyman called establishing "AI self-sufficiency" a multi-year project.
Until then, Copilot runs on OpenAI. Enterprise customers use OpenAI models. Microsoft's consumer chatbot trails OpenAI's ChatGPT in downloads. The dependency remains absolute even as the companies compete for enterprise contracts and talent.
Suleyman's manifesto reads like it's from 2023's AI safety discourse, not 2025's acceleration moment. He warns against AI systems that "exceed and escape human control." He promises models with "containment" built in. He rejects treating AI as sentient.
But listen closer. "We are accelerationists," Suleyman told Axios. "We want it to go as fast as possible." He acknowledges recursive self-improvement, where AI systems train themselves, delivers the performance gains everyone seeks. Microsoft will pursue it.
The safety positioning becomes product differentiation, not principle. When pressed about competing with less constrained rivals, Suleyman admitted: "It's a difficult one to manage."
Chinese labs don't publish manifestos about keeping humans in charge. They ship models. Nvidia CEO Jensen Huang said Wednesday that China leads the AI race. The competitive dynamics don't care about humanist values.
Buried in Suleyman's announcements: healthcare is Microsoft's first superintelligence target. The company claims its diagnostic AI beats doctors by 4x accuracy at lower cost. It's partnering with Harvard Health for Copilot responses. Suleyman calls medical AI tools "very close" to market-ready.
Healthcare offers Microsoft something OpenAI can't match: enterprise distribution through existing hospital systems running Windows, Office, and Azure. It's regulated enough to reward safety positioning. And diagnostic tools don't need chatbot personalities that might seem sentient.
This isn't about competing with ChatGPT. It's about embedding AI where Microsoft already dominates: enterprise infrastructure. The "humanist" branding helps sell to risk-averse hospital boards.
Microsoft's superintelligence push makes three strategic bets:
First, that safety concerns will return after the first major AI accident or scandal. The humanist positioning becomes valuable when Congress holds hearings about runaway AI systems.
Second, that enterprise customers will pay premiums for "contained" AI over raw capability. Banks, hospitals, and governments need defensible AI decisions more than maximum performance.
Third, that Microsoft can close the capability gap through capital. The company generated $88 billion in free cash flow last year. It can afford to match anyone's compute spending.
The hedge: if unconstrained acceleration wins, Microsoft keeps full OpenAI access through 2032. Suleyman praised the partnership even while declaring independence.
It's expensive insurance. Microsoft pays OpenAI billions annually. But it lets Microsoft claim both safety leadership and frontier capabilities. Heads, Microsoft's values win. Tails, OpenAI's models ensure Microsoft doesn't lose.
The math works until someone actually builds AGI. Then everything changes.
• Microsoft's positioning reveals the industry's real bet: AGI remains years away despite the superintelligence marketing
• The safety-versus-speed tension will define AI winners. Microsoft hedged both sides
Q: What exactly prevented Microsoft from building large AI models before?
A: Microsoft's original OpenAI deal banned pursuing AGI and capped model training size based on FLOPS (calculations per second). The restriction meant Microsoft couldn't train models beyond a specific computing threshold. Last week's renegotiation lifted these limits, allowing Microsoft to build frontier models while keeping OpenAI access through 2032.
Q: How much is Microsoft spending on this superintelligence push?
A: Microsoft hasn't disclosed the team's GPU allocation. For context: OpenAI burns $5 billion annually on compute, Anthropic raised $13.8 billion total. Microsoft generated $88 billion in free cash flow last year, giving it firepower to match anyone's spending. The company already pays OpenAI billions annually for model access.
Q: What's "recursive self-improvement" and why does it matter?
A: It's when AI systems train and improve themselves without human intervention. Suleyman admits this technique delivers the performance gains everyone wants, despite safety risks. Microsoft will pursue it while rivals like Meta and Chinese labs already use it aggressively. It's the fastest path to superintelligence but hardest to control.
Q: When will Microsoft release actual products from this new team?
A: Suleyman says frontier models are "a good year or two" away. Healthcare tools are "very close" to market-ready. The diagnostic AI that beats doctors by 4x accuracy already exists in testing. Until frontier models arrive, Microsoft's Copilot and enterprise products will continue running on OpenAI's technology.
Q: How does Microsoft's approach differ from Meta's or Google's superintelligence efforts?
A: Microsoft promises "containment" features preventing AI from appearing conscious or escaping human control. Meta renamed its entire AI division "Superintelligence Labs" without safety focus. Google's DeepMind has direct TPU access but no public safety manifesto. OpenAI claims it already solved AGI. Microsoft's betting enterprise clients will pay premiums for "safer" AI.
Apple will pay Google $1B yearly to power Siri with a 1.2 trillion parameter AI model—8x more complex than Apple's current tech. The company that owns every layer now rents the most critical one. The spring 2026 target masks a deeper dependency trap.
Sam Altman predicts AI CEOs within years while betting billions on human-centric infrastructure. His Tyler Cowen interview reveals three tensions: monetizing without breaking trust, energy bottlenecks limiting AI, and models that persuade without intent.
Palantir beat earnings but fell 8% at 250x forward P/E, triggering global risk reset. Banking chiefs gave cover for year-end de-risking while AI capex outpaces revenue visibility. When leaders wobble, concentration risk becomes system risk.
Amazon sues Perplexity to block its AI shopping agent from making purchases autonomously. The first major legal clash over whether AI assistants can act as user proxies sets precedent for the $20B question: who controls the future of online commerce?
Get the 5-minute Silicon Valley AI briefing, every weekday morning — free.
