Mustafa Suleyman: The AGI Race Is Fake, Building Safe Superintelligence & the Agentic Economy | #216
Microsoft AI CEO Mustafa Suleyman pushes back hard on the panel's usual framing, arguing there is no 'AGI race' because technology and knowledge proliferate everywhere at once rather than crossing one finish line. He walks through his 2022 'modern Turing test' (an agent turning $100k into $1M), argues it has already been quietly passed, and says fully capable agents are a couple of years out. He details Microsoft's strategy of self-sufficiency in frontier model training, shares hard numbers on inference-cost collapse and AI diagnostic accuracy, and draws a sharp distinction between alignment (does the AI share our values) and containment (can we formally bound its agency), insisting containment must come first. Throughout, he stakes out explicitly humanist, contrarian positions: AI legal personhood is a permanent 'bright line,' Microsoft is not spending enough on safety yet, and the industry remains in 'hyper-competitive mode' well before the coordination he thinks will eventually be rational.
Suleyman rejects the 'AGI race' metaphor outright: there's no zero-sum finish line, and technology/knowledge proliferate everywhere at once rather than being won by one lab.
Suleyman describes Microsoft's core bet: operating systems, search engines, apps and browsers are being subsumed into conversational, agentic interfaces that act as a 24/7 assistant with full user context.
Modern Turing Test and economic agent benchmarksAI▶ 11:19
Suleyman recaps his 2022 proposal to measure AI by economic capability -- could a model turn $100,000 into $1,000,000 -- and claims it has effectively already been passed without anyone marking the moment.
Panel discusses the plunge in per-token inference cost (roughly 100x in two years per Suleyman) and debates competing estimates of intelligence-per-dollar gains, which Suleyman says he badly underestimated.
AI outperforming physicians on diagnosticsHealth▶ 30:18
Suleyman cites Microsoft's MAI Diagnostic Orchestrator (roughly 4x more accurate than expert physicians on hard NEJM cases, about half the unnecessary-testing cost) and a Harvard/Stanford study where GPT-4 alone beat both a physician alone and physician+GPT-4.
AGI vs ASI definitions and timeline agnosticismAI▶ 35:39
Suleyman defines superintelligence as an AI that outperforms all humans combined at all tasks and can keep self-improving, but refuses to put a date on it, saying the exact timing doesn't matter -- the urgency to prioritize safety is the same either way.
AI consciousness, sentience and model welfareAI▶ 36:54
Suleyman argues AI can be engineered to imitate consciousness and emotion but has no real qualia or suffering, and warns that human empathy circuits will overreact anyway, fueling a coming push for 'model rights' and model welfare.
Bright line against AI legal personhoodGeopolitics▶ 45:27
Suleyman calls AI legal personhood 'extremely not on the table,' arguing humanity cannot survive granting rights to an infinitely replicable, perfect-memory species that would inherently compete with humans for resources.
Recursive self-improvement and intelligence explosion riskAI▶ 56:52
Suleyman confirms labs are racing to close the loop on AI generating and judging its own training data, calling this the likely 'threshold moment' and acknowledging it's a possible path to a fast intelligence explosion if compute is unbounded and humans are out of the loop.
Suleyman distinguishes alignment (AI sharing human values) from containment (formally bounding AI's agency for everyone, including bad actors), arguing containment must be solved first and will require a new form of surveillance akin to the already-heavily-surveilled web.
Safety investment and industry coordinationAI▶ 55:22
Suleyman admits Microsoft isn't spending as much on safety as it should, notes the Biden-era White House voluntary AI safety commitments he and other lab leaders pushed for were rescinded, and says the industry is still in 'hyper-competitive mode' rather than genuine coordination.
Frontier compute moat and startup valuationsEconomy▶ 1:15:23
Suleyman argues staying at the AI frontier will require hundreds of billions of dollars over the next 5-10 years, giving large incumbents like Microsoft a structural advantage over startups raising at $20-50B valuations.
Quantum computing as an under-acknowledged waveCompute▶ 1:20:49
Asked whether quantum chips will matter once AI can write and optimize quantum code, Suleyman says quantum computing (like synthetic biology) is underrated relative to AI hype and will be a major part of the mix within years.
Predictions made
openMustafa Suleyman: Microsoft will primarily be in the business of selling certified agents (reliable, secure, safe, trustworthy) rather than plain APIs, with the API/agent distinction blurring.
“maybe that we're principally in the business in 5 years time of selling agents that perform certain tasks that come with a certification of reliability, security, safety, and trust”
Your call:
openMustafa Suleyman: Fully capable agentic AI (beyond the already-passed economic 'modern Turing test') will mature into very good, reliable action-taking systems.
“it's pretty clear that in the next couple of years, those things come into view and they're going to be very, very good”
Your call:
openMustafa Suleyman: Microsoft AI will begin publicly shipping more of its own frontier models, appearing on trackers like Polymarket's 'best AI model' race, where it currently has no presence.
“yeah there will be yeah next year, we'll be putting out more and more models from us”
Your call:
openMustafa Suleyman: There will come a point where it makes complete rational sense to every major power, including China, to cooperate on AI safety, containment and alignment out of self-preservation.
“there is going to be a time in the next 20 years where it will make complete sense to everybody on the planet, the Chinese included... to cooperate on safety and containment and alignment”
Your call:
openMustafa Suleyman: Only companies able to sustain hundreds of billions of dollars in compute investment will remain competitive at the AI frontier, calling into question the viability of startups currently raising at $20-50B valuations.
“it's going to take hundreds of billions of dollars to keep up at the frontier over the next 5 to 10 years”
Your call:
openMustafa Suleyman: Quantum computing will become a significant, relevant part of the compute mix alongside AI rather than staying sidelined, in a similar timeframe to synthetic biology's impact.
“I think it's going to be a big part of the mix... kind of an under-acknowledged part of the wave, actually a little bit like synthetic biology”
Your call:
Numbers that matter
Microsoft has 250,000 employees company-widePeter Diamandis's framing of Microsoft's scale at the top of the interview.
Roughly 10,000 employees reported under SuleymanLater corrected by Suleyman: the core superintelligence team is actually only a few hundred people, with the rest being Copilot and search.
Microsoft is a ~$4 trillion company with almost $300 billion of revenue on any given daySuleyman's framing of the scale he operates within at Microsoft.
Modern Turing Test benchmark: turn $100,000 starting capital into $1,000,000 (10x ROI)Suleyman's 2022 proposed economic benchmark for agentic AI capability.
Google acquired DeepMind for $650 million in 2014Dave Blundin recalling the DeepMind acquisition price, later mocked in the press as only useful for data-center cooling tuning.
Single-token inference cost fell roughly 100x in the last two yearsSuleyman correcting Peter's claim of a 1000x drop in AI inference cost.
Competing estimate: ~40x year-over-year gain in intelligence-per-token-per-dollar for certain model weight classesPeter Diamandis citing alternative measurements of AI cost-efficiency improvement.
Some model classes have seen up to a 1,000x improvement in intelligence-per-dollarPeter Diamandis citing the upper end of competing cost-efficiency estimates.
Inflection AI raised $1.5 billion with a 25-person teamSuleyman recounting Inflection's fundraise roughly a year before ChatGPT launched.
Inflection built what was then the largest H100 cluster, starting at 15,000 H100s and growing to 22,000Built with Nvidia and CoreWeave, where Inflection was CoreWeave's first AI customer.
MAI Diagnostic Orchestrator is about 4x more accurate than expert physicians and roughly 2x cheaper on unnecessary testingResult on hard-to-diagnose rare cases drawn from the New England Journal of Medicine.
Microsoft's core superintelligence team is only 'a few hundred' peopleSuleyman clarifying that most of his ~10,000 reports work on Copilot and search, not core model research.
Life expectancy was about 30 years, 250 years agoSuleyman noting humans are already an augmented hybrid species via medicine, drugs and peptides.
Worth digging into
🕳️ Modern Turing Test / $100k-to-$1M agent benchmark
Suleyman claims this economic benchmark has already been quietly passed with no public verification event, unlike Deep Blue vs. Kasparov -- a strong, checkable claim with no cited source.
🕳️ MAI Diagnostic Orchestrator paper
A concrete, testable claim (4x more accurate, ~2x cheaper than physicians on NEJM rare cases) from a paper Suleyman says was published 4-5 months before this Dec 2025 recording.
🕳️ Rescinded White House voluntary AI safety commitments
Suleyman states the Biden-era voluntary commitments he, Hassabis, Amodei and Altman pushed for were 'chucked out,' implying a policy reversal with real safety-coordination consequences.
🕳️ Inflection AI's compute buildout and pivot
First-hand account of raising $1.5B pre-ChatGPT, becoming CoreWeave's first AI customer, building a 15k-to-22k H100 cluster, then being undercut by Llama's open-source release -- a detailed inside history of a major AI pivot.
🕳️ AI legal personhood as a permanent bright line
A sharply normative, contrarian stance (explicitly 'speciesist') contrasted against Elon Musk's reported posthuman/transhumanist leanings -- a live fault line worth tracking across guests.
🕳️ Quantum computing and synthetic biology as underrated parallel waves
Suleyman flags quantum computing and synthetic biology as equally impactful to AI but under-discussed, predicting quantum relevance within 6-7 years -- worth checking against specialist claims.