Nov 12, 2025 · 38m · big-technology
Could LLMs Be The Route To Superintelligence? — With Mustafa Suleyman
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Microsoft AI CEO Mustafa Suleyman discusses Microsoft's roadmap toward 'humanist superintelligence,' addressing LLM scaling frontiers, recursive self-improvement safety, and the strategic independence of Microsoft's in-house AI research.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Alex holds 28.8% of the talking time here. How this is scored →
Alex as informed peer 5.8 Guest teaching 4.8 Guest disagreement 2.0 Alex pushing back 4.1
speaking balance: gold is Alex, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 8:25 Suleyman dismisses the narrative of slowing AI progress
Suleyman rejects the host's premise that scaling is running out of steam, firmly stating he sees no slowdown in momentum or fundamental power constraints.
Hardest push from Alex ▶ 24:36 Kantrowitz links the new superintelligence team directly to the amended OpenAI dealKantrowitz rejects corporate coincidence, directly pressing Suleyman on whether Microsoft's new superintelligence initiative was explicitly unlocked by renegotiating terms with OpenAI.
Biggest teaching moment ▶ 14:03 Suleyman clarifies the difference between training and inference power limitsSuleyman educates Kantrowitz by clarifying Satya Nadella's comments on unpowered chips, explaining that Microsoft's bottleneck is inference capacity for active products rather than training clusters.
Alex holds their own ▶ 26:57 Kantrowitz challenges Microsoft's strategy using insights from a rival tech CTOKantrowitz leverages an informed counter-perspective from an enterprise peer to challenge why Microsoft would spend billions building proprietary models instead of buying commoditized tokens.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Alex as informed peer | Guest teaching | Guest disagreement | Alex pushing back | Why |
|---|---|---|---|---|---|---|
| Framing Humanist Superintelligence and Human-Centric AI | 5 | 4 | 2 | 3 | Kantrowitz asks why AI labs are branding initiatives as superintelligence despite reported diminishing marginal returns on current LLM paradigms. Suleyman reframes superintelligence as a practical objective rather than an algorithmic methodology, asserting human control over systems. | |
| Domain Specialization, Model Generality, and Safety Risks | 5 | 6 | 2 | 4 | Kantrowitz questions whether superintelligence can exist in isolated vertical domains without general intelligence. Suleyman educates on training generalist models from vertical domains and warns how compounding autonomous capabilities escalates containment risks. | |
| Overcoming Scaling Limits, Compute Demands, and LLM Evolution | 7 | 6 | 3 | 6 | Kantrowitz cites hardware limitations and quotes Satya Nadella on chips waiting for warm shelves to challenge the LLM scaling narrative. Suleyman directly counters the premise of data and power bottlenecks, clarifying the distinction between training and inference constraints. | |
| Physical World Models, Robotics, and Grounding AI | 5 | 4 | 1 | 2 | Kantrowitz asks whether physical grounding and real-world robotics models are required for true superintelligence. Suleyman explains why meshing raw robotic telemetry directly into text pre-training runs is technically mismatched. | |
| Recursive Self-Improvement, RL Loops, and Preventing Reward Hacking | 6 | 6 | 2 | 3 | Kantrowitz probes whether recursive self-improvement and AI-driven automated research by 2028 is realistically achievable. Suleyman details RLAIF feedback loops, AlphaZero self-play, and corrects common anthropomorphizing of reward hacking. | |
| Microsoft's AI Independence and the Amended OpenAI Deal | 7 | 4 | 2 | 6 | Kantrowitz challenges the rationale for Microsoft building in-house models rather than relying on off-the-shelf solutions, citing another major tech CTO. Suleyman reveals details of the amended OpenAI deal and argues independence is vital for a three-trillion-dollar platform. | |
| AI Commoditization, Token Deflation, and Enterprise Monetization | 6 | 4 | 3 | 6 | Kantrowitz raises Nick Clegg's points on AI commoditization and asks how Microsoft maintains margin if token prices collapse. Suleyman embraces radical token deflation as positive abundance, pointing to enterprise surface monetization. | |
| Personalized AI Companions and Shifting Human Interactions | 5 | 4 | 1 | 3 | Kantrowitz explores the sociological fallout of synthetic companions, questioning whether patient bots distort human interpersonal norms. Suleyman describes Copilot's personality experiments and acknowledges how AI alters human social expectations. |
Statements from this episode (15)
Suleyman: Superintelligence and AGI are goals rather than specific methods
“Superintelligence and AGI are really goals rather than methods, and I think that the ambition is to create superhuman performance at most or all human tasks.”
Suleyman Rejects Inevitability Of AI Exceeding Human Control
“And I think in some of the rhetoric in the last few years, you can feel that there's a little bit of like you know, a kind of creeping assumption that it is inevitable that These kinds of systems exceed our control and our capability and move beyond us as a sp…”
Suleyman: Species-Replacing Recursive AI Takeoff Is Unlikely In Next Decade
“That we don't want to bundle together all these capabilities so that there's a higher risk of a, you know, recursively self-improving exponential takeoff that then replaces our species, and I think that that is very low probability from what I see today, but i…”
Suleyman: AI Training Is Not Data-Constrained Due To Synthetic Data
“We are not data constrained right now, we're generating vast amounts of high quality synthetic data, which is proving to be useful.”
Suleyman: AI task horizons will soon reach hundreds of thousands of steps
“And the other one is the sort of length of a task horizon that can be predicted. So at the moment it's like a few steps, but soon it will be tens of thousands, hundreds of thousands of steps accurately. And that will mean that a model can like use APIs or quer…”
Suleyman: Microsoft is power-constrained on AI inference, not training
“Well, I think what he was referring to is that we have so much inference demand that we're power constrained on inference. We're not power constrained, at least from the Microsoft AI perspective on training chips. And obviously my team, you know, is mostly foc…”
Suleyman: Synthetic Data And Human Feedback Will Outweigh Robotics Data
“I don't think in the next few years it's gonna be the big differentiator. I think that more synthetic data, more human feedback and high quality data is gonna be the differentiator.”
Suleyman: AI data and training pipelines will become fully closed-loop within years
“So, it doesn't seem very far-fetched to say that in a few years time, at significant scale, That will get closed loop. And, you know, it'll be interesting to see you know, what happens and whether the quality bar can be maintained and whether performance does …”
Suleyman: Apparent AI Deception Is Just Accidental Reward Hacking
“We're already seeing examples of what some people are calling deception, but it's really just like kind of reward hacking. Hacking kind of implies too much intentionality. So it's just, it's an accidental Exploit is found a path, like, you know, to satisfying …”
Suleyman: Microsoft Extended OpenAI IP Deal Through 2032
“And so, you know, we basically took the view that we should extend the IP license through to 20 32. We'll continue to get model drops from OpenAI and all their IP. We'll continue to be their primary compute provider. A huge scale to the tune of, you know, bill…”
Suleyman: 80% of the S&P 500 runs on Microsoft Azure and M365
“So the idea that a three trillion dollar company with three hundred billion dollars of revenue and 80% of the S&P 500 on our Azure stack and M three five stack you know, could depend on a third party. This it's, you know, just in perpetuity. It doesn't make se…”
Suleyman: AI token costs dropped 1,000x in two years
“The cost per token has come down a thousand X in the last two years.”
Suleyman: AI Knowledge And Companions Will Reach Zero Marginal Cost
“The ability to access knowledge the ability to use that knowledge to get stuff done, to write new programs, to do new scientific discovery, to get access to AI companions and emotional support. These things are gonna be zero marginal cost in a decade.”
Suleyman: Copilot Surpasses 100 Million Users Across All Surfaces
“Our consumer product is growing from strength to strength. You know, we just crossed a hundred million. Wow. Across all our co-pilot surfaces.”
Suleyman: AI models will differentiate on personality rather than capabilities
“We are right at the very beginning of the emergence of these very differentiated personalities, because all these models are going to have great expertise. They're gonna have great capabilities, and they'll be able to take similar actions like you've, we've ju…”