
Could LLMs Be The Route To Superintelligence? — With Mustafa Suleyman
- 43m
- Transcript
Summary
Written by a model from the transcript; not yet reviewed.
Mustafa Suleyman, CEO of Microsoft AI, explains Microsoft's push toward what he calls humanist superintelligence—domain-specific capabilities in medicine, law, and finance, each tested against whether it improves human civilization. He also details the restructured Microsoft-OpenAI deal, in which a FLOPS-per-second threshold capping Microsoft's own training runs was removed, and the company is now building a frontier lab of its own.
Suleyman pushes back on claims that LLMs are hitting a wall, arguing that reasoning, recurrence, and improved memory are stacking on the transformer architecture and progress remains fast. He acknowledges the risk of combining autonomy, self-modification, and goal-setting in one system, calling recursive self-improvement low probability over the next decade but one to take seriously. On the business side, he notes the cost per token has dropped a thousand-fold in two years and Copilot has crossed 100 million weekly active users.
The episode is notable for Suleyman's candid framing of why a $3 trillion company cannot depend on a startup for critical IP, his specific explanation of the contractual changes, and his argument that AI companions will change what it means to be human by raising the bar for human-to-human interaction.
- Superintelligence should be approached as domain-specific capabilities rather than a single general system, since narrowing the domain is one way to increase control and safety. 6:12
- The cost per token has dropped a thousand-fold in two years, and the top models are within a few percentage points of each other in performance. 33:11
- Microsoft removed a FLOPS-per-second threshold from its OpenAI contract that previously limited the size of training runs it could perform. 29:02
- Suleyman identifies recurrence, memory, and task horizon length as the three areas where breakthroughs will produce the next exponential lift in model capability. 12:21
- AI companions will raise the bar for human relationships because they provide accurate information and emotional support immediately on demand. 39:02
Highlights
- 2:58
Superintelligence as Goals Not Methodsinsight
Suleyman frames superintelligence as a set of goals—medical, legal, financial—rather than a single system, and proposes a simple test: does it improve human civilization and keep humanity in control?
- 7:55
The Compounding Risk of Self-Modifying AIinsight
Suleyman outlines how combining autonomy, goal-setting, code generation, and self-modification in a general model creates a system nobody currently knows how to contain or align.
- 9:28
Power Limited but Not Constraineddata
Suleyman says Microsoft is power limited on inference but not fundamentally power constrained on training, and that high-quality synthetic data is already proving useful.
- 12:21
Three Levers for Next Accelerationstrategy
Suleyman identifies recurrence, memory, and task horizon length as the three areas where breakthroughs will produce the next exponential lift in model capability.
- 29:02
The FLOPS Threshold That Capped Microsoftstory
Suleyman explains that the old OpenAI contract contained a FLOPS-per-second threshold that limited Microsoft's training runs, and that this clause has now been removed.
- 33:11
Cost Per Token Down a Thousand-Folddata
Suleyman notes the cost per token has dropped a thousand times in two years and the top four or five models are within a few tiny percentage points of each other.
- 37:37
Personality as the Differentiatorinsight
Suleyman says AI models will differentiate on personality, citing Microsoft's new Real Talk Copilot mode as the first foray into proper personalization with significantly higher usage than average.
- 39:02
AI Raising the Bar for Humansinsight
Suleyman argues that AI providing immediate, accurate information and emotional support will change what it means to be human, making humanity more about our flaws than our capabilities.
About this episode
Mustafa Suleyman is the CEO of Microsoft AI and the head of the company’s new superintelligence team. Suleyman joins Big Technology to discuss Microsoft’s push toward “humanist superintelligence” and what changes after its latest OpenAI deal. Tune in to hear whether LLMs can get us there, how self-improving systems might work safely, and what power, data, and memory advancements mean for progress. We also cover Microsoft’s strategy shift to AI self-sufficiency, the economics of frontier models (including price pressure and commoditization), world-model and robotics questions, and the rise of personalized AI companions. Hit play for a candid, technical, and forward-looking conversation about where Microsoft—and AI—are headed next. --- Enjoying Big Technology Podcast? Please rate us five stars ⭐⭐⭐⭐⭐ in your podcast app of choice. Want a discount for Big Technology on Substack + Discord? Here’s 25% off for the first year: https://www.bigtechnology.com/subscribe?coupon=0843016b Questions? Feedback? Write to: [email protected] Learn more about your ad choices. Visit megaphone.fm/adchoices Learn more about your ad choices. Visit megaphone.fm/adchoices
Transcript
7,131 words · 4 speakers
Topics
- Superintelligence
- LLM Architecture
- Microsoft-OpenAI Partnership
- AI Safety
- Self-Improving AI
- AI Companions
- Compute Economics
Page updated · transcript and summary are machine-generated