May 11, 2023 · 51m · no-priors
No Priors Ep. 16 | With Mustafa Suleyman, Founder of DeepMind and Inflection
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
In this episode of No Priors, hosts Sarah Guo and Elad Gil interview Mustafa Suleyman, co-founder of DeepMind and Inflection AI, about his journey from grassroots social impact to pioneering artificial general intelligence and empathetic conversational companions. Suleyman shares insights on modular model architectures, the evolution of search interfaces, and the critical need for democratic governance amid exponential technological growth.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 15.8% of the talking time here. How this is scored →
The hosts as informed peer 4.6 Guest teaching 4.6 Guest disagreement 2.4 The hosts pushing back 1.1
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 35:30 Rejecting the consolidated single-AI world view
Mustafa forcefully rejects the mainstream Silicon Valley belief that a few massive autonomous AIs will rule the ecosystem, labeling that thesis completely misguided and fundamentally wrong.
Hardest push from the hosts ▶ 37:31 Elad highlights the risk of hyper-personalized echo chambersElad directly challenges the idealistic view of personalized companion agents by arguing they could act as extreme reinforcement engines for user delusions and polarization.
Biggest teaching moment ▶ 28:15 Mustafa deconstructs legacy search and SEO distortionMustafa educates the audience on why current search engines fail users, showing how Google's ad-incentivized ranking distorted web design into low-signal, high-friction Yellow Pages.
The host holds their own ▶ 20:04 Elad details crystallography physics to contextualize AlphaFoldElad showcases deep first-hand technical mastery as a former biologist, articulating the arduous physical chemistry of protein crystallization to underscore AlphaFold's paradigm shift.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Conflict Resolution and Lessons from the Copenhagen Summit | 3 | 4 | 1 | 0 | Elad asks an open background question regarding Mustafa's international conflict facilitation. Mustafa provides a reflective narrative on the limitations of multilateral diplomacy in Copenhagen, setting up his transition to tech. | |
| Platform Choice Architecture and Societal Behavioral Framing | 6 | 3 | 2 | 2 | Mustafa analyzes platform choice architecture and societal behavioral incentives using Facebook as an analogy. Elad jumps in with strong domain experience building early Facebook apps, contrasting micro-level developer optimization with macro-level societal impacts. | |
| DeepMind's Origins and Defining Artificial General Intelligence | 4 | 6 | 2 | 1 | Mustafa shares the founding history of DeepMind and details Shane Legg's early PhD research defining AGI mathematically across multiple problem domains. Sarah closes with a clarifying question on whether that definition still holds. | |
| Rethinking General Intelligence: Central Routing and Modularity | 7 | 5 | 4 | 3 | Mustafa challenges the industry obsession with monolithic single-agent generality championed by OpenAI and Anthropic, arguing for dynamic routers and modularity. Elad validates and builds on the technical thesis by mapping it directly to brain biology, mirror neurons, and mixture-of-experts architectures. | |
| Navigating the AI Winter and Early DeepMind Milestones | 8 | 4 | 2 | 2 | Mustafa describes the early AI winter when mentioning AGI was taboo and outlines DeepMind's pivot to AlphaFold. Elad shows deep domain expertise as a former biologist, explaining protein crystallography, salt concentrations, and x-ray diffraction to highlight AlphaFold's breakthrough. | |
| Large Language Model Breakthroughs and Founding Inflection AI | 4 | 5 | 2 | 0 | Sarah prompts Mustafa on founding Inflection AI. Mustafa explains his realization around LLM scaling at Google with LaMDA/Meena and the institutional friction that spurred him, Karén Simonyan, and Reid Hoffman to launch their own venture. | |
| Conversational Interfaces and the Evolution of Online Search | 4 | 6 | 4 | 1 | Sarah asks why Inflection prioritizes companionship over pure information retrieval. Mustafa delivers an incisive critique of Google search, describing traditional 10 blue links and SEO optimization as an archaic, ad-driven Yellow Pages conversation. | |
| Designing Pi as an Empathetic Personal Digital Companion | 4 | 5 | 2 | 1 | Sarah asks Mustafa to define Pi. Mustafa details Inflection's vision of personal AIs aligned exclusively with individual users, explaining how empathetic listening, curiosity, and conversational nuance are baked into the product. | |
| Multi-Agent Digital Ecosystems and Future Content Creation | 4 | 5 | 4 | 1 | Sarah brings up publisher anxiety regarding content generation. Mustafa firmly dismisses the popular narrative of a few consolidated super-AIs, arguing instead for hundreds of millions of user-aligned agents interacting as interlocutors. | |
| Algorithmic Filter Bubbles and Democratic AI Governance | 6 | 5 | 4 | 3 | Elad pushes on the threat of extreme echo chambers and feedback loops created by hyper-personalized agents. Mustafa agrees it is the default outcome without proactive intervention, advocating for democratic oversight and calling out tech libertarian resistance to regulation as churlish. | |
| Measuring Emotional Intelligence, Long-Term Memory, and Feedback | 4 | 4 | 1 | 1 | Sarah explores how Inflection measures emotional intelligence and shares her user onboarding experience. Mustafa outlines practical metrics around fluidity and impartiality, noting current memory constraints and future integrations. | |
| Culture and Cadence: Flat Teams and Sprint Cycles | 3 | 4 | 2 | 0 | Sarah inquires about Inflection's internal engineering culture. Mustafa describes their flat team of individual contributors (MTS), absence of middle management, and rapid six-week shipping cadences followed by hackathon sprints. | |
| Compounding Exponentials: Compute Scaling and Model Efficiency | 6 | 6 | 3 | 1 | Elad asks about model scalability limits, hardware bottlenecks, and data asymptotes. Mustafa breaks down the nine-order-of-magnitude compute surge since DQN and cites the Chinchilla scaling laws to argue architectural optimization is just beginning. | |
| The Coming Wave Book and Final Podcast Reflections | 2 | 3 | 1 | 0 | Sarah closes by asking about Mustafa's forthcoming book, 'The Coming Wave'. Mustafa explains using hard deadlines to sharpen thinking on the geopolitical intersections of AI and synthetic biology before wrapping up. |
Statements from this episode (23)
Suleyman: Global governance systems cannot keep pace with exponential technology
“The world's governance systems are not going to keep up with both the exponential challenges that we face from globalization and carbon emitting, but also like technology.”
Suleyman: Platform design choices create choice architectures reframing human connection at scale
“There's a choice architecture here. There are significant design choices which are going to incentivize certain behaviors. Obviously at that point there wasn't really ranking, but even just having a thumbs up or like the choice of, you know, which button you p…”
Hassabis originally wanted DeepMind to simulate the entire global economy
“Demis was way more in the kind of robotics land and sci-fi land. I mean, he was fully thinking that, you know, the way to manage the economy, the way to make economic decisions was to simulate the entire economy.”
DeepMind synthesized 80 definitions of intelligence into one engineering metric
“And it was his work in, I think he, I think for his PhD thesis, he put together like 80 definitions of intelligence. And aggregated those into a single formulation, which was how do we you know, the intelligence is the ability to perform well across a wide ran…”
OpenAI and Anthropic over-rotated on building single-agent systems
“I think in a weird way, it's over rotated the entire field on one aspect of intelligence, which is generality, you know, and I think open AI and then subsequently anthropic and others have taken up this default sort of mantra that like it, all that matters is …”
Central routing across specialized AI subsystems will unlock the field
“I think the key element that is going to really unlock this field Is actually going to be the router in the middle of a series of different systems which are specialized, some of which don't even look like AI at all. They might just be traditional pieces of so…”
Peter Thiel had significant early AI vision but exited DeepMind early
“Peter Thiel, you know, to his credit did actually have significant vision here, although he sold pretty early, I think, and now doesn't seem to be in the game.”
DeepMind's AlphaFold originated from an internal reinforcement learning hackathon
“That's part of the reason why we started the Alpha Fold project, because it was actually my group that was looking around for other applications of DQN-like, AlphaGo-like tools, and in a hackathon that we did one week someone stumbled across, across this probl…”
DeepMind's founding goal was cross-domain transfer learning
“If I were to sort of summarize the core thesis of deep mind, it was that it would be possible to the motivation for generality was that you would be able to learn you know, a rewarding behavior in one environment and transfer in a more compressed or efficient …”
Humanity is nearing the limits of unassisted scientific invention
“We're sort of at the limit of what, you know, the smartest humans in the world are capable of inventing. And we have, you know, very pressing, urgent global challenges, you know, from food supply to water to decarbonization to clean energy,
Transportation, you…”
Suleyman and others aggressively pushed Google to launch LLMs in 2021
“While I was there, sort of in the last year in twenty-twenty-one, I tried pretty hard to get things launched at Google. We were all kind of banging on the table, being like, come on, this is the future, and you know obviously David Luan from Adept was also in …”
Google search is an appallingly painful conversation mediated by blue links
“Conversation is the future interface. And Google is already a conversation. It's just an appallingly painful one, right? You say something to Google, it gives you an answer in 10 blue links. You say something about those 10 blue links by clicking on it. You, i…”
Google inadvertently degraded internet content by optimizing for ads and SEO
“The problem with what Google has sort of, I guess, in a way accidentally done to the internet is that it has basically. Shaped content production in a way that optimizes for ads. And everything is now SEO'd to within an inch of its life.”
Everyone will have a personal AI within the next few years
“Building on all of that, we think that Pi, I think that everyone in the next few years is going to have their own personal AI, right?”
Google should be worried because its traditional search model will fail
“You know, if I was Google, I would be pretty worried because that old school system does not look like it's going to be where we're at in 10 years time.”
Future AI ecosystem will feature billions of models, not a few
“So I really don't subscribe to this view that there's going to be like one or five AIs. I think this is like completely misguided and fundamentally wrong. There are going to be hundreds of millions of AIs or billions of AIs, and they'll be aligned to individua…”
Social platform neutrality was a lie; AI will accelerate echo chambers
“I think that the platforms were never neutral. That was the big lie. And I think that was, frankly, to me, very obvious from the very beginning. The choice architecture is a ranking. It's not a clean feed. Clearly there's billions of bits of content, so you ha…”
Silicon Valley must embrace European-style democratic oversight for AI models
“You know, the Valley probably needs to be a bit more open to the European approach. The reality is that, you know, we have to figure out as a society which bodies we trust to make decisions which influence recommendation algorithms or AI algorithms, right? And…”
Inflection's Pi will soon access real-time web content and citations
“And you know, the other thing is that quite soon it will have the ability to access real time content in the web. So it'll be able to, you know, sort of look up the weather and news and Other kind of fresh content like sports results or provide citations and, …”
Inflection AI achieves state-of-the-art performance on major non-coding benchmarks
“We run. Some of the largest language models in the world. We have state of the art performance across many of the main benchmarks with the exception of coding, because we don't have PI generate code and it's not a priority for us.”
Twelve-week OKR cycles fail; six-week cycles maximize accountability and engagement
“And I found that most people make up the second half of their okay hours anyway, and a 12 week cycle is just too long and BS. So like six weeks is actually perfect and it creates a lot of accountability and a lot of fun.”
Frontier AI training compute grew nine orders of magnitude in nine years
“Over the last decade, the amount of compute that we've used to train the largest models in the world has increased by an order of magnitude every single year. So I went back and had a look at the Atari DQN paper that we published in. And that used just two pet…”
AI scaling and architectural efficiency will advance simultaneously
“Scale, building larger models is definitely going to deliver returns. We're obviously pursuing that. We have one of the largest supercomputers in the world you know, and at the same time, we're going to see much more efficient architectures, which are going to…”