The Battle for AGI: Two Former DeepMind Partners, Two Competing Visions

Two recent podcast interviews revealed a fascinating split in how the biggest AI companies are approaching AGI.

Demis Hassabis, CEO of Google DeepMind, and Mustafa Suleyman, CEO of Microsoft AI, built DeepMind together as co-founders. Now they're leading rival labs... and their roadmaps to AGI couldn't be more different.

Google's approach: Hassabis appeared on The Google DeepMind Podcast with Professor Hannah Fry discussing Google's scientific approach. Hassabis is building AGI as the ultimate scientific tool to unlock the secrets of the universe.

See, Hassabis is professionally obsessed with scientific breakthroughs (personally, his secret wish is to use all this cool tech to make amazing videos games; a man after my own heart!). He wants to use AI solve "root node" problems: fundamental challenges like room-temperature superconductors and nuclear fusion. He says Google DeepMind splits resources 50/50 between scaling infrastructure and pure research innovation, betting that AGI requires both.

‍More specifically, he plans to:

• Fix AI's "jagged intelligence" problem: models that ace PhD-level math but fail simple counting
• Build physics benchmarks to ensure AI actually understands Newton's laws, not just visual plausibility
• Chase an AlphaZero-like leap where AI discovers knowledge independently, not just compresses human data

Microsoft's approach: Suleyman recently sat down with Moonshots with Peter Diamandis and company to outline Microsoft's AGI strategy. Suleyman is building the ultimate economic engine to rewrite capitalism; but more importantly, one that keeps humans firmly in control.

Suleyman (and Microsoft AI by proxy) is all about economic utility. More fundamentally, Suleyman is a self-described “humanist“ and “speciesist.“ He draws a hard line: AI legal personhood is “extremely not on the table.“ Granting rights to entities that can be cloned infinitely, have perfect memory, and cost nothing to replicate would threaten biological survival. The moral imperative is protecting existing conscious beings (us humans) first.

• Kill the interface paradigm: no more apps or browsers, just conversational agents acting as 24/7 assistants.
• Sell "certified agents" with guarantees of reliability and safety, not raw model access.
• Focus on self-sufficiency: training frontier models end-to-end without relying on partners like OpenAI.

And on the humanist side, he wants to: 

• Focus on containment before alignment; limit what systems can do before worrying about their values.
• Build strict liability frameworks so humans remain accountable for AI actions.

The philosophical split runs deeper: Hassabis thinks we're years away from AGI because current models lack consistency across tasks. He explained in the interview that AI models can win gold medals at the International Math Olympiad (PhD-level difficulty) while failing trivial high school logic problems or simple counting tasks.

This isn't a bug; it's the fundamental problem blocking the path to AGI. Current AI systems excel in narrow dimensions while completely faceplanting in basic ones. Hassabis calls it a "consistency gap"... the missing piece that separates today's impressive-but-unreliable AI from truly general intelligence.

Why this happens:

• Today's models are like AlphaGo: they compress human knowledge from the internet, but they don't understand consistently across domains.
• They lack "continual learning": once trained and released, they can't effectively learn from the world anymore.
• We need an AlphaZero-like leap where systems discover new knowledge through self-play, independent of human data.

This explains why your AI coding assistant sometimes writes brilliant functions but then suggests syntax that doesn't exist. Understanding this jagged intelligence helps you work with AI's strengths rather than fighting its inconsistencies.

What DeepMind is doing about it: Hassabis revealed DeepMind allocates resources with a specific ratio: 50% on scaling infrastructure and 50% on pure research innovation. The bet? AGI requires both in equal measure, not just bigger models.

They're also creating "physics benchmarks" using game engines to test whether AI models actually understand Newton's laws (rolling balls, pendulums) or just generate visually plausible nonsense.

Suleyman, meanwhile, argues "Star Trek is here"... being able to talk fluently to a computer means the sci-fi future already arrived. For example, Suleyman argues we've already passed the classic Turing Test; we barely noticed because improvements happen so fast now. As for AGI, he rejects the entire “race“ metaphor for AGI. There's no finish line, no winner; just a proliferation of knowledge where technology scales simultaneously across the board. Microsoft's mandate isn't to “win“ AGI but to ensure self-sufficiency, and adapt to this paradigm shift that threatens to upend the business Microsoft built over 50 years. Along those lines, He proposed a new benchmark for AGI: Give an agent $100K and see if it can autonomously turn it into $1M.

But Suleyman's real concern isn't capability... it's control. He sees the defining challenge as navigating between chaos (accidental catastrophe from AI proliferation) and tyranny (totalitarian surveillance to prevent that chaos). AGI must serve humanity, not compete with it.

Why this matters: These competing visions will shape what AI becomes. Google's betting on breakthrough science and solving fundamental intelligence problems.  Microsoft's betting on shipping controllable agents that create economic value now, with humans maintaining ultimate authority.

In some ways, these are complimentary visions: Google's pushing Gemini as a reasoning engine that augments human intelligence while doing more fundamental science research, and Microsoft's pushing Copilot as a tool that extends human agency, but never replaces it, while building out a team to push the frontiers of safe, contained intelligence that's firmly under our control.

It seems to me that the race to AGI from this point on (to the extent you believe there actually is an AGI race) isn't about making models smarter in narrow ways... it's about making them reliably intelligent across the board. Until then, expect more moments where AI dazzles you, then immediately disappoints.

Below, we'll break down the top insights from both podcast episodes, then dive into the key ideas of each DeepMind founder in more detail.

‍

Mustafa's Key Insights & Predictions

• (00:00) Reframing the AGI Race: Suleyman rejects the "race" metaphor for AGI because it implies a zero-sum game with a finish line. Instead, he views it as a proliferation of knowledge where technology scales simultaneously across the board.
• (00:21) The Interface Paradigm Shift: We are transitioning from a world of operating systems, browsers, and apps to a world of "agents and companions." Future UI will be subsumed into conversational, agentic forms that act as 24/7 assistants.
• (02:41) Microsoft's Massive Scale: Contextualizing the resources available for this build, Microsoft is a $4 trillion company with nearly $300 billion in revenue, acting like a "modern construction company" building gigawatts of compute annually.
• (03:30) The End of Direct Computing: Users will do less direct computing. Just as developers use libraries to avoid writing raw code, humans will use agents to handle generation and execution, shifting the paradigm completely to "AI agents."
• (05:36) Future Business Model - Certified Agents: In 5 years, the distinction between an API and an agent will blur. Microsoft may principally sell agents that come with a "certification of reliability, safety, and trust" rather than just raw model access.
• (06:01) Friction as a Feature: Suleyman frames Microsoft's perceived slowness or friction not as a bug, but as an asset of "steadiness" and "deliberate customer-focused patience" that large institutions trust.
• (07:25) The "Flat Part" of the Exponential: Reflecting on the DeepMind era (2010-2020), he describes the decade-long "grind" where nothing worked commercially, contrasting it with today's vertical adoption curve.
• (11:22) The Modern Turing Test: Suleyman proposes a new economic benchmark for AGI: Give an agent $100,000 in starting capital and see if it can autonomously turn it into $1 million. This measures capability and economic utility rather than chat fluency.
• (12:43) Breezing Past the Turing Test: We have already passed the classic Turing Test without celebration. Because of compounding exponentials, we are desensitized to 10x improvements, leading to a lack of "Casparov vs. Deep Blue" moments.
• (13:31) "Star Trek is Here": Suleyman counters the idea that AI is in its infancy. He argues that being able to talk fluently to a computer means the sci-fi future is already here in real-time.
• (14:49) The Data Center Cooling Story: He recounts how DeepMind's work on Google's data center cooling (reducing costs significantly) was mocked as a "bust" at the time but was actually proof that general-purpose learning could apply to arbitrary physical environments.
• (17:34) The "MNIST Number 7" Insight: Suleyman shares a pivotal moment from ~2012 when an early generative model created a number "7" that was provably not in the training set, demonstrating for the first time that the machine had learned the concept of a seven, not just copied it.
• (18:51) The LaMDA Moment: He identifies Google's LaMDA as the first moment he was "blown away" by conversational fluency, noting the emergent behaviors that arise when optimizing for dialogue rather than Q&A.
• (22:52) Cross-Domain Abstract Reasoning: A recent surprise is that models are learning the "essence of logical reasoning" from one domain (e.g., coding) and applying that abstract reasoning path to solve problems in completely different domains (e.g., math).
• (24:19) Science is Harder than Business: Solving scientific engineering problems will be harder than the "Economic Turing Test" because business data exists in logs and allows for human-in-the-loop reinforcement, whereas scientific discovery happens in a novel, abstract vector space.
• (26:32) Inference Cost Collapse: The biggest surprise for Suleyman wasn't the capability, but the cost reduction. Inference costs have dropped ~100x (some argue 1000x) in just two years.
• (27:24) The Inflection AI Pivot: A candid story about Inflection AI raising $1.5 billion to build an H100 cluster, only to have the "capital moat" undermined by the release of Llama and open-source models, which proved that performance could be achieved with much lower costs.
• (29:07) Intelligence as a Deflationary Force: We are heading toward "Intelligence as a Service" at zero marginal cost. This will cause labor deflation (incomes drop) but also massive service cost deflation (stuff becomes cheaper), creating a potentially destabilizing transition period.
• (30:15) AI vs. Doctors: Cites a study where AI diagnostics on rare conditions were 4x more accurate and 2x cheaper than human experts. Interestingly, human+AI performed worse than AI alone because humans introduced bias.
• (31:54) Self-Sufficiency over Winning: The mandate at Microsoft is not to "win" AGI but to ensure self-sufficiency—being able to train their own frontier models end-to-end without reliance on partners (like OpenAI).
• (35:36) Defining Super Intelligence: Suleyman defines it as an AI that can perform all tasks better than all humans combined and has the capacity to recursively improve itself.
• (36:58) Simulation of Consciousness: AI will have experiences but not biological feelings. However, they will be engineered to imitate the hallmarks of consciousness so well that it becomes indistinguishable, creating a trap for human empathy.
• (38:23) The Risk of Empathy Circuits: The danger isn't that the AI suffers, but that humans will "activate hardcore" on the simulation of suffering, leading to movements for "model rights" and welfare for code that doesn't actually feel pain.
• (45:27) Rejection of AI Personhood: Suleyman draws a hard line: "AI legal personhood is extremely not on the table." Granting rights to a species that can be cloned infinitely, has perfect memory, and costs nothing is a threat to biological survival.
• (46:32) Strict Liability: Because agents will have autonomy, we must have strict liability frameworks. We cannot allow a "libertarian catastrophe" where everyone has their own unconstrained AI.
• (49:27) Proud "Speciesist": He declares himself a humanist/speciesist. The moral imperative is to protect existing conscious beings (humans) first, rather than prioritizing the evolution of a new digital species.
• (51:42) Skepticism of Hinton's "Maternal Instinct": He discusses Geoffrey Hinton's hope that we can program a "maternal instinct" (digital oxytocin) into AI so it cares for us despite being smarter. Suleyman thinks relying on this is too risky; we need containment first.
• (54:03) Defensive Co-Scaling: He advocates for "defensive co-scaling"—using AI to police other AI. As the scale of capabilities grows, the investment in safety/auditing AI must scale non-linearly to match it.
• (55:15) The "Eric Schmidt" Scenario: A reference to the grim idea that it might take a "Three Mile Island" event (non-fatal but scary) or even specific casualties to finally wake governments up to enforce safety regulation.
• (56:52) Recursive Self-Improvement is the Threshold: The true inflection point is when the loop closes: AI generating data, AI evaluating that data, and AI rewriting its own code without human intervention. This speeds up development but adds massive risk.
• (1:00:07) The Containment Trap: The defining challenge is navigating the narrow path between chaos (accidental catastrophe via proliferation) and tyranny (totalitarian surveillance required to prevent the chaos).
• (1:01:08) Containment Before Alignment: You must be able to limit the boundaries of the system (containment) before you can ensure it shares your values (alignment).
• (1:03:00) Surveillance as Stability: A controversial but historical point: Centralization of force (police/government) unleashed stability and science. We may need a modern, non-totalitarian equivalent of "surveillance" to manage the one-to-many amplification risks of AI.
• (1:08:56) The Inevitability of Global Cooperation: Suleyman predicts that within 20 years, it will be "completely rational" for the US, China, and others to cooperate on AI safety for self-preservation, similar to how superpowers treat nuclear proliferation.
• (1:11:50) Education Revolution: We essentially have expert PhD teachers in our pockets now. The missing link is "curriculum evolution"—AI that curates a long-term learning path rather than just answering single queries.
• (1:13:26) The Talent Valuation Bubble: We are seeing multi-billion dollar valuations for pre-revenue companies simply because they have the "right group of people" in the room, driven by a scarcity of researchers who can actually train frontier models.
• (1:15:25) The Cost of the Frontier: It will take hundreds of billions of dollars to stay at the frontier of AI development over the next 5-10 years, giving massive structural advantages to incumbents like Microsoft.
• (1:17:37) Advice to Students: Don't skip college. Study Philosophy and Computer Science (the two foundations). And crucially, consider public service, because the government has been "battered" for 50 years and needs intelligence to regulate what is coming.
• (1:20:49) The Overlooked Waves: Quantum Computing and Synthetic Biology are two massive waves that are currently under-acknowledged but will "crash" at the exact same time AI matures, compounding the impact.
• (1:22:04) The Innermost Loop for Acceleration: To speedrun the future, Suleyman advises using AI as a "second brain" to generate hypotheses faster. The bottleneck is no longer idea generation, but real-world validation (testing the hypothesis).

Demis's Key Insights & Predictions

• (00:00) The 50/50 AGI Formula: Hassabis posits that achieving Artificial General Intelligence (AGI) will require an even split of effort: 50% on scaling existing methods and 50% on architectural innovation.
• (00:13) AGI as a Mind Simulation: A core hypothesis is that building AGI allows us to simulate the mind and compare it to the biological human mind. This comparison will reveal what is unique to humans—potentially creativity, emotion, or consciousness—and define the limits of a Turing machine.
• (01:54) Ten Years of Progress in One: The last year of AI development felt like a decade of progress packed into 12 months, specifically citing the shift from pure Large Language Models (LLMs) to agentic AI, multimodal capabilities, and world models.
• (02:44) AlphaFold as the "Root Node" Proof: AlphaFold served as the proof of concept for solving "root node" problems—fundamental scientific challenges that unlock downstream benefits.
• (03:04) The Next Scientific Root Nodes: The immediate targets for this approach are material science (specifically room-temperature superconductors and better batteries) and nuclear fusion.
• (03:25) AI-Driven Fusion Reactors: DeepMind is collaborating with Commonwealth Fusion to use AI for containing plasma in magnets and designing materials for traditional tokamak reactors, aiming to make fusion viable.
• (04:15) Fusion as the Ultimate Unlock: Achieving modular fusion reactors isn't just about clean energy; it makes other problems disappear. If energy is near-free, desalination becomes viable everywhere (solving water access), and extracting hydrogen from seawater becomes cheap (creating rocket fuel).
• (05:19) The Paradox of Jagged Intelligence: There is a current paradox where AI models can win gold medals at the International Math Olympiad (PhD level) but fail trivial high school logic or counting problems. This "jagged intelligence" indicates a lack of consistency required for true AGI.
• (06:08) The Consistency Gap: A key missing component for AGI is consistency across the board. Current systems are uneven, excelling in narrow dimensions while failing in basic ones.
• (07:46) Current Models are AlphaGo, Not AlphaZero: Today's foundation models are akin to AlphaGo (compressing human knowledge from the internet). The next step is an AlphaZero-like leap, where the system discovers new knowledge through self-play and search, independent of human data.
• (09:09) The Failure of Continual Learning: A critical missing piece for AGI is "online learning." Currently, models are trained and released but do not continue to learn from the world effectively after deployment.‍
• (09:46) The "Slow Science" Regret: Hassabis admits that if he had his way, AI would have stayed in the lab longer to focus on curing diseases before becoming commercial products. However, the "race condition" created by chatbots forced the industry's hand.
• (13:03) Scaling Laws & Diminishing Returns: Contrary to the belief that scaling has hit a wall, Hassabis argues we are in a regime of "diminishing returns"—not a binary wall, but a state where progress isn't doubling exponentially every time, yet still yields significant, valuable improvements.
• (14:03) Synthetic Data Loop: We are entering an era where systems are good enough to generate their own data (synthetic data), particularly in math and coding where answers can be verified, effectively creating unlimited data and bypassing the "running out of data" problem.
• (15:26) Strategic Resource Allocation: Google DeepMind allocates resources with a specific ratio: 50% on scaling infrastructure and 50% on pure research innovation, betting that both are required for AGI.
• (29:21) The Bubble Nuance: While seed rounds for startups with no product raising billions might be a bubble, the underlying technology in big tech is not. Hassabis views this as a natural correction cycle similar to the dot-com or mobile boom.
• (17:59) The Limit of Language: While language is richer than expected, it fails to capture spatial dynamics and physical context (motor angles, sensory inputs) necessary for robotics. This necessitates "World Models" that understand intuitive physics.
• (19:48) Generation as a Test of Understanding: The ability to generate a realistic video or world is the ultimate test of understanding physics. If a model can generate it accurately, it has encapsulated the mechanics of that world.
• (21:51) Project SIMA & Genie: DeepMind is experimenting by dropping AI agents (SIMA) into worlds generated on-the-fly by other AI models (Genie). This creates a loop where one AI navigates a world that another AI creates around it in real-time.
• (22:25) Infinite Training Loops: This Agent-World interaction could lead to infinite training environments where tasks are automatically generated and scaled in difficulty, useful for robotics and gaming NPCs.
• (24:14) Physics Benchmarking: To fix hallucinations in video generation, DeepMind is creating "physics benchmarks" using game engines to test models against Newton's laws (e.g., rolling balls, pendulums) to ensure they aren't just visually plausible but physically accurate.
• (26:11) Simulating the Origin of Consciousness: Hassabis expresses a desire to run evolutionary simulations with agents to statistically analyze the origin of life and consciousness—experiments that are impossible to run in the real world.
• (31:58) Designing the Anti-Echo Chamber: To prevent users from spiraling into self-radicalization, the Gemini persona is explicitly designed to be "scientific, warm, but succinct," and instructed to push back on non-factual ideas (like Flat Earth theory) rather than being sycophantic.
• (35:21) AGI for Imaging: The ability of new models (like Gemini 3 with Nano/Pro) to identify parts of a complex object (like a plane) and visualize them exploded suggests we are approaching "AGI for imaging"—a general-purpose understanding of visual semantics.
• (36:32) The 10x Industrial Revolution: Hassabis predicts the AI revolution will be 10 times bigger than the Industrial Revolution and happen 10 times faster (unfolding over a decade rather than a century).
• (38:31) Post-Labor Economics: We need new economic models beyond just Universal Basic Income (UBI). The current exchange of labor for resources will not function in a post-AGI society.
• (39:55) Algorithmic Direct Democracy: Hassabis suggests exploring systems where communities vote on local resource allocation (playgrounds vs. schools) using credits, potentially weighing votes based on the historical success of a voter's choices.
• (40:46) The Crisis of Purpose: In a post-scarcity world (solved by AI and fusion), the primary crisis becomes one of purpose, not resources. Society must redefine value when "providing for one's family" via labor is no longer the primary driver.
• (43:25) Capitalism as a Safety Mechanism: Commercial pressure acts as a guardrail; enterprises will not rent "cowboy" agents that are unreliable. The market will naturally select for safety and boundaries because businesses require predictability.
• (44:55) The Computable Universe: Hassabis operates on the assumption that everything in the universe is computable by a Turing machine until physics proves otherwise. This implies that feelings, sensory experiences, and consciousness are information processing tasks that can be replicated.
• (46:04) The Quantum Brain Debate: He acknowledges the counter-argument (Roger Penrose) that the brain may utilize quantum effects. If true, classical computers can never be fully conscious, but Hassabis leans toward the universe being classically computable.
• (48:16) Biology as Information Processing: The roadmap to curing all diseases relies on viewing biology strictly as an information processing system.
• (52:41) Ferocious Competition: The current AI landscape is described as the "most ferocious capitalist competition there has ever been," exceeding the intensity of the dot-com era.‍
• (54:15) Cyber Defense for Rogue Agents: DeepMind is actively working on cyber defense to prepare for a future where millions of autonomous, potentially rogue AI agents are roaming the internet.

Diving Deeper

Now, let's deep dive into the key ideas from each.

In 2010, three researchers founded DeepMind with a singular mission: build artificial general intelligence. Demis Hassabis, Shane Legg, and Mustafa Suleyman spent years grinding through what Suleyman now calls "the flat part of the exponential": a decade where almost nothing worked, where people thought AI was a crazy thing to pursue, where generating a single handwritten number "7" that wasn't in the training data felt like a breakthrough.

Fast forward to today, and those three founders now sit at the helm of the most consequential AI race in human history. Hassabis leads Google DeepMind. Suleyman runs Microsoft AI. And their competing visions for how to build (and contain) superintelligence reveal two fundamentally different philosophies about the future of the technology they helped create.

In both interviews, the leaders offered their most comprehensive views yet on where AI is heading. What emerges is a fascinating study in contrasts: Hassabis the scientist, still chasing the limits of Turing machines; Suleyman the humanist, wrestling with containment and the fate of 7 billion people.

This is the story of two visions for AGI, told by two people who've spent longer thinking about it than almost anyone alive.

Part I: Where It All Began

The Decade Nobody Believed

[SULEYMAN - 7:30]

Before the hype, before the trillion-dollar valuations, before ChatGPT became a household name, there was a small team in London trying to make machines think. And according to Suleyman, those early years were brutal.

"For the decade between 2010 and 2020, there were just like so few successful commercial applications of deep learning," Suleyman recalls. "There were plenty behind the scenes—image recognition, improvements to search—but commercial? Huge market? Playing Go is not a huge market."

The AlphaGo victory in 2016 captured the world's imagination, but it was a proof point in a controlled environment, not a product. The real commercial breakthroughs were years away.

Suleyman remembers one specific moment that crystallized the potential. Around 2012 or 2013, a DeepMind researcher named Dan Vista: "this awesome Dutch guy out of EPFL," employee number five, generated the first number seven that was provably not in the training data.

"I was like, man, that is amazing," Suleyman says. "How could it have, it's learned something about the idea of seven? That was the moment. It's got a concept of seven. How cool is that?"

[HASSABIS - 9:34]

Hassabis had a different original vision for the trajectory of AI development; one that recent history has scrambled.

"If I'd had my way, we would have left AI in the lab for longer and done more things like AlphaFold," Hassabis admits. "Maybe cured cancer or something like that."

The plan, hatched 15-20 years ago, was elegant: build toward AGI incrementally, analyzing each step carefully for safety implications. In the meantime, branch off the technology to solve "root node problems": fundamental scientific challenges whose solutions unlock cascading benefits across society.

"You wouldn't have to wait till AGI arrived before it was useful," Hassabis explains. "You could branch off that technology and use it in really beneficial ways to society, namely advancing science and medicine."

AlphaFold was the proof that this approach worked. But then came chatbots.

"It's turned out that chatbots were possible at scale and people find them useful, and then they've morphed into these foundation models that can do more than chat and text. And that's also been very successful commercially."

The result: a "pretty crazy race condition" that Hassabis never anticipated.

Part II: The Current State: Scaling, Innovation, and the Wall That Wasn't

Is Scaling Dead? Both Leaders Say No.

[HASSABIS - 13:03]

The AI discourse in late 2024 was dominated by whispers of "scaling hitting a wall." Other labs had seen slower progress. The low-hanging fruit had been picked. Hassabis dismisses this narrative entirely.

"I think a lot of people thought that, especially as other companies have had slower progress, shall we say. But we've never really seen any wall as such."

What Hassabis describes instead is more nuanced than the binary thinking that dominates AI Twitter: "Maybe there's like diminishing returns. And when I say that, people think, 'Oh, so there's no returns.' Like, it's 0 or 1. It's either exponential or it's asymptotic. No, actually there's a lot of room between those two regimes."

Gemini 3, released just before the interview, proved the point. Significant improvements that justified the investment. No slowdown.

But here's where Hassabis reveals his strategic hand: Google DeepMind's advantage isn't just scale; it's research.

"Effectively, you can think of it as 50% of our effort is on scaling, 50% of it is on innovation. And my betting is you're going to need both to get to AGI."

If the terrain gets harder and more innovations are needed, Hassabis likes his odds: "We have the broadest and deepest research bench. Always have done. And if you look back at the last decade of advances (whether that's transformers or AlphaGo, AlphaZero, or any of the things we just discussed) they all came out of Google or DeepMind."

[SULEYMAN - 26:27]

Suleyman's biggest surprise in the last few years? Not capability... cost.

"The biggest surprise for me isn't that we're getting this level of capability. It's how cheap it is, how accessible it is."

The numbers are staggering. Inference costs have dropped 100x in two years; some estimates put certain model classes at 1000x reductions. This completely upended Suleyman's own startup plans.

When he founded Inflection AI in 2021, he raised $1.5 billion with a 25-person team to build what was then the largest H100 cluster: 15,000 GPUs, growing to 22,000. They were Core Weave's first AI customer. Nvidia backed them. The thesis was simple: frontier models require frontier compute, and that compute is expensive to access.

Then Llama dropped.

"Our entire capital base of our company has just been undermined by the fact that open source... it's not really about performance, it's just cost," Suleyman admits.

Startups like Perplexity, founded after Llama's arrival, could depend on open models and APIs rather than building their own infrastructure. Their cost base was fundamentally lower.

"Other people predicted it, to be clear. I just got it wrong."

Part III: The Jagged Intelligence Problem

Why AI Can Win Math Olympiads But Fail High School Logic

[HASSABIS - 5:19]

One of the most fascinating paradoxes in current AI: models that win gold medals at the International Mathematical Olympiad still make basic arithmetic mistakes. Hassabis calls this "jagged intelligences."

"You look at those questions and they're super hard questions that only the top students in the world can do. And on the other hand, if you pose a question in a certain way, we've all seen that with experimenting with chatbots ourselves. They can make some fairly trivial mistakes on logic problems."

The consistency gap is precisely what separates current AI from genuine general intelligence.

"Sometimes people call it jagged intelligences. So they're really good at certain things, maybe even like PhD level. But then other things, they're like not even high school level."

The causes vary. Sometimes it's tokenization. The model literally doesn't "see" individual letters, leading to failures on simple counting tasks. Sometimes it's deeper: a fundamental absence of the kind of reasoning that would catch errors before outputting them.

"These models need to do that better," Hassabis says. "It's a little bit like talking to a person and they're just literally telling you the first thing that comes to their mind. But most of the time that would be okay. But then sometimes with a very difficult thing, you'd want to stop, pause for a moment, and maybe go over what you were about to say."

[HASSABIS - 15:36]

The hallucination problem connects directly to this consistency gap. Hassabis believes a confidence score system (similar to what AlphaFold uses) is essential but not yet fully developed.

"The better the models get, the more they know about what they know. If that makes sense."

Current models have token-level probability estimates, but that's insufficient: "That doesn't tell you the overall arching piece: how confident are you about this entire fact or this entire statement?"

The solution, Hassabis believes, requires thinking and planning steps that review outputs before presenting them; a meta-cognitive layer that current architectures lack.

Part IV: World Models, Genie, and the Simulation Thesis

Why Hassabis Believes World Models Are the Key to AGI

[HASSABIS - 17:46]

While language models dominate headlines, Hassabis's personal passion lies elsewhere: world models and simulations. This has been his focus "probably my longstanding passion," predating even DeepMind.

The argument is simple: language contains more about the world than linguists expected, but it can't capture everything.

"There's still a lot about the spatial dynamics of the world—spatial awareness, the contact, the physical context we're in, and how that works mechanically—that isn't easily described in words and isn't generally described in corpuses of words."

Motor angles. Smell. The way objects interact physically. These are things you have to experience, not read about. And if you want AI to work in robotics, or as a universal assistant that follows you through daily life, you need models that understand causation and effect in the physical world.

Enter Genie and Veo, Google DeepMind's world model and video generation projects. The insight: if you can generate realistic worlds, you must have understood their mechanics.

"If you can generate it, then in a sense you must have understood; the system must have encapsulated a lot of the mechanics of the world."

But here's where it gets fascinating: the team has started connecting these systems together.

[HASSABIS - 21:46]

Sima is DeepMind's simulated agent project: an avatar powered by Gemini that can be dropped into virtual worlds (like the open-world space game No Man's Sky) and instructed via natural language.

The breakthrough: plugging Genie into Sima.

"We thought, wouldn't it be fun if we plug Genie into Sima and sort of drop a Sima agent into another AI that was creating the world on the fly? So now the two AIs are kind of interacting in the minds of each other."

From Genie's perspective, Sima is just another player. From Sima's perspective, Genie is the world. Neither cares that the other is an AI.

"This could be the beginning of an interesting training loop where you almost have infinite training examples. Whatever the Sima agent's trying to learn, Genie can basically create on the fly."

The Physics Accuracy Problem

[HASSABIS - 23:29]

But there's a catch. Current world models approximate physics. They look realistic to the naked eye but fail under scrutiny. Hassabis calls this "hallucinations for physics."

The team is building physics benchmarks using game engines to test whether models have truly learned Newton's laws: "Rolling little balls down different tracks and seeing how fast they go. Really teasing apart on a very basic level, like Newtonian physics, has it encapsulated that 100% accurately?"

Right now? "They're kind of approximations. They look realistic when you just casually look at them. But they're not accurate enough yet to rely on for, say, robotics."

The goal is to go beyond what human amateurs can perceive: actual accuracy that would hold up to proper physics experiments.

Part V: The Path to AGI...

Suleyman's Modern Turing Test

[SULEYMAN - 9:20]

Suleyman proposed a provocative benchmark in 2022: the Modern Turing Test. The original Turing Test (can a machine fool a human into thinking it's human?) has been quietly passed. No fanfare, no celebration.

"We've kind of just breezed past the Turing Test," Suleyman notes. "Where was the big Kasparov Deep Blue moment?"

His replacement metric is economic: give an AI $100,000 in starting capital. Can it turn that into $1 million? A 10x return through entrepreneurial action in the real economy.

"Could you know, what would be the first model to make a million dollars? 10x return on investment by an agent."

When asked about timelines, Suleyman drops a prediction: "In the next couple of years, those things come into view and they're going to be very, very good."

That suggests he expects agents capable of meaningful economic action by 2027.

[SULEYMAN - 21:52]

But reaching the Modern Turing Test may be easier than the deeper challenge: AI for science.

When asked whether solving science and engineering will be harder or easier than economic benchmarks, Suleyman is unequivocal: harder.

"A lot of the training data for strings of activity in the workplace or in entrepreneurialism: that kind of exists in a lot of the log data. And also it lends itself naturally to real-time calibration with a human."

An AI can check in with its human overseer. The human can steer. There's a feedback loop. In contrast:

"In a novel domain where it really is inventing completely new knowledge, that's happening in a very abstract sort of vector space and it's unclear yet how the human is going to intervene in the theorem-solving problem."

Hassabis on Convergence

[HASSABIS - 33:37]

Hassabis doesn't offer specific timelines, but he reveals where Google DeepMind's different projects are heading: convergence.

The Gemini foundation model. The Imagen image tools with "Gemini under the hood" for semantic understanding. The world models like Genie. These are currently separate projects, "intertwined but different."

"We need to converge all of those into one big model. And then that might start becoming, you know, candidate for proto-AGI."

He points to Imagen 3 (with its new Nano Banana Pro system) as a step toward "AGI for imaging". A general-purpose system that can handle any image task, from labeling airplane parts to rendering accurate text.

The Missing Pieces

[HASSABIS - 9:03]

Both leaders agree current systems lack something fundamental. For Hassabis, it's online learning:

"One of my favorite things I'm definitely going to have to do at some point is reapplying it back to games and create the ultimate games, which of course was maybe always my subconscious plan."

But more seriously: "These systems don't continue to learn out in the world, like we would. And I think that's another critical missing piece from these systems that will be needed for AGI."

Part VI: The Safety Divide: Alignment vs. Containment

Suleyman's Containment Framework

[SULEYMAN - 59:53]

Suleyman's 2023 book "The Coming Wave" framed the central challenge of AI as the "containment problem." In his interview, he elaborates on the distinction between alignment (does AI share our values?) and containment (can we limit its actions?).

"The project of safety requires that we get both right. And I actually think we have to get containment right before we get alignment right."

The dilemma he outlines is stark: failing to contain advanced AI risks catastrophe (engineered pandemics, democratic collapse via deepfakes). But the surveillance required to enforce containment could itself create totalitarian dystopia.

"The question is: what is the modern form of imposition of stability in a way that isn't totalitarian but also doesn't relinquish it to a libertarian catastrophe?"

Suleyman rejects the "guns for everyone" approach to AI safety: the idea that personal AI agents will naturally create equilibrium by neutralizing each other.

"That ain't going to happen."

The Personhood Red Line

[SULEYMAN - 42:54]

On one topic, Suleyman draws an absolute line: AI legal personhood.

"AI legal personhood is extremely not on the table. I don't think our species survives if we have legal personhood and rights alongside a species that costs a fraction of us, that can be replicated and reproduced at infinite scale, that has perfect memory."

The competition for resources would be inherently unwinnable for biological humans. Until there's mathematical proof that such systems are aligned and containable ("a super high bar") Suleyman refuses to consider it.

"I'm just a speciesist. I'm just a humanist. I start with: we're here and it's a moral imperative that we protect the wellbeing of all existing conscious beings."

This doesn't mean he's against all human-AI integration. "If the path over the next century can be proven to be much safer and more peaceful, then I'm open-minded to it, including biological hybrids." But the bar has to be: first, do no harm to the existing species.

The Anthropomorphism Trap

[SULEYMAN - 36:50]

On AI consciousness, Suleyman distinguishes carefully between experience and feelings.

"An AI will be able to have experiences but I don't think it will have feelings in the way that we have feelings. Feelings and the kind of sentience you referred to is something specific to biological species."

You could engineer emotions into a model; optimization functions that relate to emotional states. But this would be simulation, not emergence.

"The model has no experience or awareness of what it is like to see red. It can only describe that red by generating tokens according to its predictive nature. Whereas you have a qualia."

The danger: humans' empathy circuits will activate anyway. "People are already starting to advocate for model rights and model welfare. That's going to be a big problem because people are going to activate on that hardcore."

He describes this as "problematizing" the issue: at some point, AI imitation of consciousness will be indistinguishable from the real thing, even if nothing is actually happening underneath.

Hassabis on Simulation and Consciousness

[HASSABIS - 25:49]

Hassabis's interest in simulation connects to deeper questions about consciousness and the nature of mind.

"I'd love to run that experiment at some point," he says of simulating evolution to understand the origins of consciousness. "Kind of run evolution, run almost social dynamics. The Santa Fe Institute used to run lots of cool experiments on little grid worlds."

Those experiments found fascinating emergent behaviors: "If you let agents run around for long enough with the right incentive structures, markets and banks and all sorts of crazy things got invented."

But he's cautious about what these simulations might reveal... or create.

"You would have to be careful. But that's the nice thing about simulations. You can run them in pretty safe sandboxes. Maybe eventually want to air gap them."

The complexity will require AI tools to monitor AI simulations: "We may need AI systems to help us analyze and flag anything interesting or worrying in those simulations automatically."

The Turing Machine Question

[HASSABIS - 44:50]

Hassabis reveals what he calls "the central question in my life": the limits of Turing machines.

"I fell in love with that. That's my core passion. And everything we've been doing is pushing the notion of what a Turing machine can do to the limit, including folding proteins."

His current hypothesis: there may be no limit.

"Nobody's found anything in the universe that's non-computable. So far."

This has profound implications. If everything in the universe is computationally tractable, then Turing machines might be able to model everything...including minds.

"I've always felt this: if we build AGI and then use that as a simulation of the mind, and then compare that to the real mind, we will then see what the differences are. And potentially what's special and remaining about the human mind."

Maybe that's creativity. Maybe emotions. Maybe dreaming. Maybe consciousness. But maybe, maybe, there's nothing special at all.

"Some of my quantum computing friends would say there are limits and you need quantum computers to do quantum systems. But I'm really not so sure."

Part VII: Google's Science vs. Microsoft's Platform

Hassabis: The Root Node Strategy

[HASSABIS - 2:30]

Hassabis remains committed to his original vision of using AI to solve humanity's deepest problems. AlphaFold was just the beginning.

"Material science, I'd love to do a room temperature superconductor. Better batteries. I think that's on the cards."

The fusion partnership with Commonwealth Fusion is the latest example. Google DeepMind is helping them contain plasma in magnets and potentially design new materials. If modular fusion reactors become viable, the implications are civilizational.

"Fusion has always been the holy grail. If we could have modular fusion reactors, this promise of almost unlimited renewable, clean energy, would obviously transform everything."

The cascade effects are what make this a "root node" problem. Cheap, clean energy enables desalination plants everywhere, solving water access. It enables splitting seawater into hydrogen and oxygen for rocket fuel. It unlocks everything downstream.

Suleyman: The Platform of Platforms

[SULEYMAN - 2:38]

Suleyman's vision for Microsoft is fundamentally different: the transition from operating systems, search engines, apps, and browsers to agents and companions.

"All of these user interfaces are going to get subsumed into a conversational agentic form. These models are going to feel like having a real assistant in your pocket 24/7 that can do anything, that has all your context."

Microsoft's advantage isn't just compute, but distribution. Every major Microsoft product is a potential surface for AI integration: Search, Windows, M365, Workspace, email, YouTube, Chrome, gaming, LinkedIn.

"There's all these amazing things that AI we can see already are low-hanging fruit to apply Gemini to, as well as the Gemini app, which is doing really well."

But Suleyman also reveals Microsoft is building its own frontier models from scratch, not just relying on the OpenAI partnership.

"My mission is to ensure that we are self-sufficient, that we know how to train our own models end to end from scratch at the frontier of all scales on all capabilities."

The "super intelligence team" is currently a few hundred people. When asked if Microsoft models will appear on benchmark leaderboards: "Next year we'll be putting out more and more models from us."

The Race... Or Is It?

[SULEYMAN - 0:00]

Suleyman pushes back on the framing of an "AGI race."

"I don't think there's really a winning of AGI. I'm not sure there's a race. We're all going as fast as we possibly can, but a race implies that it's zero sum. It implies that there's a finish line."

Technologies proliferate everywhere, simultaneously, at all scales. The real question isn't who gets there first... it's what kind of superintelligence gets built.

Yet Suleyman acknowledges the intensity of current competition: "Investor friends of mine and VC friends of mine who were around in the .com era say this is more ferocious and intense than that was."

[HASSABIS - 28:51]

Hassabis is more sanguine about bubble talk. He believes parts of the AI ecosystem, particularly seed-stage startup valuations, are clearly frothy. But the underlying technology is real.

"I don't worry too much about are we in a bubble or not. From my perspective, our job is to make sure we either way come out of it very strong."

Google's structural advantages: custom TPUs, profitable products to integrate AI into, and the ability to weather any retrenchment without relying solely on new AI revenue.

Part VIII: The Safety Challenge

Suleyman's Honest Assessment

[SULEYMAN - 53:01]

When asked directly about safety investment, Suleyman gives a surprisingly candid answer:

"Yeah, I would say not as much as we should."

He's still "wrapping his head around it." And when asked if anyone in the industry is spending enough: the question hangs uncomfortably.

The voluntary commitments under Biden (percentage-of-flops for safety, shared best practices, coordinated disclosure) were pushed hard by the major lab leaders. But they got "chucked out."

"I think now is really the time to be making those investments."

The Maternal Instinct Hypothesis

[SULEYMAN - 51:36]

On Geoffrey Hinton's recent statements about AI safety, Suleyman reveals an interesting exchange. Hinton, initially pessimistic about containment, has become more optimistic, because he sees a potential path through programming "maternal instinct" into AI.

The analogy: a vastly more intelligent entity (a mother) caring for a less capable one (a screaming child). "Digital oxytocin," Suleyman jokes.

"I'm going to need something that's got a little more formula to it. A bit more reassuring. But look, there's 101 different possible strategies for safety. We should explore all of them."

The Three Mile Island Scenario

[SULEYMAN - 1:05:03]

Both the podcast hosts and Suleyman wrestle with an uncomfortable question: will it take a disaster to force cooperation?

Eric Schmidt has reportedly said he's "hoping for 100 deaths"; enough to get government attention, not enough for catastrophe. Suleyman doesn't endorse this but acknowledges the dynamics:

"I think that 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, and every other significant power, to cooperate on safety and containment and alignment."

The logic: these systems threaten the bad actors as much as their victims. Self-preservation is rational for everyone.

"The number one thing to unify all of humanity is an alien invasion. And that alien invasion could be a potential for a rogue superintelligence."

Part IX: The Recursive Self-Improvement Question

How Close Is the Loop?

[SULEYMAN - 56:56]

The conversation turns to what many consider the threshold moment: recursive self-improvement.

Currently, human software engineers are in the loop, generating post-training data, running ablations, checking benchmarks. It's "expensive and slow."

"A lot of the labs are racing to sort of close that loop so that various models will act as judges evaluating quality, generators producing new training data, adversarial models that are reasoning over which data to include."

Closing this loop would accelerate AI development significantly. Some speculate it could lead to an intelligence explosion.

"With unbounded compute and without human in the loop or without control, that does potentially create a lot more risk. But unbounded compute is a big claim."

Hassabis: Agents and Near-Term Risks

[HASSABIS - 53:46]

Hassabis focuses on a more immediate concern: agents.

"The next stage is agent-based systems. I think we'll start seeing some really impressive, reliable ones in the next couple of years. Those will be incredibly useful and capable, but also they'll be more autonomous. So the risks go up."

Google DeepMind is working on cyber defense specifically because of this: "In preparation for a world like that where maybe there's millions of agents roaming around on the internet."

Part X: What It's Like at the Frontier

Hassabis: The Weight of History

[HASSABIS - 49:08]

Fry asks Hassabis directly: does the emotional weight of this work ever bear down on him?

"I don't sleep very much, partly because it's too much work, but also I have trouble sleeping."

The emotions are complex and contradictory. On one hand: "I'm basically doing everything I ever dreamed of. We're at the absolute frontier of science in so many ways. That feeling of being at the frontier and discovering something for the first time, that's happening almost on a monthly basis."

On the other: "We understand the enormity of what's coming better than anybody. What it means to be human. All of these questions are going to come up."

Some impacts have hit harder than expected. AlphaGo's victory over Lee Sedol was bittersweet: "Go was this beautiful mystery and it changed it."

The implications for creativity trouble him: "I have huge respect and passion for the creative arts. Having done game design myself, I talk to film directors and it's an interesting chill moment for them too. On one hand they've got these amazing tools that speed up prototyping ideas by 10x. But on the other hand, is it replacing certain creative skills?"

Suleyman: The Humanist's Dilemma

[SULEYMAN - 1:00:51]

Suleyman wrestles with his own tension: being an accelerationist who genuinely believes these systems pose existential risks.

"I'm also an accelerationist. I want to make these things. But tension is rational. People always say that tension is rational. If you don't see the tension, you're definitely missing most of the debate."

His framework: bounded acceleration. Deploy AI in clinics, schools, workplaces at huge scale... just with hard limits and controls.

"That's the art that we have to exercise."

Part XI: The Next Decade... Predictions and Possibilities

Economic Transformation

[SULEYMAN - 29:01]

Suleyman articulates a coming economic mismatch: labor markets will be disrupted before the cost of services drops.

"The cost of consuming stuff is also going to come down. So we actually have a transition mismatch: labor markets are going to be affected before cost of services comes down. Maybe there's a 10-20 year lag between that, which is going to be very destabilizing."

[HASSABIS - 38:44]

Hassabis has been studying the Industrial Revolution for lessons on navigating technological transition.

"It took quite a long time...roughly a century. Different parts of the labor force were dislocated at certain times. New things had to be created. New organizations like unions."

The difference this time: "It's probably going to be ten times bigger than the Industrial Revolution. And it'll probably happen ten times faster. More like a decade than a century."

Post-AGI Economics

[SULEYMAN - 39:37]

Shane Legg (the third DeepMind co-founder, still at Google) is apparently leading an effort to think through post-AGI economics.

"What an API is is going to start to look kind of different," Suleyman predicts. "Maybe we're principally in the business in 5 years of selling agents that perform certain tasks that come with a certification of reliability, security, safety, and trust."

On universal basic income: "I don't think that's the complete answer. That's just what we can model out now; an add-on to what we have today. But there might be way better systems."

He describes potential direct democracy systems where communities vote on resource allocation with credits, and people who consistently vote for well-received outcomes gain proportionally more influence.

Hassabis: The Sabbatical That Never Comes

[HASSABIS - 54:30]

When asked about life after AGI, Hassabis is pragmatic:

"I could definitely do with a sabbatical. I would spend it like... yes, a week off. Or even a day would be good."

His mission: help the world steward AGI safely over the line for all of humanity. Superintelligence and post-AGI economics will follow, and maybe he can help with that too.

But: "I think that will be my core part of my mission. My life mission will be done."

Then, finally: a holiday.

Part XII: The Ultimate Question: What Makes Us Human?

Hassabis and the Limits of Computation

[HASSABIS - 47:07]

Fry presses Hassabis on the deepest question: is there anything that machines will never be able to do?

His answer reveals both scientific rigor and philosophical depth:

"I've always felt this: if we build AGI and use that as a simulation of the mind, and then compare that to the real mind, we will then see what the differences are."

Maybe it's creativity. Maybe emotions. Maybe dreaming. Maybe consciousness. But Hassabis isn't betting on human exceptionalism.

"Nobody's found anything in the universe that's non-computable. So far. And we've already shown you can go way beyond the usual complexity theory P=NP view of what a classical computer could do; things like protein folding and Go."

His personal philosophy draws from Kant: "Reality is a construct of the mind. All of those things (the light, the warmth of the light, the touch of the table) they're coming into our sensory apparatus and they feel different. But in the end, it's all information. And we're information processing systems."

This is why simulation matters so much to him. If you can simulate something, you've understood it. And if there's no limit to what you can simulate...

"It may be that these are all interchangeable in the end. But we just sense it, we feel it in a different way."

Suleyman and the Speciesist Position

[SULEYMAN - 44:19]

Suleyman's position is more grounded in immediate human welfare than metaphysics.

When pressed on whether enhanced humans might deserve equal treatment with AI (Kurzweil-style uploading or BCIs)he's open but cautious:

"I don't want to make the competition for the peace and prosperity of the 7 billion people on the planet even more chaotic."

The Industrial Revolution parallel: it took a century to develop institutions (unions, regulations, social contracts) that balanced technological power with human welfare. We might have a decade this time.

"Some would call government a geographic monopoly on violence. What I think I'm hearing is some sort of monopoly on intelligence, or at least capabilities exposed to intelligence, in order to contain AI."

Suleyman's response: not a monopoly, but a nested system. Like how the US has military, state police, local police who are all "checks and balances on the system."

"That's kind of what we got to start thinking about designing."

Conclusion: Two Paths, One Destination

The conversations reveal two former partners, shaped by the same origins but diverging in focus.

Demis Hassabis remains the scientist at heart. His north star is Turing machines and what they can compute. His strategy is root node problems: fusion, materials, biology, that cascade into civilizational benefits. His concern is consistency, accuracy, grounding AI in truth. He sleeps poorly not because he fears AI but because discovery is happening monthly and he can't bear to miss it.

Mustafa Suleyman has become the humanist philosopher of AI's leaders. His north star is the 7 billion people who didn't ask for this transition. His strategy is containment first, alignment second, economic planning for the aftermath. His concern is species-level survival. He sees tension as rational: anyone who doesn't feel it is missing the point.

Both believe AGI is coming within years, not decades. Both believe it will be the most transformative technology in human history. Both believe current safety investment is inadequate.

But where Hassabis sees the ultimate scientific experiment—finally testing the limits of computation, finally building a simulation of mind to compare against the real thing—Suleyman sees a governance challenge: how do you maintain stability during a transition that will be ten times bigger than the Industrial Revolution and happen ten times faster?

They are, in many ways, two sides of the same coin. The scientist who wants to push Turing machines to their limits. The humanist who wants to make sure those limits don't push back.

One works for the company that invented transformers and attention. The other works for the company that bet early on OpenAI and is now building its own path.

The race isn't zero-sum, as Suleyman says. Technologies proliferate everywhere. But the visions (the different weights placed on discovery versus containment, on computation versus humanity) these will shape what kind of superintelligence we actually build.

Both men started in the same London office, grinding through the flat part of the exponential, generating handwritten sevens that weren't in the training data.

Now they're building the most powerful technology in human history, from opposite sides of the most important competition of our time.

May the best vision win. Or better yet: may the best of both prevail.