Demis Hassabis on AGI, Robots Scale Production, and Elon’s $1T Mars-Shot Comp | EP 253
Summary
SpaceX’s proposed compensation package turns moonshot execution into an all-or-nothing governance contract. Elon Musk would receive 200 million 10-to-1 super-voting shares if SpaceX reaches both a million-person Mars colony and a $7.5 trillion valuation—roughly a $500 billion payout—plus 60.4 million restricted shares for bringing 100 terawatts of space compute online. The panel’s investor framing: founder control can look dangerous, but “all the really world-changing, crazy-sounding stuff comes from that structure.”
The Musk–OpenAI trial is as much about control of economically transformative AI as it is about nonprofit promises. Musk seeks $150 billion in damages, restoration of full nonprofit status, and the removal of Sam Altman and Greg Brockman; Brockman’s disclosed diary says, “The true answer is that we want Elon out,” while testimony also showed Musk’s team negotiating for-profit equity and xAI distilling from OpenAI models. With Polymarket showing a 33% Musk win probability, Blundin argued Musk “doesn’t need to win to win”—slowing OpenAI’s recruiting, morale, and momentum may be enough.
The panel could not agree whether AGI already exists, illustrating why headline capability claims remain hard to underwrite. Diamandis reported Demis Hassabis’s view that there is a 50/50 chance another breakthrough—perhaps in world models—is still needed, while Alexander Wissner-Gross dates AGI to GPT-3 in summer 2020 because compressing general human knowledge produced general task performance. Steven Kotler called today’s systems “the narrowest technology in the world,” citing worse creative writing and missed interdisciplinary neuroscience connections; Wissner-Gross and Diamandis replied that recursive improvement can come through algorithmic experimentation, massive sampling, and selection rather than humor or literary mastery.
Humanoid robotics is moving from demonstrations into factory arithmetic. Figure increased output from one robot per day to one per hour and targets 100,000 units by 2030; 1X targets 10,000 this year and 100,000 in 2027, while Musk projects one million Optimus units by 2030 and Musk and Brett Adcock both contemplate up to 10 billion humanoids by 2040. Blundin’s capital-allocation call was blunt: “software has a limited life left” because AI writes it so well, while robotics and data centers may offer a decade or more of buildout.
The investable robotics debate is not only how many machines get built, but whether humanoid is the winning form factor. Salim Ismail argued repetitive jobs favor purpose-built wheeled machines, drones, or appliances, while Wissner-Gross countered that elder care and environments designed around human bodies, including nuclear facilities, require humanoids. Wissner-Gross expects humanoids to overtake wheeled robots in the early 2030s, yet thinks the larger prize is “post-humanoid” machinery—potentially many trillions of direct cellular nanorobots by 2040.
China’s worker-protection ruling suggests AI adoption and employment guarantees may advance together rather than trade off cleanly. A Hangzhou court reportedly rejected cutting a worker’s monthly pay from 25,000 yuan to 15,000 yuan after AI automated part of his role, reasoning that AI adoption was a business choice rather than an unavoidable shock. The panel argued China’s shrinking workforce, aging population, and desire to preserve enthusiasm for AI could let it protect jobs without materially slowing automation—making the ruling a signpost in the coming rewrite of the social contract.
GLP-1 economics are already rivaling frontier AI, with manufacturing capacity—not demand—described as the binding constraint. The episode’s chart put 2025 Ozempic and Mounjaro revenue at 2.4 times OpenAI and Anthropic’s, while retatrutide trial data showed 37 pounds of weight loss versus six for placebo over 40 weeks, cholesterol down 27%, triglycerides down 41%, liver fat down 80%, and A1C falling from 7.9 to 6.0. With approval projected for mid-2027, Demis Hassabis called this drug lineage the likeliest route to longevity escape velocity, perhaps in the early 2030s.
By 2028, AI may feel ambient and indispensable—but scarce compute could divide consumer possibility from enterprise access. Predictions included systems with permission to read messages, calls, calendars, recordings, and wearables; persistent AR; perfect memory; real-time coaching; and a generation conducting 70–80% of its conversations with AIs. Yet Friedberg expects enterprise demand to absorb available data-center capacity, warning that today’s affordable access to the best foundation models is a temporary “beautiful moment in time” before a bottleneck lasting into roughly 2030–31.
Deep dive
1. SpaceX turns executive compensation into a Mars contract
Diamandis laid out the package as unprecedented even by Musk standards: 200 million shares carrying 10-to-1 voting power, contingent on both a million-person Mars colony and a $7.5 trillion SpaceX valuation. At that valuation, the shares would be worth roughly half a trillion dollars—“not a Mars landing, not a photo from Mars, not a mouse on Mars.”
The second award is 60.4 million restricted shares for deploying 100 terawatts of space compute. Diamandis emphasized the scale jump: Musk had been discussing 100 gigawatts, while the compensation threshold is 1,000 times larger.
Blundin’s investor case for super-voting stock was that it preserves the founder’s ability to pursue projects conventional boards would reject. He traced the structure through MicroStrategy, Google, and Meta, concluding that satellites, Moon missions, and Mars ambitions often emerge because “those founders are the ones” still empowered to act.
Wissner-Gross saw a new corporate form taking shape alongside shareholder-focused C corporations and public-benefit corporations: companies explicitly organized to achieve moonshots and compensate leaders accordingly. His challenge was expansive—give every S&P 500 CEO comparably ambitious outcome-based goals.
2. Exponential organizations need a navigable present, not a five-year fantasy
Ismail rejected the idea that Musk’s command-and-control style is incompatible with an exponential organization. The founder holds the massive transformative purpose, delegates execution, and intervenes on engineering questions; once teams commit, “they get it done.” If Musk hits the milestones, Ismail said, investors should be “laughing all the way to the bank.”
Founder leadership was treated as the usual prerequisite, though Ismail offered Gucci as an exception: after its CEO committed to turning it into an exponential organization, the company reportedly grew tenfold over three to four years. Diamandis countered that the aspiration is closer to 1,000x or a million-fold change.
Ismail’s operating model came from TEDx. Announcing 20,000 events in five years would have alienated the team; setting the purpose “Ideas Worth Spreading,” defining community rules, and instrumenting progress produced 20,000 events instead of the roughly 2,500 implied by linear planning—and at essentially zero cost.
The corresponding dashboard combines a massive transformative purpose with OKRs and a one-year operating plan measured in real time. The five-year vision stays outside daily management because people “can’t cognitively grasp that”; real-time metrics become the rudder.
3. The OpenAI trial exposes the cost of postponing governance
Diamandis summarized Musk’s claims as breach of charitable trust and unjust enrichment, with requested damages of $150 billion, returning OpenAI to full nonprofit status, and removing Altman and Brockman. Judge Rogers reportedly blocked Musk’s team from turning the case into a trial about AI extinction.
Brockman’s diary supplied the sharpest language: “We truly want the B corp. The true answer is that we want Elon out.” It continued, “If three months later we’re doing a B corp, then it was a lie,” and anticipated “a nasty fight” over any for-profit conversion.
The contrary evidence mattered: Musk’s team negotiated for-profit equity for him in 2017; he admitted not reading the conversion’s fine print; xAI had distilled its models from OpenAI’s; and Altman offered him equity that Musk rejected as a “bribe.” The record was damaging in both directions, not a clean morality play.
Wissner-Gross’s lesson was structural: a nonprofit may unexpectedly “catch the car” and discover its moonshot is economically transformative. Anthropic followed a similar sequence from alignment lab to revenue, models, capital, and a for-profit public-benefit corporation; his prescription was to settle the capital structure at inception.
4. Winning the case may matter less than slowing the rival
Ismail argued that nonprofit status could be appropriate in a future post-capitalist society, perhaps within “two, three years” by his estimate, because chasing money could corrupt the mission. The correct structure, he said, is whichever one most cleanly advances the purpose.
Polymarket assigned Musk a 33% chance of winning, down from roughly 50/50 weeks earlier. Blundin argued the binary misses Musk’s strategic objective: disclosures, recruiting friction, damaged morale, and lost momentum can hurt OpenAI even without a favorable judgment. “He doesn’t need to win to win.”
Friedberg magnified the stakes around super-voting control and self-improving AI. He called it the “trial of the millennium” because capability and political power may concentrate in a few hands; he noted that Altman had contemplated a California gubernatorial run and that technology founders are not merely apolitical startup operators.
Kotler’s refusal was itself useful pushback: if he is not actively working on a problem, he does not like discussing it because “then I’m just gossiping.” Diamandis conceded the gossip label but maintained that the litigation sits at the center of an ecosystem capable of shaping humanity for centuries.
5. AI’s literary failures do not settle its capacity for self-improvement
Kotler’s daily experience led him to say, provocatively, “I think AI is fake”—meaning radically more limited than the panel’s claims. As a writer, he spends more time teaching and correcting it than watching it write; as a neuroscientist, he sees it miss obvious cross-disciplinary gaps despite access to the field’s literature.
His mechanism-level objection was that current systems are convergent-thinking engines, while major breakthroughs often require divergence. He also said ChatGPT’s writing had deteriorated since work began on their book, with several models performing worse rather than steadily improving.
Wissner-Gross separated literary performance from recursive research: AI can test algorithm parameters, transfer functions, and architectures iteratively at enormous speed. Diamandis’s thought experiment was that Claude 4.7 might lose once to Kotler, but could generate a billion attempts and place a selector on top. Kotler conceded: “That’s a fair point.”
Diamandis turned the criticism into an empirical challenge: build a rigorous benchmark encoding Kotler’s standards and demonstrate the regression. Hassabis said a well-constructed evaluation would attract frontier labs because it would become an optimization target—turning subjective dissatisfaction into measurable capability pressure.
6. Robot factories are crossing from prototypes to volume curves
Figure AI reportedly raised capital at a valuation above $30 billion, increased production from one robot per day to one robot per hour, and targets 100,000 robots between now and 2030. Diamandis used the footage to expose a new viewer reflex: deciding whether an impressive robot video is real or a deepfake.
At 1X Technologies, a new 58,000-square-foot Hawthorne facility is meant to support 10,000 Neo robots this year and 100,000 in 2027, with shipments expected later this year. Diamandis said 1X’s CEO Bernt had promised him a Neo by the fall.
Blundin’s Gigafactory insight was that much of production is already automated; humanoids need not fabricate every tiny component. Automation equipment arrives almost like an Amazon package, and a robot performs the remaining human-shaped steps needed to install it. “The bottleneck is the humanoid actions.”
That changed the panel’s investment map. Blundin expects the window around the hottest AI software companies to close as AI commoditizes software creation, while robotics and data centers could keep compounding for a decade or more. Blundin’s near-term job-security pick was robot repair, while Diamandis noted the surrounding insurance and service ecosystems.
7. Humanoids may win human spaces before stranger bodies supersede them
Musk predicts one million Optimus robots by 2030. Diamandis said Tesla may eventually be known for Optimus rather than cars. Kotler challenged whether such declarations are partly capital raising and pre-selling before cheaper Chinese robots arrive, comparing grand promises with Meta’s still-unrealized metaverse pitch.
Ismail’s engineering objection was sharper: repetitive tasks suit specialized forms—wheels, drones, dishwashers—not adaptable humanoid bodies. “At least give it another pair of arms,” he argued; a human shape is an arbitrary constraint when purpose-built machinery could perform the task better.
Wissner-Gross supplied the strongest counterexample: elder care and disaster work occur in environments designed for human bodies. Nuclear facilities, doors, stairs, beds, and the people being assisted make a humanoid form functional rather than cosmetic.
Wissner-Gross compared the ramp with 1.6 billion cars already on roads, 100 million cars manufactured annually, and iPhone output growing from 1.3 million in year one to 250 million a year. He expects humanoids to pass wheeled robots in the early 2030s and, with little remaining room for a contrary prediction, pass humans by the end of the 2030s—but considers direct cellular nanorobots, numbering in the many trillions by 2040, the more consequential post-humanoid form.
8. China treats AI displacement as a social-contract problem
In the Hangzhou case described, Zhao earned 25,000 yuan per month before AI automated part of his role; his employer offered a demotion to 15,000 yuan. The court reportedly sided with him because adopting AI was a discretionary business decision, not an external shock making the employment contract impossible.
Diamandis framed the unresolved economic question: if an employee uses AI to do ten times the work—or performs the job while the employee exercises—does the created value accrue to the worker, company, or state? Existing labor law prices human labor, while the emerging system manages tasks, agents, and radically lower coordination costs.
Ismail called the ruling one of the surreal signposts that reveal a transformation, like a Kentucky coal museum using solar panels. The deeper issue is rebuilding a social contract over the next 10–20 years as institutions designed around human employment confront AI labor.
The panel rejected the concern that protection would necessarily handicap China in an AI race. A declining workforce and aging population leave room for both employment guarantees and aggressive automation; worker-friendly rules may also preserve China’s reported 85% AI optimism, versus roughly 25% in the United States.
9. AGI may be zero breakthroughs away—or still badly defined
Diamandis reported Hassabis’s updated view as a dramatic narrowing: more than a decade ago he expected at least five major breakthroughs, while now he assigns a 50/50 chance that one more—perhaps in world models—is necessary. His stronger bet remains foundation models “because of how successful they’ve been.”
Wissner-Gross went further: AGI arrived no later than the GPT-3 paper in summer 2020. His claimed discovery was that general human knowledge could be compressed into a model and then applied across a diverse range of human-level or near-human-level tasks; transformer refinements since then are incremental.
Kotler immediately called that “an insane statement,” while Diamandis noted at least 14 competing AGI definitions and multiple facets of intelligence: emotional, spatial, musical, linguistic, and contemplative awareness. The panel can raise hundreds of billions for a capability it still cannot consistently “define, measure, or task.”
Blundin reframed the definitional dispute around economic sufficiency. Researchers such as Hassabis and Dario Amodei may hope there is more fundamental science left, while entrepreneurs ask whether today’s architecture can already manufacture goods, treat disease, care for elders, and build abundance. It need not reproduce the last uniquely human skill to transform the economy.
10. Automated science turns model capability into physical discovery
At Lyra Scientific, Diamandis described “science factories” where AI proposes theories, designs experiments, and robots pipette, move materials, gather data, and update the theory. Running that loop 100 or 1,000 times faster than a graduate student could accelerate physics, chemistry, biology, and materials research.
Kotler pushed back on the phrase “solve physics” as an overclaim. Diamandis kept the practical middle: definitional arguments matter, but curing cancer and solving previously inaccessible problems is sufficient reason for excitement.
The episode’s entrepreneur specimen came from dentist Ashley Gaunt. In one afternoon, an AI questioned her into discovering a preventive-health idea, helped brainstorm it, vibe-code an initial app, and create a path from concept to monetization. “I didn’t even realize I had an idea until I started my discussion with my AI.”
Blundin expects that experience to become common and suggested frontier labs may actively surface opportunities because their models can see gaps in training data, applications, and markets. The tool is not merely executing an idea; it may help users identify the idea and their role in pursuing it.
11. De-extinction creates approximations—and a much larger resurrection thesis
Colossal announced the bluebuck as its sixth species, targeting a return in 2028 from a confidential pipeline of 15. The named program also includes the woolly mammoth, thylacine, dodo, moa, and dire wolf, which Diamandis described as already back.
Kotler’s enthusiasm came with two warnings. Species-loss rates are reportedly 1,200 times baseline, so restoring a handful must not trivialize the environmental crisis; moreover, reconstructed animals splice DNA fragments across time and are hybrids—“creatures that have never existed before on this planet”—with uncertain consequences.
Diamandis placed de-extinction inside a more exotic program: use technology to reconstruct every past species and eventually every individual human, whether biologically, computationally, or through simulation. A park containing 100 once-extinct animals would be only the early, comparatively modest expression of that ambition.
12. GLP-1s and AI are competing for factories, capital, and human attention
The episode’s chart said Ozempic and Mounjaro generated 2.4 times the 2025 revenue of OpenAI and Anthropic. Salim Ismail called the comparison “apples and oranges”; Wissner-Gross defended it as biological software versus digital software—capital extending productive human health on one curve and AI labor on the other.
Both categories were described as sold out, with supply rather than willingness to pay setting revenue. GLP-1 demand could support as much as $500 billion, but physical production limits sales; meanwhile, reduced eating may already be lowering US food-shipping truckloads and creating second-order logistics disruption.
Ismail challenged the productivity euphoria around AI’s rapid feedback loops. Dopamine and flow can create ego inflation: “It doesn’t mean that the quality of your thinking matches the quality of your feeling.” His organization’s only piece of branded merchandise carries the warning, “Never trust the dopamine.”
Retatrutide supplied the health-side acceleration: over 40 weeks, 37 pounds lost versus six with placebo, cholesterol down 27%, triglycerides down 41%, liver fat down 80%, and A1C down from 7.9 to 6.0. Diamandis compared Mounjaro with GPT-5.5 and retatrutide with AGI, projecting FDA approval by mid-2027.
13. Longevity escape velocity may arrive through an evolving drug class
Ismail said he had used retatrutide for several months and felt better, particularly because of improved liver function after decades of alcohol-related damage. His account was personal experience, not a generalized clinical claim.
Diamandis expects new generations of drugs to move more quickly from design through testing and availability, with the FDA potentially replacing some phase-three work with expanded phase-two trials. The commercial implication is direct: a therapy credibly offering ten extra years—or treating cancer, cardiovascular disease, and inflammation—would meet enormous uncapped demand.
Hassabis judged the evolving GLP-1 class the likeliest path to longevity escape velocity, perhaps in the early 2030s or sooner. His market signal was Eli Lilly’s trajectory toward roughly $1 trillion, placing longevity alongside AI companies and SpaceX among the few industries producing businesses of that scale.
14. AI in 2028 could be ambient for consumers and capacity-constrained for everyone
Diamandis’s consumer vision begins with permission: AI reads calls, messages, email, calendars, recordings, cameras, and wearable data. His own “Skippy” already has broad access; the endpoint is “ambient AI” or “automagical AI” that adjusts lighting, music, health guidance, and surroundings to the user’s condition and desires.
Wissner-Gross thought the more important 2028 question concerns enterprises. He argued state-of-the-art reasoning is finding stronger demand among organizations able to pay for it, while scientific discoveries—not smart-home conveniences or another Apple-like device category—remain the less obvious frontier. Ismail likewise expects AI-native enterprise operating systems and each person to function like a small company with an AI team.
Blundin introduced a newly launched holodeck—screens on every wall, immersive sound, conversational world-building, music, movies, and vibe coding—showing what consumers may want. Friedberg identified the constraint as compute: enterprise demand is absorbing data-center supply, and he does not expect relief by 2028, perhaps not until “2030, 2031 after the Terafab.”
The speed round made the experience personal: AR glasses that reinterpret a Manhattan street as 1905; autonomous transport at 20 cents per mile; perfect long-term memory; and AI coaching that detects fear and adjusts behavior in real time. Friedberg expects some people to conduct 70–80% of their conversations with AIs.
Wissner-Gross extended “merging with the machines” from wearables to ingestibles. Extrapolating the computer size historically required for one gigaflop, he projected that by the mid-2040s—around the period associated with Ray’s predicted technological singularity—a capable computer could approach the size of a human eukaryotic cell, enabling cell-scale AI nanomachines.