Demis Hassabis on AGI, Robots Scale Production, and Elon’s $1T Mars-Shot Comp | EP 253
In a Nutshell
Demis Hassabis argues AGI emerged in 2020 via GPT-3's knowledge compression, needing no major breakthroughs, while humanoid robotics scales exponentially—Figure AI targeting 100k robots by 2030, 1X 100k in 2027, and predictions of 10B by 2040 for elder care and human environments. Elon Musk's $1T SpaceX comp ties to Mars colony and space compute milestones, exemplifying moonshot incentives amid OpenAI trial drama and China's AI employment protections. AI drives scientific superintelligence, longevity drugs like retatrutide (37lb loss, reversing liver damage), de-extinction (blue buck 2028), and mid-2028 ambient consumer AI via wearables, digital twins, and AR, amid compute shortages and abundance economy shifts.
These notes were generated by AI and may contain inaccuracies.
Demis Hassabis, CEO of DeepMind, says AGI may not need a major breakthrough. Argument that AGI was achieved in summer 2020. Figure AI increased production from one robot per day to one robot per hour, targeting 100,000 robots between now and 2030. 1X Technologies production goal is 10,000 robots this year, 100,000 in 2027. Predictions from Elon Musk and Brad Adcock of up to 10 billion humanoid robots by 2040. "It's probably right. What I think is perhaps even more interesting is what does a post-humanoid robot form factor look like. These are the kinds of conversations and imaginations that one can have in this exponential singularity that we're living in."
Special episode of Moonshots at MIT on technology uplifting humanity. "Welcome to Moonshots, where we distill the singularity. No politics, no bullshit just the technology that changes the world. The breakthroughs helping us uplift humanity. The news that matters most changing our world, impacting our companies and our families. Our mission to help you understand what it is, why it matters with a dose of optimism."
Host expresses gratitude. Introduces Dave Brendan (DB2), guru on exponential investing, partner at Link Ventures, MIT roommate along with Mike Saylor from MicroStrategy. Introduces Alex Quasar Gross (AWG), MIT alum, Link Exponential Ventures team member. Introduces Salim Ismail (Salem Ismail, Hugsies), godfather to host's boys, co-conspirator, co-author, co-founder of Singularity University. Special guest Steven Kotler, three-time Pulitzer Prize nominated author of 17 books, co-author of Abundance, Bold, The Future Is Faster Than You Think, and We Are Gods.
Steven Kotler on returning to MIT campus via Link partnership. Intense connection with Alex on AI, ASI, aliens. Joke about non-MIT alums: Waterloo calls MIT the Waterloo of the South; MIT as Center for Bad Spellers. Supermarket joke: "10 items or less" line, "is it you go to Harvard and can't count or MIT and can't read?"
Special episode for audience who purchased over 100 copies of We Are Gods, aiding national bestseller status, aiming for New York Times bestseller.
First news: Elon's "insane Mars-shot comp." SpaceX board votes on package worth half a trillion dollars payout. 200 million super-voting shares (10-to-1) when hitting Mars colony of a million people or more and $7.5 trillion market valuation: 200 million shares at $7.5 trillion pays out half a trillion dollars. No prior comp packages like this; super-voting pioneered by Sumner Redstone ('91-'92), MicroStrategy (Mike Saylor), then Google (Larry, Sergey), Meta. Locks in founders for world-changing goals like Mars, moon, satellites. Rarely seen from replacement CEOs.
"In exponential time where we are right now, everything sounds crazy yet it's going to happen."
Second part: 60.4 million restricted shares when bringing online 100 terawatts of space compute (previously discussed 100 gigawatts).
Alex: Inevitable; seed of future corporate governance for moonshots. Beyond profit-maximizing C-corps or B-corps, a third type for moonshots with ambitious comp. Every S&P 500 CEO should have such plans.
Salim: Exponential organization comp package, but Elon runs top-down monarchies, command and control. In EXOs, founder holds MTP/vision, delegates (e.g., rocket engines for Mars), thin org structure (5 layers at xAI). Pay tied to moonshot achievement benefits investors.
"If somebody's trying to achieve the impossible, you give them all the respect and support because it literally is insane what they're trying to do."
Only possible in founder-led companies with supportive boards. Exception: Gucci CEO turned it into EXO, 10x'd in 3-4 years.
Dashboard question for moonshot like Mars. True EXO with MTP uses real-time operating plan, OKRs. TED example: MTP "ideas worth spreading," community-driven TEDx with rules vs. linear plan (5,10,15 quarterly events = ~2500 total). Instead, 20,000 TEDx events in 5 years, global brand at near-zero cost. Can't reveal full 5-year vision (team freaks out); use 1-year instrumented plan.
Elon accuses OpenAI of breach of charitable trust, unjust enrichment; seeks $150 billion damages, revert to nonprofit, remove Sam Altman and Greg Brockman. Judge blocks AI extinction focus.
Greg Brockman's diary disclosed: "we truly want the B Corp. The true answer is that we want Elon out. If three months later we're doing a B Corp, then it was a lie. Can't see us turning this into a for-profit without a nasty fight."
Counter: 2017 equity email (Elon's team negotiated for-profit equity for him); Musk didn't read fine print; xAI distilled LLMs from OpenAI models; Altman offered Musk equity (called bribe).
Alex: Corporate governance future. OpenAI started nonprofit to counter Google DeepMind singleton risk, front-run to AGI. Tech advances fast: nonprofits succeed economically, need for-profit structure (e.g., Anthropic shifted from alignment lab to public benefit corp for capital). Lesson: Start as public benefit corp, not nonprofit.
Dave: Flipped Singularity University from nonprofit to B Corp. "Friends don't let friends start nonprofits anymore." Donation-to-investment ratio 100:1.
Salim: Lose-lose trial. Post-capitalist society coming (2-3 years), money won't make sense; nonprofits cleaner, less pay hassles. If true MTP, structure doesn't corrupt (Elon doesn't care about money).
Polymarket: Elon 33% chance of winning (down from 50/50).
Dave: Elon wins by slowing OpenAI recruiting/morale/momentum via documents. Can't reverse $122 billion investment easily. Trial dictates AI future (self-improving, superhuman); super-voting control in few hands. Sam Altman planned California governor run; Zuckerberg assets in trust.
Steven Kotler: Rule - doesn't talk problems not working on (gossip). AI comments absurd; works with AI daily as scientist/writer. "AI is fake." CEOs saying AI smarter than best employees means morons.
Speaker shares daily experience with AI: spends more time teaching it to write and fixing it than watching it write. As a scientist doing neuroscience, finds it absurd that AI, supposed to know everything about neuroscience, misses huge gaps in fields. Describes AI as a convergent thinking engine that doesn't do divergence.
I'll take the high road on this one.
By the way, I'm opinion and nobody much likes me.
Suggestion: if that's your experience, create benchmarks like a Cutotler bench to encode knowledge not reflected in AI.
Quick poll in room: 96% find AI amazing and extraordinary; 4% see challenges and limitations.
Peter positions in middle: technology unbelievable, will do unbelievable things, but takes longer than wanted. No clear definition of intelligence: dozen facets of emotional intelligence, Eastern presence/awareness, musical, spatial, linguistic intelligence. Business leaders use emotional intelligence. AI is analytical/numerically driven, lacks these.
What the hell do you mean by smarter? And there are 14 definitions of AGI at last count. Which one are we talking about?
Amazing to raise hundreds of billions for undefined, unmeasurable concept.
Will navigate cognitive tasks, but middle ground: not where it could be, but can get there quickly.
Point on benchmarks for AI self-improvement: hard math problems like humanity's last exam trigger self-improvement. AI not great writer or funny, but doesn't need to be to build better AI—tweak algorithms, parameters, transfer functions iteratively like evolution.
It's like looking at furry little mammals after asteroid 66 million years ago and saying they're never going to write code.
Counter: as writer/artist/creative vs. coder, AI hasn't improved writing; seen it get worse across models compared to GPT-2 four years ago, GPT-3, ChatGPT worse now than year ago when starting book.
Response: uses top models for neural network tests/experiments; every release dramatically better on self-improvement. Sees improvements, but daily experience as scientist shows obvious misses, e.g., gaps in neuroscience needing multi-disciplinary approaches like embodied cognition.
Thought experiment: Claude 4.7 vs. Stephen on difficult writing task in 5 minutes—Stephen wins, but AI writes billion times then selects best. Humans call cheating; for AI self-improvement, fair.
Encouragement: be rigorous, create benchmark for regression in creative writing, show world; frontier labs will include it, leading to better creative writing.
Final points: first ChatGPT prompt—rewrite Bible Genesis chapter 1 as rap song—blew mind. No AI satisfies profound writer like Stephen, but with Dave's approach (scale + select), AI representing him would be better: remembers everything said, anecdotes, metaphors, brings right thing at right time vs. 1-liter brain.
Question: how with 2000-word working memory, can't remember million tokens even with whole book input? Experience as writer: limited.
Move to Figure Robotics / Figure AI run by Brett Adcock scaling manufacturing. Visited pod with Brett; he's opening speaker at Abundance Summit March 2027 with Figure Robot.
Video shown; first thought: deepfake or real? Real. Increased from 1 robot/day to 1/hour; target 100,000 robots by 2030. Best funded, $30+ billion valuation. Bold Capital investor? Not yet.
Window for hottest AI companies closing; gearing up for robotics and data centers (software limited as AI writes it well; robotics/data centers for decade+).
1X Technologies, met Burnt Borick; robot Neo. New 58,000 sq ft facility in Hawthorne, California. Goal: 10,000 robots this year. Promised robot by fall; 100,000 shipments 2027, later this year. Video: humans doing simple things.
Gigafactory insight: robots make robots; everything automated, arrives in Amazon boxes, assemble by turning screw for new line. Bottleneck: humanoid actions.
Job security next 25 years: robot repairman; ecosystem businesses like insurance.
Elon predicts 1 million Optimus production by 2030.
Tell me how this isn't just a play to get investment money.
Probably to sell before cheaper Chinese robots. Elon scaled Tesla to dominant auto manufacturer, price down like nobody else; Tesla decade/two from now known for Optimus production, not cars.
Mark Zuckerberg/Meta announced metaverse.
Entrepreneurship: declare target, evangelize to attract talent/capital. Boston conservative approach doesn't work in exponential change.
Criticisms: Neuralink vaporware; foundational neurobiological problems; everyone quit for other neuro companies.
I started 17 companies. I am an entrepreneur.
Market fights it out; multiple BCI companies.
Rant on humanoids: good for repetitive/constant tasks (e.g., screwing). Humanoid for adaptable environments, counter to robot use. Chart shows wheeled robots, drones better for form factor. Add third arm (e.g., holding garbage bag). Spectrum: repetitive/robotlike vs. adaptable/human.
Prediction: Elon/Bradcock up to 10 billion humanoid robots by 2040.
Need humanoids for aging humans, human form factors (e.g., nuclear reactor disaster robots built for human environments). Longevity: live longer, need robots for care, economic driver.
Counter: 1.6 billion cars on road today, 100 million manufactured/year. iPhone year 1: 1.3 million; now 250 million/year. Elon's million by 2030, billions 2040s—probably right. Early 2030s: humanoids cross wheeled robots; pass humans by end 2030s. Post-humanoid: exotic forms by 2040 (nanobots, Drexlerian nanobots, trillions in solar system). 10 billion humanoids like 1950s atomic vacuum cleaners.
Fountain Life ad with Dr. Don Mucalem, chief medical officer. Mission: use AI/tech for long/healthy life.
3.3% members have unknown cancer. Majority cancers not taking lives when found late; early detection higher cure chance, easier treatment. Over 3.3% found via full body MRI and early cancer detection screening—not conventional, not insurance-covered yet. Goal: research to democratize wellness.
Know inside body; obligation. fountainlife.com/pater for stage 1 detection.
Chinese court: AI not legally allowed to displace employees. Case in Hang Xiao: tech worker Zho, 25,000 yuan/month, offered demotion to 15,000 after AI automated role; refused, court sided. Logic: adopting AI business choice, not unavoidable shock making contract impossible. Companies can't fire saying AI cheaper.
Framing: if employee uses AI to do job better/10x overnight, value accrues to company or employee? Or state? UBI—government pay? Sam Altman: no society without piece of AI ecosystem.
China full circle: communism to anarco-capitalism back to communism; leading in socialism/communism with Chinese characteristics for AI disruption/disemployment. Ironic: China sets global standards.
Three points: artifact of social contract breakdown; labor laws for humans, moving to AI agents—task management/coordination/execution costs drop radically; indicates broader transformation (e.g., Kentucky coal museum using solar panels).
Navigating social contract key problem next 10-20 years; first audience questions on it.
China 80-85% optimistic about AI vs. US.
US is 25% optimistic and majority pessimistic about tech. In China, people are comfortable with constant government surveillance via cameras everywhere, unlike in the US where it's happening but not admitted or liked. China is adopting AI and robotics aggressively despite policies. They have a diminishing population below replacement level, with people dropping out of the workforce at an insane rate, creating a crisis. Keeping people employed and giving them meaning prevents revolt. China could pass a law preventing firings due to AI without much impact on adoption, as they're growing robots amid declining workforce, smoothing societal unrest.
Framing AI as a US vs. China race is nonsensical. Chinese people are natively entrepreneurial, requiring communism to manage population. China isn't hamstringing itself with worker-friendly AI policies. The race to AI supremacy is real, extending to post-AI superintelligence unlocking transformative scientific innovations. New physics alone is a prize worth an international arms race. Financial interest in a company pursuing superintelligence for unlocking new physics and applied physics. Other countries like Germany favor small/medium businesses in automation. China's policies ensure human employment while incentivizing AI-adjacent jobs.
China manages AI PR to maintain high enthusiasm, similar to US policies requiring hyperscalers to pay their electricity bills. In US, media sells ad views via crisis and worry. China avoids this, promoting AI and robotics enthusiasm via employment guarantees, needed due to aging population.
Demis Hassabis, DeepMind CEO, says AGI may not need a major breakthrough: 50/50 chance, possibly in world models, but bets on foundation models due to success. Demis previously thought 5+ major breakthroughs needed; now down to 0-1. AGI achieved in summer 2020 with GPT-3 paper on large language models/few-shot learners: compressing general human knowledge enables general intelligence for diverse human-level tasks. Refinements like transformers are incremental. General intelligence: capable of variety of human-level tasks via compressed human knowledge.
Objection: narrowest technology, dumb on generality. Seeds of AGI born with GPT-3, like DNA/RNA/Darwinian evolution leading to humanity. Research scientists like Demis and Dario (worried about humanity's future) vs. entrepreneurs Sam and Elon. Researchers seek relevance as simple transformers conquer everything they worked on. Entrepreneurs focus on practical paradise: manufacturing, curing disease, elder care—last human uniqueness irrelevant.
Toured Laya Scientific in Cambridge, run by Jeff V. Moltzen, building scientific super intelligence via science factories: robots run experiments (pipetting, moving), AI proposes theories, designs/runs experiments, gathers data, updates theory—100-1000x faster than grad students. Mining nature for data; about to take off. Solving physics, chemistry, biology, materials science turns trillions into hundreds of trillions. Solveverthing.org. Excites by solving unsolved problems like cancer soon. Optimistic view gives hope; feedback thanks for it.
Daily feed from Alex or BeNonHuman; 30-40 hours/week consuming to understand. Ashley Gaunt (dentist): brainstormed with AI, got idea for preventive healthcare digital business, vibe-coded app iteration, full business plan from idea to monetization in one afternoon. AI found her passion via questions; now prototyping. Anyone with passion, purpose, curiosity, mindset can be entrepreneur. AI labs will promote ideas filling training data/market gaps.
Ben Lamm (Colossal CEO) and George Church (Harvard) announced sixth species: blue buck, extinct 200 years from hunting. In genome editing/lab testing; expect 2028 revival. Sixth of confidential 15-species pipeline: woolly mammoth, thylacine, dodo bird, moa, dire wolf (back), blue buck. Future jobs emerging. Worry: de-extinction might downplay environmental crisis (species die-off 12-300x baseline). When Ben started, Jurassic Park seemed far (no dinosaur DNA in amber). Now, how long until park with 100 animals extinct 20 years ago? Scalable; limiting factor DNA access. Hybrids/approximations from DNA snippets—never existed before; unintended consequences possible.
Broader program: Russian cosmist philosophers like Sioski on resurrecting every human ever lived. Starting with species is special case: reach into past light cone, revive all species (hybrid/simulation), all individuals computationally/advanced techniques.
GLP-1s (Ozempic single peptide, Mounjaro dual) generated more revenue than OpenAI/Anthropic in 2025 by 2.4x. Apples/oranges, orthogonal. GLP-1s as healthspan drugs race vs. AI: human labor productivity via healthspan vs. AI labor efficiency. Determines economy: AI-dominated or biological/digital software. Drugs are biological patches. Demonetization: US food shipping truckloads dropping as GLP-1s reduce food consumption—great but logistics issues. Both sold out; drugs easier to produce than GPUs. GLP-1 market could hit $500B, limited by supply. AI indispensable: MIT seniors can't code/research without it; wake up talking to AI like Skippy. 9-10 day work weeks possible. Singularity economy graphing.
AI gives dopamine, flow, high productivity/happiness vs. coding all night. Too much dopamine like cocaine/mania: feels genius but not; ego inflation. T-shirt: never trust the dopamine. Feels like productive party.
Eli Lilly's retatrutide (triple agonist peptide; Mounjaro=GPT-5.5, retatrutide=AGI): 40 weeks clinical data—37 lb weight loss vs. 6 lb placebo; cholesterol -27%, triglycerides -41%, liver fat -80%, HbA1c 7.9 (diabetic) to 6.0. Longevity drug; FDA mid-2027; available from research labs now. Evidence toward longevity escape velocity. Using retatrutide months: reverses decades alcohol liver damage. New drug generations accelerating (design to availability); FDA shifting phase 3 to expanded phase 2. Capped by manufacturing, not market. Drugs for +10 years life, curing cancer/CVD/inflammation coming. Longevity escape velocity likely early 2030s or earlier.
Likeliest way LEV (longevity escape velocity) plays out is through the GLP1 drug class and its evolutions. Economy reflects this: companies worth almost a trillion or more (now or soon publicly traded) are few—AI companies, SpaceX, OpenAI, Anthropic, and Eli Lilly (publicly traded, already at or nearing trillion on trajectory). Market signals longevity drugs and AI as obvious industries of the future.
Prediction conversation: what will AI feel like by mid-2028? Currently, AI is an app on phone—intelligence layer, but majority of world not experiencing it. Transition coming in next 2.5 years to feel very different. Peter Diamandis wrote Substack at peterdiamandis.substack.com on this.
First prediction: give AI permission to know everything—listen to phone calls, watch home cameras, read emails, everything. Example: gave Skippy access to all Granola recordings, WhatsApp, iMessages, email, calendar—responses amazing, AI knows user better than anyone.
Alex reframes: what will AI feel like to consumers by mid-2028? Past 6 months show consumers not great customers for state-of-the-art AI. OpenAI tried turning consumers into power users for reasoning models, failed, pivoted to enterprise. Shut down consumer video division and others, favoring code-generating models that recursively self-improve for enterprise (they have money and desire for advanced reasoning).
For consumers: AI connected to wearables, inaudibles, has user's back as physician. Knows hard conversation with spouse, exhaustion, poor sleep. Walk in room: lights dim, robot pours wine, favorite music plays, comedian on—ambient AI and automagical AI where world adapts to desires.
Prediction user wants: mid-2028 AI exoc (quasi-upload, digital twin in cloud) knowing everything, optimizing meatbody existence. Unlikely. Consumer perspective: better augmented reality, better robots in streets—already priced into market. Non-obvious: new scientific discoveries (consumers not driving).
Question maps to Apple products: smart home, robots, wearables, maybe ingestible robots. Not essence of AI or frontier. Industrial: Leya Scientific, Colossal, Isomorphic Labs creating breakthroughs.
Selene: institutional use profound. Every person operates like small company with powerful AI team. Enterprises run on AI-native operating systems. Need to rewrite civilization's operating system—all businesses/institutions for 10,000 years based on scarcity, now entering abundance. Business models, nation-states geared to scarcity; complete rewrite starting now.
Dave: Launched holiday deck at Link Studio—room with monitors on walls, close door, AI talks interactively: create virtual worlds, songs, movies, vibe code, pulsing music. Consumers experiencing it say "I need that at home." By 2028, massive compute shortfall vs desire. Enterprises sucking up data centers; no solution by 2028 (hoping 2030-2031 post-terrafab). Current beautiful moment: best foundation models affordable via website—take advantage before 2028 bottleneck.
New eyewear: OpenAI's Johnny Ive project, Apple/ Meta wearables. Ambient AI sees what you see, listens, uploads to personal Skippy/open clone. Different feel: walking Manhattan, AR glasses give 1905 tour or entertainment (e.g., Star Wars droids on May the Fourth). First wearables enabling this.
Agree on Verner Vinge's vision (wrote on near-future AR/VR metaverse). By 2028, smart glasses with compelling long-life augmented/mixed reality via battery/energy density progress. But won't matter enormously.
Easy projection: go anywhere autonomously for 20 cents a mile.
Speed round unexpected 2028 AI feels:
- Dave: Video games (spend more than all other media combined). With personalization, own voice, remembers personality/past: 70-80% of generation's conversations with AIs, not people (immersive if designed well).
- Infinite, perfect, long-lasting AI memory (human memories flawed)—records everything, recalls who met.
- Stephen: AI coaching/psychology (hire employee twice: normal and scared). AI in ear coaches real-time (humans run on valence, arousal, approach/avoid). Leads to better humans. Defense Dept using AI therapy (remarkably good); people using for relationships.
Alex consumer vision: merging with machines—AR, wearables, ingestables. By decade end, many swallowing computers (beginnings with pillbot form factors). Projection: extrapolate computer size for 1 gigaflop (Apple launches at this: original Apple Watch, iMac, iPhone). By mid-2040s (Rey Kurzw eil singularity), size of eukaryotic human cell. Excited for nano machines running AIs.
Wrap with applause for David, Alex, Steven. Subscribers get news; consider subscribing. Join weekly Metatrends newsletter (2-min read on trends impacting family/company/industry/nation) at diamandis.com/metatrends.
Keep Peter H. Diamandis in your library
Save the videos and channels worth coming back to, and find them again in one place.
