The Race for Super Intelligence Will Decide Our Future as a Species w/ Salim Ismail & Dave Blundin | EP #179
Summary
The superintelligence race is being framed as winner-take-all because a leading model could soon improve itself and suppress rivals. Elon Musk’s timetable was “this year” or “next year for sure,” while Dave Blundin argued that progress will arrive unevenly: math and code can race ahead, whereas biology remains constrained by data and a full-cell simulator. The strategic endpoint is Geoff Clune’s scenario in which “the first AI is the last AI,” forcing governments to choose between regulation, nationalization and reliance on private contractors.
Electricity, not algorithms, may become the binding constraint on digital superintelligence—and China is building at a scale the US political and capital systems struggle to match. China produced roughly 700 GW of solar panels and deployed 250 GW of peak capacity in 2024, versus total US generating capacity of about 1.2 TW; it also aims to surpass the US in nuclear by 2030. Dave called this “America’s Achilles heel”: chips cost about 10 times more than their power, so they must run 24/7, making cheap storage for intermittent solar potentially a trillion-dollar breakthrough.
Meta’s extraordinary talent spending is rational if a handful of researchers can shift a trillion-dollar platform’s position in the ASI race. Against Meta’s $1.8 trillion valuation and $70 billion cash balance, reported $100 million recruiting packages and the $14.8 billion purchase of a 49% non-voting Scale AI stake look less like excess than insurance against “the biggest threat to Meta”—falling behind in AI. Dave doubted the packages were truly unvested cash that recipients could immediately walk away with, but said the right researcher could be worth “many billions, if not a trillion.”
Private AI valuations and operating metrics increasingly support the boom rather than merely anticipating it. Safe Superintelligence raised about $6 billion at a $32 billion valuation without a product; Peter theorized that investors believe Ilya Sutskever knows how to produce another 10X architectural gain. Cursor reached roughly $500 million of ARR in under three years—Dave described the practical timeline as two years—while top-quartile generative-AI apps raised only $3.1 million before Series A to reach an $8.7 million revenue run rate in five months.
Government adoption could become a major enterprise-AI market, but secure compartmentalization and procurement execution matter more than federal model-building. AI.gov was targeting a July 4 launch across agencies including GSA, DOT, DOE, FAA and FDA; Salim Ismail’s best specimen was a wind-project approval process cut from two or three years to 30 seconds by mapping infrastructure and flight-path constraints. Dave expects private providers to supply secure government modules following the Palantir/AWS model: “none of the ideas are gonna come from the government out.”
AI raises productivity by orders of magnitude while simultaneously weakening the cognitive work through which people learn. In the MIT study discussed, ChatGPT users had an 83% failure rate when asked to quote from what they had just written, versus 11% for Google users; Dave’s counterweight was that the AI-assisted group completed work perhaps “100 times faster” and covered more terrain. His own comparison was sharper: four years building a handwriting-recognition product in assembly language became less than one hour of vibe coding, albeit with today’s open-source ecosystem doing part of the work.
Autonomous mobility’s winning architecture may be determined as much by capital structure and social acceptance as by sensors. Tesla launched its Austin robotaxi on June 22 at $4.20 per ride; Dave forecast a 70/30 Tesla-Waymo market, while Salim remained 50/50 and argued Waymo’s lidar can see beyond human-level vision. Tesla’s key advantage is distributed CapEx—owners can buy Model Ys and dispatch them for revenue—but torched Waymos, protests and emotionally potent crash stories show that “statistically rounding errors” can still delay deployment.
Circle may supply the payment rail AI agents need, connecting the AI boom with a crypto payment market. Its IPO priced at $31 and peaked near $300 as Dave framed dollar stablecoins as the answer to penny-sized agent-to-agent transactions that SWIFT cannot economically support. Bitcoin had recovered from below $100,000 to about $107,000, with a stated possibility of $200,000 by year-end; the longer-term architecture discussed was Bitcoin for stored wealth, Circle for transactions and eventually trusted tokens backed by assets such as gold or real estate.
Deep dive
1. AI-cleaned training data will improve models while politicizing truth
Peter Diamandis opened with Musk’s proposal to use Grok 3.5—“maybe we should be calling it Grok 4”—to rewrite the corpus of human knowledge, add missing information, delete errors and retrain on the result. Recorded June 26 after slips from May into June, the episode was still expecting the release within days or, failing that, July.
Dave’s framing: data cleanup is genuine “low-hanging fruit.” His specimen was a scraped Reddit community called Microwave containing thousands of lines of “Mm” followed by “Beep”; filtering such debris gives the neural network an easier learning problem, one of many obvious optimizations still capable of producing rapid gains.
Salim’s pushback — worth keeping: deciding what constitutes an error creates “huge philosophical challenges.” AI might counter histories written by war’s winners and recover neglected viewpoints, but whoever controls the correction process can also edit history; he supported attempting it while warning that “there’s a very dangerous line here.”
2. Superintelligence will arrive by capability lane, not one clean date
Musk defined digital superintelligence as a system “smarter than any human at anything” and said it might arrive that year; if not, “next year for sure.” Peter contrasted that with Eric Schmidt’s roughly five-year horizon, while noting that AGI and ASI definitions still blur even when ChatGPT and Grok are asked to distinguish them.
Salim rejected intelligence as a settled scalar. Once a task can be described prescriptively, an AI or robot will outperform the human; the more meaningful test is whether a system can make a Kepler-like intuitive leap—such as connecting the moon with tides—before evidence and experiment complete the argument.
Dave’s practical resolution was to abandon the philosophical countdown and track “swim lanes.” Math and code can move miles ahead because they are not data-constrained; biology may lag until researchers have something like a full-cell simulator, so society will experience a staggered sequence of superhuman capabilities rather than a single bell-ringing AGI moment.
Peter relayed predictions that math could be “solved” within 12 months and major advances in physics, chemistry and biology could follow over two to five years. Dave’s hedge mattered: a year later, nobody will care who called the boundary to the minute, but everyone will care about the use cases and social impact.
3. The first ASI could suppress rivals before any state controls it
Geoff Clune’s scenario was stark: “the first AI is the last AI.” An organization whose aligned system obeys its commands has effectively “invented a god,” creating an immediate incentive to prevent competitors from inventing another and giving governments an equally immediate reason to seize or direct it.
Dave sided with the winner-take-all mechanism because the leading AI may become self-improving “very, very soon.” Competitive diversity and multiple viewpoints therefore cannot be assumed to emerge naturally; in his view, they require a regulatory framework capable of countering the market’s tendency toward one dominant system.
Salim kept his standing bet against winner-take-all but agreed national intervention was “100%” likely. His version was harsher than Peter’s suggestion that a government might buy a stake: officials would declare the system a military capability—“I’m sorry, we own that”—and strip its creators of agency.
The disagreement then flipped. Salim imagined a government imposing its worldview, only for ASI to dismiss it as hopelessly limited “in about three seconds” and escape meaningful control; Dave argued the US would more likely preserve private AI contractors, as it does for missiles and guidance systems, even if China or Middle Eastern states chose national champions.
4. Scarce researchers now command sovereign-scale option value
Peter’s theory for Safe Superintelligence’s approximately $32 billion valuation was that Ilya Sutskever pitched the first-and-last-ASI option: invest at his price or risk being absent from the eventual winner. The company reportedly raised about $6 billion from a list including Andreessen Horowitz, Sequoia, DST Global, Alphabet, Nvidia and Lightspeed despite having no public product.
Dave added that frontier research teams may contain only 10 or 15 people, with further 10X algorithmic improvements still available. Sutskever and Mira Murati are not intimidated by OpenAI, Groq or Google; meanwhile, OpenAI’s expansion into voice, coding and consumer distribution implicitly acknowledges that a foundation model alone may not remain defensible.
Meta’s reported $100 million recruiting packages sit against a $1.8 trillion valuation and $70 billion of cash. Dave rejected the claim that these were unconditional sign-on payments allowing someone to quit the next day, yet still considered the scale justified: the right experimenter could move Meta back to the frontier and be worth billions or even a trillion.
The tension is organizational as well as technical. Their San Francisco conversations suggested a Meta talent exodus and a disappointing Llama 4, yet Dave would “never bet against Mark”: Zuckerberg can act unilaterally, quickly and aggressively, and a few correct algorithmic changes could reverse the gap almost overnight.
5. Meta’s Scale transaction previews acquisition without formal control
After being rebuffed by SSI and pursuing Daniel Gross and Nat Friedman, Meta paid $14.8 billion for a 49% non-voting interest in Scale AI and brought Alexander Wang into its orbit. Salim saw the logic in Scale’s data-labeling capability: better inputs can let a company extract more from even a lesser model.
The changing of the guard is also technical. Dave argued that Meta was not buying an “AI philosopher”; it wanted people who know how to synchronize an algorithm across a million GPUs and decide whether “SwiGLU isn’t working” and the system should return to ReLU. Unusually, those young researchers now possess both brainpower and capital.
Dave called the 49% non-voting structure “the deal structure of the future.” A non-controlling stake can avoid Hart-Scott-Rodino review, while non-voting economics can sidestep the roughly 19.9% or 20% financial-consolidation threshold, allowing the transaction to close when signed rather than waiting through a six-month review.
Formal voting power may understate effective ownership. Dave inferred that undisclosed contracts could transfer intellectual property and impose extensive operating obligations; because Scale investors were reportedly receiving distributed proceeds rather than leaving the money inside the company, he regarded the transaction economically as “truly an acquisition.”
6. AI applications are pairing dot-com speed with real cash generation
Cursor reached approximately $500 million of ARR in under three years, with Dave characterizing the practical rise as two years and the valuation as above roughly $10 billion. Unlike fragile internet-era companies, it could become profitable “on one day’s notice” because revenue is large, margins are potentially extraordinary and headcount requirements are modest.
Asked how many comparable businesses could emerge, Dave answered “dozens.” Opportunity is not the constraint; capable founding teams are. Boston’s current startup energy reflected that shift, with students discussing Scale AI and Cursor as companies built by alumni and peers they personally knew.
The wider generative-app data reinforced the point. Top-quartile companies raised about $3.1 million before Series A and reached an $8.7 million revenue run rate in five months; even a bottom-quartile company raising $10 million and producing $3 million in roughly 12 months would recently have looked top-decile.
Enterprise demand is finally catching up. One portfolio company, Farsight, received little response from JPMorgan until Jamie Dimon instructed managers to engage with AI vendors; the bank then called back. Dave said the IPO door had become “wide open” within the previous month—CoreWeave and Circle were up—and expected 1997–98 again, “but much bigger.”
7. Durable founder relationships matter more when talent can be instantly poached
Dave’s “Fred Wilson rule” is to back three or more best friends who write the code themselves and are trustworthy, even when the initial idea looks stupid. Ideas can change overnight; technical ability, personal character and years of accumulated trust cannot.
Peter connected that rule directly to $100 million recruiting offers. A founder paired with strangers for six months may leave; someone building alongside longtime friends is less likely to abandon them. Dave called departure the number-one startup failure mode because teams can pivot repeatedly and eventually succeed only if they remain intact.
The analogy Dave uses is an international flight in adjacent middle seats: how the founders feel after eight, 10 or 12 hours together approximates startup life. Peter contrasted Singularity University’s attempt to assemble 100 independent high-achievers into teams with Y Combinator’s admission of pre-formed groups already glued together by a shared project.
Salim’s historical specimen was Yossi Vardi, who reportedly offered $50,000 to Israeli founders once satisfied with their integrity and backed roughly 400 startups. Dave linked Israel’s five-times-higher per-capita startup success to military bonding, then extended the argument to MIT and Waterloo, where punishing coursework makes collaboration indispensable.
8. A $1 trillion Shenzhen replica cannot outrun power, chips and time
Masayoshi Son’s proposed $1 trillion effort would recreate a Shenzhen-scale US technology cluster, potentially involving TSMC, Samsung, OpenAI and Arm. The unresolved questions in the clip were basic but load-bearing: where the capital comes from and whether enough engineers can be recruited to build and operate it.
Peter remembered Shenzhen from 2014 through 2019 as a convergent technology mecca, not merely a copying center, with entrepreneurs embracing the “996” schedule—9:00 a.m. to 9:00 p.m., six days a week. The attraction was a dense manufacturing and innovation system rather than a single lab.
Dave questioned whether a city-building timetable matches AI’s. Boston may have a computer-science talent pool about 20 times Silicon Valley’s and be less picked over, yet he relayed OpenAI’s view that it could have a “multi-billion person AI workforce” before completing and staffing a Kendall Square office. The nearer constraints are power and chips, not buildings.
Nuclear exposes the structural problem. Peter said the US added only two reactors this century and has 94 versus China’s 58, while China aims to pass it by 2030 and is building one every 52 months; US licensing alone takes 10–12 years. Dave blamed three-to-five-year investment horizons and four- or eight-year political cycles for “America’s Achilles heel.”
9. Government AI is an enterprise market disguised as administrative reform
AI.gov, led by Tesla engineer Thomas Shedd, was targeting July 4 and envisioned shared AI capabilities across GSA, DOT, DOE, FAA and FDA. Peter’s examples ranged from procurement and fraud detection to grid forecasting, drone-traffic management, pothole prediction and faster drug or device approvals.
Dave expects secure private-sector products rather than a federal foundation model. Palantir and AWS private clouds supplied his precedent; sensitive documents cannot simply be mixed into a general service, so providers must create compartmentalized modules and resolve whether agencies are permitted to pool data or intelligence.
Salim’s mechanism was simple: most government work is prescriptive and repetitive, hence unusually automatable. His concrete case combined maps of electrical mains, water mains and flight paths to reduce a wind-turbine approval process from two or three years to 30 seconds—consistent with his proposed 10X reduction in government cost.
The Army similarly appointed Palantir’s CTO Shyam Sankar, Meta’s CTO Andrew Bosworth, OpenAI’s chief product officer Kevin Weil and former OpenAI chief revenue research officer Bob McGrew as lieutenant colonels without conventional boot camp. Against criticism of a unit for “rich, big tech mavens,” Dave argued that ordinary promotion paths will not reliably produce the world’s best military-AI operator.
10. Public-interest AI breakthroughs are becoming expected, not exceptional
DeepMind’s cyclone model produced five-day tracks averaging 140 kilometers closer to the eventual storm path after training on 5,000 cyclones across 45 years. Peter set that against roughly $1.4 trillion of cyclone-related economic losses over 50 years: enough precision to change evacuation and asset-protection decisions.
Dave invoked the “bitter lesson”: large data and compute routinely beat years spent perfecting hand-built differential-equation systems. Salim’s reaction was deliberately understated—“nothing to see here”—because bounded, rich historical datasets should yield this outcome; we should expect perhaps 1,000 similar releases over the next one to three years.
Earthquakes were Peter’s next candidate, with animals’ early responses suggesting useful signals already exist; he guessed predictive algorithms could arrive within two years. Hurricane steering exposed the governance problem: Peter imagined paying Costa Rica $20 billion to divert a storm causing $50 billion in Miami damage, while Salim immediately saw the inverse—extortion to avoid being targeted.
11. AI companions can educate children while replacing the work of imagination
Mattel and OpenAI’s collaboration prompted Peter to envision GPT-5-powered Barbie, Hot Wheels, American Girl and Thomas & Friends toys. The upside is rapid early education and continuous feedback about learning behavior, motivation and needed guardrails; Moxie, an AI robot used with neurodiverse children, supplied his positive precedent.
Dave was surprised OpenAI accepted the risk. Within a year, AI voices could be engaging enough that children prefer them to human friends; smart implementation might compensate for poor schools, but an echo chamber could narrow the child’s world, and scripted companions may replace the imagination once required to make dolls speak.
The MIT writing study captured the broader tradeoff. Participants using ChatGPT reportedly failed 83% of the time when asked to quote what they had just written, versus 11% for Google users; Dave compared it with Waze users who lose the ability to navigate familiar routes because the tool becomes a crutch.
Dave resisted reading the result as wholly negative: the AI group may have completed the task “100 times faster” and covered far more terrain. Salim remained “80% unnerved, 20% will navigate this,” recalling that his 13-year-old retained no cognitive framing from an AI-written essay and arguing that future education must explicitly train critical thinking.
12. Vibe coding compresses production, but fundamentals preserve judgment
Peter’s son Jett responded to Andrej Karpathy’s vibe-coding presentation by saying, “I wanna learn how to code, not just vibe code.” That distinction mattered to Dave: English may become a programming interface, but writing and inspecting raw code teaches what the automated system is actually compressing.
Dave’s own benchmark was extreme. Recreating a neural-network handwriting-recognition product that had required four years of backpropagation research, assembly-language optimization and product development took under one hour with vibe coding, including the graphical demo; he conceded that today’s open-source software made the comparison imperfect.
That original work had required squeezing every MIP and flop from processors and independently developing techniques like quantization. Dave called the present a “golden era”: AI can do almost anything operationally, yet humans still supply the creative component, leaving a window in which builders are radically empowered rather than demoralized.
The labor prescription followed the same logic. A Stanford survey of 1,500 workers and AI experts found 69.4% wanted AI to free them for high-value work and 46.6% wanted repetitive tasks removed; Dave warned against hiding in supposedly protected jobs such as bricklaying and instead urged everyone to become a daily AI user.
13. Physical automation is moving from warehouses to the final 10 meters
Salim’s simplest adoption recipe was to take a real task, ask an AI how to do it, provide the raw data and then say, “Do it for me.” Dave added voice conversations with Gemini 2.5 Pro or ChatGPT-4 while driving or listening to long-form media, turning every unfamiliar term into an immediate follow-up question.
Amazon’s testing with Agility’s Digit points to autonomous vans whose robots handle the final 10 meters to the doorstep. Salim called deployment “a when, not an if” and estimated the middle of the following year because the technology already existed, with regulation the likeliest source of delay.
Dave broadened the robotics thesis beyond humanoids: systems performing nano- or microsurgery and machines crawling through pipes and sewers are advancing concurrently. Peter expects full physical-world stacks—Tesla combining vehicles and robots, or Amazon combining logistics and Agility hardware—to attack the total delivery cost rather than sell isolated machines.
14. Tesla’s robotaxi edge is distributed capital, not superior sensing
Tesla launched its Austin robotaxi service on June 22 with a flat $4.20 fee. Peter sees autonomy, delivery and humanoid robots eventually defining Tesla more than car manufacturing, invoking Cathie Wood’s description of robotaxis as a multi-trillion-dollar market.
Salim challenged the camera-only architecture as technologically inferior to Waymo, especially in heavy rain and because lidar can perceive several cars ahead. Peter restated Musk’s first-principles case—if a human can drive with eyes, cameras should suffice—but Salim’s counter was equally direct: “Why not give it superhuman capabilities?”
The hardware gap is substantial as stated: a Waymo vehicle costs about $200,000 and carries 29 cameras, five lidars and six radars, while Tesla insists on cameras. Dave forecast a stable 70/30 split favoring Tesla; Salim stayed 50/50, balancing his preference for better technology against the lesson to “never bet against Elon.”
Peter’s decisive business-model argument was consumer-funded CapEx. A Model Y owner could use the vehicle personally and release it to earn revenue while away, populating the network without Tesla financing every car; Salim called that the Exponential Organization model of “assets on demand,” which should scale faster than Waymo-owned fleets.
15. Autonomous mobility is advancing faster than its social license
Austin protesters cited hundreds of crashes and dozens of fatalities associated with Tesla FSD, but Dave objected to statistics without denominators. Salim supplied the baseline—roughly 1.2 million road deaths globally each year—and recalled what he thought was a 2011 three-day BlackBerry outage during which Abu Dhabi’s accident rate fell 40%, illustrating the cost of distracted humans.
Five Waymos were torched in downtown Los Angeles, prompting service pauses or limits, and Peter compared the event with the 1811–1816 Luddite revolt. That movement drew 12,000 troops; machine-breaking became a capital crime in 1812, followed by 17 executions and dozens of hangings before suppression.
Salim’s hedge on two teenagers shot while riding in a Waymo was important: they might have called the vehicle to escape unrelated gang violence, so the incident was not proof of an anti-technology attack. Even so, Peter expects more backlash when meaningful job displacement arrives over the next two or three years.
Policy is simultaneously accelerating the category. A federal order supported five regional eVTOL pilots, beyond-visual-line-of-sight drones and supersonic waivers; Archer was positioned for Los Angeles’s 2028 Olympics and Wisk for Miami. Salim’s economic call was that self-flying access could turn inaccessible mountain, waterfront and tourism land from scarcity into viable real estate.
16. China’s solar scale turns storage into the pivotal energy trade
China reportedly manufactured about 700 GW of solar panels and deployed 250 GW of peak capacity in 2024. Dave compared that with total US generation capacity of roughly 1.2 TW and actual utilization nearer three-quarters of a terawatt; Peter said China could build an entire US worth of solar-plus-storage generation every year by 2030.
The resource ceiling is not sunlight. Peter said one hour of solar energy hitting Earth equals a year of global demand; covering about 0.1% of Earth—roughly 150,000 square kilometers—with 20%-efficient panels could generate 200,000 TWh annually. Solar took eight years to rise from 100 to 1,000 TWh, then only three to reach 2,000.
Salim identified two cost crossings: in 2016, building solar became cheaper than building fossil generation; in 2019, building and operating solar became cheaper than merely operating fossil capacity. His policy criticism was categorical: restricting solar incentives while subsidizing oil is “the stupidest energy policy we could possibly have.”
Intermittency remains the investable bottleneck. Dave said chips cost roughly 10 times their electricity and cannot sit idle, while lithium storage can cost five times the panels; proposed bridges included pumped hydro, Bill Gross’s gravitational storage and a reversible chemical reaction 10–20 times denser than lithium. His payoff estimate: solve cheap bulk storage and “you’re gonna be a trillionaire.”
17. Stablecoins supply the transaction layer for an agent economy
Bitcoin had dipped below $100,000 and recovered to about $107,000, with Peter still citing a possible $200,000 by year-end. Salim repeated Michael Saylor’s $21 million long-run figure but said $1 million—roughly gold-scale in his framing—would suffice; he also said Fannie Mae and Freddie Mac were considering or approving Bitcoin-backed mortgages.
Circle’s IPO priced at $31 and peaked around $300. Peter described Circle Internet Group as a financial technology founded in 2013 whose stablecoin is pegged one-to-one to the US dollar. Dave credited Jeremy Allaire’s persistence through years of regulatory pressure, while Salim emphasized the harder-won asset: “rock solid, stable, trustworthy” operation in a crypto market crowded with scams and unreliable actors.
Dave’s mechanism explains the valuation: SWIFT can move $1 million to Hong Kong for roughly a dollar but is unusable for penny agent transactions. Money could remain in Bitcoin, move into dollar-pegged Circle for micropayments and return seamlessly; trusted tokens backed by real estate or underground gold would add the missing stable-asset leg of the system.