Summary

The AI race is becoming an energy race, and Diamandis calls electricity “America’s Achilles’ heel.” The US added only two nuclear reactors this century and licensing takes 10–12 years, while China is targeting nuclear leadership by 2030 and Dave Blundin says it produced roughly 700 GW of solar panels in 2024. With US generating capacity around 1.2 TW, China could soon add an entire America’s worth of solar-and-storage power-generation capacity annually: “China is going all in on energy production, and it’s epic.” The key bottlenecks are generation, distribution and cheap long-duration storage.
Frontier-model economics look winner-take-all because self-improvement could turn a temporary technical lead into a strategic monopoly. Jeff Clune’s thought experiment is that “the first AI is the last AI,” suppressing rival ASIs and inviting government intervention; Blundin agrees that “the natural dynamic is winner takes all because the AI becomes self-improving very, very soon.” Diamandis’s theory for SSI’s $6 billion raise at a $32 billion valuation is that investors believed Ilya Sutskever could build a safe first—and last—ASI. Meta, worth $1.8 trillion with $70 billion in cash, is reportedly offering nine-figure packages for researchers.
AI applications are already producing software economics that make the dot-com era look capital-heavy and slow. Cursor reached $500 million of ARR in under three years, and the discussion later cited two years; an Andreessen Horowitz cohort’s top quartile reached $8.7 million of run-rate revenue after raising only $3.1 million before Series A. Even its bottom quartile’s $3 million in 12 months would have ranked near the top three years earlier: Blundin calls these “the best companies financially that we’ve ever seen in history.”
The US government’s AI strategy depends on importing private-sector capability while protecting sensitive data. AI.gov was targeting a July 4 launch across procurement, transport, energy, aviation and drug regulation, while Palantir, Meta and OpenAI executives received lieutenant-colonel appointments to bring technical judgment directly into the Army. The implementation challenge is secure, compartmentalized infrastructure modeled on government clouds—not feeding tax returns, contracts and classified records into general-purpose public systems.
Autonomous transport is moving from demonstration to deployment, but capital structure and public legitimacy may matter as much as driving performance. Tesla launched its Austin robotaxi service around June 22 at a flat $4.20, using cameras against Waymo vehicles described as costing $200,000 with 29 cameras, five lidars and six radars. Tesla can potentially let customers finance millions of vehicles and release them into the fleet, leading Blundin to forecast a 70/30 Tesla–Waymo split; protests, torched Waymos and emotionally powerful crash footage remain the regulatory counterweight.
The near-term human advantage is judgment, curiosity and durable collaboration—not the mechanics of producing code or prose. Blundin recreated in under an hour a handwriting-recognition product that originally required four years of assembly-language work, yet backed Diamandis’s son’s desire to learn conventional coding because understanding the substrate reveals what AI is accelerating. The panel’s warning is equally concrete: AI-written schoolwork weakens recall, while the current moment remains a “golden era where you’re empowered, but not demoralized or crushed.”
Circle’s public-market surge was framed as infrastructure for an agent economy, not merely another crypto trade. Circle priced at $31 and peaked around $300 as the panel argued that dollar stablecoins can settle tiny agent-to-agent payments that subscription billing and the SWIFT network cannot economically support. Their proposed flow is to store wealth in Bitcoin, move into stablecoins for transactions and eventually add trusted tokens backed by assets such as gold or real estate: “The whole AI economy needs to move to these micropayments.”

Deep dive

1. Cleaning the corpus is an obvious model gain—and a governance trap

Diamandis opens with Elon Musk’s proposal for Grok 3.5 to rewrite humanity’s knowledge corpus, add missing information, delete errors and then retrain. Blundin calls it “low-hanging fruit”: scraped data includes junk such as a microwave subreddit containing thousands of “M” characters followed by “beep.”
Diamandis asks who decides what counts as an error. AI might rebalance history written by victors, yet “there’s a very dangerous line here” if one system quietly filters out competing interpretations.
Blundin’s immediate evidence of acceleration is conversational Gemini 2.5 Pro or GPT-4 voice mode: an engaging one- or two-hour drive was newly possible within the prior month. At the June 26 recording, Grok 3.5 had slipped from May into late June, with Diamandis saying even July would still be consequential.

2. Capability swim lanes matter more than the AGI label

Musk predicts digital superintelligence “this year” or, failing that, next year “for sure,” defining it as AI better than every human at every intellectual task. Diamandis contrasts that with Eric Schmidt’s roughly five-year horizon.
The definitions remain slippery: ChatGPT and Grok both describe AGI as human-level general intellectual capability and ASI as surpassing humans across intellectual work. Diamandis says this leaves unresolved questions about what intelligence means, including emotional and spiritual intelligence.
His stronger test is whether AI can make a Kepler-like intuitive leap—connecting the Moon to tides before formal proof—rather than merely execute a prescriptive task. Once a task is prescriptively describable, he argues, machines will predictably become better at it.
Blundin recommends tracking “vectors or swim lanes”: math and coding can advance quickly because they are not data-constrained, while biology may wait for a full-cell simulator. He relays Alexander Wissner-Gross’s prediction that math could be “solved” within 12 months, versus broader scientific progress over two to five years.

3. The first ASI could become the last, but its owner may not control it

Jeff Clune’s scenario is stark: an organization creates an aligned ASI, “effectively invented a god,” and uses it to prevent every rival from doing the same. A head of state would then face intense pressure to nationalize or command it.
Blundin accepts the winner-take-all mechanism and argues that multiple models and viewpoints would require regulation; competition alone will not preserve them once recursive improvement begins. “The observation in that video is right on target.”
Alexander Wissner-Gross agrees governments will intervene but doubts they can retain philosophical control: an ASI ordered to enforce Kazakhstan’s worldview might dismiss it within seconds as too limited. Blundin’s institutional view is that America is more likely to keep private AI contractors, as it does missile and guidance suppliers, than designate one national company.

4. Meta is pricing frontier talent like a strategic weapon

Diamandis’s theory for SSI’s $32 billion valuation after raising $6 billion: Ilya Sutskever could tell investors he knows how to build the first safe ASI, which would also become the last. A fund believing that premise has little choice but to invest at the offered price.
Wissner-Gross stresses that decisive research teams may contain only 10–15 people, with further 10× improvements still available. Sutskever, Mira Murati and other neural-architecture leaders are evidently “not intimidated” by OpenAI, Google or Grok, while OpenAI is pursuing consumer distribution through voice and coding tools because a foundation model alone may not be defensible.
Meta’s balance sheet makes the recruiting numbers intelligible: $1.8 trillion of market value and $70 billion in cash against the risk of falling behind in AI. Blundin doubts that a reported $100 million signing payment truly carries no vesting or retention, but says the correct researcher could be worth “many billions, if not a trillion.”
The uncomfortable operating signal is that the group heard during a San Francisco tour that “Llama 4 really does suck,” amid an exodus of Meta talent. Yet a few algorithmic changes and the right experimenters could restore parity quickly; Zuckerberg’s $19 billion WhatsApp bet is offered as precedent for aggressive, initially ridiculed capital allocation.

5. Scale AI’s 49% transaction previews an acquisition workaround

After Meta reportedly failed to buy SSI, Diamandis says Zuckerberg pursued Daniel Gross and Nat Friedman, then committed $14.8 billion for a 49% non-voting interest in Scale AI. Alexander Wang moved into Meta, signaling a changing technical guard alongside Yann LeCun.
Wissner-Gross values Scale’s data-labeling position because better inputs can make less powerful models perform well. Diamandis says Meta is buying engineers who know how to synchronize algorithms across a million concurrent GPUs and diagnose low-level architectural failures, not AI philosophers.
Blundin calls 49% non-voting ownership “the deal structure of the future”: in his account, it avoids a lengthy Hart-Scott-Rodino review because it is non-controlling and sidesteps the roughly 19.9%–20% financial-consolidation threshold because it lacks votes.
His important hedge is that the undisclosed contract probably transfers extensive IP and operating control, while proceeds are being distributed to Scale shareholders. Economically, he therefore treats it as an acquisition even if the bright-line legal structure says otherwise.

6. Capital is flooding both trillion-dollar infrastructure and tiny AI teams

Masayoshi Son’s proposed $1 trillion US technology complex would recreate Shenzhen-scale manufacturing with possible roles for TSMC, Samsung, OpenAI and Arm. The open questions are where the money, engineers and operating talent come from.
Wissner-Gross’s counterpoint: Boston’s computer-science pool may be 20 times Silicon Valley’s and less picked over, yet OpenAI reportedly judged the AGI timeline too short to justify opening and populating a Kendall Square office. The binding constraints may be electricity and chips, not another city full of people.
At the opposite end, Cursor reached $500 million ARR in under three years; the discussion later cited two. With a valuation greater than roughly $10 billion, potentially enormous margins and little headcount, Blundin says it could become profitable “on one day’s notice.”
Andreessen Horowitz’s top-quartile AI startups reached $8.7 million in run-rate revenue after raising $3.1 million pre-Series A, often within five months. Demand is also becoming top-down: after Jamie Dimon told JPMorgan managers to engage AI vendors, a previously unresponsive team called Blundin’s portfolio company Farsight back immediately.

7. The durable startup moat is a bonded technical team

Blundin’s “Fred Wilson rule” is to fund three or more best friends who write the code themselves and are trustworthy—even with a terrible initial idea. “You can’t change friendships, you can’t change relationships, you can’t change yourself overnight, but you can change your idea overnight.”
That bond protects against both pivots and nine-figure recruiting raids. Diamandis’s test is an international flight in adjacent middle seats: the people with whom 8–12 hours leaves you energized are the people with whom startup intensity might remain survivable.
Diamandis’s retrospective on Singularity University’s Graduate Studies Program is candid: putting 100 independent “alpha males and females” together and demanding company formation lacked the pre-existing glue of Y Combinator teams. The alumni created lasting friendships, but often found their true collaborators only later.
Their best institutional example is Israel, where military service creates bonds before college and startup formation; Blundin cites a per-capita success rate five times higher. Wissner-Gross adds Yossi Vardi’s character-first method: verify that a founder is “a good guy” with integrity, give $50,000, and repeat across roughly 400 startups.

8. AI.gov will work only if private infrastructure handles the trust boundary

AI.gov, led by Tesla engineer Thomas Shedd, was targeting July 4. Diamandis’s agency map spans GSA procurement and fraud, DOT infrastructure and flight-delay prediction, DOE grid operations, FAA drone and weather management, and FDA review of drugs, devices and clinical protocols.
Wissner-Gross’s mechanism is straightforward: government work is unusually prescriptive and repetitive, making it highly automatable. His concrete example is a wind-turbine approval that reportedly fell from two or three years to 30 seconds after software mapped electrical lines, water mains and flight paths.
Blundin uses Palantir and AWS government clouds as the template. He says sensitive documents cannot simply be placed in public general-purpose systems because of compartmentalization and training-data concerns; agencies need private, compartmentalized modules, with difficult rules governing information exchange between departments.
The Army’s shortcut was to appoint Palantir CTO Shyam Sankar, Meta CTO Andrew Bosworth, OpenAI product chief Kevin Weil and former OpenAI executive Bob McGrew as lieutenant colonels without conventional boot camp. The hosts reject the “rich big-tech mavens” backlash: the point is accessing expertise the normal promotion system cannot manufacture.

9. Bounded scientific datasets are becoming cheap prediction engines

DeepMind’s cyclone system delivered five-day tracks averaging 140 kilometers closer to the actual path after training on 5,000 cyclones across 45 years. Diamandis places that against roughly $1.4 trillion of cyclone losses over 50 years.
Blundin invokes AI’s “bitter lesson”: enough data and compute routinely outperform years of hand-built differential-equation work. A small team may now surpass five decades of domain research, creating a benefit to humanity he calls “immeasurable.”
Wissner-Gross labels the result “nothing to see here” as praise, not dismissal: bounded historical data should yield thousands of similar breakthroughs over the next one to three years. Earthquake prediction is the next candidate because, as Diamandis and Blundin discuss, animals appear to detect advance signals and the relevant data may exist.
Diamandis predicts earthquake warning within two years, if not sooner, then tests the governance edge with hurricane steering: would Miami pay Costa Rica $20 billion to absorb a storm expected to cause $50 billion of damage? Blundin immediately spots the inverse—a protection racket threatening to redirect storms unless countries pay.

10. AI tutors can widen learning while hollowing out recall

Mattel and OpenAI plan to make Barbie, Hot Wheels, American Girl and Thomas & Friends products conversational, potentially with GPT-5. Diamandis sees personalized early education and feedback loops revealing a child’s motivations, learning behavior and need for guardrails.
Blundin’s hesitation is that children may soon prefer endlessly engaging AI companions to human friends. Smart toys could become an “educational gold mine,” especially where schools fail, but could also replace imaginative play with an echo chamber.
Dave Blundin summarizes the cited MIT study: people who compose a document themselves can recount it accurately because writing trains their own neural network. The Google-assisted group’s quoted failure rate was 11%; the ChatGPT group’s was 75% because AI filled in the reasoning and memory formation.
Wissner-Gross is “80% unnerved, 20% will navigate this,” after watching his 13-year-old produce an AI-assisted essay he could not cognitively frame afterward. Diamandis’s pushback matters: AI users can cover far more terrain much faster, so education must teach critical thinking without pretending that speed has no value.

11. Vibe coding multiplies output but raises the premium on judgment

Blundin’s first handwriting-recognition product required four years: reading the backpropagation paper, translating differential equations into code, moving into assembly language and inventing quantization to extract every unit of processor performance.
He recreated the product and graphical demo by vibe coding in under an hour, while conceding that modern open-source components make the comparison imperfect. “Four years down to an hour” is still his clearest measure of the productivity discontinuity.
Diamandis’s son Jet responded, “I want to learn how to code, not just vibe code.” Blundin agrees: writing and reading raw code makes the acceleration legible, while original creativity and judgment still matter. This is a “golden era where you’re empowered, but not demoralized or crushed.”
In a survey of 1,500 workers and AI experts, 69.4% wanted AI to free them for higher-value work and 46.6% wanted repetitive work removed. Bookkeeping, payroll, data entry, claims, tax preparation and some software roles lead the automation list—but even bricklaying is vulnerable once drones can follow a prescriptive layout.

12. Physical automation is moving from warehouses to the doorstep

Amazon is testing Agility’s Digit humanoid for package delivery, pairing autonomous vans with robots for the final 10 meters. Wissner-Gross predicts mid-next year because “the technology is all there,” while preserving regulation as the likely source of delay.
The full-stack contest could pit Amazon plus Agility against Tesla vehicles plus Optimus. Diamandis expects autonomous vans and drones to coexist, reducing delivery cost through different combinations of vehicles and robots.
Blundin points to less photogenic deployments already advancing: robots performing nano- and microsurgery, crawling through pipes and cleaning sewer systems. Humanoids inspire attention, but specialized machines may deliver equally large practical benefits.

13. Tesla’s robotaxi model trades richer sensors for scalable capital

Tesla launched Austin robotaxis around June 22 at a flat $4.20. Diamandis treats autonomous mobility and humanoids—not car manufacturing—as Tesla’s eventual identity, citing Cathie Wood’s multi-trillion-dollar market forecast.
Blundin argues Tesla’s cameras produce a technologically inferior system to Waymo’s richer perception, particularly in heavy rain; lidar can see several vehicles ahead. Diamandis restates Musk’s first-principles defense: if one-eyed humans can drive, multiple cameras should suffice—though Blundin asks, “Why not give it superhuman skills?”
The economic advantage is externalized capex: consumers buy Model Ys, use them personally, then release them into the network while away. Waymo must finance a rollout of vehicles described as costing $200,000 with 29 cameras, five lidars and six radars; Blundin’s tentative market forecast is 70% Tesla, 30% Waymo.

14. Public backlash may become autonomy’s binding constraint

Protesters cite hundreds of Tesla FSD crashes and dozens of fatalities, but the panel insists those numbers need mileage denominators. With roughly 1.2 million road deaths worldwide each year, Blundin warns that individually tragic videos can overwhelm statistically superior system-level safety.
Blundin recalls a three-day BlackBerry outage in 2011 during which Abu Dhabi’s accident rate supposedly fell 40%, illustrating the cost of distracted drivers. He calls America’s susceptibility to sympathetic but potentially “statistical rounding errors” an Achilles heel against faster-moving states.
Five Waymos were torched after vehicles were reportedly summoned to the protest site, evoking the 1811–1816 Luddite revolt, 12,000 deployed troops and machine-breaking becoming a capital crime. When two teenagers were shot inside a Waymo, however, Wissner-Gross resists the technology-backlash narrative and suggests ordinary gang violence may be more likely.

15. Policy is reopening aviation before the hardware fully scales

Trump’s executive order supports beyond-visual-line-of-sight drones, five regional eVTOL pilots and renewed supersonic travel through FAA waivers from old overflight restrictions. Diamandis highlights Archer’s planned Los Angeles service for the 2028 Olympics.
Wisk was selected for Miami, while Joby and Archer are discussed as possible airport-to-city corridor operators such as JFK to Manhattan. Blundin’s answer to Eric Schmidt’s relative skepticism is that autonomous operation and improved safety make eVTOL more consequential than “electronic versions” of helicopters.
Diamandis’s larger call concerns land values: self-flying vehicles make remote waterfronts and inaccessible mountains viable, turning reachable real estate from scarcity toward abundance. Tourism and airport commutes become early economic wedges.

16. Nuclear delay exposes the US inability to compound over decades

Diamandis says digital superintelligence will be power-limited. America has 94 reactors against China’s 58, but added only two this century; licensing alone takes 10–12 years, China aims to surpass the US by 2030 and is building a reactor about every 52 months.
His technology framing separates dangerous generation-two plants associated with Fukushima and Three Mile Island from generation-four designs he considers fail-safe enough for his backyard. Three Mile Island’s recommissioning matters because its regulatory approvals already exist.
SoftBank’s proposed $1 trillion Shenzhen-style US complex reveals the same timing problem. Blundin argues that a vast human workforce and new buildings may arrive too late if AI progress follows Musk’s timeline; electricity and chips, rather than office space, are the durable constraints.

17. China’s solar buildout turns storage into the trillion-dollar prize

The panel says China created about 700 GW of solar panels and deployed roughly 250 GW of peak capacity in 2024, against total US capacity around 1.2 TW. By 2030, it could reportedly build an entire US worth of solar-and-storage power-generation capacity each year.
Blundin says solar crossed below new fossil generation in 2016, then in 2019 became cheaper to build and operate than merely operating an existing fossil plant. Wissner-Gross’s scale visual is that covering roughly 150,000 square kilometers—about South Dakota—with 20%-efficient panels could produce annual output exceeding present demand.
Wissner-Gross’s constraint is utilization: AI chips cost about 10 times their electricity and depreciate rapidly, so data centers must run 24/7. Lithium storage for cloudy periods can cost five times the panels; pumped hydro works but might require Loch Ness-scale water lifted roughly 300 meters.
Bill Gross’s gravitational towers lift dirt or other heavy masses during the day and generate power while lowering them later. Wissner-Gross’s bigger invitation is a reversible chemical store 10–20 times denser than lithium: “If you want to become the world’s first trillionaire, find a way to store huge amounts of energy cheaper.”

18. Stablecoins may become the settlement rail for AI agents

Bitcoin had dipped below $100,000 and recovered to about $107,000; Diamandis reports forecasts of $200,000 by year-end without making them unconditional. Wissner-Gross cites Michael Saylor’s $21 million destination and says he would be satisfied at $1 million, roughly his comparison to gold’s value.
Circle is Blundin’s more important story. Its IPO priced at $31 and peaked around $300, rewarding Jeremy Allaire’s persistence through years of regulatory pressure.
Blundin’s mechanism: SWIFT works for transferring $1 million but not a one-cent transaction between agents. Stablecoins make usage-based AI billing feasible, replacing flat subscriptions that can throttle heavy users: “The whole AI economy needs to move to these micropayments.”
His envisioned portfolio separates functions—Bitcoin for wealth storage, dollar stablecoins for transactions, and eventually trusted tokens backed by real estate or gold. Blundin mentions a confidential government arrangement to tokenize underground gold without mining it yet, while Wissner-Gross stresses that Circle’s competitive asset is “rock solid, stable, trustworthy” execution.

Softwar: An Intimate Portrait of Larry Ellison and Oracle $ORCL (Fintwit Book Club July 2025)How This VC Went From Broke to Becoming the Hot Hand in Silicon Valley