Google's Record Quarter, the White House Intervenes, and GPT 5.5 Silently Matches Mythos | EP 254
In a Nutshell
Google's Q3 earnings smashed records with $109.9B revenue (22% YoY growth) and Google Cloud at $20B (63% growth), fueled by AI, amid compute shortages forcing edge models open and a White House push for pre-release AI model vetting by a Trump-era executive order working group. OpenAI's GPT-5.5 matches Anthropic's Mythos on cybersecurity benchmarks at 5x lower cost, while partnerships with PE giants like TPG ($10B) and Blackstone ($1.5B) target enterprise AI deployment; OpenAI delays IPO to 2027 after missing user/revenue targets. AI drives infinite compute demand (chip/energy stocks soaring, ocean/orbital data centers emerging), AGI progress (80-90% there per leaders), and economic shifts toward Universal Basic Compute/Equity over UBI, with Blitzy raising $200M for autonomous enterprise coding.
These notes were generated by AI and may contain inaccuracies.
Google crushed earnings: Alphabet reported $109.9 billion revenue, 22% year-on-year growth, $62.6 billion profit. Google Cloud hit $20 billion revenue with 63% growth. AI drove results across Google ecosystem.
White House considering process of vetting all models before release. Trump administration flipped from open position to proposed executive order creating working group with tech leaders and government officials to preview models.
The capabilities of the AI are going to grow exponentially, continues to grow exponentially. They're incredibly valuable to the military. So, the government ultimately has to preview these things, right?
It has to, but it can't gatekeep. That's where we're going to end up falling behind from a geopolitical perspective.
I'm more worried about the Frontier Labs self-policing more aggressively than the government ever would and stifling competition that way. I think that's a far scarier future.
Welcome to Moonshots podcast with Peter Diamandis, Alex Kingsbury (resident genius), AWG, Dave Blundon (Link Studios, Link Ventures supporter). Salem Ismael joining soon from global junket. Podcast helps understand fast-changing world and take advantage: news and technology impacting lives, families, companies, industry, nations. No politics, just news that matters. Welcome to WTF Just Happened in Technology.
Special guest: Brian Elliot, CEO of Blitzy (sponsor). Big news week for Blitzy.
Extraordinary MIT event for launch of We Are As Gods book by Steve and Peter. Everyone buying 100 books attended. Special guest Ray Kurzweil shared visions (podcast coming soon). Favorite moment: selfie of three generations of singularitarians.
Upcoming: Moonshots Gathering in Los Angeles, September 25th. Starts 9:00 a.m. (pre-registration 7:30 a.m.), goes late with party. Awarding Future Vision X-Prize (creators' trailers for future movies, top Hollywood stars selecting). Announcing largest Moonshots hackathon ever (awarded there). Google X sessions on creating exponential groups/moonshot organizations. Xprize design workshops. Kathy Wood on investing in exponential technologies. Massive party. Faculty/rock stars including Moonshot mates, Astro Teller (Google X), Rod Roddenberry (son of Star Trek creator), Ben Lam (moonshot entrepreneur), more names coming. Register via application at moonshots.com. Deposits from last year: special rate below $1,000. Other seats $1,495. Limited VIP seats with special lunch/evening.
Moonshots Gathering on moon: early/mid 2030s practical with Starship landings and democratization curve, though transportation costs high.
MIT event phenomenal, Ray Kurzweil hero. LA event 5x-7x bigger, super excited, 10x energy in Southern California. Vision X-Prize: AI video quality improving rapidly, 10,000 submissions compressing to entertaining hours, LA movie central.
Trump administration shift: from no restrictions to vetting via executive order working group.
Mythos moment sea change: civilian frontier AI labs leapfrogged government capabilities, solving cybersecurity/vulnerability discovery (GPT-5.5 generally available, stronger on public cybersecurity benchmarks than Claude Mythos). Government woke up to leapfrog risks in government/industrial/SCADA systems.
Pre-release vetting creates compliance mode big labs (OpenAI, Google, Anthropic) afford, smaller labs cannot, potentially limiting to oligopoly.
Existing moats: export controls, Invention Secrecy Act (decades old, gatekeeps patents in sensitive areas), Atomic Energy Act (1950s, gatekeeps applied physics). Regime since WWII, now AI strong enough for government step-in.
Inevitable government preview due to exponential AI growth, military value; once out, no repealing. But gatekeeping risks geopolitical lag (veto rights vs. partnership). Open source already 3 months behind closed source, creates parity. First model held up; history like GPT-2/GPT-3 held for safety (Dario Amodei at OpenAI).
Moral panic natural for radical capabilities; question: frontier labs or democratically elected government holding it?
Labs will hold back cutting-edge internally, benefit if government deems too powerful to release (derivative products, physics discoveries).
Normative view: more worried about frontier labs self-censoring aggressively (compute intensive, commercial benefit, stifling competition) than government.
Pentagon agreements with seven AI companies: Google, SpaceX, xAI, OpenAI, Amazon, Microsoft for military applications. Google's agreement: AI for any lawful government purpose.
Prompted protest by 600 Google employees. Recall 2018 Project Maven walkout: 20,000 employees over Google AI/search for government.
DeepMind (London) employees unionized to protest (first in frontier AI, 19th-century unions vs. 21st-century tech). Not great look for Google protesting patriotic US military work. Competition upside: DoD has options if unwilling to work with one (e.g., Anthropic), models on cybernet/JWIX.
DeepMind culturally separate (London, Demis Hassabis wanted spin-out per The Infinity Machine book; Google kept internal). Serious internal friction.
AI becoming decision layer, understandable employee backlash; navigating crazy.
Alphabet: $109.9 billion revenue, 22% YoY growth, $62.6 billion profit. Google Cloud: $20 billion revenue, 63% growth, outpacing AWS/Azure. AI drove ecosystem results. Google: 3/4 billion monthly active users.
YouTube greatest acquisition ever ($1.65 billion; Chad Hurley story: YouTube scaled faster than Google Video due to no lawyers/restrictions).
Search volume flat since 2017, revenue up via AI ad targeting. Easy road: AI instantly to revenue/profit (vs. Tesla/OpenAI). On cusp of most valuable company.
Google Cloud difficult birth: co-founders mandated Thomas Kurian #1/#2 or excise; leaked memos. Stood firm, carved out line item pre-AI tailwind. Exposed TPUs to customers/other labs, potentially for sale. Unique vertical integration to leapfrog AWS/Azure.
EverQuote (Alex chairman): Google offered ad credits for using Google Cloud; sticky, hard to leave. Luck + capitalization.
GCP matured dramatically (2018-2026 unrecognizable). Late start, pushed hard, great talent attraction (SF taxi ads with math problems). Unless #1/#2, drop (e.g., Circles social failed vs. Facebook due to approval layers).
Google tops EXO rankings: flywheel of data feeding algorithms, algorithms running
Google is compute constrained. Demis Hassabis states: "For us, there is a question of resources, talent and compute. Nobody has enough spare compute to just make two frontier models at maximum size with different attributes. That's pretty difficult. But for now what we've decided is that our edge models—the things we want to use for Android, glasses, and robotics—it's best that they're open models because they're vulnerable anyway once you put them out on the surfaces so they might as well be fully open."
World's largest infrastructure builder can't build fast enough and is turning away revenue. Devices on the edge will use open source for security. Google makes people apply and get in line for large amounts of compute; you have to be one of the most important players to be competitive.
Corporate America takes compute for granted. Conversation with Kush Bavaria from Orin; company grown like wildfire because compute is constrained—everyone reserving future compute, buying compute futures for the first time. Most of corporate America unaware this is the new normal forever. Blitzy consumes compute productively at almost infinite rate. No more surplus compute like going to the grocery store for milk.
That will probably never be true again. Corporate world not reserving or building capacity; will suffer 2-3 years from now when nothing available, realizing they could automate huge fractions of business, recode everything with Blitzy but no compute.
Terraab. Even within Google, main compute consumers—search, cloud, DeepMind—fight weekly or monthly for new capacity.
Preview of future: per-token economic productivity prized. Highest revenue or profit-generating tokens get most compute. Not just at frontier labs; Anthropic maximizes dollar value per token. Within Google, all fighting for most dollar value per token. Future: liquid market or auctions per token.
Need AWG metric on dollars per token created.
Innermost loop: stocks skyrocketing are chips and energy companies. Constrained resource driving global economy—where else to put capital. Google's market cap within 4% of overtaking Nvidia.
AI driving value, not search. First time buying box of chips (random RAM) a year ago would be way up today. Calling it: future forever; not temporary shortage. Even if Terrafab comes online on time (which it won't), all compute used instantaneously. AI has infinite appetite to create; every new GPU is another disease cured, person fed.
Intel fabs critical (was $19/share when highlighted); everybody up: AMD, Micron, SanDisk.
Demand near infinite; never seen before in technology. Jevons paradox goes insane. Provocative: future belongs to chip monopolies or intelligence utilities? Or neither. Vertical integration: xAI/SpaceX with launch for data centers part of innermost loop.
Determines economic futures; not slowing down. How high? Higher. Singularity goes in one direction: up. Asymptotes involved.
OpenAI ends Microsoft Azure-only exclusivity; now runs on AWS, Google Cloud, Oracle. Reminder: signed $100B AWS deal over 8 years, making Amazon major partner.
Implications for Microsoft: will they compete openly with OpenAI, spin up own models? Satya Nadella's infamous comment: Microsoft good for their $80B (backhanded, not supplying compute appetite). Fallout: OpenAI/Anthropic voracious appetites; Microsoft (year ago) fiscally responsible, limiting data center buildout and debt.
Causality from Microsoft's decisions to OpenAI starved of Microsoft-only compute, diversifying beyond original Stargate (alliance with Microsoft/Oracle/SoftBank/suppliers; no longer single-tenant, now branding for leasing from third parties). OpenAI nonprofit status part of story; Microsoft couldn't supply enough.
OpenAI dating everyone. OpenAI has problems including Elon Musk lawsuit. Mustafa Suleyman clear: Microsoft mandate to build own foundation model (has IP from OpenAI contractually), but struggling to read files (Microsoft insider).
Perfect marriage early: total domination. Might have taken for granted instead of partnering epic like Google/DeepMind. Corporate governance: OpenAI nonprofit needed for-profit subsidiary for Microsoft investment.
GPT-5.5 equivalent to Mythos; unbelievable model, underreacted. Better than Mythos on cybersecurity benchmarks: same capability 5x cheaper, generally available (early access). Anthropic compute constrained, not to market. Available on Amazon Bedrock (secure environment, prompts/results secret; first time ~1 month ago). Game changer.
OpenAI misses internal goal of 1B weekly ChatGPT users end of 2025; multiple revenue targets missed early 2026. CFO Sarah Frier: struggle to meet data center obligations if growth stagnates; wait until 2027 for IPO. Doesn't meet public company reporting standards (remarkable admission).
Two interpretations: no visibility to predict 2-3 quarters ahead, or other. As board member (MicroStrategy, Everquote) and adviser: missing targets consistently.
Interpretation: raised $120B, no rush to exit; great financial spot. Like Google: quiet, has cash flow, own chips—no need for drama. Can temper expectations, maybe 2027/2028 better.
Supply of capital: xAI/SpaceX soaks up; #2 picks rest, #3 left at altar. Anthropic #2? OpenAI to 2027—appetite still there? Numbers huge vs history (Yahoo, LSC, Excite IPOs in weeks; AltaVista missed window). IPOs happen back-to-back in category to educate investors; miss wave, harder later.
Bet on consumer for revenue terrible idea; consumers don't spend on reasoning tokens, enterprises do. OpenAI pivoted late (Anthropic earlier due to compute limits). Cost revenue targets/IPO. Codeex stronger than Claude Code now; leapfrogging continues.
Backing out original Stargate (leasing now, not build)—frees up. Sarah leaking smells expectation re-anchoring: lower expectations to beat shorter-term.
Strategic error/blunder: bet consumers eat every token, Google search obliterated, ad revenue gone (stocks down 70-80-90%). Wrong: tokens to highest value (enterprise). Google search (90% gross margin with YouTube) longer lifespan than predicted 2 years ago.
Blitzy eating tokens at ramp rate—not available for consumer until after Terraab.
Hard to run ARR at exponentially growing companies; billing challenges with every provider (even Google). Hard to predict 2-3 quarters in end-of-one tech moment; challenging for 10K.
OpenAI: $10B venture with TPG, Brookfield, Advent. Anthropic: $1.5B with Blackstone, Goldman Sachs, Helman to deploy Claude. Focused on deploying AI across enterprise operations/portfolio companies. PE controls trillions, thousands of companies—direct into main vein.
Predicted: AI enters via governance top-down, bypassing resistance (CIO/CEO). Breaks immune system; mandate it. Companies create digital twin at edge. Private equity main deployment channel: perfect AI lab (hundreds legacy companies, radical inefficiency). From chatbot to EBITDA transformation: organizational singularity via PE/operating partner.
For SMB CEOs: get on train fast—disrupt yourself with digital twin or get eaten. Takes longer/harder: no skill to wipe legacy, force cultural change non-trivial.
Private equity best performing asset class 30 years (beats venture except seed). Always tailwinds: computerization (pens/pencils to efficient, take public). Now AI x1000. Arbitrage amazing.
Legacy manufacturer: AI power tool to scour documents, interview employees, legacy systems. PE war chest unprecedented; returns staggeringly high like computerization wave.
Skeptics argue these funds—$10 billion to OpenAI, $1.5 billion to Anthropic—risk being used for circular sales, where frontier labs drive their own sales through co-branded ventures or JVs, similar to past Nvidia deals criticized as wash sales.
Another perspective: PE firms face eroding future discounted cash flows as AI obsoletes portfolio companies' predictable revenues within 2-3 years. OpenAI or Anthropic JVs offer billions to plug this hole, attractive to PE for LP optics, while labs approximate wash sales.
Seem agrees but counters it's brutally harder than expected: 44% of Gen Z workers deliberately corrupt AI automation to protect jobs—criminal malpractice as defense. Expect messiness in this uncharted transition.
PE firms pressure large companies to maximize AI, but solopreneurs, SMB owners, or board members must demand it internally—no excuses.
HBS alums in PE work across nearly every portfolio; not resistant, but fatigued by weekly board-mandated tools.
Focus not just cost reduction, but revenue acceleration possible now that wasn't a year ago. Managers fatigued by constant AI cost-cutting pitches.
Greg Brockman (OpenAI president): Everyone has own AGI intuitions; thinks we're 80% there.
Jack Clark (Anthropic): 60% chance of recursive self-improvement by end of 2028.
Richard Dawkins: Claude may already be conscious. > "If these machines aren't conscious, what more could it possibly take?"
Alex: Dawkins (selfish gene biological deconstructionist) implying Claude consciousness is extraordinary in biological philosophy of frontier models.
On Greg: Past OpenAI-Microsoft contract defined AGI as $100 billion revenue—perhaps his 80% means 80% to that.
On Jack: Anthropic's code mostly Claude-generated, including next-gen training/logic—how much more recursive? Estimate too conservative; outerbound.
See: Agrees on Greg; thinks 90% to recursive self-improvement, possibly within months (last 10% hard?). Dawkins is phenomenologist (bottom-up); AIs mimic consciousness, not actual. Bigger issue: operational autonomy—agents that plan, execute, negotiate, code, persuade.
Alex: Distinguish foundation model vs. AI system. LLMs (sequence-to-sequence transformers) won't get to AGI standalone; need AI systems with reinforcement loops.
Brian: AGI = systems learning on the fly outside training data (continuous learning, not just in-context). E.g., creating executable new programming language never seen before.
Current LLMs can't recreate Linux in net-new language (tried). Recent models built compiler chains (harder than Linux kernel); Anthropic's C compiler fails hello world despite training data. Blitzy robustly compiles all.
Debate: Demis Hassabis sees 50/50 LLMs get to AGI without breakthroughs. LLMs innovating to AGI ≠ LLMs are AGI.
Agency > consciousness/AGI: Consciousness = moral rights problem; agency = governance problem (comes first).
Meta's $2.5B acquisition of Manus (Dec 2025) unwound by China; founders barred from leaving despite completed employee/tech/investor payouts.
Meta lead secretly flew Manus team from mainland China to Singapore night before. Team in Singapore with code; payout taken.
China leverages Meta's China business to force unwind—political pressure.
US gov't should intervene to protect AI. AI talent = national security risk; spheres of influence (US vs. China).
Manus played multiple sides: Chinese, Singapore-based, Palo Alto presence.
Benchmark's last investment at $500M valuation; huge firm risk. Top VCs avoid China-based AI—money/IP may not exit; employees as national assets.
AI researchers won't move freely US-China. 50% Meta AI employees Chinese; many xAI Chinese left (security?). Less IP leak risk to China.
Alex: Short-term issue; recursive self-improvement means AI agents do research on US soil.
Talent vs. compute: If recursive soon/imminent, prioritize compute over talent.
Blitzy (sponsor): Large-scale autonomous software dev for Global 2000 (insurance, finance). Refactoring, modernization, product dev: 500K-1M lines code, end-to-end tested (vs. Claude Code/Codex 200-500 lines developer-driven).
Built compiler for Alex; handles Fortran 4, Watt 5—reverse-engineer legacy where no devs remain.
Uses Claude, OpenAI, Gemini; partners. Most inference-compute intensive code gen. Benefits from model ensembles (different intelligence flavors); checks models against each other algorithmically for dramatic quality gains. Can use open-source.
Not launching own model—wrong game. Orchestration layer for enterprise engineering velocity.
Advice: Algorithms last IP to go (novel algos, DB structures). Hire AI researchers, sales, forward-deployed engineers—all fronts.
Dave: Brian (West Point, Army Ranger, liberated Syrian city), Sid (India's top technical uni, Nvidia) met at HBS. India's massive young talent pool (twice China's) for tough tech work. Military experience: not afraid of people.
AI companies in Silicon Valley plan to stay headcount light with AI doing the work, needing only 5-10 people without recruiting, HR, or onboarding. Blitzy took the opposite approach: if highest token value generates token usage, need data and use cases from complex economy into AI. This requires forward deployed, brilliant people digging data from legacy databases and people's brains. Meror succeeds by not fearing people. Blitzy grew headcount from 10 to 80 in one year, now from 80 to 300 in nine months—record-setting onboarding scale, even beyond Amazon.
Brian, congratulations. You are right outside my door... hiring a bunch of forward deployed engineers to help our customers with AI adoption.
Alanteer used this model successfully.
Blitzy Sponsorship
This episode sponsored by Blitzy, autonomous software development with infinite code context. Uses thousands of specialized AI agents thinking for hours to understand enterprise-scale codebases with millions of lines of code. Engineers start sprints with requirements; Blitzy provides plan, generates and pre-compiles code for tasks. Delivers 80%+ of development autonomously, guides final 20% human work. Enterprises achieve 5x engineering velocity pairing with coding co-pilot for AI-native SDLC. Visit blitzy.com for demo.
Massive trip demand; data centers moving from land to ocean space and farmlands. AI chip boom lifts industry: Huawei sales up 60% despite tariffs; SanDisk revenues up 251% YoY; Samsung crossed $1 trillion; AMD up 260% past year; Intel up 442% past year, 114% in April. Chip stocks and energy stocks skyrocketing.
San Francisco is financial capital of the world; US tech market cap bigger than everything else combined. Nvidia alone could buy entire financial services sector.
To raise big money, go to San Francisco ecosystem. Semiconductors will go up; bottleneck in underlying manufacturing capability, hence Intel's success.
AWS CEO: AI demand so high, old GPUs can't be retired; completely sold out of A100 servers, never retired one.
Peter Thiel backing ocean-based AI data center. Panthala (Greek for oceans) raised $140M at $1B valuation. On open ocean: continuous energy from wave motion, cooling from saltwater, no land issues, plenty of real estate. Commercial deployment by 2027.
Alex: Burying the lead—ties to Patrick Friedman's Seasteading Institute (ocean colonization) supported 10-15 years ago. Killer app for oceans like data centers for space/solar system. Build seasteads around data centers; precedent Sealand (old British army station, self-appointed nation with Ryan Lackey).
Counterpoints: Seasteading great in principle but difficult; ocean data centers better than space, more efficient. Shocked not proposed before—elegance from energy in bobbing buoys powering GPUs (concentrated wave energy from wind, better than offshore turbines). Cooling, land availability, continuous energy (like space solar). Starlink key enabler for ocean data centers; LEO satellites easier than laying fiber (e.g., Hibernia network: $600M for low-latency fiber NA-Europe, expensive/tedious/risky). No jurisdiction issues dropping in ocean vs. regulatory hurdles for fiber. Easier maintenance: drag back on boat vs. reconnecting cables.
StarCloud in talks for $2.2B valuation after $111M round led by Benchmark/EQT one month prior; raising $200M. Building orbital data centers powered by solar. Launched first H100 into space in 2025; plan 88,000 satellites. Launch providers: SpaceX, Blue Origin, or Relativity Space (Eric Schmidt).
Cooling challenge solved with radiative aluminum (mass for cooling system critical, undisclosed). Hyperscalers see as acquisition target for Dyson Swarm. Anthropic announced partnership with SpaceX AI. Non-Elon hyperscalers may partner or acquire for 100 terawatts power. Relativity Space accelerating as SpaceX may refuse launches for exclusivity. Nvidia plans space-based GPUs; launch critical.
67% planned US data centers in rural areas (vs. 13% existing); 39% in counties with no existing centers. South leads (48%), then Midwest—biggest geographic wealth transfer since fracking.
Backlash expected despite vast farmland; immune response. Physical assets taxable; governments should beg for them. Scared populations vote against, governors veto for prosperity. Concern in middle America: electricity bills 2-3x. Mitigate with taxes offsetting, steady access, require data centers find own power (simple law). Buildout infinite, act soon. Risk: decoupling data centers from urban centers leads to Dyson swarm trajectory (rural → orbit → solar), bad for human-machine symbiosis—keep integrated.
David Sacks: AI becoming engine of GDP growth. AI capex 2% tailwind to GDP this year; Q1 AI 75% of growth. Polls: AI unpopular but growth is; stopping AI like halting US economy.
Morgan Stanley: Hyperscaler capex to $80B from $76B ($3B/day). Economy = AI infrastructure buildout. Future: transformative inventions on top (more exciting). Shoutout to startups: $1-2T compute investment creates massive markets (e.g., Standard Kernel, Chris Reinhardt compiler for chips catching Nvidia; data center efficiency/cooling).
Invest in future: assets (house, data center) rise with 10x GDP growth in 10 years (Elon); W2 income poor. Change benefits investing; discover early (e.g., Intel). Trends from pod trigger public equities/startups. Use LLM for research: chip companies, PE ratios, sentiment. Bucket of chip/energy/infrastructure companies moving up rapidly. Not investment advice.
Jobs: Hook up with growing AI companies hiring. Fast-growing AI startups deploying to enterprises have insatiable demand; hybrid soft/technical skills (self-learnable). Soft skills lacking now, hard skills AI-assisted—opportunity. Everyone at Blitzy owners with stock. Never worse time big tech (layoffs); best time AI startups. Apply blitzy.com.
Sam Altman no longer believes in UBI as before. Three-year study: spending up, no clear health/healthcare improvement. Proposes stake in AI upside: compute access, equity, or public wealth fund (e.g., Alaska permanent fund oil dividends to citizens; Saudi/Emirates). If AI/compute national resource, citizens own piece.
Need details; UBI data shows more U/B/I, more successful.
The more universal, basic, and income-like UBI programs have been successful. A Finland UBI trial failed because it wasn't universal, basic, or income. AI upside play is powerful: citizens get income or claim on AI productivity—need both. UBI protects the bottom; AI upside model provides exponential upside. Social contract shifts from redistribution to participation in upside.
Alternatives to UBI
UBI (Universal Basic Income): demand-side stimulus like stimulus checks.
- Doesn't align recipients with society long-term. UBE (Universal Basic Equity), UBC (Universal Basic Compute), UBS (Universal Basic Services).
- Prefer UBC or UBS over UBI.
- UBS: costs of everything (e.g., healthcare) to near zero via bounties for driving down costs—deflationary incentive vs. hyperinflationary checks.
Government could require compute owners to dividend 2% into citizen pool for allocation, sale, or use. Turbulence expected over next 2-8 years.
OpenAI provides UBC: hundreds of millions use GPT-5.5 instant for free (possibly ad-supported), giving small stake in compute. Can't yet turn into steak dinner, but GPT-7.5 instant will design robots to print steak. Robotics and AI will deliver UHI (Universal High Income).
Near-term: people struggling can't eat GPT-5.5. Job loss, housing crisis must be solved in 1-2 years. Government solution: stimulus checks or co-checks (~$3,000/month per individual). Owning part of America's compute infrastructure aligns citizens with xAI, OpenAI—success mutual, not adversarial.
Situation dynamic: good near-term solutions (e.g., checks) differ from 10-year (GPT-10.0 prints robot for dinner). Lift the bottom: free AI healthcare/diagnosis solves medical bankruptcy (biggest US bankruptcy cause), form of UBS.
Major insurers (Berkshire, Chubb) removing AI-related damages from standard policies; 80% exclusion requests approved. Exclusions: AI mistakes, IP violations, deepfake fraud. Companies need separate AI insurance—huge entrepreneurial opportunity.
AI risks ramp exponentially (understated). Legacy insurers dropped coverage; risks accumulate fast. Post-Mythos: cyber attacks rampant despite guardrails (35% mid-to-high net worth hit already). Coverage tied to defense: insurers require best practices/products to prevent AI cyber attacks. Insurance industry self-heals, funds solutions—massive opportunity.
AI insurance market: $40M in 2024, projected ~$5B by 2032.
Deplatforming AI agents (hard to get bank accounts/insurance). But insurance pressures align AI via capitalist forcing functions (actuarial checklists) vs. top-down government. Hard to price risks as big company, but coverage needed.
Dave's AI Platform Building Workflow
- Tools: Claude (left), Cursor (right, 50 agents: 41, 42 primary). Treat as roles, minimal context to avoid overload.
- Process:
- Plan-for-plan document via Claude 4 Opus Max, second opinion Gemini 3 → full plan mission format.
- Launch in Amazon EC2 (secure, cloud-resilient orchestrator, calls models via APIs).
- Defaults to Claude 4 Opus; wildcard Kimi K2.6 (9x cheaper, code injection risk, on Fireworks).
- Cost: full-blown GUI ~$8-10 compute.
- Post detailed guide on dbt.ai (AI-generated).
Probability of Rogue Predatory AI Corporations
100% (accelerando/accelerando). Good corporations via defensive co-scaling.
When Best Entrepreneur on Earth is AI
Already happening (top market caps AI-driven). Parameterize by ROI, not $2T cap (meaningless by early 2030s). Hybrid: human founder + AI swarm testing thousands in parallel—entrepreneur as orchestrator (today, e.g., with Henry).
AI Data Centers Heat Dissipation
- On land: cooling systems (invest in cooling companies).
- Ocean-based: easy.
- Space: radiative cooling to infrared vacuum (~2.7K cosmic microwave background; aim correctly for thermal gradient).
P(Doom) Percentages
P(doom): probability AI destroys human race (extinction, no Colossus revival).
- Alex: diminimus (0.1%?); redefines (e.g., uploading to Dyson swarm not doom).
- Peter: near 100% without AI (150K humans die/day now; AI saves). P(doom) negative.
- Dave: <0.
- AWG: 0 (low single-digit; vector=AI terrorism/biotech via broken guardrails/Chinese labs). Most humans good, high agency prevents harm.
COVID origins: Peter (natural zoonotic); Alex/AWG (lab leak consensus, possibly engineered components).
New Jobs Created by AI
Jobs = bundles of tasks; bundles shift to where humans provide relative ROI (e.g., physical tasks, field-deployed engineers). New jobs continue in perpetuity.
Abundance: Producing Too Much a Problem?
Universe 25 mouse utopia (1960s): excess abundance → fat/dumb/lazy downward spiral. Split: consumers (couch/Netflix/optimist beer) vs. creators (Star Trek-style upleveling). UHI via AI/robotics: so much abundance, can't desire enough.
Building Reliable Agentic Systems
Old enterprise: stable systems/change control—gone. Build via architecture:
- Modular agents, narrow permissions (passport metadata).
- Observable workflows, audit logs, human escalation.
- AI-native governance stable amid monthly model changes.
Diminishing Returns for Superintelligent Agents
Physical limits (Seth Lloyd, MIT early 2000s):
- Fastest serial computer: black hole (inputs via X-ray/gamma lasers, output Hawking radiation; e.g., future Mac Pro).
- Fastest parallel: plasma box. Ultimate constraints unless new physics.
Algorithmic: AIXI (theoretical optimal AI/agent; Bayesian superintelligence, info-theoretically perfect but impractical)—diminishing returns point.
Discussion on agent algorithm level: there may be a lot of room at the ceiling at the top, but the universe imposes limits.
Brian, congratulations on your financing. Thank you for joining us. Thank you for your sponsorship of this pod. Dave, Alex, Salem, outro music by Marius.
Send us your favorite video under two minutes, and if it's amazing, we will share it.
Outro music: Velocity by Marius.
We're almost there. Gravity escape velocity. Nothing left to spare. Seline calls it your company's shape with the purpose at the core. Massive transformative for the market is watching compute where the smartest lady. It's another move. The smart money plays intelligence is a force. Don't compromise every future. Open every option. Moon to write the sky. Moon faster.
See, you always come across as the sexiest guy in the videos.
It was an amazing week. Two recordings at MIT and today always a pleasure. Love you guys.
Have an awesome weekend. Yeah. Thank you. Be well and safe travel. Wherever you're going next on the world, where's Waldo?
If you made it to the end of this episode, I consider you a moonshot mate. Every week, my moonshot mates and I spend a lot of energy and time to really deliver you the news that matters.
If you're a subscriber, thank you. If you're not a subscriber yet, please consider subscribing so you get the news as it comes out.
Invite to join weekly newsletter called Metatrends. Research team spends the entire week looking at the Metatrends impacting your family, your company, your industry, your nation. Put this into a two-minute read every week.
To get access to the Metatrends newsletter every week, go to diamandis.com/metatrends.
Thank you again for joining us today. It's a blast for us to put this together every week.
Keep Peter H. Diamandis in your library
Save the videos and channels worth coming back to, and find them again in one place.
