The OpenAI Internet Browser Has Arrived: ChatGPT Atlas w/ Dave Blundin & Alexander Wissner-Gross
Peter Diamandis hosts Dave Blundin and Alex Wissner-Gross (Salim Ismail traveling) for a news-dense episode that opens with X-Prize Visionering 2025, where three new prize concepts (a $250/month 'abundance' basic-services prize, a fusion prize, and a landfill-sorting robot prize) all got funded in one night. The main event is OpenAI's ChatGPT Atlas browser, which the panel frames not as a product but as a distribution channel for super intelligence in an escalating data/platform war with Google; related threads cover computer-use agents winning web games, DeepSeek's image-token OCR breakthrough, and Google's Genie 3 interactive world models. A long compute-and-energy segment tracks Meta's $27B debt-funded data center, Oracle's 16-zettaflop cluster, Anthropic's 1-million-TPU deal, Tesla's unified AI5 chip, StarCloud's space data centers and the long-run Dyson-swarm question, Google's Willow quantum chip, and a Trump administration push to take equity in US quantum firms. The back half covers America's nuclear cost crisis, a new zero-federal-funding DOE fusion roadmap, Amazon's SMR bet and delivery-glasses data-collection play, and closes with a lighthearted detour into alternative-respiration research that pivots into Ray Kurzweil's 2033 nanobot and longevity-escape-velocity prediction.
ChatGPT Atlas browser launch & the OpenAI-Google browser warAI▶ 14:16
OpenAI launches the Atlas browser with chat-everywhere, browser memory, and local agent mode; the panel debates whether this is a product or a distribution channel for super intelligence, and whether Google can simply copy it.
X-Prize Visionering 2025: abundance, fusion, and trash-sorting robot prizesOther▶ 1:39
Peter recaps X-Prize's Visionering event, where $3.5M was raised in one night and all three finalist concepts (a $250/month basic-services prize, a fusion prize, and a landfill-sorting robot prize) got funded into development.
Computer-use agents (CUAs) winning web gamesAI▶ 21:22
Alex demos Atlas's agent mode winning a single-player web chess game (asking the site itself for hints), calling it the best CUA performance he's seen and a preview of agentic AI for everything, not just games.
Anthropic's life-sciences team wants Claude to be a superhuman research assistant fluent in lab tools; Dave highlights Lila Sciences' 24/7 robotic 'dark lab' science factories as a more advanced version of the same idea.
Wikipedia traffic decline & the shift to AI-generated content/GEOEconomy▶ 29:00
Wikipedia traffic is down 8% year-on-year; the panel argues traffic isn't disappearing but shifting to AI-generated, ad-funded, GEO-optimized content as people increasingly just ask an AI instead of visiting source sites.
DeepSeek OCR and the path to universal multimodal tokensAI▶ 39:14
DeepSeek OCR encodes whole page images (not text tokens) and decodes them back to text, pointing toward models that perceive visual formatting the way humans do and a future of 'universal tokens' spanning modalities.
OpenAI automates Wall Street junior banker workEconomy▶ 45:01
OpenAI hired 100+ bankers at $150/hour to train AI on M&A, LBOs, and IPOs; the panel expects this vertical-by-vertical automation of white-collar work everywhere, but sees it as creating entrepreneurial opportunity, not a single winner-take-all platform.
Google Genie 3: interactive AI-generated worldsAI▶ 48:23
Google previews Genie 3's text-to-world generation; the panel expects world models to merge with foundation models, enabling immersive education (walking into ancient Greece), personalized gaming, and enterprise invention/design use cases.
Meta borrows $27B at 6.8% for a Louisiana data center, Oracle plans a 16-zettaflop/800,000-GPU supercomputer, and Anthropic will scale to 1 million Google TPUs by 2026, as credit markets join equity markets in financing the AI infrastructure buildout.
StarCloud's space data centers and the Dyson-swarm questionSpace▶ 1:00:40
StarCloud plans space-based data centers powered by solar and cooled by radiating heat into the vacuum; Alex frames this as the opening move in a possible multi-century trajectory toward a Dyson swarm around the sun.
Tesla's unified AI5 chip and Amazon's delivery glasses as embodied-AI training dataRobotics▶ 1:07:29
Elon announces the AI5 chip as a single architecture for both data centers and robots/cars; Amazon's new delivery-driver smart glasses are read as a parallel play to capture physical-world telemetry to train delivery robots and future consumer wearables.
Google's Willow quantum chip and Washington's equity push into quantum firmsCompute▶ 1:14:08
Alex explains Google's Willow quantum-chaos result as a real but not yet economically transformative quantum speedup, while the Trump administration moves to take equity stakes in IonQ, Rigetti, D-Wave, Quantum Computing Inc., and Atom Computing.
Nuclear and fusion energy buildout for AI data centersEnergy▶ 1:25:05
US nuclear construction costs have risen ~1,000% since the 1970s versus falling costs in China; the panel covers DOE opening weapons-grade plutonium to private reactors, a new zero-federal-funding fusion roadmap, and Amazon's SMR bet with X-Energy, all racing a 2030 power deadline for AI.
Predictions made
openPeter Diamandis: OpenAI's banker-automation effort could eliminate between a quarter and half of the junior headcount across Wall Street.
“this effort by OpenAI could eliminate between a quarter and half of the junior headcount across Wall Street within two years”
Your call:
openAlex Wissner-Gross: Humanity has effectively begun the long-term construction of a Dyson swarm around the sun, unless something changes.
· due: unspecified (decades to centuries) · ▶ watch
“mark this point in time, we're at the beginning unless something changes of the construction of a Dyson swarm”
Your call:
openAlex Wissner-Gross: Fully capturing the sun's light to unbind/disassemble Jupiter would take approximately two centuries, implying a roughly 200-year timeline for solar-system-scale compute buildout.
“what I'm hoping for is that sometime in the next few years, we will achieve a definitive breakthrough speed up for quantum acceleration of AI”
Your call:
openDave Blundin: The US needs to add 100 gigawatts of new power, about a 10% expansion of the total US power supply, just to run already-committed GPU production.
“I have to imagine by 2030 we're going to have figured out more energy-efficient compute um 10x or 100x more efficient”
Your call:
openPeter Diamandis: Ray Kurzweil's prediction that nanobots capable of high-bandwidth brain-computer interface and cellular/subcellular repair (enabling longevity escape velocity) will arrive.
“his prediction is nanobots by uh the early mid 2030s, so 2033”
Your call:
openAlex Wissner-Gross: The DOE fusion roadmap's private-sector-driven timeline will deliver early fusion pilot plants and then production fusion power plants.
“the third quote unquote long-term calls for actual operation at production of generation power plants between 2030 and 2035”
Your call:
Numbers that matter
Abundance X-Prize goal: food, water, housing, electricity, and bandwidth for a community for $250/monthWinning X-Prize Visionering 2025 concept aimed at easing turbulence from AI-driven job disruption.
X-Prize Visionering raised $3.5M in capital in one night; every $1 invested in a prize historically drives $60 of R&D (a 60x return)Funding mechanics behind X-Prize's grand-challenge model.
Fusion X-Prize: $500,000 committed to develop the prize concept, against roughly $10B already invested across 37 venture-backed fusion companiesFusion-focused experts argued a prize was still needed to build public/government support.
Uber pays 50 cents to $1 per 2-3 minute microwork task, processed within 24 hoursDrivers train AI during downtime, echoing Mercor's data-labeling business model.
OpenAI hired over 100 bankers at $150/hour to train AI on M&A, LBOs, and IPOsEarly step toward automating white-collar finance work.
Meta's SPV is borrowing $27 billion at 6.8% interest to fund a multi-gigawatt Louisiana data centerMeta is redirecting all cash flow plus new debt into AI infrastructure.
Oracle's planned AI supercomputer: 16 zettaflops, scaling to 800,000 GPUsPart of the Stargate Abilene buildout Oracle is financing and operating.
Anthropic plans to scale to 1 million TPUs on Google Cloud by 2026Anthropic's compute deal with Google deepens as it diversifies away from Nvidia GPUs.
Earth receives about 8,000x more solar energy per day than humanity consumes, yet that's still under 1% of the sun's total outputFraming for the feasibility of space-based solar-powered data centers and a future Dyson swarm.
US nuclear reactor construction costs have risen roughly 1,000% since the 1970s while China's costs have steadily declinedAttributed to lost manufacturing experience curve and regulatory/litigation overhead.
DOE will let private nuclear firms use 19 tons of weapons-grade plutonium from old warheads as reactor fuelAimed at reducing reliance on Russian uranium, though fuel cost is a minor factor overall.
DOE's new fusion roadmap has $0 in federal funding but $9 billion of private investment behind itRoadmap targets pilot plants by 2028-2030 and production power plants by 2030-2035.
D-Wave delivered an approximately 8,000x return from its lowest point to today after Dave took it public via SPACCited amid news of Trump administration equity interest in US quantum firms.
Intel's stock roughly doubled (about $20 to $40/share) after a US government equity deal; quantum computing stocks rose 10-15% on reports of a similar planIllustrates how government equity stakes trigger follow-on private investment.
Worth digging into
🕳️ Lila Sciences' "dark labs"
24/7 robotic science factories that design and run their own experiments overnight could be how biology fully automates; founded by George Church and Jeffrey Vanmolson.
🕳️ DeepSeek OCR's image-token architecture
Encoding whole pages as image tokens instead of text tokens suggests a path to 'universal tokens' spanning text, image, audio, and video, which could change how all frontier models ground multimodal understanding.
🕳️ Oracle's 16-zettaflop Stargate Abilene cluster
Alex calls million-GPU superclusters a new computing form factor comparable to the PC or smartphone; whether Oracle actually delivers at this scale is a major test of the current AI-infrastructure buildout.
🕳️ DOE's new fusion roadmap
Zero federal funding but $9B of private investment, with pilot plants targeted for 2028-2030, is a concrete test of whether fusion finally beats its decades-long '30-50 years out' reputation.
🕳️ StarCloud's single-H100-in-orbit mission
First real test of whether space-based data centers (solar power plus radiative cooling) are commercially viable, a possible first step in a multi-century Dyson-swarm trajectory.
🕳️ Ray Kurzweil's 2033 nanobot timeline
If accurate, this is the gating technology for both high-bandwidth BCI and longevity escape velocity, two of Peter's biggest recurring themes.