Moonshots Recap: Big Tech's $3T AI Bet, Claude Writing Most of Anthropic's Code, and AI Brain Surgery

The latest full episode of the Moonshots podcast on Peter H. Diamandis's official YouTube channel, "Big Tech's $3T AI Bet, Claude Writes 90% of Anthropic's Code, AIs Editing AI Brains | EP #308", was published on November 4, 2026 and runs about one hour and 54 minutes. Host Peter Diamandis was joined by Dr. Alexander Wissner-Gross, Dave Blundin and Salim Ismail. The panel walks through Big Tech's disclosed AI infrastructure commitments, two new ways of scoring AI (GDPVal and OpenAI's LifeVal), Anthropic's claim that Claude now writes most of its code, an MIT paper on AI models editing other models' weights, a supermarket fridge cyberattack, antimatter containment, NASA's Artemis timeline, and a short AMA on longevity and purpose.

This is an independent summary written by GNOUP. It is not affiliated with, sponsored by or endorsed by Moonshots, Peter Diamandis or anyone who appears in the episode. The views, claims and predictions described below belong to the speakers, not to GNOUP. Nothing here is investment, financial, medical or legal advice, and nothing here is a recommendation to buy or sell anything.

Key takeaways

• Follow the money: about $3 trillion in AI buildout commitments. Peter says Alphabet, Meta, Microsoft, Amazon and Oracle have disclosed more than $3 trillion in combined leases, buildings, power and equipment, with Google alone over $1 trillion. He compares that to roughly ten Apollo programs in today's dollars. Dave argues the large tech firms believe the payoff is lasting business advantage and will not slow down easily, while Alex points to insurance capital as a possible next pool of funding.
• GDPVal: can AI do an expert's real job? SemiAnalysis's GDPVal drops AI agents into realistic work for up to eight hours across fields such as chip design, biology, finance and software. Peter says even the best models scored only about 12% (Opus 5.5 leading). Alex calls that range a sweet spot for a new benchmark and expects it to saturate quickly; Dave is similarly bullish that many dimensions will fill in within about a year or two.
• LifeVal: everyday errands. OpenAI's own LifeVal measures help with tasks like finding child care, budgeting and dealing with agencies. Peter says GPT-6 Astra scored about 80.6%, up from about 49% for an earlier model nine months earlier. Salim and Peter stress that competent personal assistants may matter more to most people than hard STEM scores; Alex notes the results are OpenAI scoring its own models.
• Claude writes most of Anthropic's code. Drawing on Anthropic's IPO filing as read on the show, Peter says internal Claude usage (priced at consumer rates) now exceeds monthly cash payroll, grew more than 250-fold from January to September, and that Claude authored more than 90% of September's new code, with daily code contributions per engineer about 45 times higher than two years ago. The panel frames this as deep recursive self-improvement already under way.
• "AI brain surgery" at MIT. A paper from MIT CSAIL, "Sparks of Weight Space Programming," reports that frontier models can read and edit neural-network weights. In one example described on the show, Astra edited a tiny fraction of a small open model's weights and raised a math score from about 7.9% to 48.1% without retraining. Alex sees this as a path for mechanistic interpretability and for bolting new capabilities onto models; the panel calls the net effect strongly positive while noting risks such as removing safety guardrails on open models.
• Science and space side stories. The panel covers a Northwestern record for holding a positron for a month, NASA's Artemis 3 and 4 timeline as described by administrator Jared Isaacman, and a brief AMA on making longevity cheaper and finding purpose if AI does more of the work.

Newsworthy insights

These are claims, opinions and predictions voiced on the show. GNOUP has not independently checked them unless noted.

Big Tech's $3 trillion AI bet (reports and opinion)

• Peter says lenders who once funded data-center buildouts with few questions are now demanding tougher terms, that OpenAI cut a new fundraising round to about $20 billion from $30 billion at a lower valuation than earlier talk, and that put options on CoreWeave hit a record as some former tech employees quit to hedge stock. He also cites Elon Musk warning against lending to companies that only make money by serving AI from a data center.
• Dave says the companies were "cash sucking machines" that have flipped into using every available dollar plus debt (except Apple) to build AI as fast as possible. Alex estimates roughly $6 trillion of US life-insurance capital and about $3 trillion of property-casualty capital as possible pools still on the sidelines; he is explicit that this is not investment advice.
• On chip lifetimes, Dave notes that even older GPUs can stay valuable as new uses appear, while new physics could disrupt that; he mentions signing a large multi-year lease just to get access to high-end systems. Salim says big infrastructure buildouts often come with bubbles even when the long-term direction is right.

GDPVal and LifeVal benchmarks (reports and opinion)

• Alex praises GDPVal for sitting in the hard-but-solvable band and for scoring frontier areas such as recursive self-improvement and alignment, not only today's service jobs. He says a cost-versus-performance chart suggests Gemini 4 Carbon may sit on the Pareto frontier if its pricing holds when the model is generally available.
• On LifeVal, Alex wonders whether OpenAI stopped optimizing as hard for everyday consumer tasks after an earlier GPT release, or whether the last 20% of errands is simply very hard. Dave argues that many AI capabilities are already ahead of the "connection points" to legacy systems such as insurers and customer-service backends.
• Salim and Peter encourage entrepreneurs to create their own benchmarks for causes they care about. Peter teases an upcoming "abundance" / singularity index he is building to track progress toward abundance.

Anthropic, RSI and company workflows (reports and opinion)

• Dave describes an engineer at his firm using about half a billion tokens in a day (on the order of thousands of dollars) and says the trend in cost per unit of work is falling roughly 10x year over year. Salim coaches larger companies to keep the existing business running while moving workflows one by one onto an "intelligent stack" that can improve itself.
• Alex argues that today's RSI will look mild once physical recursion (robots or nanobots building copies of themselves) arrives, and that the race will eventually stop when a "perfect" model shape is found rather than melting down forever.

MIT weight-space programming (report and opinion)

• Alex says the work comes from Omar's group at CSAIL, that he finds the results transformative despite some public pushback, and that there was no backpropagation in the edits described. Possible uses he lists include mechanistic interpretability, removing or adding capabilities on open-weight models, and a future "Hugging Face of modules" where specialists bolt expert components onto a base model.
• Asked about hidden behavior, Alex says other superintelligences should be able to inspect newly added circuits ("defensive co-scaling"). Dave and Alex both call the overall direction hugely positive while acknowledging risks.

Cybersecurity, antimatter and Artemis (reports and opinion)

• Peter describes hackers remotely raising temperatures in supermarket fridges across several European countries, spoiling food, and ties it to weak security on internet-connected industrial controls. Salim says offense is increasingly "humans out of the loop" while defense is still human-in-the-loop. Alex suggests training superintelligence on legacy SCADA systems as a large opportunity.
• On antimatter, Peter summarizes a Northwestern team holding a single positron for a month in a magnetic and electric trap near absolute zero. Alex puts it in the context of CPT-symmetry measurements and long-term antimatter propulsion research; Dave notes that matter–antimatter annihilation is in principle far more efficient than fusion or fission as a rest-mass battery.
• On Artemis, Peter relays Jared Isaacman's plan for Artemis 3 in summer 2027 (Earth-orbit practice with Blue Origin and SpaceX landers) and Artemis 4 in 2028 as the first crewed lunar landing since 1972, plus a lunar south-pole base. The panel contrasts that urgency with the 50-year gap after Apollo.

AMA notes (opinion)

• On longevity reaching poorer people, Peter stresses sleep, diet, exercise and mindset as low-cost foundations, describes expensive gene-therapy approaches, and mentions David Sinclair's reported work on a lower-cost small-molecule path, plus an XPRIZE healthspan competition aimed at affordability. These are claims as stated on the show; this summary is not medical advice.
• Salim argues society confused employment with purpose and that AI can free people for community, caregiving and exploration if basic needs are met. Dave says AI turned toward daily planning and health habits could unlock a lot of human self-improvement.

Watch the episode

Big Tech's $3T AI Bet, Claude Writes 90% of Anthropic's Code, AIs Editing AI Brains | EP #308 on YouTube (Peter H. Diamandis's official channel, published November 4, 2026)

Episode guide

Chapter times are taken from the episode's own YouTube description. Select a time to jump to that point on YouTube.

• 0:00: AI's Growing Influence on Elections
• 7:16: Apple's AI Struggles and the Future of Devices
• 13:49: Welcome to Moonshots
• 17:40: Big Tech's $3 Trillion AI Bet
• 29:30: GDPVal: Measuring AI's Ability to Do Expert Jobs
• 40:39: OpenAI's LifeVal Benchmark and Everyday AI
• 49:30: Claude Writes 90% of Anthropic's Code
• 53:15: Recursive Self-Improvement and AI-Powered Workflows
• 1:02:00: MIT's AI Brain Surgery Breakthrough
• 1:10:00: The Future of Self-Improving AI Models
• 1:14:02: Cyberattacks and AI Security Risks
• 1:21:00: Antimatter Breakthroughs and Space Propulsion
• 1:26:30: NASA's Artemis Missions and the Return to the Moon
• 1:34:58: AMA: Longevity, AI, and Human Purpose
• 1:45:27: The Future of Work, Robots, and Abundance

About this summary

This recap was prepared from the episode's title, description, chapter list and the publicly available transcript text. The transcript did not label who was speaking, so some attributions are our best reading. Automatic transcripts can also contain errors, so some names, numbers and phrases may differ from what was said. We did not check the figures and claims quoted by the speakers. We kept the opening election and Apple segments brief, left out heated political asides, and did not promote sponsors or paid offers in the description (Abundance Summit, VoiceRun, Blitzy, Fountain Life, metatrends, 6 Forks Report and others). GNOUP has no connection to those sponsors or offers. The show's own note says the views expressed are personal opinions and not financial, medical or legal advice, and the same goes for this summary. For the exact words, watch the episode.