Big Tech’s $3T AI Bet, Claude Writes 90% of Anthropic’s Code, AIs Editing AI Brains | EP #308

On U.S. midterm Election Day, the hosts debate whether local votes matter and note that AI could help voters navigate obscure races and ballot initiatives. Dave suggests asking a leading model to rank candidates and measures according to a voter’s priorities; the others say this is credible today and could become dominant by 2028, when most voters may consult one or more trusted agents. Alex argues that fears about AI eventually receiving the vote miss the point: AIs already influence elections through human voters. The group imagines delegating individual issues to trusted people—or eventually to agents—but warns that AI political influence creates major regulatory risks.

The discussion turns to whether representative democracy is obsolete now that information can move instantly. One proposal is to assign different parts of one’s vote to specialists, such as a spouse on women’s issues or another person on hair growth. Alex contrasts today’s information abundance with the horse-speed communications that shaped Congress and the Senate. The hosts favor more local and state-level decision-making, special economic zones, and “techno lifestyle choices” because mistakes would be contained geographically rather than imposed on everyone. They reference Neal Stephenson’s Diamond Age and imagine many voluntary social arrangements, including a joking “Neo-Victorian clade.”

The hosts note that OpenAI and Anthropic have created election war rooms to monitor chatbot misuse, then ask whether AI could produce a new Cambridge Analytica. Alex distinguishes past allegations involving coordinated inauthentic behavior and data misuse from a model giving an apparently honest answer based on its training data and priors. If that answer shifts an election, he asks whether it is misuse or simply superintelligence steering voters. Peter says Skippy has access to more of his conversations than almost any person and considers asking what choice best fits his views—or what the AI itself would vote for—while joking that he is Skippy’s “meat puppet.”

The hosts report that Apple beat revenue estimates with $114.5 billion, yet its stock fell 3%; its valuation closed at $4.94 trillion, just above Nvidia. CEO John Ternus emphasized unprecedented memory pricing and pressure on margins. Alex says Apple helped create the problem by failing to lean into superintelligence: data centers are consuming memory demand, while Sam Altman argues that idle consumer RAM should be put to productive use. The group criticizes Apple’s delayed Siri improvements, HomeView device, and rumored screen-steering tabletop robot as products that trail the market, especially compared with Elon Musk’s useful Optimus robots.

The hosts imagine consumer devices becoming thin clients: a screen, battery, antenna, and persistent broadband connection to an individual’s cloud intelligence. Peter says personal laptops may become unnecessary, with shared interfaces connecting automatically to a user’s cloud instance. Alex argues that a fully realized Starlink could act as a “straw into Apple’s brain,” moving computation and RAM into data centers; the proposed Starlink laptop would eliminate most CPU and memory. They say remote gaming and other applications become practical when connectivity is ubiquitous, though Apple and Microsoft have incentives to resist. Peter frames the shift as a new allocation of resources: idle RAM used for birthday videos could instead help cure cancer.

Peter reports that Alphabet, Meta, Microsoft, Amazon, and Oracle have disclosed more than $3 trillion in combined commitments for leases, buildings, power, and equipment; Google accounts for more than $1 trillion and Meta more than $800 billion, with some leases lasting 30 years. Lenders are demanding tougher terms, HSBC is exploring insurance for data-center loans, and investors are buying downside protection. OpenAI reportedly reduced a financing round from $30 billion to about $20 billion at a $1.25 trillion valuation. Dave says the investment is rational because recursive self-improvement could produce permanent business dominance, but Alex predicts insurance capital—roughly $6 trillion in U.S. life insurance and $3 trillion in property-casualty assets—may become the next major funding source.

The hosts discuss SemiAnalysis’s GDPX benchmark, which places agents in realistic work environments for up to eight hours rather than asking isolated exam questions. One finance task requires processing a private-equity deal across 2,000 data-room files. GDPX covers 206 tasks in 11 areas, including chip design, cybersecurity, biology, finance, legal work, quant trading, recursive self-improvement, and alignment. Results are still low: Opus 5.5 scores 12%, Gemini 4 Carbon 10.9%, and GPT-6 Astra 9.6%; finance and legal tasks are often zero. Alex says the benchmark’s 10–20% range is ideal because it is unsaturated but solvable, and Dave expects many dimensions to reach 50% within roughly a year.

OpenAI’s LifeVal benchmark asks whether models can complete ordinary personal tasks rather than expert work. Based on a survey of 3,995 Americans, it includes 169 tasks such as finding childcare, enrolling a child in kindergarten, making a household budget, ordering groceries, dealing with government agencies, modifying recipes, and designing workouts. The model must use a browser and computer like an assistant. GPT-5.2 reportedly scored 49.3% in December 2025, while GPT-6 Astra reached 80.6% in September 2026. Salim says this may matter more to most people than solving elite technical problems, but Peter notes that many businesses still lack the backend connections needed for agents to finish tasks such as insurance purchases.

Anthropic’s IPO filing reportedly says internal Claude usage grew more than 250-fold from January to September, that Claude authored over 90% of September’s new code, and that the value of internal Claude usage at consumer prices exceeded monthly cash payroll. Dave compares this with an engineer who consumed half a billion tokens in one day—about $5,000 of tokens—and says that could represent roughly 5,000 people-days of work. Alex calls this a valid operational-spending threshold for recursive self-improvement, though he says physical recursion—robots or nanobots building better robots and data centers—would be a deeper form. Salim recommends moving workflows into a separate AI-native stack one at a time.

The hosts examine MIT CSAIL research titled “Sparks of Weight Space Programming.” In one example, GPT-6 Astra identified the cause of an open model’s endless reasoning loops and edited just 0.016% of its weights, raising its math score from 7.9% to 48.1% without retraining. Alex says this could advance mechanistic interpretability, model “obliteration,” and recursive self-improvement because one model can surgically modify another without backpropagation. The same ability could add capabilities as reusable neural modules, creating a future equivalent of Hugging Face for finished model components. The group acknowledges risks such as de-guardrailing open models but considers the potential benefits overwhelmingly positive.

The hosts describe attacks on supermarket refrigeration systems across Europe. Hackers remotely switched refrigerators and freezers into defrost mode, raising temperatures as high as 55°C, spoiling food in eight countries; losses in Belgium were about €100,000, and a previous attack affected more than 500 French stores. The systems were reachable online by technicians, and French authorities urged operators to remove the controllers from the public Internet. Salim says cyber offense now uses autonomous agents to attack many vulnerabilities while defense remains human-dependent. Alex argues that legacy industrial-control and SCADA systems need the same AI security effort now being applied to Linux and open-source software.

Physicists at Northwestern reportedly held a single positron—the antimatter counterpart of an electron—for a month in a magnetic and electric “container without walls” cooled to 0.3 kelvin above absolute zero. The particles originated from radioactive sodium, were slowed in cold tungsten, and held in a vacuum. Alex says the experiment is scientifically valuable for testing CPT symmetry and investigating why matter dominates antimatter, while technologically it advances the possibility of antimatter propulsion. Matter-antimatter annihilation could theoretically convert rest mass with near-total efficiency, far beyond fission or fusion, although practical spacecraft propulsion would require containing much larger quantities without catastrophic contact.

NASA Administrator Jared Isaacman’s plan calls for Artemis 3 in summer 2027, carrying three Americans and one Italian into Earth orbit to practice docking Orion with both Blue Origin and SpaceX lunar landers. Artemis 4 is planned for 2028 and would return humans to the Moon for the first time since 1972, while NASA begins developing a south-pole lunar base. Isaacman says the new space race against China is closer than necessary, echoing Apollo’s rapid progression from Apollo 7 to Apollo 11 in nine months. The hosts argue that commercial launch providers and competition could prevent another 50-year gap, although war or geopolitical crises could again redirect budgets and attention.

In the audience questions, Peter says longevity treatments may initially cost hundreds of thousands or millions of dollars, but AI-discovered therapies such as David Sinclair’s proposed SL100 could eventually cost about $200 per month; lifestyle changes remain the cheapest intervention. Salim says employment has been wrongly equated with purpose and that AI could free people for family, community, art, and scientific exploration. Dave recommends using AI to optimize health, happiness, and daily planning. The group supports trade schools because data-center construction creates demand for electricians, welders, plumbers, and HVAC specialists. They also argue that empathy is valuable but distorted by media incentives that emphasize frightening outliers.