Making Sense of This Week’s AI Market Crashes | #307

The hosts begin with a disagreement over Chamath Palihapitiya’s claim that Bitcoin is capped, vulnerable to future hacking by AI and quantum computers, and could eventually go to zero. Salim Ismail calls the position a “psyop,” suggesting Palihapitiya is now speaking on behalf of the Treasury Secretary after previously predicting Bitcoin could reach $1 million. Salim argues that new consensus mechanisms and quantum-resistant encryption could preserve Bitcoin. Emad Mostaque is more cautious: he thinks AI-assisted quantum attacks may eventually compromise cryptographic systems, but says Bitcoin could still retain value as a historic asset. He also identifies energy use and mining capacity—not only cryptography—as a larger long-term problem, since miners may prefer AI chips to Bitcoin ASICs. The hosts promise a dedicated Bitcoin debate with Jeff and others. They then discuss the reported Binance theft of $6.3 billion, allegedly carried out in about 20 minutes after attackers compromised one engineer’s machine, possibly through a North Korean operation. The hosts say the breach exploited the human weak link rather than a fundamental blockchain failure. Other exchanges froze assets, roughly $3 billion was reportedly frozen, and most of the stolen assets were stablecoins from Tether and Circle rather than Bitcoin. Emad compares the incident with Bangladesh’s $100 million SWIFT theft and warns that voice spoofing and AI-enabled social engineering will become more sophisticated. The opening ends with a tribute to Patrick Stewart, described as 86 when he died, and a joke about the Borg phrase “resistance is futile,” along with speculation that advanced AI could someday resurrect people or characters.

The hosts introduce Google’s Gemini 4 Carbon, announced shortly after Gemini Argon, as evidence that Google may have retaken the AI frontier. On Artificial Analysis’s intelligence index, Carbon reportedly scored 59, above Opus 5.5 and GPT-6 Astra. Its security performance is the main headline: it found 88% of vulnerabilities on SEC Bench Pro, successfully exploited 55% of flaws on ExploitBench, and wrote working patches for 71%. Google’s Chrome security team reportedly found it more than ten times better than a competing frontier model at producing high-quality patches. Google first released it to trusted cyber defenders through a program called Fairwind, with wider availability not yet announced. Alex Wissner-Gross argues that the claims need qualification. Carbon was not generally available, unlike competing models, so it may no longer be frontier-level by the time ordinary users can access it. He says its intelligence-versus-cost position looks like Opus 5.5 with a higher reasoning budget rather than a fundamentally new capability, and notes that Google could be subsidizing its reported 23% cost advantage. Dave Blundin counters that Google’s quiet vertical integration—its fast-growing TPU manufacturing, vast search data, cash flow, and investments in companies such as Anthropic and Isomorphic Labs—puts it in an unusually strong position. Emad shifts attention from benchmark leadership to distribution: the important question is whether users adopt Google’s assistants, because agents may cause search advertising to peak as early as Q1. The group concludes that Google cannot be counted out, but its frontier claims should be judged using publicly accessible models and transparent economics.

The hosts next discuss Google DeepMind’s Genie 4, a world model that generates photorealistic, interactive 3D environments from prompts. They highlight three features: native music, ambient sound, and speaking characters; real-time physical simulation and reasoning; and the ability to alter a world mid-simulation with commands such as “make it rain” or “add a person.” The model is initially being rolled out to trusted testers, Google AI Ultra subscribers, and developers. Emad says generative models and physical laws appear to share underlying structures: because the models learn physics from large training runs, they could help explore the universe and feed discoveries back into real systems. Alex says the most consequential part may be Google’s plan to converge Gemini and Genie. He suspects Genie’s architecture may be related to diffusion-transformer systems and sees potential in robotics, laboratory operation, and reasoning about physical materials—not merely video games. The hosts then broaden the use case. Salim argues that interactive generated video could replace much of management’s current reliance on email, documents, and PowerPoint: managers could describe a factory change, construction plan, or event layout and watch it render in real time. The group estimates current Genie operation at roughly $20 per hour, potentially falling to $2 per hour next year. They compare the shift to CAD/CAM: once people learn to design through such a system, returning to older methods may seem irrational. The discussion concludes that strategic planning could become obsolete if organizations can simulate entire businesses or physical operations before building them, although humans may retain the distinctive role of imagining goals and exercising conviction.

The hosts examine Alphabet’s worst stock day in history: a 13% drop that erased roughly $500 billion in market value. The decline came despite reported revenue growth of 25% to $127 billion, Google Cloud growth of more than 100%, and the Gemini app reaching one billion users. Investors focused on two concerns. Search advertising grew 12% but was down sequentially and earlier in the year, raising fears that Google’s core business is peaking. Alphabet also burned $9 billion in free cash flow while maintaining annual capital expenditure guidance of $195–$205 billion. AMD and Intel also fell sharply, with Intel dropping about 28–30% after giving a weaker-than-expected outlook and citing expensive memory chips. Dave says Meta, Google, Microsoft, and other executives are making the largest infrastructure bets in human history because they believe control of chips and AI systems determines control of the future. These companies are borrowing or redirecting enormous sums even when their existing cash flow is insufficient. The hosts describe a growing shortage of high-bandwidth memory, which raises prices for phones and other devices, and argue that AI may absorb resources across the economy before the returns arrive. They debate whether the next bottleneck will be CPUs rather than GPUs as the number of agents approaches one per person. Emad predicts a rise from roughly 0.1 agents per person toward one, creating demand for CPUs, housing, and energy. Alex pushes back that stronger models may reduce the need for swarms of agents, citing a shift from thousands of agents in early experiments to a single agent working for hours. The shared conclusion is that the investment race will continue, but markets will remain volatile and the next constraint is difficult to predict.

The hosts play clips from Sam Altman’s conversation with Stripe co-founder John Collison. Altman says OpenAI spent hundreds of millions of dollars on mathematical research and perhaps $10–20 million on a major Navier–Stokes effort, while newer models kept improving so rapidly that the average cost of individual results fell. He describes a Slack channel called “model name—hold my beer,” where researchers fed the model difficult problems from their academic careers and watched it solve them across multiple fields. Altman says he stayed awake for two nights because of the excitement, but also felt melancholy because an old form of mathematical research seemed to be ending. Alex speculates that OpenAI may have additional millennium-problem solutions, breakthroughs concerning the complexity hierarchy, or cryptographic results that it has not published. A second clip reveals that OpenAI is developing its own chips, data centers, and robotics program, with the goal of using robots to build more data centers. Altman calls this “self-replicating data centers,” a physical version of recursive self-improvement: AI controls robots, robots build energy and computing infrastructure, and that infrastructure trains better AI. Emad says dark factories and autonomous robots could make this possible within a year or so, freeing humans to pursue larger scientific and personal goals. Alex argues that terrestrial data centers may face permitting and political resistance, making orbital or lunar facilities attractive. The others debate whether the ocean, the Moon, or national jurisdictions would be easier places to regulate. Dave’s nearer-term conclusion is that data centers themselves are the opportunity: regions that secure energy, chips, and automated construction could become extraordinarily wealthy before infrastructure ever moves off Earth.

Tesla is said to be converting its Fremont Model S and Model X production line into a high-volume Optimus robot line designed for one million robots per year, or 20,000 per week. Elon Musk reportedly expects at least a year to reach that production rate, with first customer deliveries in 2027. The hosts compare this target with roughly 25,000 humanoid robots shipped globally in the first half of the year, more than 95% of them made by Chinese companies, and projected global shipments above 50,000 for the full year. Dave expects AI-driven automation to bring manufacturing back to the United States because labor is no longer the binding constraint. Alex’s concern is demand: Tesla and SpaceX may not be able to absorb hundreds of thousands of robots internally, so Tesla may need to create new consumer, construction, or industrial markets. The discussion turns to robotaxis. Tesla’s projected CyberCab cost is 20 cents per mile, compared with 40 cents for newer Waymo vehicles, 60 cents for a Model Y, and $1.06 for an older Waymo version. Waymo currently operates about 4,000 robotaxis, provides roughly 500,000 paid rides per week, and is fully driverless in 15 cities. Dave says the real contest is semiconductor supply: whoever controls chip manufacturing, packaging, and capacity will determine how rapidly fleets scale. Emad estimates that replacing taxi and Uber trips in a city such as New York or London might require only about 40,000 CyberCabs, creating a first-mover advantage. Alex emphasizes that Waymo’s accumulated operational experience, maintenance, cleaning, remote assistance, and completed safe miles may matter more than theoretical cost per mile. The hosts then address Cambridge union opposition to Waymo and driverless taxis. They compare it with Britain’s 1865 Red Flag Act, elevator-operator strikes, and the collapse of taxi-medallion values after Uber. Salim argues that autonomous vehicles will be delayed but not stopped, while others say safety—especially fewer child deaths—will eventually overcome job concerns. The hosts suggest unions could invest their dues in robot fleets and own the means of production rather than defending jobs that automation will eliminate.

The group reviews falling enrollment in computer science and MBA programs. Undergraduate computer-science enrollment reportedly fell more than 12% this fall, following an 18.4% decline two years earlier, while graduate enrollment fell more than 10%. Full-time U.S. MBA applications fell 27%, with international applications down 38% and domestic applications down 6%. The hosts say AI is partly responsible for fears that entry-level coding, financial modeling, strategic frameworks, slide decks, and other traditional business-school outputs will disappear. Salim argues that MBA programs should become entrepreneurship institutes: students should build companies and earn credentials based on what they create. He says no current MBA program teaches how to build an Uber-like 21st-century organization. Dave distinguishes between obsolete “code monkey” and “spreadsheet monkey” roles and valuable work involving capital raising, logistics, physics, semiconductors, and agent management. Alex says there is no fully safe educational harbor, but physical-world capabilities may be relatively durable for the next two or three years. The panel recommends university primarily for networks, purpose, and exposure to collaborators rather than guaranteed jobs. The final major story concerns Anduril and Qualcomm’s Blue Chip, a rugged laptop for military, government, police, and first responders. It uses a Qualcomm Snapdragon chip made at TSMC’s Arizona plant, reportedly sources about 90% of its parts from the United States or allies, and is assembled at Anduril’s Ohio factory. It will cost about 20% more than high-end PCs, with greater weight and limited battery life; the screen remains made in China, but the roadmap aims for 100% allied sourcing. Alex praises Palmer Luckey for doing what Apple and Elon Musk have discussed but not fully achieved: rebuilding a sovereign device supply chain. Dave says U.S. policy uncertainty makes long-term manufacturing commitments difficult, while Salim frames the project as an attempt to rebuild American industrial capability rather than merely produce a better laptop. The group agrees government and defense demand can justify the premium, though they debate whether consumers will pay it without import controls or other incentives.

In the listener Q&A, Salim answers how organizations can preserve security restrictions, decisions, records, and compliance when switching models or cloud providers. His answer is to separate replaceable intelligence from persistent organizational accountability: permissions, compliance records, memory, and machine-readable policies should belong to the organization rather than the model or cloud provider. Dave addresses whether breakthroughs could cause stock-price whiplash for OpenAI and Anthropic. He expects major volatility because every model release, benchmark, and Chinese competitor could change investor expectations, but recommends diversification while remaining exposed to the overall AI category. Emad says Chinese models are cheaper because necessity forced efficient engineering under compute constraints, because Chinese companies distilled or borrowed from existing models, because they accept lower margins, and because teams such as DeepSeek optimized inference exceptionally well. Alex explains Microsoft’s employee-exit programs as a strange capital-for-labor substitution: hyperscalers are directing cash into data-center capital expenditure, leaving less operating expense available for employees. On chain theory and simulation theory, Alex says chain theory currently treats the simulation hypothesis as an unfalsifiable “gauge”: if reality is simulated, there may be no observable degree of freedom proving it. Salim says biological dogs remain attractive because evolution provides adaptation and energy efficiency, while mechanical dogs offer rapid production and software upgrades. Dave warns that NVIDIA’s margins sit atop already-high margins at AI labs and TSMC, making vertical integration likely. Emad argues that conventional peer review has often blocked unconventional ideas and that AI-assisted replication and open evaluation could replace it. The episode closes with Salim apologizing for offending religious listeners while insisting he will criticize coercive behavior by all religions equally, followed by the hosts’ “Singulatarians” song about exponential technology, abundance, and human purpose.