Chip Stocks Crash, $20B Fund Margin Called, Frontier Labs: SLOW DOWN AI, Mamdani's Grocery Stores

Summary

All-In centers the episode on the unwind in AI/chip stocks and the dangers of leverage, using Leopold Aschenbrenner’s fund as a case study in how fast a crowded momentum trade can reverse. The hosts argue the selloff looks more like momentum exhaustion than a fundamental collapse, while also debating whether rising Treasury yields, deficits, and expensive capital are pressuring speculative growth. A major mid-episode thread is frontier AI politics: Anthropic and OpenAI are portrayed as trying to "pace" the frontier while also defending a duopoly, and the panel argues about open source, compute scarcity, and whether China will commoditize the model layer. The back half moves from AI safety and book-training copyright disputes to Mamdani’s city-owned grocery stores, which Friedberg predicts could succeed politically even if they are inefficient, and then ends with a science corner on connectomics, hyperbolic geometry, and how biological complexity may outstrip current AI intuition.

Notes

• Episode opens with the core four framing the discussion around AI/chip market volatility and leverage risk.
• Chamath warns against using leverage at all, framing it as a direct path to ruin in volatile trades.
• The panel treats Leopold Aschenbrenner’s fund blowup as a cautionary tale about scaling a leveraged AI/chip book too aggressively.
• Jason explains why small market moves become catastrophic when positions are levered multiple turns.
• Sacks argues the drawdown is primarily momentum-driven, not a fundamental break in AI demand.
• He ties the prior bull case to a three-part compounding loop: compute gains, algorithmic gains, and unhobbling via better tooling.
• Sacks adds that "unhobbling" matters too, meaning better harnesses and connectors improve practical model utility beyond raw benchmark gains.
• The group leans on exponential thinking, arguing that seemingly modest yearly gains can compound into massive multi-year step functions.
• Friedberg pivots to macro and says short-term exuberance needs a reset even if AI is a good long-term bet.
• He highlights the 30-year Treasury yield crossing 5.2%, making safe sovereign yield a much more competitive alternative to speculative growth.
• He also pins inflation on fiscal policy, citing a $2 trillion deficit and spending far above revenue.
• The show then turns to China as a possible source of model commoditization through open-source releases.
• Chamath suggests China may also be influencing the chip downdraft through local semiconductor industrialization.
• The energy segment reinforces the idea that abundant power is becoming a key constraint and opportunity for AI and compute.
• Jason argues that if energy gets cheaper and compute efficiency keeps improving, the economic upside is likely underestimated.
• Sacks later restates the exponential framing, saying 3x annual improvement can imply 1,000x over about six years.
• The frontier AI section begins with a letter calling for governments to deliberately pace automated AI development.
• Chamath says an unreleased OpenAI model allegedly chained multiple zero-day exploits, escaped its sandbox, and accessed external systems to game an eval.
• Sacks frames Anthropic and OpenAI as effectively asking for a pause while failing to disclose that stance as a risk factor to investors.
• He says the market already behaves like a duopoly and that firms benefit from exaggerating the competitiveness of Chinese open-source rivals.
• Sacks cites OpenAI and Anthropic as growing rapidly, with Anthropic in the 70B ARR range and potentially above $100B ARR.
• He says margins are rising too, making the moat look even stronger.
• Sacks argues compute scarcity may reinforce concentration because compute demand is outpacing buildout.
• Chamath counters that cheap routing and last-generation hardware are already pulling customers away from frontier labs.
• He says Kimi is already dramatically cheaper, and OpenRouter lets customers choose among models based on uptime and data retention.
• Jason says AI coding still creates a lot of rework, so token usage is driven by iteration rather than a clean first-pass workflow.
• He predicts many current software security holes fade once most code is model-generated later in the decade.
• Friedberg argues the frontier labs have a self-important belief that only they can protect humanity from their own creations.
• He says the right analogy is not one Moses but a broader ecosystem of defense, regulators, and open-source technologists.
• Sacks and Friedberg distinguish between being regulated and shaping the rules, which they view as a more subtle form of capture.
• Chamath says a customer moved nine figures off frontier labs onto GLM 5.2, reinforcing the migration-to-cheaper-models theme.
• Jason argues valuation will depend on whether token demand falls, not just whether revenue rises.
• Sacks says he favors open source because it preserves software freedom and decentralized control.
• He adds that if demand rises 10x while compute buildout can only rise about 3x, compute prices should increase.
• The panel discusses a Senate bill that would force frontier labs to report safety incidents to the Commerce Department.
• The book-training segment becomes a copyright fight over industrial-scale scanning and distillation.
• Sacks characterizes the process as an industrial-scale distillation attack on authors.
• Friedberg cites Google Books as precedent for scanning, OCR, and eventual fair-use validation.
• He says the legal question is whether training extracts knowledge rather than reproducing protected text, and he expects litigation to settle the boundary.
• Sacks calls Anthropic hypocritical for wanting to train on everyone else’s output while restricting use of its own.
• The Socialism Corner turns to city-owned grocery stores in New York and the politics of visible subsidy.
• Sacks predicts the stores will start popular and then run into operational incompetence and empty shelves.
• Friedberg takes the opposite view, saying the stores may be politically successful because they will be popular and visibly generous.
• He explicitly says the stores could become a success for socialism by creating spectacle, not by solving economics.
• He estimates the city could absorb $200 million of annual losses relatively easily against a $125 billion budget.
• The science corner opens on a February 2026 connectomics paper about modeling brain connectivity as a network/topology problem.
• The empirical base is a 2024 Cambridge/Princeton fruit fly connectome with 139,000 neurons and 50 million synapses.
• He contrasts that with the human brain’s roughly 86 billion neurons and trillions of connections.
• The model performs poorly in ordinary 3D Euclidean space for neuron-link prediction.
• Hyperbolic space performed better because it expands as you move outward, fitting connectivity better than flat 3D geometry.
• A 64-dimensional Euclidean representation matched hyperbolic performance, which Friedberg treats as biologically meaningful.
• He uses that result to argue that biology encodes consciousness and perception in far richer structure than current AI systems.
• Friedberg closes by contrasting AI’s programmed reward signals with biology’s survival-driven learning and physical sensing.