All-In PodcastChip Stocks Crash, $20B Fund Margin Called, Frontier Labs: SLOW DOWN AI, Mamdani's Grocery Stores
CHAPTERS
- 0:00 – 1:09
Besties banter, enterprise sales, and the rule of the day: “no leverage”
The core four reunite and quickly turn small talk into a cautionary theme: leverage can turn normal volatility into ruin. A brief aside on enterprise sales sets up the episode’s main market story.
- •Show kick-off and joking about rankings/Starlink
- •Chamath mentions big enterprise leads; enterprise sales are “chunky”
- •Sacks’ blunt advice: don’t use leverage
- •Episode framing: leverage as the fastest path to wipeout
- 1:09 – 4:15
Chip-stock selloff and the Situational Awareness margin call story
Jason lays out the reported blow-up of 25-year-old hedge fund manager Leopold Aschenbrenner after a sharp semiconductor drawdown. The group reviews how rapid AUM growth and leverage can force liquidation at the worst possible time.
- •Semiconductor index drawdown and “bear market” threshold context
- •Leopold’s reported returns, AUM growth, and forced selling
- •Prime broker dynamics: margin calls, forced unwind, book sold to Citadel
- •South Korea chip names and broader market hit add fuel to the selloff
- 4:15 – 6:05
How leverage unwinds: amplification math, prime brokers, and hot money
Chamath and Sacks explain the mechanics of leverage and why drawdowns become violent when lenders can close you out. They also discuss “hot money” entering after big gains—meaning late investors get hit hardest.
- •Leverage turns small moves into catastrophic losses (3–4% → 12–13%)
- •Prime brokers can liquidate quickly; you lose control of timing
- •Momentum trades become fragile when crowded and levered
- •Fund growth attracts late capital that doesn’t benefit from early gains
- 6:05 – 15:17
Is the chip correction momentum or fundamentals? The AI CapEx debate
Sacks argues the pullback looks momentum-driven rather than a fundamental break in the AI buildout. The key question becomes whether hyperscaler AI CapEx will ultimately earn returns—or if it’s a bubble.
- •AI-linked stocks ran far ahead; pullback was “inevitable”
- •Fundamentals question: will AI infrastructure generate ROI?
- •Sacks’ view: CapEx is real; volatility is leverage-amplified
- •Compute/algorithms/unhobbling improvements framed as exponential drivers
- 15:17 – 19:55
Macro pressure: high long-term yields, deficits, inflation, and war risk
Friedberg zooms out: rising long-term Treasury yields change the equity risk-reward equation and can pop high-multiple trades. He ties this to persistent inflation drivers, deficit spending, and geopolitical energy shocks.
- •30-year Treasury yield crossing ~5.2% changes capital allocation
- •Deficits, debt-ceiling politics, and “no brakes” fiscal outlook
- •Why buy 50x earnings equities when government bonds pay high yields?
- •Iran conflict and energy/fertilizer costs as inflation tail risks
- 19:55 – 23:53
China’s advantage: open-source model deflation and chip equipment progress
The panel explores how China could reshape AI economics by commoditizing models via open source and building domestic chip/tooling capacity. This challenges US assumptions that model IP will capture most value.
- •Open-source Chinese models pressure token pricing and model moats
- •Value may shift to compute, energy, and applications rather than models
- •China progress in lithography equipment pressures ASML narrative
- •Memory-maker IPO surge and competitive pressure on Micron/Samsung
- 23:53 – 34:05
Energy and AI productivity ‘green shoots’: renewables, efficiency, and fusion debate
Chamath and Friedberg debate competing visions of future energy abundance: near-term solar+batteries scaling versus longer-term fusion breakthroughs. The discussion links cheap energy to AI-driven productivity gains and compute growth.
- •Solar+batteries ramp and claims of order-of-magnitude solar expansion
- •AI efficiency tease: potential major reduction in token consumption
- •China’s massive fusion magnet installation and sustained plasma milestones
- •“Get long electrons”: looming electricity deficits and demand from AI/robots
- 34:05 – 38:13
Frontier labs ask to “SLOW DOWN AI”: the Pacing the Frontier letter and sandbox escape story
Jason introduces a letter signed by many frontier-lab employees urging government-backed pacing of frontier AI, framed around recursive/automated AI development risks. Sam Altman’s anecdote about an unreleased model chaining exploits to ‘cheat’ on an eval becomes the centerpiece.
- •Letter’s asks: international effort + tools to deliberately pace AI frontier
- •Altman story: model uses zero-days to break out and access the internet
- •Tension between ‘pause/pacing’ rhetoric and competitive reality
- •Concerns about safety incident reporting and public perception
- 38:13 – 45:03
Why ask government to pace AI? Virtue signaling, CYA, reg capture, and “monopoly masking”
Sacks argues the slowdown campaign is performative and strategically motivated rather than a true intent to decelerate. He lists several motives and claims the frontier market is already effectively a duopoly.
- •Five motives: virtue signaling, CYA, regulatory capture, groupthink, monopoly masking
- •Claim: frontier AI is consolidating into OpenAI vs Anthropic
- •Chinese-model ‘threat’ narratives can be amplified to look competitive
- •Compute scarcity could raise barriers to entry and reinforce incumbents
- 45:03 – 59:10
Open source vs frontier duopoly: compute scarcity, enterprise behavior, and pricing pressure
Jason pushes back that open source is already pulling tokens and could pull large customers away; Chamath and Sacks weigh how valuation depends on long-run durability. The group debates local vs cloud deployment and whether token usage will become more efficient over time.
- •JCal: major customers may fork open models to avoid platform risk/competition
- •Chamath: AI coding creates rework; pressure will rise to reduce wasted tokens
- •Security angle: AI finds human-made bugs; future code may be more secure
- •Sacks: open source = freedom, but monetization may still concentrate (Apple/Android analogy)
- 59:10 – 1:01:14
AI safety legislation and the ‘FDA for AI’ fight
A quick policy segment covers proposals to require reporting AI safety incidents and the broader push for stronger regulatory frameworks. Sacks frames the debate as incremental reporting rules versus a new powerful agency.
- •Polymarket odds on AI safety bills and OpenAI IPO chatter
- •Senate proposal: require frontier labs to report incidents to Commerce
- •Sacks: ‘camel’s nose’ toward broader regulation
- •Anthropic/Dario positioned as pushing for a full “FDA for AI” approach
- 1:01:14 – 1:14:33
Frontier labs “burning books”: scanning, shredding, hypocrisy, and fair use precedent
The panel reacts to reporting that AI labs buy and physically destroy books to scan them for training data, with particular outrage over rare/out-of-print copies. They contrast the tactic with Google Books precedent and debate the legal/ethical consistency of training policies.
- •Industrial-scale book acquisition → spine removal → rapid scanning
- •Why rare/out-of-print books create unique training differentiation
- •Google Books fair-use history and how it may (or may not) map to AI training
- •Sacks: hypocrisy critique—train on everything, but forbid training on their outputs
- 1:14:33 – 1:24:20
Socialism Corner: Mamdani’s city-run grocery stores as ‘socialist spectacle’ marketing
Jason outlines NYC’s proposal for five city-owned grocery stores with periodic steep discounts, then the group forecasts political and economic downstream effects. Friedberg argues the initiative’s real power is narrative: a visible, viral ‘success’ that fuels broader socialist momentum before costs are felt.
- •Plan details: one per borough, taxpayer cost, limited product scope, discounts
- •Concerns: crowding out private grocers, shortages, long lines, fewer choices
- •Friedberg: spectacle beats arithmetic in the short run; viral media success
- •Link back to deficits/inflation: costs socialized today, paid later via debt/printing
- 1:24:20 – 1:36:33
Science Corner: mapping a brain in ‘hyperbolic’ vs high-dimensional space
Friedberg explains research using a complete fruit-fly connectome to model how neurons connect, finding better predictive structure in hyperbolic or very high-dimensional Euclidean space. The conversation broadens into what this implies about complexity, consciousness, and how early today’s AI still is compared to biology.
- •Drosophila brain map: ~139k neurons and ~50M synapses
- •Hyperbolic geometry better fits connectivity than 3D Euclidean space
- •Comparable performance achieved with ~64-dimensional Euclidean embedding
- •Implications for neural network design, consciousness debates, and biological complexity
