Gutenberg Lab — Morning Paper

AI Investment Boom and Bust: Leopold Aschenbrenner's $20B Fund Collapses Under 3.5x Leverage as China's Open Source Strategy Threatens US AI Productivity Thesis

Speakers

Jason Calacanis × Chamath Palihapitiya, David Friedberg, David Sacks

Bottom line

Even brilliant investors betting on AI's exponential growth can be destroyed by leverage and short-term volatility—being right in the long run doesn't matter if you can't survive the short run.

Three-Line Summary

Former OpenAI researcher Leopold Aschenbrenner's $20B hedge fund was margin called and liquidated after semiconductor stocks crashed 20%+. Using 3.5x leverage, the fund grew from $225M to $45B peak in two years but was wiped out when South Korea's market fell 40% in 40 days. Meanwhile, China is deploying open source AI models (Kimi, DeepSeek) at 90% lower cost, threatening the US AI productivity thesis that many see as the only escape from fiscal crisis.

Three Key Points

1. The Destructive Power of Leverage: Leopold's fund used 3.5x leverage. Chamath's math: "A 34% move is amplified 12-13x... a 25% move becomes a 75% loss." In South Korea, 1.2M leveraged trading accounts were margin called, with ~1M (3% of population) fully liquidated. Despite 5-year gains of 850% (Micron) and 875% (NVIDIA), a 20% short-term correction erased everything.

2. 5.2% Treasury Yields Create Structural Headwinds for AI Stocks: 30-year Treasury yields crossed 5.2% (first time since 2007), equivalent to ~10% pretax return. Friedberg's question: "Why pay 50x earnings for semiconductor stock when I can make 10% pretax owning federal bonds and go to the beach?" With $2T annual deficits and persistent inflation, risk-free rates now compete directly with AI equity valuations.

3. China's Open Source Strategy Threatens to Delete Value from Model Layer: China is releasing open source models (Kimi, DeepSeek) at 90% lower cost, attempting to "delete" value from the model layer. Friedberg: "If value shifts entirely to compute infrastructure and application layers, and China controls energy production + chip manufacturing, we end up losing much of that value." This undermines the core assumption that AI productivity gains will rescue the US from fiscal crisis—what Friedberg calls "the 30-year timeline."

Editorial Perspective

This episode validates a pattern we've tracked across platform transitions: the capital allocation trap. Whether in the early internet era, the App Store economy, or blockchain speculation cycles, exponential growth conviction justifies leverage—and destroys even the most brilliant investors. Leopold's "OOMs" (Orders of Magnitude) framework may be fundamentally correct, but markets are voting machines in the short run and weighing machines only in the long run (Buffett). The lesson remains timeless, even in the AI age: leverage is the only way smart people go broke.

Source: All-In Podcast "Chip Stocks Crash, $20B Fund Margin Called, Frontier Labs: SLOW DOWN AI, Mamdani's Grocery Stores" (2026-07-31)

https://allinchamathjason.libsyn.com/chip-stocks-crash-20b-fund-margin-called-frontier-labs-slow-down-ai-mamdanis-grocery-stores

AI Disclosure: Produced with AI assistance; facts and analysis reviewed by our editorial team.

• --

AI Doesn't Steal Jobs — It Steals the Joy of Doing Them: Reclaiming Sensory Contact with the Physical World

Guests

David Pierce (The Vergecast host) × Ian Bogost (writer at The Atlantic, professor at Washington University)

Bottom line

Smartphones are only part of the problem; the real threat is "dematerialization" — a century-long systemic stripping of everyday physical contact, driven by economic incentives and automation since the Industrial Revolution.

Three-Line Summary

In his new book *The Small Stuff*, Ian Bogost argues that modern life has systematically removed sensory, physical engagement with the material world — a process he calls "dematerialization." Smartphones are only one factor; sensor-driven public restrooms, economic disincentives to own property, and automation have been eroding tactile experience for a century. The remedy is not digital abstinence but expanded attention to sensory opportunities already present in daily life — the warmth of folded towels, the crunch of gravel underfoot, the feel of ideas forming in your fingertips as you type — a state he calls "gratification," distinct from distant-future "happiness."

Three Key Points

1. AI threatens embodied experience, not just output

Bogost argues "ideas live in my fingertips, not my brain" — what AI threatens is not the product of work but the moment-to-moment sensory delight of doing it. Even unwanted tasks (deleting emails) provided tactile engagement. Automation has historically never delivered promised leisure time, only "more work, and that work is also usually worse." [00:29:29–00:30:47]

2. Digital devices can support physical-world attention

ASMR videos (e.g., 20-minute towel-folding clips) model focused attention to ordinary objects, showing that digital media can encourage sensory engagement. A powered-off phone is still a "delightful sensory object" (textured case, heft). Algorithmic feeds deliver "vicarious gratification" — a tomato-sauce video reminds you of the sensory experience of cooking. [00:24:39–00:28:46]

3. Tickets as the exemplar of dematerialization's trade-offs

Physical tickets involved loss anxiety, the friction of cardstock rubbing together, repeated pocket-checking rituals, and memento value. QR codes eliminated the anxiety but also all sensory texture. Bogost calls this a "small loss" — individually trivial, but cumulatively producing the feeling "I don't feel like I'm a real human being in the world anymore." [00:05:01–00:09:21]

Editorial Perspective

This argument shifts the focus of tech criticism from "screen time" to "total sensory contact." Bogost's distinction between "gratification" (moment-to-moment sensory pleasure) and "happiness" (achievement-driven fulfillment) clarifies what optimization culture has overlooked — the texture of warm towels, the sound of gravel, the feel of ideas forming in your fingertips. His case study of using five computers to recreate a traditional doorbell sound in his 1909 house demonstrates that technology and sensory experience need not oppose each other. Yet the unresolved question remains: can this equilibrium hold when capitalism has "every possible incentive to dematerialize us as much as possible to just turn us into a warm bag of credit cards"?

Source: The Vergecast, "The life-changing magic of touching stuff" (2026-07-07)

(unknown)

AI Disclosure: Produced with AI assistance; facts and analysis reviewed by our editorial team.

• --

AI solves PhD-level math but can't write a good tweet — Taste Labs trains models on expert preference data to fix the "taste problem"

Guests

Jason Calacanis (host, angel investor) × Thais Castello Branco (Founder/CEO, Taste Labs)

Bottom line

As AI commodifies creation, human curation becomes more valuable, not less — but the industry must build infrastructure to inject taste into AI outputs.

3-Line Summary

Taste Labs addresses why AI succeeds at objective tasks (code, math) but fails catastrophically at subjective ones (design, writing). Foundation models are trained to predict the "most likely answer," which by definition produces average, forgettable results. The company curates preference data from approximately 1,000 expert "tastemakers" to pull models away from the mean, while building infrastructure (APIs, design indexes) so applications can inject taste into AI outputs.

Three Key Points

1. The taste problem is structural, not fixable by scale alone: Foundation models optimize for "most likely answer," which works for objective domains but fails for creative ones. Great design is out of distribution — the opposite of what LLMs are trained to produce. Thais argues you need a separate layer of expert preference data to pull models away from the mean [00:06:21]

2. The curator class has been gutted without replacement: Professional critics (film, food, books) were eliminated by newspaper and magazine collapse. Social media promised to replace them but delivered chaos. Jason admits: "I haven't replaced that function yet" [00:12:44]

3. China froze autonomous vehicle permits to protect jobs, not safety: Bloomberg reported China suspended new AV permits in April 2026. Jason argues it's about preventing mass driver unemployment protests, not safety concerns. Predicts Western governments will face same pressure; proposes licensing fees on AVs to fund driver retraining [01:07:28]

Editorial Perspective

Taste Labs is betting on the inverse of early internet history, when human directory editors lost to search algorithms. This time, the algorithm itself has a structural flaw — regression to the mean — creating space for human curation to regain value. Leopold Aschenbrenner's 4x leverage collapse illustrates how the gap between technological acceleration and human judgment can be fatal, whether in investing or product design. The question is whether "taste infrastructure" can scale before AI slop drowns the internet.

AI Disclosure: Produced with AI assistance; facts and analysis reviewed by our editorial team.

Source: This Week in Startups "Why AI has no taste and how to fix it (w/ Thais Castello Branco) | E2319" (2026-07-31)

https://4a885955-6823-4b22-a3e1-a526c25516a5.libsyn.com/why-ai-has-no-taste-and-how-to-fix-it-w-thais-castello-branco-e2319

• --