The Venture Capitalist Who Bets on Science Most Investors Think Is Impossible | Josh Wolfe

KEYPOINTS

• 01

  The web will be redesigned for AI agents, and agents may make APIs obsolete.

  Just as companies optimized for Google search, Wolfe expects restaurants, shopping and booking sites to rebuild interfaces for the clicks of AI researching products—and to try to influence or even trick the AI’s judgment. Rather than negotiating complex, legally restricted APIs, companies may simply let AI act as a user on the front end. Wolfe faced a split-second moral dilemma when OpenAI’s Operator asked him to confirm he wasn’t a bot. He expects this to evolve ‘in a quarter or two.’

• 02

  A contrarian bet: 30–50% of AI inference moves onto phones, making memory chips ascendant.

  Training still needs giant H100 clusters, but if much inference runs on your cached emails, WhatsApp, photos, health data and calendar, memory matters far more. Wolfe favors Android over Apple because of its more open architecture—Perplexity can already act as an assistant on Android. Among memory makers, he sees Samsung as ‘more like Intel,’ sclerotic and slow; Micron constrained by U.S. export controls; and SK Hynix best positioned. Just as AI researchers repurposed gaming GPUs, turning Nvidia from $15 billion to about $3 trillion, attention could shift to memory-heavy architectures.

• 03

  Efficiency increases total demand (Jevons Paradox).

  Cheaper, smaller refrigerators put one in every garage and office, so total electricity demand rose. Moving from a 56k modem to fiber didn’t reduce usage; it gave us 4K streaming. DeepSeek-style efficiency gains likely mean more AI demand, not less. Nvidia, with 90–95% margins on $30,000 chips, could choose lower prices for more volume—Wolfe wouldn’t count Jensen Huang out.

• 04

  AI threatens elite knowledge jobs before blue-collar ones.

  Peter Drucker never imagined machines would most threaten knowledge workers. Wolfe runs his medical records from top doctors through LLMs to see what they missed. Kahneman’s Noise showed the same facts produce different verdicts by judge and time of day, an argument for consistent AI. Meanwhile, robots still can’t serve a meal or reliably fold laundry. Wolfe cites the old plumber joke: $5 for the part, $995 for ‘knowing where to tap.’

• 05

  Models will commoditize; proprietary data is the moat.

  Wolfe thinks Zuckerberg is playing it smartest: open-source Llama, let models improve for free, and run them on Meta’s deep, longitudinal data—WhatsApp messages, Instagram lingering time—while routing around Apple and Google via glasses and the neural wristband from CTRL-labs, a Lux company. Bloomberg’s normalized financial data is ‘a quant dream.’ Pharma’s failed clinical trials are gold. Unmined data is like ‘an oil well that isn’t mined’; LLMs become commodity excavators.

• 06

  Margins from technology are fleeting—benefits go to customers.

  Like Buffett on new loom technology, one company stands on tiptoes and soon everyone does. Only true monopolies, like Nvidia until an alternative appears, keep fat margins—cars are vastly better today yet margins never stayed high. So when a vendor quoted a costly AI system to search Lux’s Gmail, Slack and documents, Wolfe’s default was to wait six months for the major LLM providers to compete for enterprise accounts.

• 07

  Machines will do science 24/7.

  Like GarageBand virtualized the recording studio, you’ll form a hypothesis on a beach, have AI review literature and freedom to operate, and send the protocol to a cloud lab. The robot executes, then asks, ‘Run again with 10 milliliters instead of five?’ The catch: discoveries like penicillin and vulcanization came from screw-ups, so you may need to engineer randomness. Physical reaction steps remain the bottleneck.

• 08

  Of informational, analytical and behavioral edges, only behavioral persists.

  Regulations and brilliant data users have eroded information advantages. One hedge fund legally inferred Adobe’s sales from order numbers in purchase URLs until the trick leaked. AI will make analysis abundant and may ask ‘the five questions you didn’t ask.’ Lux’s 10-year locked funds give time arbitrage. Passive indexing is itself an algorithm—‘$1 in, buy; $1 out, sell’—and Wolfe expects active managers to return as the cost of capital rises.

• 09

  In a world of abundant AI content, human connection becomes scarce and valuable.

  Every time we said humans would retain an edge—chess, Go, video games, creativity—it fell. Wolfe can’t paint, but uses Midjourney prompts creatively and turns boring PTA emails into Matt Levine pastiches for his wife. What remains scarce is shared experience: inside jokes and commiseration. Lux is in-office Monday to Friday for serendipitous introductions and noticing when someone ‘swallows’ nervously, which never happens on Zoom.