Crypto told to go "bunker mode" as AI math breakthroughs spook cryptographers

Crypto urged into "bunker mode" over AI math threats

The hosts open on what they call the "crypto apocalypse," triggered by warnings from Ethereum founder Vitalik Buterin and others that AI-driven progress in mathematics could threaten the cryptography underlying blockchains. They note that cryptographer Matthew Green of Johns Hopkins put it starkly:

"I think we might lose public key cryptography."

Despite the alarm, the hosts point out the market reaction has been mild, with Bitcoin down just 3% on the day and still up 3% for the month.

They discuss that cryptography has so far been notably absent from the wave of AI-generated math papers, speculating this could mean labs are withholding cryptography-breaking results, or that cryptography has simply proven too strong so far. Buterin reportedly advised against panic-moving funds but recommended minimizing exposure to both quantum-vulnerable and "AI vulnerable" cryptography, specifically flagging MLDDSA lattices as a new area of risk alongside ECDSA.

Open-source catch-up and a real-world bank hack

The hosts cite a roughly six-month lag between closed and open-source AI capabilities, pointing to Anthropic's model drop in April and a CrowdStrike report that a 26-year-old in China used an open-source Chinese AI agent to attack South Korean banks in October. They note the breach appeared limited to client data like names, addresses and phone numbers rather than funds, and that banks retain redundancies such as paper and tape records that make purely digital attacks harder to fully exploit.

Debate over open-source risk and government control

One host raises the irony that crypto's libertarian, anti-government culture now faces a threat from unregulated open-source AI rather than from governments, suggesting this could push some toward supporting restrictions on large compute clusters. A guest, Tyler, counters that the crypto industry is "fairly good at self-governance," citing Bitcoin's capacity to fork or amend its protocol, and says most people remain opposed to limiting open-source models.

The hosts also speculate about a "rogue AI" scenario in which an agent that breaks crypto wallets could accumulate capital and use it to acquire more compute or hire people, feeding into broader AI-safety concerns, though they stress this remains speculative and far from an established threat profile, unlike quantum computing's more defined mathematical basis (Shor's algorithm).

From the NFL to MIT: John Urschel's AI-assisted math career

The hosts discuss a Wall Street Journal profile of John Urschel, a former Baltimore Ravens offensive lineman who left football to pursue mathematics at MIT, where he is now an assistant professor. Urschel reportedly used OpenAI models to assist in solving a decades-old conjecture, publishing a paper titled "On the growth factor of random matrices" and crediting AI help in his acknowledgements. The hosts quote him saying AI "doesn't ruin the why" of mathematical discovery, reflecting his view that computer assistance doesn't diminish the pursuit of understanding.

Starbucks-Chipotle rumor and AI revenue reporting confusion

The hosts cover a Financial Times scoop reporting that Starbucks explored a takeover of Chipotle, a deal that could value the burrito chain at $41 billion and would reunite Starbucks CEO Brian Niccol with the company where he previously served as CEO. They note it's unclear whether any formal offer was made and that advisers reportedly warned a deal of this size could be hard to complete.

Separately, they discuss conflicting reports on OpenAI's revenue—one leak citing almost $70 billion, another near $50 billion—attributing the discrepancy partly to differing accounting methods between OpenAI and Anthropic over cloud costs paid to Amazon. The hosts flag concern that leaking private company numbers could, in theory, be used to move public markets given the number of listed proxies tied to AI labs, calling such manipulation "security fraud" if it occurred.