Weekly Roundup 09/11/26 (Liquid hack, AI x-risk, Silvergate revisited, can AI leak MNPI?) (EP.739)

The episode begins with Matt Walsh discussing alarming claims from employees at Anthropoc, who warn that their AI technology could potentially lead to catastrophic outcomes for the planet. He cites a 10% chance of such disastrous results, as suggested by some experts, while expressing skepticism about the authenticity of Anthropoc's messaging, which he believes may be strategically crafted for public relations, especially as senior Democrats push for AI regulation ahead of the midterms.

Walsh predicts that both OpenAI and Anthropoc may be seeking regulatory frameworks to limit competition in U.S. frontier markets, although he notes that these regulations might not affect international labs. He presents two interpretations of this regulatory push: one cynical, aimed at eliminating competition, and another where the companies genuinely perceive their technology as dangerous. He questions the coherence of employees believing they can save the world from AI while still working at Anthropoc.

The episode also highlights the Liquid network hack, where approximately 4,000 Bitcoin were stolen due to a bug in the transaction logic that allowed the attacker to create fake LBTC, which was then converted into real Bitcoin using a service called Sideswap. This hack resulted in an estimated loss of 47 to 50 million dollars, raising significant concerns about the security of confidential transactions and Bitcoin side chains.

In a separate discussion, Alan Lane, former CEO of Silvergate, asserts that the bank did not fail but underwent an orderly winddown following a significant deposit run. He claims that Silvergate was a victim of a coordinated political attack and regulatory pressures, arguing that history will vindicate the bank and its leadership. Lane emphasizes that the worst allegations of criminality against Silvergate have been proven untrue, yet the government effectively made its business illegal without any regulatory takings claim being brought against it.

The episode also touches on OpenAI's reported solution to the Navier-Stokes problem, a significant issue in fluid dynamics that could enhance turbulence theory and climate models. There are rumors that Anthropic was close to a solution before OpenAI, and a mathematician utilized ChatGPT in their work on the problem, raising concerns that OpenAI might have internalized scientific discoveries made on its platform without proper attribution.

Concerns are raised about the risk of AI models leaking material nonpublic information (MNPI) through user-uploaded documents, particularly as lawyers may inadvertently upload sensitive information into AI systems. This situation could lead to legal and ethical dilemmas, prompting the SEC to investigate insider trading cases linked to AI-generated trading advice, highlighting the urgent need for regulatory guardrails to prevent insider tipping through AI platforms.

While discussions around AI risks are crucial, Walsh emphasizes that they should not overshadow the technology's potential benefits. He argues that no technology is universally good or harmful; it can be used for both beneficial and harmful purposes. The episode concludes with a warning from DOOMers about existential risks from AI, suggesting that even low-probability risks deserve significant attention. Predictions indicate that AI is expected to cause real harms soon, with a forecast that something negative involving AI will occur within the next six months, including the potential for an AI swarm to disrupt a medium-sized country's power grid for two months. The primary concern should focus on the harms caused by AI in the hands of individuals with malicious objectives, rather than the misalignment of the models themselves, emphasizing the need to mobilize AI to defend against harmful uses.