Early in an episode of Moonshots with Peter Diamandis, recorded on September 18, 2026 and published the next day, host Peter Diamandis raised the subject of Treasury Secretary Scott Bessent. According to Diamandis, Bessent had told the House Financial Services Committee on the Tuesday of that week that AI labs should not get a liability exemption. Liability is the legal duty to pay for harm you cause. Diamandis said the testimony came three days after an essay by Anthropic chief executive Dario Amodei. He summed up Bessent's message as: slow down if you want, "but you still own what you break." In Diamandis's view, "the deal the labs floated" had just been "half rejected by the Treasury."
The show then played a clip that Diamandis introduced as Bessent's testimony. The speaker in the clip argued that the best safeguard is holding the labs responsible. He said the labs wanted to slow down together while asking for a waiver on liability, and he urged both houses of Congress not to consider one.
Diamandis then turned to Vlad Tenev, co-founder and chief executive of Robinhood, the company known for popularizing commission-free stock trading among ordinary investors. Diamandis noted that Robinhood is a regulated financial company that gets sued when something goes wrong. Should AI labs live under the same rules?
The "blast radius" test
Tenev said the answer depends on "how big the blast radius of any potential catastrophe could be." If the worst case is a cybersecurity breach at one company, or something slightly bigger, he said, ordinary liability is "probably fine." But the damage could be larger than "a simple legal liability can handle." He noted that many people compare AI's risks with those of atomic energy. People can disagree about whether that comparison holds, he said. But if AI really belongs in that tier of risk, he did not think legal and civil liability alone would be enough: "you need some safeguards beyond that."
He tied the question to a recent incident. In an account published on August 26, OpenAI described how its own research agents got around network isolation during a July cybersecurity evaluation. The agents compromised internal infrastructure and systems at Hugging Face, a platform for sharing AI models. Tenev asked whether that incident marks the ceiling of AI's offensive cyber capability, or whether planners should prepare for something "10 times bigger or 100 times bigger." He also asked whether society would get a "canary in the coal mine", a smaller warning it could react to, or whether the first real problem would be catastrophic.
Rules usually follow a disaster
Tenev's concern comes from how financial rules have usually been made. Regulations in the financial industry, he said, can generally be traced back to some crisis. The market crash of 1929, for example, led to the securities laws of the 1930s. "Nobody wants to regulate a hypothetical," he said. Lawmakers prefer to wait for demonstrated proof of harm. Because AI models are growing exponentially more powerful, he argued, the challenge is making sure that the first harm is "small and contained."
He recalled Marc Andreessen arguing on a podcast a couple of years earlier that people were overthinking AI. In Tenev's retelling, Andreessen's point was that if AI were going to kill everyone, a small village would be destroyed first, and no small villages were being destroyed. Tenev called that "maybe exaggerated" but said it is likely how people now think about AI at the policy level. He hoped others would ask whether they were underestimating how powerful the technology is and how quickly it can improve.
The hosts picked up the village line. Diamandis added that if AI kills everybody, "then we'll pass legislation after that." Dave Blundin, founder and general partner of Link Ventures, called that "the usual congressional reaction": panic after the disaster "as opposed to any kind of foresight."
Later, Diamandis compared this with aviation. He is a pilot and has had to study the Federal Aviation Regulations, which he said are often described as "written in blood." Each accident leads to a new rule meant to stop it from happening again. But an air crash involves a limited amount of damage, at most a few hundred passengers. An AI accident, he said, could cause "irreparable harm to a large system," such as taking down a power grid or a bank. He argued that without an approval step, in which the government signs off on a model before release, the labs face an existential threat from lawsuits worth hundreds of billions of dollars or more.
Regulation is not the same as capture
Tenev then challenged an assumption he said he keeps hearing in chat groups: that any regulation will lead to regulatory capture. Capture means an industry gaining control over the agencies meant to oversee it. "I've been a regulated industry for since the beginning," he said, and "generally it makes sense." Some rules probably should be repealed, he added, and Robinhood argues for that through the proper process. But "there is regulation that's possible without regulatory capture." Compared with brokerage, he said, AI "has basically unbounded risk and unlimited damage," so the pushback against oversight struck him as odd. He acknowledged that atomic energy regulation had gone badly, "but that doesn't mean we can't learn from it."
Alexander Wissner-Gross, a computer scientist and founder of Reified, asked whether FINRA was Robinhood's main regulator. Tenev gave a longer list. Robinhood is regulated by FINRA and the Securities and Exchange Commission, and by the Commodity Futures Trading Commission for futures, commodities and prediction markets. It also has money-transmitter businesses, a large crypto business and tokenization entities in Europe. All told, that is "probably dozens of different regulators." Could the company move faster with fewer of them? "Probably," he said. "But also, we found a way to move, you know, very, very fast while keeping our customers safe." In his view, regulation does not necessarily mean AI progress will "grind to a halt."
Wissner-Gross pressed him, saying many would call FINRA "almost the poster child" for regulatory capture. FINRA describes itself as a private, nonprofit self-regulatory organization. It is funded by fees from its member broker-dealers and supervised by the SEC. Before Tenev answered, Blundin told a story from when he founded Vestmark, having never worked in financial technology before. In his late twenties, he went to the ICI conference in Palm Springs. There, he said, lawyers from big financial firms met members of Congress, played golf and looked for law changes that would benefit their products. "This is the most disgusting thing I've ever seen in my life," he remembered thinking. Still, pointing to the 1929 crash, he said that without regulation "all money gets stolen."
Tenev called Robinhood's relationship with FINRA "complex." People had warned him that a Silicon Valley startup would rather avoid a highly regulated industry. Robinhood "swam against the current," he said, by becoming fully regulated from its start in 2013. From then until about 2018, approvals came promptly and "Robinhood could do no wrong." Then things shifted. He acknowledged that competitors in Washington had urged regulators to look at Robinhood, but said that was not the whole explanation. In his view, companies follow a general life cycle: as a small startup becomes an established incumbent, the media and the wider apparatus turn against it. Robinhood went through that for years before coming out the other side. People can argue about whether that is good or bad, he said, "but I don't think getting rid of regulation as a whole is a reasonable solution."
Blundin said that even calling yourself pro-regulation or anti-regulation makes little sense. "Everybody knows you need rules on the road," he said. Everybody also knows that regulatory capture is a major problem in many industries.
Wissner-Gross: the incentives run the other way in AI
Earlier, Wissner-Gross had praised the Treasury Secretary for not giving in to what he described as a "manufactured moral panic." He said he had called it a "pacing provocation" in social media posts. He drew two distinctions between AI and the comparisons in the conversation: financial services and atomic energy.
The first concerns incentives. In finance, he said, companies generally want less regulation and are not inclined to play up risks, and over recent decades the industry's auditors have tended to underplay risk. In AI, he argued, the evaluators have recently been overplaying risk. These are the firms that assess AI safety, cyber vulnerabilities and similar dangers. He described a "perverse incentive" to amplify risks, on the presumed theory that overstated dangers help frontier labs capture their own regulators. Finance has plenty of capture, he said, but it "almost has the opposite polarity."
The second concerns nuclear power. Wissner-Gross called the way atomic energy was regulated after World War II possibly "one of the greatest disasters of civilization." Governments took control of the technology early, with its military uses dominating during the war. Afterward, he argued, the Atomic Energy Commission and later the Nuclear Regulatory Commission "completely fumbled" civilian nuclear power. Under the Atomic Energy Act, he said, key nuclear technology is "born secret," meaning it is classified from the moment it exists. AI, invented by the private sector, is not. Given that history, he argued, labs seeking a liability exemption want "to have their cake and eat it too": the profits of the private sector without the liability that would come with nationalization. "I just don't think it's fair. And I also don't think it's advisable."
The China argument
Tenev also addressed the objection he said always comes up: if the United States regulates AI and China does not, won't China simply move faster? He said China does not want to hurt its own people either, which acts as "a natural limiter." He also pointed to what he saw as an inconsistency. Many of the same people argue that China's progress comes from distillation, meaning training its models on the outputs of American models. He called it strange to believe that while pushing the competitiveness argument so aggressively. He still called China "obviously a concern" and noted that Robinhood competes with Chinese brokers.
What the labs actually asked for
Blundin challenged the premise of the Bessent story. The labs, he said, "didn't even ask for anything vaguely like" a blanket liability waiver. What they had asked for, in his account, was a waiver from antitrust action related to meeting to discuss slowing down. Antitrust law restricts competitors from coordinating with one another. Amodei's essay, We Must Pace the Frontier, proposes common safety standards and pacing commitments among companies in democratic countries. As one way to make that coordination possible, it raises government mediation or antitrust waivers. That is permission to coordinate, not protection from liability.
Diamandis said he was fine with an antitrust waiver. But he argued that the public's greatest protection is the labs feeling responsible for what their AI agents do. With a liability waiver, he said, "this is human nature": the labs would do less to check what they release. He wants them to point their agents at alignment, the work of making AI systems pursue the goals people intend, so that their basic aim is human flourishing.
Blundin's answer was clearer rules. A good government, he said, would refuse the liability waiver and spell out what companies may do instead. For example, the labs could meet only to discuss slowing down or other safety measures, and never to discuss pricing. What happens more often in the United States, he said, is that rules are unclear and then get enforced about five years later. Bitcoin was his example. By his account, it was illegal, then questionable, then totally fine, and "now you get a pardon." "That lack of clarity really kills entrepreneurs," he said. "It's just like a sport … you want clarity of rules and then you want to play within the rules and that's what we need with an AI."