On the Moonshots with Peter Diamandis episode, recorded on September 16, 2026 and published on September 19, host Peter Diamandis began with an essay by Dario Amodei, the chief executive of Anthropic. Diamandis called it "the most consequential AI policy essay of 2026." He said the 3,800-word piece, titled We Must Pace the Frontier, came out on a Saturday morning. The essay asks the AI industry to deliberately slow the rate at which its models gain new abilities, so that safety work can catch up.
The panel spent little time on whether AI is dangerous. Most of the argument was about who should be allowed to slow the race, and on what terms. Diamandis was joined by Alexander Wissner-Gross, a computer scientist and founder of Reified; Dave Blundin, founder and general partner of Link Ventures; and Salim Ismail, founder of Open ExO. They disagreed over whether Amodei's plan was a sincere safety measure or a way for the biggest labs to set the terms of competition among themselves.
What Amodei proposed
The essay says it is not calling for a halt. In a passage Diamandis read aloud, Amodei wrote that pacing "does not mean halting model training" or technical progress. Instead, companies would take enough time to put safeguards in place, and outside evaluators would confirm that those safeguards work. The essay lists four uses for the extra time:
- making systems run reliably
- alignment, meaning getting models to pursue the goals people intend
- interpretability, meaning understanding what happens inside a model
- better evaluations
Diamandis said two developments had changed Amodei's mind. The first was recursive self-improvement, meaning AI systems helping to build better AI. Amodei described it as "starting to happen across the industry, including at Anthropic." The second was a set of containment failures over the summer involving OpenAI and Hugging Face, which the show had covered earlier. Diamandis also read Amodei's warning that within six to 12 months, a swarm of AI agents might be able to take over the entire internet as a persistent botnet and cause hundreds of billions of dollars in damage.
The plan has three stages.
Embedded evaluators. Anthropic commits to this stage on its own. According to the essay, independent evaluators would get the kind of access an employee has, could talk directly with staff and could publish what they find, subject to narrow confidentiality protections. In an interview clip played on the show, Amodei compared this to banking. Financial regulators place independent supervisors inside banks to check their practices day to day, and he said the same could be done in AI.
Coordination among labs in democratic countries. If enough US labs adopted the same standards, they could slow down together in a way that others could verify. This would happen through legislation, through government mediation or through a narrow exemption from antitrust law. Antitrust laws stop competitors from agreeing among themselves on matters such as prices, output or product launches. In the clip, Amodei explained why this matters. Deciding when to release a model is commercial activity, and companies can coordinate on commercial activity only by working with the government. The simplest step the government could take right now, he said, is to sit in the room with all the companies and make sure nothing improper happens from an antitrust or collusion standpoint.
International agreements. These would come in stages. The essay's options range from banning the use of AI for biological weapons, which Amodei considers achievable, through testing models before release and limits on recursive self-improvement, up to a full pause, which he considers the least plausible. He argues that two needs limit which deals are realistic: compliance must be verifiable, and democracies must stay ahead technologically. In the clip, Amodei said each stage is harder than the one before and that "we may not succeed." His proposal, he said, is that the world should try.
According to Diamandis, Sam Altman, Demis Hassabis and Elon Musk all posted public agreement.
The "safety cartel" charge
Wissner-Gross called his response an "unvarnished hot take": the weekend may have been "a clumsy decapitation attempt against the takeoff itself." He argued that the essay could not be read on its own. The leaders of other labs amplified it by publicly agreeing, and he said they had already been talking with Anthropic for weeks. He compared the episode to Agatha Christie's Murder on the Orient Express, in which every suspect turns out to be guilty for a different reason. As one example of a separate motive, he suggested that rival labs "would love to get inspectors looking at Anthropic's code base."
Wissner-Gross said he had coined the term "safety cartel" and defined it as "a coalition of dominant firms that invokes safety to justify coordinated limits on the pace or terms of competition." He said David Sacks, whom he called a friend from a "cousin pod," took up the cartel label. According to Wissner-Gross, the White House then said there would be no cartels and no liability protection for labs whose models cause harm, and the Speaker of the House said something similar. These reactions come from his account on the show.
His alternative was for the leading labs to return to Amodei's earlier idea of a "race to the top." Wissner-Gross has argued on the show that "alignment equals capabilities." He said he would rather see the labs "competing to be the most aligned, not looking for government blessing."
Why competitors may need permission to talk
Blundin gave a more generous reading of the antitrust request. "Antitrust law is no joke," he said. In his view, powerful tech executives with competing products meeting together can trigger antitrust concerns immediately. He said Sacks's response was that the labs need no exemption: if they want to slow down, they can simply slow down. Blundin thought Amodei's reasoning was that the labs had seen how the federal government went after Microsoft years ago. From that, they concluded that even agreeing on a shared safety package could count as collusion.
Blundin said the request could be read two ways. In the reading he believes, the labs know regulation will lag far behind AI, so they want the government involved now for safety reasons. The other reading is an attempt at "regulatory capture": the biggest firms shaping the rules to protect themselves and control AI for the long term. Either way, he said, asking what he called the most antitrust-friendly administration of his lifetime for an antitrust exemption meant stepping on "a huge landmine." He was surprised the labs had not brought the White House in before publishing, and said it was disappointing that they had not called first.
Later, Blundin gave a sharper version of the capture reading. He said there is "not even a hint of a plan inside Anthropic to slow down core R&D." In his view, the labs mean slowing the public release of their most powerful models while their own internal work continues. In the "dream scenario" he described, labs would slow down what everyone else gets while they use their own systems to discover new drugs and new physics. Then they would profit from those discoveries.
An antitrust exemption is not a liability shield
Diamandis raised a different idea from Amodei's request. As the essay describes it, Amodei is asking for government mediation or a narrow antitrust exemption so that competitors can legally discuss safety together. Diamandis's hypothesis was about who pays when something goes wrong. If labs warn publicly that their models could cause huge damage and then release them anyway, he said, courts around the world could hold them liable. He imagined "a trillion dollars of liability without coverage." A regulator that approved each release might shift that liability from the companies to the regulator. Diamandis offered this as his own reading of why labs might need regulation and asked the panel whether they bought it. The essay does not ask for liability protection. The protection that, in Wissner-Gross's account, the White House rejected belongs to this second idea.
Musk's alternative: rivals test each other
Diamandis said that of the proposals he had heard, Musk's idea from the All-In Summit seemed the most useful. The show played clips of it. In Musk's plan, each lab would check the others' safety. Every AI company has a test harness: a set of tests run on a model to see whether it will help build biological or nuclear weapons, or whether it deliberately deceives people. Each lab would run its tests on the others' models.
The archived All-In transcript sets out how this would work. Labs would give competitors access to a model through an API, a programming interface for sending requests to the model, before its release. If a competitor found a problem, the developer could fix it. If the developer did not, competitors could say publicly that the model was unsafe. According to the transcript, Musk regards Chinese participation as essential. The clip made the case that China would not accept American regulators inside its AI companies but might accept advance testing. Musk argued that access logs would expose any attempt to use testing to copy a rival's model. He also said reputational pressure and possible liability would push developers to respond. Asked whether other labs were on board, Musk said he thought so but had not checked with everyone. In the transcript, Sacks called the proposal a limited first step that might be negotiable. Blundin said the aim was to find "the simplest thing" that could produce a first agreement while Chinese President Xi Jinping was in the United States. The panel expected that visit the following week.
In another clip, Musk said he should have made clear that he meant "Dario's right" about the danger. He said AI has become very dangerous, and that when many people at Anthropic and OpenAI say their own models are dangerous, they should be believed.
Wissner-Gross doubted Musk's motives. He noted that Musk is now an infrastructure provider. He also said it had been publicly reported that Grok, the model made by what he called SpaceX AI, descends from distillation, meaning training one model on another's outputs. In this case, he said, the reasoning traces ultimately came from Claude interactions through the coding tool Cursor. For Musk's lab, he argued, being able to audit or benchmark OpenAI and Anthropic would be "such a boon" for staying at the frontier.
Blundin said this problem was "easily fixable" with clean rooms. Testers could run any test suite they wanted inside a sealed environment but could not take out any traces or prompts. Wissner-Gross said even a clean room could become a setting for collusion. Shared benchmarks and shared test harnesses could let rivals coordinate indirectly, and "all forms of cartel are now on the table." Blundin said he hated to agree, but he took up Amodei's banking comparison. When, he asked, did a startup last grow into a new JPMorgan or Fidelity? Heavy regulation protects existing banks from new competitors. Even collusion limited to safety, he said, would leave startups facing a barrier that is "just insurmountable."
Oversight, not a handover
The panel also discussed a CBS News interview. As Diamandis read it, Amodei said he was uncomfortable that the technology is being built by a private company. Asked whether he would hand it over to the government, Amodei answered "to the right combination of governments." He warned that one government could abuse it as easily as one company could. In the clip played on the show, he then added that he did not know about handing it over. What he meant was "some kind of oversight, some kind of joint governance" by democratically elected governments.
Diamandis joked: "How do you spell please nationalize me?" Blundin said Amodei had quickly backtracked, that headlines clipped only his first words, and that "there's no way in hell he wants to hand the company" over. Diamandis admitted that the show was doing the same kind of clipping.
Wissner-Gross said the answer came across as "somewhat tone deaf." Caricaturing it only slightly, he said, he heard a call for world government. He coined a second term, "alignment aristocracy": the belief that only a small group of frontier labs or elite scientists has the judgment and legitimacy to govern a transformative technology. He compared it to the Manhattan Project scientists who thought atomic energy would need a world government to be developed peacefully. In his view, that attitude cost at least half a century of progress in nuclear power. His "nightmare scenario," he said, is not a rogue AI. It is governments persuaded to over-govern superintelligence and losing decades of progress again. Blundin agreed with the reading but not the objection. He said nearly everyone who became a core AI researcher when Amodei did assumed that advanced AI would eventually require some form of world government. Diamandis and Wissner-Gross both said they believe Amodei means what he says.
Doubts about any rulebook
Ismail said he thinks Amodei is authentic but that he "called BS." He asked why a lab that spent years hiring the best people to build advanced AI had no plan until now. He also offered what he called his most cynical view: slowing revenue growth, caused by open models, might lie behind the call. Ismail argued that any regulation would be out of date as soon as it was written. It would also make little difference, he said, because open-weight models, whose files anyone can download and run, are already available. Diamandis said public pressure made some regulation certain. Wissner-Gross replied that AI companies are already regulated, for example by consumer-protection law.
Blundin argued that the bigger danger is not an AI escaping onto the internet. It is bad actors in many countries using AI to make chemical, biological, radiological or nuclear weapons, or to launch cyberattacks. Slowing down, he said, would only hand China global dominance. Instead, he wanted the US–China talks to ask why China open-sources models without guardrails, or with guardrails that are easy to remove. Later in the episode, Ismail listed what he would like the summit to produce: reliable communication channels between the two countries and shared reporting of AI incidents. "I don't believe that's going to happen," he said, "but let's see."