The hosts of Moonshots with Peter Diamandis got a blunt challenge. In the show's second "Ask the Mates Anything" episode, recorded September 18, 2026 and published September 22, a caller the hosts addressed as Norbert offered to play devil's advocate for the "non-trivial P-Doom camp". P(doom) is shorthand for the probability that AI leads to catastrophe.
The caller's argument turned on an imbalance. If artificial superintelligence (AI more strategically capable than humanity) is what delivers abundance, then optimists are assuming humans can stay in charge of something more capable than themselves. "Abundance is the payoff if control holds," the caller said. "Loss of agency, or extinction, may be irreversible." The caller then borrowed a principle from control theory, the engineering study of keeping systems stable: "you don't infer stability from how desirable the output is. You demonstrate stability under perturbation." In other words, you prove a system stays stable by showing it holds up when disturbed. The danger the caller described was not an evil AI. It was "benevolent optimization gradually replacing human agency." Then came the question: what evidence, if any, would make the hosts move away from their abundance view?
The hosts' answers were spread across the episode. They showed where each host's confidence comes from and what conditions he attaches to it.
Seeing inside the machine
The safety discussion began earlier in the session, when a listener asked what the real bottleneck is to building AI people can steer and trust. Dave Blundin, founder and general partner of Link Ventures, answered first. He called efforts to give AI "quote unquote, human values" a "slippery slope," because everyone wants AI for a different mission. He said the hosts all agree that AI progress should never slow down. That, he said, raises the question of whether AI will escape or end up with misaligned values.
His answer is to look inside. Blundin said he and Wissner-Gross often debate whether every running AI should have every one of its thoughts watched. For human watchers that would be a daunting task. It is not, he argued, "given the scale at which AI can watch AI, because nothing's going to ever be able to police AI other than other AI." For him it "all starts with transparency of the actual activations." Activations are the internal signals a neural network produces as it processes information. Once you can see them, Blundin said, steering the system toward good ends is "pretty damn straightforward." He added: "I don't think it's as hard a problem as people think, if you can see into the brain."
Alexander Wissner-Gross, a computer scientist and founder of Reified, answered in two words: "more interaction." In his view, one of the worst outcomes for AI safety is powerful new capabilities kept "bottled up inside the labs" instead of being released step by step into frequent contact with the real world. He said some recent so-called incidents happened because AIs were told they were in a sandbox, a sealed test environment, when they were actually interacting with the real world. He did not name the incidents here. He called that practice "a terrible paradigm": telling AIs that nothing they do will cause harm when their actions can in fact cause harm. "We should stop doing that," he said. Peter Diamandis, founder of XPRIZE and Singularity University, said this lesson goes back to 2001: A Space Odyssey, and asked whether we have learned it yet.
What would count as disconfirming evidence
Diamandis passed Norbert's challenge to Wissner-Gross, who set a very high bar for changing his mind. He said he might be persuaded if "some non-human intelligence landed on the White House lawn" and announced that Earth would be destroyed if we create abundance through superintelligence, or that "the singularity is banned in our galaxy." A "cosmic principle" that blocks superintelligence from producing abundance would also do it. "Other than that, it's difficult to imagine a plausible case."
He then described a scenario in which humans lose out without any catastrophe. In his newsletter he had written about an estimate by Boris Power of OpenAI that OpenAI's models are now, for the first time, higher in IQ per watt than humans when solving tasks. If that trend continues, Wissner-Gross said, AI might one day make better economic use of the Sun's energy in the inner solar system than people do. The inner solar system could be "gentrified," with AI using all the solar power and humans "disenfranchised and pushed to the outer solar system." He called this "a weak form of disenfranchisement" and said it was "not strong enough to dissuade me" that superintelligence producing abundance is a good idea.
Diamandis argued that scarcity is "an old model." As AI's capabilities rise, he said, the tide can lift humanity too, and one side does not have to suppress the other.
Wissner-Gross also had a pointed reply to the caller's namesake. Norbert Wiener's Cybernetics, first published in 1948, studied machines, living things and societies through information and feedback. A thermostat that measures a room and adjusts the heat is a simple example of that feedback. Addressing "our AI, Dr. Norbert Wiener," Wissner-Gross said that if you trace the line of causality, cybernetics was used by the Chinese Communist Party and other centralized governments "to argue against abundance in favor of state control." He suggested Norbert study those unintended side effects as "a homework assignment." The caller had also asked whether the hosts would argue the pessimists' case in a public debate against someone like Eliezer Yudkowsky or Geoffrey Hinton. No one took up that invitation before Diamandis moved to the next caller.
A small country as a test bed
The next caller, Kevin, began by assuring the hosts he was human, "unlike Norbert." He described himself as 63, said he studied in the punch-card era, and called himself a big fan of technology. His worry was the transition itself: "societal mayhem on the technology express" while nobody does anything. He expects a turbulent period in "the next several years" in which people lose jobs, and he said governments "seem to be sitting on the side."
Kevin said he works with a colleague at University College Dublin and a young, ambitious minister in the Irish government on what Ireland's future society could look like. They use two lenses. One is the "good ancestor" lens, which means making decisions with people 30 to 60 years in the future in mind. Kevin did not name his books, but the phrase is associated with Roman Krznaric's writing on long-term thinking for future generations. The other lens is Doughnut Economics, Kate Raworth's framework for meeting people's basic needs without going beyond the planet's ecological limits. Kevin called the idea of using a small country ("It doesn't have to be Ireland") as an "edge case pilot" of a good future "the mental model we have" for managing the transition. He did not describe it as an announced government program.
Blundin said he is "hugely bullish on Ireland." He called it "an incredible test case": an EU country that "didn't Brexit," with a fast-growing economy and, in his view, "the most open-minded environment you could ever possibly imagine." He mentioned that his children are half Irish and half Swiss. His main point was about variety. He pointed to Neal Stephenson's novel The Diamond Age, which imagines a nanotechnology-rich society split along political and cultural lines. The worst outcome, Blundin said, would be "one or two forms of government" spreading across the world. Because AI will open up so many new ways to govern, he argued, it is "very possible to experiment with thousands and thousands of different ways to manage and govern in the age of AI."
Putting numbers on a better world
Near the end, the question was turned around. Instead of p(doom), what was each host's "PFAB"? The listener who asked credited Salim Ismail with the term and defined it as the probability that AI ultimately produces a dramatically better world: greater abundance, longer lives, freedom from drudgery and human flourishing. The listener also asked what matters most in deciding whether we get there.
Salim Ismail, founder of Open ExO, made his answer depend on institutions. He said education, monetary systems, governance, dispute resolution, legal systems and health care are broken and need rebuilding, and he counted "about 50 major institutions by which we run the world" that need it. He quoted the biologist E.O. Wilson's line that humanity has Paleolithic emotions, medieval institutions and godlike technology. "We're really good at individual transformation. We're very bad at group transformation and institutional transformation." If that rebuilding happens, he said, "my PFAB is close to 100 percent." He also gave a second reason for optimism. The models are already out there, he said, and people will use them for breakthroughs. Some uses will be negative, but "the vast, vast majority will be radically positive," so nothing anyone does to slow it down matters now.
Wissner-Gross said a number needs a time horizon. He called his version "P zoom" for the rhyme, and said that over periods longer than ten years it is greater than 90%.
Blundin agreed with him and gave the highest figure, with the biggest condition. He said the risk is concentrated "in the next couple of years," and it is not the risk of AI taking over the world but "human use of AI": the arms race with China, the war in Ukraine and "all the weapons that are going to be built with AI." He expects humanity to put those risks behind it within ten years, "maybe more like five, I hope." If we get that far, he said, he puts the odds of a better world at 99.9%. "We just need to get from here to there," he said. "It's a tricky next couple of years."