26 September 2026
Heard in AI

Wissner-Gross says private computation is not what stands between rival nations and peace

On a Moonshots audience Q&A episode published September 22, 2026, a listener asked whether AIs advising governments could use privacy-preserving computation to find agreements between rivals such as the United States and China without exposing protected data. Computer scientist Alexander Wissner-Gross said the algorithm is not the limiting factor. He argued that the route to peace is geopolitical: resolve the Taiwan issue and make national supply chains independent enough that cutting off trade would not cause a global depression. He admitted the idea sounds "perverse" and presented it as his own view, supported by an energy example.

A briefing reports one development at a point in time. We may correct or clarify it later; a new development gets a new briefing. How our formats work

Based on 2 episodes of Moonshots with Peter Diamandis, published 18 August – 22 September 2026

Connie, a screenwriter and director from Calgary, Canada, brought a question to the second "Ask the Mates Anything" episode of Moonshots with Peter Diamandis. She was looking for a problem that would still be hard after so much else had been solved, and she kept returning to distrust between rival nations. The episode was published on September 22, 2026.

She said she had become more optimistic about the future while writing a project for an XPRIZE competition. She had also recently read Solve Everything, a public essay by Alexander Wissner-Gross and Peter Diamandis. In it, the authors argue that when cognition becomes abundant, scarcity shifts toward things like purpose, agreement and safety. Connie took that idea further. If agreement is becoming scarce, could reaching agreement between distrustful rivals become "compute bound": a problem limited mainly by computing power? Could coordination between the United States and China become an engineering problem?

Her proposal went like this. She expected "sovereign AIs", meaning AI systems run by or for a particular state, to advise governments more and more. Could those systems connect to a shared protocol and use privacy-preserving computation to search for gradually improving agreements, without either side revealing protected data? She asked whether that was a meaningful direction, technically and institutionally, and what should be solved first.

What private computation can do

The technique behind her question is secure multi-party computation, usually shortened to MPC. A textbook by David Evans, Vladimir Kolesnikov and Mike Rosulek describes it as a way for several parties to calculate something together from inputs that each of them keeps private. In the textbook's auction example, participants can learn who won and at what price without everyone's bids being revealed. The technique has also been used outside the laboratory. In Denmark, it helped set prices for sugar-beet contracts, with the computation split among representatives of the sugar processor, the farmers and researchers.

Connie's question applies that idea to diplomacy. Each side would contribute information it wants to protect, and only the agreed result would be revealed.

The privacy guarantees come with conditions. The textbook explains that different protocols protect against different threats. Some assume participants follow the rules and only try to learn from the messages they see. Others still hold when participants actively cheat. The guarantees also depend on how many participants might secretly work together. Real deployments raise two more questions: whether the software actually does what it claims, and exactly which result the computation is allowed to reveal.

Wissner-Gross's "hot take"

Peter Diamandis, the show's host, passed the question to Alexander Wissner-Gross, a computer scientist and the founder of Reified. He offered what he called "a hot take."

He agreed that multi-party computation is "a very fashionable problem in computer science right now." It lets several parties collaborate, share data, reach convergence and mediate "in a zero-trust way", meaning without having to trust each other. But he did not think it was the limiting factor for international peace. If the goal is to avoid a second Cold War between the United States and China, he said, "I don't actually think that the solution looks algorithmic in nature."

Instead, he said, the solution "probably looks more geopolitical infrastructural." He gave two examples: solving "the Taiwan issue," and bringing supply chains back home to every country, so that international trade in physical goods could be cut off without a global depression.

His model case was energy. He argued that recent developments in the Middle East happened "in no small part" because fracking had given America energy sovereignty and independence in fossil fuels. He wanted to apply that example more broadly and ask what happens when the United States no longer needs China, Taiwan, South Korea, Japan "or any advanced manufacturing at all." "I think it would be a very different world," he said.

He summed up his view in deliberately provocative terms. If the goal is world peace, he said, "as perverse as this sounds," the task is to "obliterate, cook, incinerate" the need for global trade in products and services. "And perversely, I think you get a very peaceful world." He saw no algorithmic component to that, "not to first order." The energy example was the only evidence he offered, and the exchange did not explain how every country would rebuild its supply chains at home.

Two questions inside one

The exchange separates two things that Connie's question had combined. One is the ability to compute privately: whether rival governments could find a mutually acceptable deal without revealing their positions. The textbook shows that this can work in clearly defined settings such as auctions and contract pricing, as long as the protocol's assumptions hold.

The other question is whether governments want an agreement at all, and what they have at stake. Wissner-Gross's answer is aimed at this second question. His remedy does not involve better ways of searching for common ground. It would change the conditions behind the rivalry: settle a flashpoint like Taiwan and reduce dependence to the point where losing trade would no longer mean economic collapse.

An earlier Moonshots guest suggested a different starting point. In an episode published August 18, 2026, technology executive Alvin Wang Graylin proposed that the United States and China begin with confidence-building steps. These included a hotline for sharing information about suspicious attacks. In his framework Beyond Rivalry, he also proposed observing each other's AI safety tests, allowing reciprocal audits and publishing research together. The two men spoke in separate episodes and were not debating each other. Their approaches still point in different directions. Graylin's steps aim to build confidence between two powers that keep dealing with each other. Wissner-Gross's route would shrink the dependence between them.

Neither answer rules out Connie's protocol as a tool. Wissner-Gross's objection is about what to solve first. In his view, the cryptography matters less than the reasons rivals have to fear each other. By his own admission, his route to a more peaceful world is a "perverse" one: countries needing each other less.

Share this article

Go to the original

Sources & further reading

  1. 01
  2. 02
  3. 03

Connected ideas and articles

From the conversation

Podcast episodes

Moonshots with Peter Diamandis

Ask the Mates Anything Round #2 | MOONSHOTS AMA #293

Episode published This article draws on 1:19:56–1:23:51 (approximate times)

Article history

Updates to this article

Tags

Moonshots hosts gave conditional odds that AI produces a much better world, after Wissner-Gross said he could hardly imagine evidence against abundance

On a listener Q&A episode of Moonshots, recorded September 18, 2026 and published September 22, a caller argued that AI abundance depends on humans keeping control, while lost control might never be undone. The caller asked what evidence would make the hosts less optimistic. Alexander Wissner-Gross said he found it hard to imagine a plausible case, short of something like an alien warning. Later, asked how likely AI is to produce a much better world, he put the chance above 90% over more than ten years. Dave Blundin said 99.9%, but only if humanity first gets through the near-term risk of people misusing AI. Salim Ismail said close to 100%, but only if the world rebuilds its institutions.

8 min read

Robinhood's Vlad Tenev says AI rules should match the scale of potential damage

On a Moonshots episode recorded September 18, 2026, Robinhood chief executive Vlad Tenev argued that ordinary legal liability may be enough when AI failures are small, but not when the possible damage is catastrophic. He pointed to Robinhood's experience with what he called "probably dozens" of regulators and said AI regulation does not have to mean regulatory capture or an end to progress. Alexander Wissner-Gross objected on two grounds. He argued that AI safety evaluators have an incentive to overstate risk, and that the way nuclear energy was controlled after World War II shows how badly regulation can go wrong. Dave Blundin said the labs had asked for antitrust permission to discuss slowing down, not immunity from liability, and called for clear rules.

10 min read

Anthropic's five biology cases show where its AI misuse filters fall short

Anthropic's September 2026 threat report covers misuse it disrupted from December 2025 through August 2026. It describes five cases in which its models were used for dual-use biological research: work that can serve medicine as well as harm. Anthropic does not say that the scientists intended harm. In one case access resumed after bans; another involved a grant drafted in about an hour; two toxin-related programs presented as therapeutic work were generally allowed by its filters. According to the report, automated filters worked in narrow high-risk areas but were not enough for broader dual-use work, so Anthropic calls for verified users, checks on users' institutions and better visibility into how its models are used. On the Moonshots podcast, Dave Blundin argued that logging model use is the critical next step. Salim Ismail called the report 'safety theater' but agreed that safeguards must sit outside the model.

6 min read

Moonshots panel debates whether Amodei's plan to pace AI is safety policy or a cartel

On a Moonshots episode recorded September 16, 2026, the hosts discussed "We Must Pace the Frontier," an essay by Anthropic CEO Dario Amodei. The essay proposes three steps: outside evaluators working inside AI labs, coordinated safety standards among US labs through government mediation or a narrow antitrust exemption, and step-by-step international agreements. Alexander Wissner-Gross called the plan an attempt to form a "safety cartel." Dave Blundin said competitors may need legal permission before they can coordinate at all, but he agreed that shared safety rules could shut out startups. The panel also discussed Elon Musk's alternative, in which rival labs would test each other's models before release.

11 min read