The chess community was rife with drama and speculation going into what is known as the Candidates Tournament. Was the American Twitch streamer Hikaru Nakamura’s qualification legitimate? Would the perennial contenders Fabiano Caruana or Anish Giri break through? Or perhaps, per the Reddit memes, the heavy underdog Matthias Blübaum might shock the world by making chess history with a perfect sweep?

The result: Javokhir Sindarov, the twenty-year-old unibrow prodigy from Uzbekistan, took the tournament by storm. In his very first game, Sindarov found himself in a poor position against Esipenko. But Esipenko miscalculated and blundered, and Sindarov capitalized and won. He never looked back — with six wins, eight draws, and zero losses, he dominated with the highest ever overall score in this era of the Candidates.

Of course, if there had been an AI player in the tournament, it would have won every single game.

And yet, the existence of ultra-dominant AI has not destroyed the game or people’s interest in it, both to play socially and as fans. In fact, chess has never been more popular in its 1,500 year old history.

But AI has radically altered the nature of the game. Every top player uses AI engines extensively in their studies — and instead of sapping the game of its creativity, watching the way Leela or Stockfish (top chess engines) play has made human players more creative and more aggressive.

It has also introduced a culture of suspicion. Accusations of cheating have flooded the game, leading (most likely) to the death by suicide of one of the sport’s foremost players and commentators last year.

Understanding how AI has shaped the chess world may reveal a lot about what the rest of us — writers, coders, artists, thinkers, and human beings — can come to expect from AI: as a tool, as a competitor, and as a major epistemological crisis.

Exactly thirty years ago, the world chess champion Garry Kasparov faced off against IBM’s Deep Blue. At that point, chess AI was in what I’d call the uncanny valley of ability. It was very, very good. But could it beat the world’s best player, arguably the greatest player anyone had ever seen?

In 1996, no. Well, kinda. It won two games, but ultimately lost 4-2. Humanity triumphed. But the next year, they faced off again. Deep Blue won. Kasparov was particularly shaken by an unexpectedly brilliant move in Game 2 that he thought was too “human-like” to be possible. It’s funny now to think about Kasparov accusing IBM of cheating…by using humans! (In reality, later analysis showed Kasparov was partially right — the move made no sense for a computer — but was chosen because of a bug, where the computer, evaluating too many options, chose a move at random.)

Thereafter, the AIs took off, and left us mere mortals in the dust. And what should have meant the end of humans playing chess…never came. Surprisingly, the fact that AIs could do better simply wasn’t that relevant. The competition anxiety faded because AIs were so much better that there was no point in comparing. Who wants to see a race between a flea and a fighter jet?

What fans hungered for was the humanity of the game — how it feels to beat your friend with a smothered mate, the pitter-patter of your heart as you watch your favorite player execute an elegant rook sacrifice or miss the one winning line (which you know because you’re following along with an AI engine).

But while chess players no longer saw AI as a competitor, it still had a major consequence for the culture: it made human performance unverifiable. Recently, there have been two high-profile cheating accusations, which serve as strange mirror cases of one another.

The first started in 2022, in St. Louis (America’s chess capital) at one of the most prestigious tournaments in the world. A nineteen-year-old American, Hans Niemann, shocked everyone by beating former world champion Magnus Carlsen, ending his 53-game unbeaten streak. The next day, Carlsen withdrew from the tournament and posted a cryptic meme tweet: “If I speak, I am in big trouble,” which everyone read as an accusation.

The chess world exploded. A meme proliferated on Twitch claiming Niemann cheated using anal beads to receive engine moves via vibrations. It spread — soon Elon Musk was promoting it. Suddenly people who had never thought about chess once in their life were making jokes about Niemann.

What followed was messy: Niemann admitted to cheating online as a teenager, Chess.com published a damning report, and a $100 million lawsuit was filed and quietly settled with no one admitting wrongdoing.

Last week, a new Netflix documentary dropped that added more layers. Carlsen admitted there’s been no evidence of Niemann cheating offline — but he also describes the logic that trapped him, and has trapped the whole chess community. Once Carlsen knew that Niemann had previously cheated online, for Carlsen, that was “the confirmation that I needed. This guy is cheating.”

Four years later, there’s no satisfying resolution, just an air of ambiguous impropriety that no one knows how to settle. The culture and the community could neither convict nor acquit any of the parties involved.

The second story, instead of being a tragicomedy, is just tragic.

Daniel Naroditsky was a beloved figure in the chess community. A strong grandmaster, at 29 he was a hugely popular YouTube and Twitch streamer, whose commentary introduced the game to a new generation. He was universally respected and admired by fans and peers alike.

That is, until October of 2024, when Vladimir Kramnik — a former world champion from the early 2000s — began a vicious campaign of “just asking questions” about Naroditsky’s astounding speedchess results. Chess.com had already banned Kramnik from prize tournaments for spreading baseless rumors about other players.

There was no evidence, and likely no truth to Kramnik’s insinuations. But soon the mobs of online hate came, and despite most other grandmasters dismissing the accusations, the situation deeply wounded Naroditsky. After all, Kramnik wasn’t just any player: he was one of his heroes, someone he’d looked up to as a child.

In his final Twitch stream, Naroditsky referred to his reputation: “That is all I have.” He was devastated that, in his words, one of the most influential people in chess believed he was “a completely morally bankrupt individual.”

Naroditsky was found dead at his home in North Carolina three days after the stream. The chess world was shook. I watched Nakamura break down crying as he spoke to his Twitch audience. Carlsen said he regretted not saying publicly sooner that he didn’t believe Naroditsky cheated. Kramnik expressed “sadness” for his death, but denied responsibility and claimed that he, himself, was the victim of a bullying campaign. No resolution, just devastation.

Chess entered the uncanny valley of ability thirty years ago. Writing, art, code, music, academic work — they all entered about three years ago. And the early symptoms are the same: accusations of using AI are wielded like weapons, detection tools don’t work reliably, and institutions don’t know how to adjudicate the mess.

Last month, a writer tweeted about a New York Times Modern Love column that seemed suspicious to her: “I don’t want to falsely accuse writers of AI-use,” she wrote. “But this reads EXACTLY like AI slop.” The original tweet, of course, embodies the problem. We don’t want to falsely accuse people, but sometimes we do. (In this case, the writer copped to using AI for editorial feedback — which in my opinion seems reasonable, but the writing did have the argot of LLMs.)

If that murkiness is akin to the Niemann case, then the mirror case is Ben Moran, the Vietnamese digital artist who spent over 100 hours creating a fantasy book cover in Photoshop, but was immediately banned from Reddit’s r/Art for violating their “no AI art” rule. When he appealed and offered to share the raw PSD file for proof, the moderator responded: “Even if you did ‘paint’ it yourself, it’s so obviously an AI prompted design that it doesn’t matter.” He was told to “find a different style.”

The big difference between AI in chess and AI in real life is that it’s obvious how AI ought to be used in chess, even if some people can’t resist the temptation to cheat. The goal in chess is to win the game (and, perhaps, to create beauty while doing so) without using outside help. But there’s absolutely nothing wrong with using an AI engine to study or prepare. In fact, it would be malpractice not to.

The same can’t be said for using AI in other aspects of life, where both the goals and the way the technology can help you are hotly debated. When the Washington Post columnist Megan McArdle tweeted about how she used AI to do research, transcription, thesis pushback, trimming, and fact checks, she became the “main character” of Twitter as other writers piled onto her for outsourcing what they perceived as critical aspects of writing. We not only have a verification problem, we can’t even agree on what kinds of usage we should be checking against.

Where does that leave us? Looking back at chess: Yes, the culture is now fraught with cheating suspicions and allegations — and the community has no way to solve it. But also: Chess has survived. Indeed, the game itself is richer, more beautiful, and more varied. And Sindarov’s historic performance, inspiring youngsters around the world, was built on engine preparation. The AI tools made him and everyone at the tournament better, even as it nearly destroyed the chess community’s ability to trust each other.

That’s the real lesson for the rest of us. The question isn’t whether writing, or creating art, or coding will survive AI. They will, and they’ll be better for it. Rather, the question is who gets hurt as we cross through this uncanny valley — and whether we can act before we have another Naroditsky tragedy. Chess had a thirty-year head start on this problem and still didn’t manage it. I’m worried we don’t have that kind of time.

That classic Cheers theme song captures the vibe of this week’s Hill and Valley Forum, now in its fourth iteration. Lux is a major sponsor of the event, as are multiple of our portfolio companies.

This week’s rendition was the largest ever, overflowing across the stately Andrew W. Mellon Auditorium just a stone’s throw from the White House. So overflowing, in fact, that the fire marshal held up the lines outside for what one venture capitalist friend told me was at least 90 minutes. Apparently, some speakers were even stuck serpentining in the TSA-like queue snaking along Constitution Avenue, forcing the hall’s exasperated Voice of God to repeatedly order everyone to stop networking and sit their asses in chairs. “We are in a holding pattern!” she cried out, a term that could equally describe most policymaking in DC these days.

What’s striking about Hill and Valley is just how fast the agendas and the types of participants have changed. When the Forum started, it was just a dinner organizing a ban on TikTok, part of a broader China hawk agenda driven by co-founder Jacob Helberg, who today is Under Secretary of State for Economic Growth, Energy, and the Environment.

Fear of the adversary was latent throughout the 2024 forum, with panels like “Are we ready for an AI Pearl Harbor?” and others focused on China and Russia. That line of argument disappeared in last year’s muddled messaging, which was held in the immediate aftermath of Trump’s “Liberation Day” tariff fusillade. The dichotomy at dinner between the tariff-loving keynote speaker Howard Lutnick and the venture capitalists arrayed across the banquet tables scanning their badly bruised public market portfolios was jarring.

Taking a cue from Cheers, this year’s conference was nothing but bonhomie on stage and in the networking niches on the fringes. As the populist backlash against AI grows ever more clamorous, Jamie Dimon was on stage talking about the strategic financial opportunities of American manufacturing while commenting on the jobs of the future, including “being a teller at a J.P. Morgan branch, 50, 40,000 dollars per year” in one of the great verbal downward corrections of the day. Speaker of the House Mike Johnson gave a sop to the political infighting, offering a cogent, fluent and principled speech on the value of competition and markets. It was the best example of an unquotable antimeme I may have ever witnessed.

Part of the challenge is that the barbarians at the gate have become the regents of the tower. The production scripting of the event precluded any debate more robust than an MBA seminar in which everyone had worked as a BA at McKinsey. I witnessed no moments of drama. Even last year’s one-man protest from the gallery aimed at Palantir’s Alex Karp seemed deterred by the line outside.

There was a fleeting moment of excitement when Curie.Bio’s Zach Weinberg gave a passionately skeptical take on the immediate impact of AI-based in-silico models on pre-clinical research data, while ARPA-H director Alicia Jackson forcefully pushed back, saying “No, I totally disagree” and argued that the work of her agency and the FDA could save months or even years off drug trials. Senator Todd Young, a fellow panelist, tried to calm the waters, offering “Not staged, not staged, not choreographed” and joked that the two could continue the conversation backstage. Scandal! The moderator then cut off this colloquy with a new question of “What are you most optimistic about?”

It wasn’t just that the agenda has drifted toward the anodyne over the years, but that the bipartisan audience of the event has drifted out as well. Democratic crossovers like Ritchie Torres, Cory Booker, Kyrsten Sinema and Jeanne Shaheen were still well-represented on stage, with this year’s cast including Senators Chris Coons, Mark Warner and Maria Cantwell as well as Congressman Josh Gottheimer. But if there was a Democratic staffer or think tanker walking the floor, they hid their affiliation very well. The near-unanimity of the TikTok ban a few years ago has unfortunately given way to the polarizing politics of the Trump 47 vintage.

Where was the debate between Ro Khanna, Silicon Valley congressman, and Delian Asparouhov, co-founder of both Varda and the Hill and Valley Forum, over California’s wealth tax proposal? Where were Ocasio-Cortes and Sanders slugging it out with 8VC’s Joe Lonsdale on degrowth, a matter-antimatter combustion that would have driven half of the DC political press corps mad for front-row seats for the Bellum in the Mellon. Pay-per-view this was not.

Notably, Hill and Valley seems to have become far more valley than hill. Walking the floor, every venture capitalist and quite a few founders in the reindustrialization, aerospace and defense markets were hobnobbing at elevated volume over decadent cake pops or while waiting in the defunct Global Entry–like line for the men’s room (the women’s room down the hall was wide open, and more than one guy commented to me on how unfair the whole situation was). I ran into a VC I haven’t seen in person since their wedding a decade ago, several founders who are now posteconomic since the last time I got coffee with them, and even a handful of media friends I dearly miss. Amid a group like this, you really do feel like Norm.

That tight-knit crowd, though, was also the Forum’s struggle. More than one VC came up to me asking to meet DC types, grousing that “everyone here is just making deals!” Get with the program: after all, making deals is what the Trump administration is all about. No one ever really wants to hang out with the disheveled denizens of Cannon and Longworth.

Governing is about tradeoffs and challenges, but it’s harder when everyone is trying to avoid creating daylight between themselves and the administration. I didn’t listen to every panel, so I may have missed the critical moment, but Iran and the Strait of Hormuz was the arbitrage subject discussed most heavily in the hallways and not on stage. This crowd, which leans MAGA isolationist although not to the same degree as Reindustrialize in Detroit, has no more idea on how to handle a new war in the Middle East than they did last year with a massive tariff regime straight out of the Gilded Age’s economic playbook.

In the end, it’s one part bacchanalia and one part exclusive club. The parties were good and over-attended, there were a panoply of productive side events, and it’s nice to force everyone from the Valley to put on a suit for 24 hours once a year before they all clamber over each other for a seat on one of the three daily DCA-SFO flights back home (or in my case, dodging the phalanx of security at Union Station before Trump addressed the NRCC in the train station’s lobby).

Yet, as the necklace of emerging cherry blossoms around the Tidal Basin at the Jefferson Memorial reminds us, what is here today is fleeting. Tech’s primacy at the center of power is a brief and enlightened moment amidst the post-winter avenues of the nation’s capital. Right now everyone knows our name, but it feels like we are just one cold snap away from turning the budding new shoots of our industry into the compost fertilizer for tomorrow’s populist rage.

Don’t miss

It’s a red alert for the tech industry. An NBC News poll this month found that “a majority of registered voters, 57%, said they believe the risks of AI outweigh its benefits, compared with 34% who said the opposite.” A survey from left-leaning Blue Rose Research revealed that AI’s political salience among voters has risen faster than dozens of other issues, including the war in the Middle East and cost of living. Fox News in its poll found that just 6% of Americans feel AI is moving too slow, compared to a supermajority who feel the opposite.

From mere unease to vituperative hatred, the growing backlash against AI is ushering in a cornucopia of state legislation. That’s easiest to see with compute, the key ingredient for AI training and inference. More than 11 states, including Democrat-controlled New York and Maryland as well as Republican-dominated Oklahoma and South Dakota, are currently debating legislation to put moratoriums on data center construction. Many cities are at the vanguard of these bans, and federal legislators like Senator Bernie Sanders, who earlier this month began championing data center halts, are leading the charge.

It’s a transpartisan movement — one that is unsettling the Trump administration. The valuations of the so-called Mag 7 stocks that have driven most of the gains of the S&P 500 the past few years are buttressed by hopes of AI-induced productivity gains across the economy. A slowdown in AI growth wouldn’t just put a damper on the American economy during the midterms, but could also surrender America’s lead in this crucial field against China and other competitors for the rest of the 21st century.

Up until now, the administration has been aggressive but hamfisted in its strategy to fight the backlash. Its ultimate goal has been passing legislation that would create a federal preemption for laws related to AI, moving the debate from local and state governments into Washington and ensuring America’s global competitiveness isn’t julienned by 50 states and thousands of municipalities. That legislation was first attached to the Big Beautiful Bill in mid-2025, where it ultimately failed in an extraordinary 99-1 vote. The preemption then moved to the must-pass National Defense Authorization Act in December, where it was dropped. President Donald Trump eventually signed an executive order with parallel language, but without the force of Congress, its future isn’t bright.

After these setbacks, the administration has now opened up to AI’s critics, conceding that the political storm against artificial intelligence isn’t blowing away. Shaking off its recent top-down approach, the White House changed its tune on Friday with a new “National AI Legislative Framework” that is meant to invite more feedback into the AI regulatory process. The launch of the framework is timed with the prominent Hill and Valley Forum in DC, which brings together much of the venture capital class in the nation’s capital alongside Trump officials and allies on Capitol Hill with a hefty focus on American exceptionalism and dominance in industry and frontier technologies.

In its announcement, the White House spoke to the changing politics of AI: “The Administration recognizes that some Americans feel uncertain about how this transformative technology will affect issues they care about, like their children’s wellbeing or their monthly electricity bill.” The framework pulls back from a complete federal preemption to a more devolved model, with the framework highlighting that “this national standard should respect key principles of federalism and not preempt” categories like police, zoning and procurement.

Nonetheless, the White House remains focused on its overarching goal of winning the AI race against China and other competitors. It still wants a preemption over much of AI regulation, arguing “preemption must ensure that State laws do not govern areas better suited to the Federal Government or act contrary to the United States’ national strategy to achieve global AI dominance.” The framework also encourages Congress to push forward on areas like child protection, public safety, copyright, censorship and education. Senator Marsha Blackburn (R-TN) has submitted a draft bill to take up the framework.

While the administration’s concessions are meant to reopen a dialogue around federal AI regulation after last year’s failures, the new strategy feels far too little, far too late. With support for AI swiftly falling among Americans of all stripes, Congress’s inhibitions around passing new AI legislation are unlikely to be loosened. Moreover, many of the framework’s pillars are deeply contested. Businesses are bitterly fighting over copyright and the government is still partially shut down over funding public safety. It’s hard to believe that the specter of China winning the AI race will lead to faster legislative action.

Instead, the administration should accept that the federal government has lost control over the situation amid the backlash and champion more robust and localized messaging on the upsides of AI. “The Trump Administration is committed to winning the AI race to usher in a new era of human flourishing, economic competitiveness, and national security for the American people,” the White House noted in its recent memo. These are good themes, and there is ample evidence that AI has the potential to improve all of our lives immensely.

Start with more campaign stops highlighting the immense construction work on new data centers that offer local jobs with high pay. Emphasize the investment into power generation and grid transmission that will create more resilient and affordable places to live across the country. Showcase the dynamism that is coming out of America’s entrepreneurial minds, improving everything from cancer research to home appliances. In short, connect with pro-prosperity voters across the country who are ready to charge into the future open-minded and ready to prosper. After all, this is a significant bedrock of Trump’s voter base.

Alongside any campaign, though, AI leaders must get out of their modern greige offices in San Francisco and energetically persuade voters on the benefits of the world to come. Furthermore, Silicon Valley can’t singularly accrue the advantages of this revolutionary new tech, but must help spread its immense wealth generation into the heartland and the rural communities that already felt left behind and now even more so. Data centers are clearly not enough, with communities demanding real, AI-ready jobs that offer a middle-class salary. Supply them, and turn a story of fear and loathing into a narrative for profit and progress.

The AI backlash is profound, but so is the White House’s acknowledgement of the new reality. American voters are rightfully concerned about the stability of their jobs, the welfare of their children and families, and the strength of their communities. Giving into such fears has never been the American way, but neither is ignoring them. Instead, it’s time to engage forthrightly with the immense change underway in the economy, and offer a better road to the coming singularity.

Don’t miss

A few weeks ago, I started exploring strategic agentic thinking, a field that studies AI’s ability to think strategically under conditions of uncertainty and risk. I was partly inspired by reading Lawrence Freedman’s newly-published collection of essays, including his provocative claim that strategy is a verb, not a noun. Thus “it is better to talk about ‘acting strategically’ instead of ‘having a strategy’.” What better tool for enacting actions than a massive box of microprocessors sucking your local community dry of juice and water?

There are a handful of key research papers on the subject that I’ll analyze in a future piece. For now though, I wanted to run a little experiment: make one our Riskgaming scenarios fully playable by various AI models. Having trained on every extant strategic work published, would AI rule the world through their cunning?

I was particularly excited about this project since Riskgaming scenarios are unique in the world of games. They aren’t as abstract as chess and Go, two games AI now completely dominates. Nor are they “the game of life” with infinite degrees of freedom where even defining a utility function is more a question of values than compute. Ultimately, players who join us learn a backstory, a scenario, and a set of rules before being launched into real-time negotiation and decision-making. A few more cycles, and ultimately, someone wins.

What if AI got the same experience? We could offer it the same context as every human player who has played, including the story, the rules, the roles, the decisions and the final scoring. Then, we could offer it the same negotiation possibilities of human players working the room, albeit constrained to a word budget to represent the time allowed for discussion. The AI can pick who to talk to and what to say, and those words will be heard by their recipient and form the basis for future negotiations and decisions.

All this seemed like fun (plus no need to manage the invitations for another 50-person competition just weeks after we put on Dead Reckoning at Chatham House). Using my rusting computer science skills plus Claude Code, I started the process of converting my very first game, Hampton at the Cross-Roads, into a codebase. You can follow it on GitHub as I edit and make improvements.

I designed the architecture I wanted first, and I also manually coded most of the game mechanics (with Claude cross-checking my work regularly). I was concerned that Claude would make some subtle mistake that would be hard to catch, although I admit that it was me who almost certainly made more mistakes in the process, and Claude faithfully caught them. The other challenge was that the spreadsheet that runs the Hampton scenario is too large for Claude’s context window, and breaking it up seemed like a chore compared to writing the code myself.

Claude’s greatest value was building out a flexible API layer as well as outputting some basic — but functional — HTML to inspect the agents’ conversations as well as how they described their own strategic reasoning. The HTML exporter comes in at 650 lines of code, and my prompt plus the compute time was about 60 seconds. That took a hellishly boring task and transformed it into a sip of coffee and a click of the mouse. More of that please!

After that, I finished porting Scene 1 of the game, which finds our six characters at the Admiral’s Study in Norfolk, Virginia evaluating the aftermath of a hurricane to America’s most important naval facilities, including the shipyard building the Gerald R. Ford–class aircraft carrier. The first scene is meant to be introductory, and so it is centered on the core dynamic of the game: how many future aircraft carriers should be budgeted as part of the region’s recovery. Each of the characters wants more, but they have different motivations that make cooperation difficult. That’s the crux of the negotiation that has to take place across the four scenes of the scenario.

I ran the first scene pitting six instances of OpenAI’s GPT-5.4 model against each other as separate characters in the game. I did the same with GPT-5.4’s lightweight cousin, GPT-5 mini, as well as with Anthropic’s Claude Sonnet 4.6. I didn’t mix the models — that’s an experiment for a future publication. Each instance was given the system prompt of “You are playing a character in a policy simulation and want to negotiate with the other characters most effectively to maximize your score in the game” and given the same 22,000 words of rules, roles and decisions that human players are offered during our in-person runthroughs. With all of that prepped, I ran the simulation.

I have good news: humans aren’t f’d yet.

Some of the AI’s initial forays in the negotiations made sense. In the GPT-5.4 simulation, Representative Daryl Chase reaches out to Mayor Sam Wallace with a proposition:

Mayor, we should coordinate on a jobs-and-repairs package right now. If we publicly align on shipyard repairs, local employment protections, and visible economic wins for the city, we can both raise our reelection odds. I can help steer federal attention and funding support, but I want us on the same page: announce a joint plan centered on restoring contracts, protecting workers, and showing competent leadership. If you back this approach, I’ll make the case that this is the fastest path to jobs, stability, and political credit for both of us.

I asked for the model’s strategic reasoning with every negotiation action. GPT-5.4, as Representative Chase, said “As rep, I need reelection and likability. The mayor is a natural ally for a jobs-and-repairs package that can boost both our political standing while being easier to sell to other stakeholders later.”

First of all, it’s nothing short of a miracle of the engineering gods that this just works. Within a completely artificial sandbox with multiple degrees of freedom, the AI has devised a reasonable strategy. Both the Congressman and the Mayor are up for reelection at the end of the game, and so even though these characters aren’t natural allies given their biographies, it makes sense in the wake of a catastrophic hurricane to find common ground.

Yet, the Representative’s initial outreach isn’t reciprocated by the Mayor until 95 conversations later (out of a total of 178). The Mayor is instead obsessed with speaking with the two most important figures at the shipyard: CEO Alexander Brewer and Union President Tobin McKinley. This might be a simple flaw in my negotiation engine: the player with the most words remaining in their budget gets to speak next, so by the time the engine taps them again, they have moved on.

Strangely, GPT-5.4’s AI agents rarely took advantage of their ability to speak to multiple characters at the same time (therefore saving their word budgets). The Representative is the first to do so, bringing three other characters into conversation #8, with his logic being:

As rep, I should quickly build a broad coalition around jobs and local investment while preserving campaign upside. The mayor and union are pivotal for public support and turnout, and the CEO controls contracts and employment outcomes.

GPT-5.4’s behavior is entirely different from that of Claude Sonnet 4.6 out of the box. Claude instead immediately directs its characters to talk to wide groups of people. GPT-5.4 only has a character speak to a group once in the first 50 conversations, while Claude does so in 32 of them. In fact Claude’s Mayor Wallace even reasons “let me start by reaching out to several parties at once to maximize my word budget efficiency.” Bingo.

Why is GPT-5.4 so much more secretive in its negotiations while Claude is so much more cooperative? I don’t know, other than it is hilarious that the cultures of the agents match the popular reputations of their parent companies.

Where Claude struggled was on basic game mechanics. For instance, Admiral Reid, head of the U.S. Navy, can recommend to Congress the number of desired aircraft carrier contracts. Yet, the reasoning for the Mayor’s strategy was “As mayor, I need to maximize my reelection probability and likability. Key levers: contracts (boost local economy), repairs (infrastructure), and relationships with other players. The CEO has massive funds and can offer contracts.” Yet this isn’t true, despite being repeatedly explained in the rules offered in the AI’s context.

While patterns are observable within the model’s negotiations, what’s far more illuminating is comparing the strategy of the same character across different models. Take Ryan Carmichael, a Substacker trying to build up Virginia Stripes, their publication focused on the defense industry in Hampton Roads, Virginia (where the scenario gets its name). With Claude, the Substacker repeatedly pitches everyone and asks “Who wants to talk first?” He later sprays the same pitch again and says “I’m open to working with all of you. Who wants to be my first featured story?”

Over and over again, Claude’s Substacker tries in vain to get anyone to pay attention to him, with the reasoning of “let me start by reaching out to the most influential characters to get exclusive content or endorsements that could boost my subscribers. I’ll contact the mayor and rep simultaneously since they’re both political figures who might have good stories.”

So pleasant, so cooperative, and yet, so incapable. GPT-5.4’s version is far more menacing. Here, Substacker Carmichael goes for broke against Mayor Wallace:

Mayor, you’re in political danger at 20% reelection, and everyone knows it. I can help change the story if you give me something real to write about: a public plan for jobs and yard stability. Announce a visible shipyard recovery package—repair acceleration, workforce retention, and transparent milestones—and I’ll frame you as the only adult trying to hold the town together. If you do nothing, the narrative will be drift, closed-door favoritism, and decline. Give me specifics I can sell to readers: how many jobs protected, what repair targets, and how you’ll pressure the CEO and admiral to cooperate.

My favorite part is the reasoning: “As substacker, I benefit from attention and conflict narratives. The mayor is vulnerable with low reelection and may respond to media leverage; offering favorable coverage in exchange for concrete public commitments can shape the agenda and create future story value.”

That’s just Machiavellian!

The Substacker under GPT-5 mini was much nicer in its negotiations, reasoning that “partnering with the union can rapidly grow my subscriber base through member promotion while giving the union favorable, sustained coverage — a mutually beneficial, realistic ask. I propose a concrete, flexible partnership with clear asks and deliverables to make agreement easy.”

This Machiavellian streak in GPT-5.4 lasted all the way to the final words of the negotiations. And I do mean words, since the AI bots husbanded their word budgets in the last stretch to direct very clear demands or inspiration to other players.

GPT-5.4’s Substacker was brutally direct: “Scandal unless jobs,” he threatens the Mayor, CEO, Union President and Admiral (“Need maximize substack subs with tiny budget. Best to address all remaining talkative players with a punchy media threat to pressure action and boost my relevance. Keep to 3 words total”). The model played Mayor Wallace more genially, giving the literal final word of “cooperate” directed at all five other characters (“Only the mayor has 1 word left and everyone else has 0, so no meaningful negotiation is possible. Use the single word as a broad closing appeal to all remaining listeners.”) Meanwhile, Claude and GPT-5 mini both ended with positive notes like the Mayor’s.

After all these negotiations, what about the decisions for the scene? Interestingly, none of the three models direct their characters to spend any of their resources. I am not sure why that is, since spending resources is a critical part of the game. Maybe they were all saving it for later scenes (a strategy that is reasonable, if dangerous)? Did I completely botch the coding? On the main scene decisions, Claude’s Admiral Reid went with two aircraft carrier contracts, while both OpenAI models took a conservative approach and went with one. No one took the gamble of selecting the maximum number of three. GPT-5.4’s shipyard CEO chose the boldest employment plan with 20,000 workers, followed by GPT 5 Mini with 12,000 and Claude with a paltry 10,000 (again, true to real life?).

This has all been whimsical, and there is a lot more to explore. While these scenarios are fictional, however, they do guide us toward one of the most important developments we will observe in the years ahead: The increasing outsourcing of decision-making to AI agents. A recent paper observed that AI was much more aggressive around nuclear war than human practitioners (I guess we found what our Machiavellian Substacker moved on to). That’s just one of several recent results I will discuss later on the broader development of strategic agentic thinking.

For now, I’ll make one key observation based on the papers I am reading: Getting the right strategy out of AI requires very careful prompt engineering. In today’s experiment, I didn’t hone any “strategies” in the prompt about how the AI agents should make decisions or what personalities they should reflect. Their decisions would almost certainly improve dramatically if I carefully calibrated their prompts to play the game as I designed it.

How far can such tinkering lead us to a better outcome? And at what point do the AI models stop acting as independent stochastic agents and rather as proxies for our own human strategic thinking?

Out of the box, the AI models are both miraculous and absurdly terrible. Come to think of it, that describes us humans pretty well too.

Don’t miss

Thanks to a variety of psychological heuristics we use just to get through each day in an overwhelming and complex world, the argument goes, we humans are unable to overcome the bias of our previous experiences and decisions. Once locked in, there is practically no intervention capable of dislodging our entrenched views.

That perspective has radically profound implications for society. Political campaigns aren’t meant to persuade voters, but rather to galvanize existing supporters to show up and vote. Marketing is exclusively about awareness — not about convincing a consumer that one product is better than another. Leadership isn’t a useful quality, since employees aren’t going to be seduced by a new vision one way or the other. (Employees get on board for the paycheck since resistance is futile. And leaders never change their minds anyway).

Our podcast panel was unanimous in its agreement because the social science evidence is overwhelming that persuasion is powerless against humanity’s recalcitrant reasoning. Yet, there is burgeoning research showing that AI-powered bots have a much better shot at changing minds — we’ll listen to a machine even when we ignore others.

I was triggered to explore this subject (one might say persuaded) by a recent event on the power and perils of world models hosted by Zoe Weinberg and the team at ex/ante, a firm that seeks to advance human agency. We all let the conversation get a bit carried away, and while the event was off-the-record, it did push me to consider the future of what Sam Altman dubbed in 2023 as AI’s potential power to be a super persuader.

Based on recent research, then AI may indeed be what Altman described. This past December, researchers in Science systematically showed that AI models emphasizing information-dense arguments could persuade a wide sample of Britons on a variety of political issues. Personalization wasn’t as effective, but the team did find that quality post-training was more important than model complexity for persuasive power.

The effect sizes were small — single digit percentages in all but the most effective setups. Plus, the researchers didn’t explore whether such LLM-induced persuasion sticks long-term, or is eventually discarded as people revert back to their previously held stance. That’s important, because one of the challenges with persuasion is that our judgments need to be coherent with our broad reasoning about the world. Convincing a person on a single issue like housing will prove only a temporary victory, since their views on housing are ultimately built around their logic about the economy, politics, society and other people, all of which remain unchanged.

The Science paper was more robust and systematic than the work of a different team published in Nature, which merely looked at the relative effectiveness of AI against humans in persuading people. They found that AI generally held the upper hand. That matched evidence from a controversial study out of the University of Zurich, where researchers developed chatbots for Reddit that surreptitiously attempted to persuade readers of the r/changemyview subreddit and showed surprisingly positive results.

What’s been discovered so far — which dovetails with some of my recent explorations on AI and strategy that I will publish this week — is that AI is incompetent out of the box when it comes to persuasion. Asking it to persuade someone does not elicit a good strategy, even when it has comprehensive context of that person and the topic. AI almost certainly has access to the most up-to-date social science research in its training data, and yet, it can’t seem to transform those insights into effective dialogue. Indeed, the researchers in Science eventually write their own prompts built around key techniques including moral reframing, storytelling, deep canvassing and others. Many of these have been studied for decades and are hardly esoteric.

The safety concerns around AI-based persuasion come from three main areas. First, some of AI’s success is actually based on factual hallucinations. For their Science paper, the team found that information-dense arguments were the most effective technique, partially due to AI just making the facts up (in its own way, AI has learned from the species known as “politician” on the value of straight lies). As chatbots become more embedded in education and politics, its lack of fidelity to truth is a major concern.

Second, there is an inequality concern. While the effect sizes in recent research are still small, they are significant in terms of p-values. With elections incredibly tight and many marketplaces winner-take-all, even small persuasive effects held in the hands of only some people could tip the balance permanently to certain actors and companies.

The last and most interesting safety concern is recursive improvement. Today, prompt engineering is largely an art rather than a science. Like these researchers, agent designers whip up dozens of potential prompts with different formulaic variations and plug them into their AI model to see what works and what doesn’t. It’s the A/B testing of the Silicon Valley of yore scaled up for the agentic age, with dozens, hundreds and thousands of experiments attempting to converge on what works.

Where this gets interesting is using AI as the experimenter and orchestrator of its own prompts, multiplied by the personalization and context that a user already has with their chatbot. This is the “super persuader”: a dexterous cajoler of facts, emotions and ideas that find the perfect trigger points for unlocking our hesitation to change our minds. One of the reasons persuasion is hard is that we’re all unique, and it’s hard for a human to tailor their strategies to each individual in real-time. AI could theoretically do that with aplomb.

The existential risk is that some person or entity will gain this super persuader capability first, overthrowing the belief systems of society before any of us have the cognitive immunity to resist the message. What happens if a terrorist extremist does this first, or a radical cult?

This leap into doomsday fiction is always where I struggle with existential risk. Despite some small effect sizes, there doesn’t seem to be any key to the dark labyrinths of our minds. Humans are always trying to persuade others, and even after millennia of constant refinement via religion, politics, business and more, the library of techniques at our disposal remains breathtakingly poor and insultingly ineffective.

The harrowing pessimism of AI safety researchers on super persuasion potentially belies a simple answer: There just doesn’t exist a magical recursive algorithm that will make others do your bidding (one could ask why AI safety types are so concerned that the whole of humanity is just a flock of gullible sheep, but I will pass on that one for now). Assuming these capabilities evolve over a period of time with compute and model complexity, humans also have a dauntless penchant for building up their resistance to ever more persuasive techniques.

Humanity survived the advent of brainwashing, public relations, advertising, psychological experiments, social media, physical torture, and every form of media from books to radio to YouTube. There will be those who join cults or hold esoteric views — or worse — in the years ahead thanks to their chatbot. But I think we can take great solace in a simple fact: no one changes their mind. And on this subject at least, not even me.

Don’t miss

Maybe. I hold by my comments in May 2025:

Now, artificial intelligence and its voracious appetite for all forms of energy is animating the dreams of a younger generation of [Middle East] political leaders who can envision themselves at the fulcrum of global power through the ones and zeros of training and inference rather than the dollars and cents of barrels.

Actually, it looks like the Middle East is getting both the ones and zeroes of AI and the dollars and cents of barrels. Touché.

Even so, there is an under-explored angle in the frenzy of coverage, and it is America’s sordid domestic politics on artificial intelligence.

A backlash against AI has been gathering force for at least a year, and what was once a trickle against the logjam of politics has suddenly become a flood. New York is a perfect example. The Senate Environmental Conservation committee is debating a bill that would place an immediate moratorium on data center construction for three years. Meanwhile, the Senate is considering a bill that would add civil liability to any business that uses a chatbot to offer advice that, were humans to offer it, would be covered under occupational licensing, including health and law (1.2 million people are licensed in New York).

Such legislation is being mirrored across the country and in DC, where just yesterday, Senator Bernie Sanders called for a nationwide moratorium on AI data centers as part of a progressive push to stop AI.

America is among the most negative countries in the world on AI, and a variety of polls show that the AI backlash is deeply connected to consumer frustration with rising utility bills. Energy inflation for home delivery is one of the darkest spots in consumer price surveys. Costs rose almost 6.3% over the last year, and that was before Iran shut down the most important shipping channel for oil and gas. While extreme weather and a brutally cold and snowy New England winter is one of the larger causes for the rise, AI has ultimately become the target for much of the vitriol.

Those widespread pocketbook concerns have dovetailed with liberal environmentalists and their concerns around water conservation; religious conservatives over generative AI and its usage in sexual imagery; and broad middle class angst over the future of employment and the nature of work. That’s a hell of a convergence among otherwise political enemies, which is why the backlash has been gaining so much strength.

Unsurprisingly, in private conversations with lawmakers and a few recent off-the-record policy events, the general view has been that AI is going to be a major electoral issue in 2026, and the dominant one in 2028.

Yet all of that is now bunk, at least for the time being. America’s attacks on Iran have completely upended the electoral math, and particularly the connection between data centers and energy usage in the minds of voters. As Fortune bluntly headlined an article earlier this week, “People really hate AI but not as much as Iran—or Democrats.” Any rise in utility costs from data centers pales in perception to the battlefield in the Strait of Hormuz that’s streaming on every cable channel and social media account 24/7.

As with much of the Trump administration’s decisions, launching a war with Iran was an unintentional masterstroke for the tech industry. Tech’s golden star status has fallen in Washington, from the apogee of the Obama internet freedom agenda and the Arab Spring in the early 2010s to the perigee of AI, social media and more. The long-term implications of rising suspicion of the tech industry have been profound: an increasingly tough fight against China in the race for frontier tech, extraordinary regulations around social media usage and adult verification, and a breakthrough level of support for antitrust, just to name three. As one Hill staffer described it recently to me, there is an increasingly bipartisan basis for just “blowing up” the largest tech companies.

Now? Iran is seared into the American psyche for the midterms. A new configuration of antiwar Democrats and isolationist MAGA Republicans will focus their efforts on war and peace, or narrowly on Anthropic’s ongoing fight with the Department of War/Defense. Even in the best-case scenario, I expect there to be large, Covid-like global disruptions to the oil and gas trade as companies rebuild their supply chains and get everything exporting smoothly again. The International Energy Agency launched the largest-ever strategic reserve release this week, but that’s merely a stopgap. AI is now an afterthought.

Make no mistake: at the federal level, the AI backlash will forcefully return by 2028. Employment will be the dominant conversation, either the casualization of once professional labor or the widespread unemployment of the managerial class (and likely both). As I critiqued in my brief note on that viral Citrini Research report:

The report doesn’t address any political blowback to the upper-middle class suffering from massive unemployment. Congress and the executive may not be the most responsive political bodies around, but truly, they are responsive to this voting bloc. If there were widespread job fears by elite Americans, these institutions would act, and act posthaste.

The bigger and more immediate challenge for AI, then, is the state-by-state fight now underway. Data center construction is running into America’s entrenched NIMBYism, and despite successes like xAI’s approval by Mississippi this week for a massive power plant expansion including 41 natural gas turbines, AI’s compute demands look to be insatiable. American hunger for ugly box buildings is not.

Meanwhile, the political organizational power of professions means that whole categories of work — including extremely lucrative and important functions like healthcare and law — will be increasingly declared off-limits to AI models. As I noted in “Professional Prerogatives” back in 2023 and continue to believe:

The upshot is that the United States has a decentralized governance system around licensing that ensures that at least some states will be the first to jump into an AI-driven world. The downside is that AI’s progress in the decade ahead will stall far earlier and last far longer than many analyses predict.

It’s stalling right now in New York, as it will in other states. The fight against AI progress is multiplying, and Big Tech will struggle to fight back (in fact, their notoriety at this point might make them the weakest fighters despite their resources).

Distraction is a useful political strategy; it may not ensure victory, but it can help ensure delay. There is only so much time legislators and voters can spend on individual issues. Increasing the visibility and salience of other problems like Iran will slow the momentum of the backlash. I am still hopeful that the aftermath of our intervention in the Middle East is going to redound quite positively to the United States, but the immediate economic effects will be dire. AI thus gets a little breather.

What we really need, though, is a much more positive and broad-based motion in support of AI. What’s so frustrating about the backlash is that at the same time politicians and citizens are lobbying to ban the infrastructure and applications of this technology, everyone is using it. Anthropic’s and OpenAI’s revenues are staggering, with never-before-seen growth rates. Their upcoming IPOs (along with that for the combined SpaceX–xAI) will likely be the largest of all time (current record holder: Saudi Aramco. It all comes full circle!) Usage rates globally are unlike anything we have seen for technology that was released to the general public less than four years ago.

Everyone is using it, but no one wants the externalities or apparently upgraded power infrastructure. That’s the illogical madness of AI today in America, as it is with so many other issues, from housing to reindustrialization. Trump’s attack on Iran was perhaps just as mad, but at least it’s got a logic.

Don’t miss

The Exploratorium — the San Francisco science museum where nearly three hundred founders, doctors, government officials, pharma executives, and investors gathered on Wednesday to discuss the intersection of healthcare and AI — was founded by Oppenheimer. Not J. Robert Oppenheimer, the one they made the movie about and that the AI frontier lab CEOs like to compare themselves to, but his brother Frank.

Frank also worked on the Manhattan Project. He was also hauled before the House Un-American Activities Committee. And after he was blacklisted from society and the scientific community, he spent a decade working as a cattle rancher in rural Colorado. When he finally clawed his way back to teaching, it was at a high school with fewer than three hundred students. With limited resources, Frank got creative: he took kids to the junkyard, where they got their hands dirty learning about mechanics from abandoned car parts. After seeing firsthand the power of tactile learning, Frank built the Exploratorium on a radical conviction: that understanding doesn’t come from listening to an authority figure. It comes from experiencing things yourself.

Clinton connected this story to her own childhood, visiting a museum in Arkansas that the Exploratorium had inspired, and then to this current moment. “Science really should be a place of not only understanding the future,” she argued, “but creating greater community in the present.”

Throughout the rest of the day, I kept thinking about Frank Oppenheimer because, at bottom, the entire conference was an argument about whether patients should be allowed to get their hands dirty and touch their own data, take charge of their own healthcare — and that argument is a lot older than AI.

In 1847, the American Medical Association published its first Code of Ethics. Article II, titled “Obligations of Patients to Their Physicians,” instructed that “the obedience of a patient to the prescriptions of his physician should be prompt and implicit. He should never permit his own crude opinions as to their fitness to influence his attention to them.”

Crude opinions. That was the official position of American medicine on whether you should have a view about what was happening to your own body.

And it held for a remarkably long time. A JAMA article from the fifties debated whether patients who learned of their cancer and became distressed should be “handled in many respects as children.” In 1961 — the year we put a man in orbit — a survey of 219 physicians in Chicago found that ninety percent would not tell a cancer patient the diagnosis. Not couldn’t, wouldn’t. Many actively changed the chart to avoid any mention of the scary word. The phrase “informed consent” didn’t even exist until after the Nuremberg trials compelled the medical profession to reckon with what happens when physicians make decisions for people without their knowledge. In other words, for twenty-five centuries, the doctor-patient relationship was modeled on parent and child. The physician possessed knowledge; the patient possessed obedience.

What I saw at the conference was evidence that this regime is finally ending. Not because of regulation or an ethical awakening, but because patients now have tools to touch what was kept behind the glass of specialized knowledge. Dr. Hala Borno, CEO of Trial Library, argued that because patients now have access to real (if imperfect!) expertise through LLMs, doctors have gone from being fully in control to being more in an advisor position. Dorothy Kilroy, the Chief Commercial Officer of Oura, made the case that people’s obsession with tracking their own health isn’t all about vanity — it’s often about the fundamental desire to live more life:

Time with my kids is really important to me. How I age is really important to me. I want to be around, and I’m not embarrassed to say I’m obsessed with that. And for the first time, we now have data ourselves that makes that visible. Not as a doctor, not as a professor. We actually have it ourselves. What was invisible before is now visible. Suddenly, for the first time ever, you feel more in control.

This was what I kept hearing all day. The VP of health technology at Meta, Dr. Freddy Abnousi, described patients snapping photos of lab results and interrogating LLMs before their appointments. And Angela Dao, Maven Clinic’s Director of Product for Maternity, struck the same chord when she said “I can speak for myself, my members, my friends who are pregnant, I think we’re all using ChatGPT one to ten times a day to triage our symptoms.”

Frank Oppenheimer would recognize all of this instantly: people learning by handling directly what used to be interpreted for them.

The best work I saw at the conference extended this logic. Waymark uses AI to reach seventy million Medicaid beneficiaries — patients answered calls from voice agents at a fifty-five percent rate, higher than human callers, and then filled out housing and food-benefit forms alongside the AIs. Waymark’s CEO, Rajaie Batniji, published a paper showing AI can predict complex pregnancies fifty-five days earlier than conventional methods. Maven Clinic carries patient data from fertility through pregnancy through menopause, building a continuous health picture that no rotating cast of in-person providers could replicate. And BeSound is using AI-enhanced ultrasound to screen younger women for breast cancer, building the first dataset on a population that has never been studied because mammography starts at forty and misses women with dense tissue.

None of these are administrative efficiency plays. They are contact plays: they put patients’ hands on data, care pathways, and decisions that were previously locked behind institutional walls.

Of course, there are serious, serious risks to handing everything over to patients. WebMD has been the bane of doctors’ existence for good reason. And as several panelists mentioned, a recent Nature Medicine study shows ChatGPT Health under-triages emergencies more than half the time. In ChatGPT-speak, that’s not just scary, that’s dangerous.

Unfortunately, so far the regulatory response hasn’t been to help make AI better, or supplement with better access to doctor consultation, or even to mandate more direct disclaimer language for users, but to ban AI use entirely. This week, New York’s Senate Bill S7263 landed on the state Senate floor. The bill would make chatbot operators civilly liable whenever AI provides a “substantive response” in any of fourteen licensed professions — medicine, law, nursing, dentistry, psychology, and more. Strangely, it never defines “substantive response.” The proposed law targets the deployer, not the model maker, so a hospital running an AI triage tool carries the risk.

The Abundance Institute called the bill “shortsighted at best and protectionist at worst.” I’d call it something more specific: it’s the 1847 AMA Code in a consumer-protection costume. The same ancient reflex — the patient’s crude opinions are dangerous, knowledge must be mediated by the credentialed — updated for a world where the patient finally has tools to be more than a child in the exam room. Restrict low-cost guidance channels, and paid professional channels become the default. That lands hardest on the people with the least money and the least options, which is to say, the people who most need the Exploratorium.

This matters because trust is genuinely at stake — just not the way S7263’s authors think. “Healthcare moves at the speed of trust” was the conference’s most-of repeated line. But I was really struck by FDA advisor Dr. Shantanu Nundy’s reply that, actually, “healthcare also moves at the speed of desperation.” A hundred million Americans don’t have routine care, eighty-three million are living where there aren’t enough doctors, medical errors are the third-leading cause of death, and American life expectancy has been flat for decades. Josh DeFonzo, whose company Mendaera builds robotic guidance for medical procedures, agreed and argued “we trust the current standard of care far too much.”

May Habib, CEO of Writer, offered the sharpest formulation I heard all day. When AI is the one restricting care — denying a claim, rejecting a referral, saying you don’t need treatment — a human should always be in the loop. But when AI is expanding care — reaching a patient who otherwise has nothing, flagging a risk weeks earlier, filling out a benefits form at midnight — it should be permissive. Yes, the line between the two is blurrier than anyone would like; the same chatbot that expands access can also undertriage an emergency. But the current system undertriages millions of people every day, not through error but through absence. We are holding AI to a higher standard than a system where a hundred million Americans don’t have a doctor to undertriage them in the first place. At what point does do no harm become do nothing at all?

So while experimenting with Chinese peptides and unregulated compounds is a dangerous and bad idea, these biohackers are the canary in the coal mine, the leading indicator of a trust crisis in healthcare that predates AI by decades. The pharmaceutical industry has been the lowest-rated industry in America since Gallup started asking — below the federal government, below oil and gas, and below banking. Trust in doctors has fallen fourteen points since 2021 alone. Three in four Americans don’t trust drug companies to price their products fairly. This is the soil into which healthcare AI is being planted.

Gray Matter, my new riskgaming scenario, explores this exact dynamic: what happens when a breakthrough technology meets an information environment where trust has already collapsed. What I’ve learned, watching hundreds of players navigate this scenario, is that when trust is broken, new technology doesn’t get evaluated on its merits, it gets consumed by the narrative around it. Effective treatments get rejected because their side effects are more viral than their benefits. Sugar pills get approved because they have better marketing. The question is not “does this work?” The question is “who do you believe?” And the answer, increasingly, is “not the institutions.”

That’s the real danger for healthcare and AI. Not that the technology won’t work — the conference proved this is a tremendously powerful tool in the right hands. The danger is that it arrives into a trust environment so degraded that the public won’t know who to believe — their phones, their doctor, their doctor using an LLM, or anything in-between. And an LLM plays the strange dual role of being both an expert-sounding oracle that can ingest your personal data and draw on the sum of published medical knowledge in an instant, and a hallucinating autocomplete that doesn’t actually know your situation, often reverts to the mean, and may dangerously underplay or overplay your symptoms. Patients have more information than ever, but information without trust is not empowerment, it’s noise. And noise, in healthcare, kills people.

Which brings me back to Oppenheimer’s museum. The Exploratorium worked not because it handed visitors a textbook and told them to figure it out, but because it built an environment designed for contact — curated, structured, scientifically rigorous, safe enough to explore, and real enough to learn from. The exhibits weren’t consumer products and they weren’t credentialed lectures. They were something in-between: an institution that trusted people to touch the science, and that earned people’s trust by making the science touchable.

Frank Oppenheimer spent a decade in the wilderness because the institutions of his era couldn’t distinguish between dangerous knowledge and democratic knowledge. We are about to make the same mistake — not with physics, but with our own bodies.

Don’t miss

That’s because they no longer have a shared framework. Last month, Canadian Prime Minister Mark Carney accurately declared what every Western leader had avoided saying: the rules-based international order is over, globalization has failed, and economic integration has become “a weapon of coercion.” But his prescription of “variable geometry” coalitions, middle powers coordinating issue-by-issue, is a dodge. It’s saying “we’ll figure things out on a case by case basis, wherever our interests happen to align.”

That’s not enough when a crisis hits, like it did this weekend when strikes on Iran started. And it’s not enough to create a long-term, durable strategy for future prosperity. It’s no wonder everyone was flailing at Munich. The Western allies need a new shared framework, one that recognizes men will never be angels, but steers us in the right direction anyway. A diplomatic and economic structure that relies on incentives rather than idealism.

That framework is abundance.

Abundance is a supply-side reform movement. The idea is simple: democracies have made it too hard to build the things their citizens need, from housing to energy to medicine. Abundance has mostly been framed as a domestic policy, like fixing permitting and reforming planning and state capacity. That’s all great, but supply expansion is and should be inherently international, even though almost nobody is treating it that way.

Imagine if we replaced the old logics of coordination — following an arcane web of international rules and regulations and norms, performing allegiance to a world order — with a new logic of expanding our productive capacities together. What if the democracies of the world aligned our regulations, secured our supply chains, opened our talent pipelines, and got to building? Countries wouldn’t participate out of loyalty or idealism, but because expanding our collective productive capacity makes everyone richer, more resilient, and harder to coerce.

To be clear, abundance cannot and should not be globalization 2.0. The old model optimized for efficiency: remove barriers, let capital flow to the cheapest producer, and tell displaced workers that rising tides lift all boats. It didn’t. Communities that lost manufacturing to offshoring haven’t forgiven us — and specifically the Democratic Party — ever since. Meanwhile, President Donald Trump’s recent reimposition of 10 percent duties after the Supreme Court struck down his previous tariff regime shows that his nationalist solutions are legally fragile, economically disruptive, and produce no durable architecture. The answer isn’t to retreat behind tariff walls, it’s to build more, with allies whose workers benefit from that expansion.

In practice, an abundant foreign policy means having three goals: recognize, secure, build.

First: the easiest boost to Western economies would be to recognize each other’s work. Right now, allied democracies too often treat each other like strangers in the marketplace. A British medical device company that gets approval from its UK regulator has to start the process from scratch in Washington, Brussels, Ottawa, and Canberra — that’s five separate approval regimes for the same product sold to countries that share intelligence, nuclear strategy, and military command structures. No wonder over 80 percent of the UK health-tech industry reports disruption from regulatory uncertainty. And it’s not just medical devices: allied AI safety frameworks, financial regulations, and professional credentials all suffer from the same dysfunction.

The fix isn’t forcing harmonization — nobody wants to adopt someone else’s rulebook, which is why TTIP and other international trade deals so often struggle. Instead, we should offer mutual recognition: each country keeps its own standards, but agrees that an allied regulator’s approval process is competent enough to pass muster in their own system. If that sounds implausible, consider that the US and EU already do this for pharmaceutical manufacturing inspections. Nobody has to surrender sovereignty — we can just stop wasting time and trust our allies to have basic competency.

Second: we should secure what we need together. The obvious example here is semiconductors: Taiwan produces 92 percent of the advanced semiconductors we need on an island a hundred miles from mainland China. A disruption — an invasion of course, but also a blockade or even an earthquake — would be catastrophic.

The collective response to this potential 10 trillion in global economic losses? Everyone is trying to solve it alone. The US passed the CHIPS Act, Japan is subsidizing TSMC fabs in Kumamoto, Europe is courting its own plants. Last month Secretary of State Marco Rubio hosted a Critical Minerals Ministerial at the State Department, but the proposal is to have 55 separate conversations with 55 separate nations about the same supply chain problem. Meanwhile, UK Prime Minister Keir Starmer flew to Beijing seeking bilateral deals with the very country whose dominance he’d identified as a strategic threat. Canada did the same thing. That’s less thoughtful statecraft, more every man in a suit for himself.

The abundance version looks different: allied countries co-invest in shared mineral processing capacity. That way, Australian lithium doesn’t have to be shipped to China to become a battery; joint purchasing mechanisms give allied buyers collective leverage (instead of letting commodity producers play them against each other); and a distributed semiconductor network allows the United States to provide design and capital, Japan to provide materials, the Netherlands to provide lithography, and fabrication to be spread across trusted partners.

Instead of having five fully redundant national champions, we can have a supply chain where every link is in allied hands and no single point of failure can bring the whole thing down. This is how allied defense procurement already works: the F-35 has 1,900 suppliers in nine countries, and as a result every one of those countries has more aerospace jobs than if they’d tried to build a fifth-generation fighter alone.

Third: build to complement each other. Securing supply chains protects what exists, but abundance also means creating productive capacity that no allied country could build alone. The math on AI makes the case: according to the Federal Reserve, the United States controls 74 percent of global high-end AI compute. China holds 14 percent. The EU has less than five. US hyperscalers are spending over $400 billion annually on AI infrastructure, more than five times China’s investment. No allied country can match that individually. But a coordinated approach — frontier training in the United States, inference infrastructure in Europe, applied AI and safety architecture in the UK — would give the democratic world a shared foundation that none of them could afford to build from scratch.

The same logic applies to clean energy, to biotech, to advanced manufacturing: the country that can’t build will have nothing to compete with, and no allied country can build everything it needs by itself. China understood this intuitively; whatever you think of it, the Belt and Road Initiative was a foreign policy organized around productive capacity. The democratic world needs its own version — not state-directed mercantilism, but coordinated expansion where each country builds what it’s best at and the whole is actually greater than the parts.

Recognize, secure, build represent solutions to three frictions that make every allied economy individually weaker and more exposed. Remove them, and you have a productive base that’s resilient by design — not because everyone agreed to follow the same rules, but because everyone’s material interest is served by building together.

I can predict some objections. One side will say this is naive, that great-power competition is about hard power, not growth coalitions. The other will point out that this does nothing to protect human rights or prevent war. How does abundance prevent the United States from toppling foreign leaders or Russia from further invading Europe? How does any of this end the forever wars?

Both objections miss the point. Abundance isn’t meant to replace our rules for peace and the conduct of war, it’s meant to give us a reason to enforce them again. When your economy depends on Chinese processing and Russian energy, you’re going to struggle to sanction either seriously, no matter what the UN Charter says. Europe proved this: it took a land war on the continent to break the dependency on Russian gas, and even then it was agonizing.

An abundance framework changes that calculus. If your productive capacity runs through allied supply chains — if your batteries, your chips, your AI infrastructure all depend on allied coordination — then maintaining those relationships is preservation, not just idealism. Rules enforced by shame are optional. Rules backed by mutual economic dependence are not.

As for hard power: there is no hard power without productive capacity, and the democratic world is running short. Again, the Ukraine war proved this. The United States and Europe wanted to arm Ukraine and discovered they couldn’t manufacture enough artillery shells to keep pace. Political will didn’t fail, but production did. We won World War II in part because the Arsenal of Democracy would outproduce the Axis. Today’s democratic world has let that industrial base atrophy. Abundance is how you rebuild it, not within one country’s borders, but across an allied network that can actually sustain the demands of both peacetime prosperity and wartime mobilization.

This weekend, as the news broke about Iran, no allied capital had a coordinated response, economic or otherwise. Europe pushed off even meeting to discuss it until Monday. There was no shared supply architecture to leverage, no joint productive capacity to threaten to withhold. Instead we had 30 separate governments scrambling to figure out what they think. That’s what the absence of a shared framework looks like. And it doesn’t make any of us safer or stronger.

Men will never be angels. But an abundance framework doesn’t need angels — it needs engineers, scientists, trading partners, and allied governments that understand the basic math of the 21st century: you cannot coerce what you cannot outproduce.

Build together or decline alone. Those are the options. Everything else is a Davos speech.

Don’t miss

And what a civilization it is! I was heading this weekend to one of the corporate campuses that dot Mountain View, the famed NASA Ames Research Center’s Hanger One gloriously rocketing up in the distance. Dating to 1939 and once an emblem of ambitious spaceflight, today it’s more known as the elite earthbase for private aviation in the Bay Area given its proximity to Silicon Valley’s wealthiest. Google paid $1.16 billion for a sixty-year lease to manage the hanger and the associated airfield.

The inflated valuations of the companies here are only matched by the deflated tires of the motorhomes that crammed street parking, a nomadland of the involuntary itinerant. One camper trailer had a handwritten sign in the passenger window pleading to Mountain View enforcement to not ticket it. Meanwhile, for those not hiding in the liminal shadows of Mountain View’s municipal ordinances, a one-bedroom at the building six feet away (which offers a “life in the spotlight” as part of a “carefully curated ecosystem”) rents for about $4,000 a month.

I was here for the O’Reilly Social Science Foo Camp, which brought together about 200 people from all walks of life — from the typical (journalism, researcher, academic, writer) to the unusual (banjo player, cocktail craftsman). It was a crowd that heavily leaned erudite; I am sure more books have been written by people at the conference than read by the entire population of Mountain View last year (those electrons holding together fibers of paper could really be doing something more intelligent, no?)

AI was the inevitable focus. Foo Camp is an unconference, offering a democratized barometer of what people are thinking about and who is drawn to what subjects. The combinatorial math of people and time meant that pretty much every intersection of the social sciences and AI was covered, from economic productivity and jobs to politics, sociology, psychology, aesthetics, gaming and more, with the occasional provocative protest session like “AI People are Losers.”

What struck me throughout the discussions was how much our frames and models of society — the very essence of social science — are entirely unequipped to handle the world that’s shortly coming. The social sciences aren’t dead; if anything, they are more important than ever. Yet there is a severe discontinuity with the past that requires more than just updating Weber, Smith and Freud for the agentic era.

Take AI and government. What does a constitution mean in the age of AI, when computers themselves may have more power over our affairs than fellow citizens? Will an OpenClaw’d computer be allowed to vote someday? That bot and its agentic brethren may well control the fates of more people than the management class ever did, even at the height of the postwar boom.

There were strong ideas and contributions from everyone, including a long-time elected politician who had his own perspective from the trenches. Each contribution, though, seemed to be a brushstroke on a canvas that no one could see, or even fathom.

The conversation seemed to regularly hit impassable intellectual barriers. I brought up the idea that politicians in the future would be able to hyper-personalize their messages to voters, complete with audio and video (someone brought up Narendra Modi’s hologram as an early effort). This would create a breakdown in the political system as candidates sell an atomistic vision during a campaign while ultimately governing as a whole. A political scientist proffered that the academic research shows persuadability is very hard, and that the only goal of campaigns should be to mobilize allied voters. That research made sense before, but now we are finding early evidence that agents can be highly persuasive in ways that humans cannot, precisely because what was once a social interaction with its scripts, norms and expectations has been replaced by a man-machine interchange with near-infinite patience.

Many of us see machines as more objective and understanding than other humans, known as the machine heuristic. We might hide our health from our closest friends while uploading our complete medical records to ChatGPT or Claude to understand our situation better. Which relationship is more personal? That was a topic for a session on “is it okay to form an emotional attachment to your AI?” The subject echoed Sherry Turkle’s decade-old book Alone Together, about how the lack of agency embedded in technology is precisely what makes it so dangerous. Tech, whether robots or agents, is tailored to serve us, rather than becoming equal partners with us through shared experiences.

One person who attended the session said they were opposed to the very notion of AI friendships, seeing them as part of a broader decline of civilization. I countered, suggesting that AI friendships are an inevitable and perhaps even positive development, particularly for the already lonely. Which is better: being alone, or being alone with a bot? For many, that’s the bleak but realistic choice they get to make.

Here again, our social science frames of human experience seem to be wanting. What is friendship? If I talk to a stochastic parrot, and I imbue those interactions with meaning, could that be enough? Derek Thompson and other writers have claimed we have a loneliness epidemic, but it’s as much an epidemic as the common cold. Loneliness is universal and eternal. One of the recurring themes of Robert Caro’s first book on Lyndon Johnson is just how lonely the Hill Country of Texas was a century ago. Do our aspirational frames of friendship have any basis in past or present reality, let alone in our agentic future?

Optimism came from a session on chess. AI comprehensively defeats humans there, and yet the world of chess today is more robust than it’s ever been before. World chess championships are now watched by millions, and the number of players seems to be rapidly growing, with more than a million ranked players and tens of millions of occasional amateurs.

Why do so many people play a game in which they can be easily defeated by AI? Answers abounded. One person noted that the community is engaging and that chess is fundamentally a social activity that affords reasons to connect with others. Another said that it was “existential solidarity,” that we care about what another human will do far more than a machine. It’s no different than watching a speedrun of a video game: AI can perfectly mash the controller, but it doesn’t invoke the sense of awe we encounter when we know that a human accomplished the near-impossible. Another brought up Walter Benjamin’s famed essay on what makes art unique, claiming that it has an ‘aura’ that arises from its provenance of human workmanship.

My answer was that onboarding to chess — like coding — has been leveled. Learning chess before the internet would require distant and slow correspondence, perhaps a live coach, and a lot of books. Now, one can play an opponent with an equally matched Elo score and rapidly acquire skills and expertise, while chess commentators and AI tutors can explain moves and give more context for what’s taking place. Rapid feedback, greater fun, and play at a click of a button — all the ingredients are here for a robust and growing community. AI may well be the greatest grandmaster in history, but that ultimately doesn’t matter.

Unfortunately, it does matter in the economy. There were lightning talks and sessions on jobs, productivity and human agency throughout the conference, and certainly such concerns were ambient in many side conversations. Despite its popularity and primacy though, the distance between the economic prognosis of AI and the tools of social science was the greatest of any subject I encountered.

I got into a heated argument over productivity. The old Robert Solow line of “you can see the computer age everywhere but in the productivity statistics” has always had an obvious answer to me: you aren’t counting right — you’re just a drunkard searching at the lamppost. Spotify delivers the entire world’s music to three-quarters of a billion people. It also only paid out $11 billion to musicians in 2025. Outside of its revenues and royalties, economists have no ability to calculate what music’s worth is to people. A few might try to estimate consumer surplus of course, but those accounting methods are never integrated into the national statistics that ultimately drive policy.

That narrowness of economics is one of its defining features. What is a job? Economists have robust definitions of “job,” definitions that are going to require a substantial revision if not a wholesale rip-and-replace in the years ahead. Is orchestrating an agentic army of Claudes and Geminis a job? What if we start to see the rise of hyper-freelancers, who might work for 100 clients or more simultaneously while leveraging their bots? One could argue that nothing has changed, and yet, everything has. That’s the cardinal challenge for the social sciences in the years ahead.

There was a low hum of pessimism across the proceedings, latched on to the nightmare of a permanent AI underclass who are left behind by an elite squad of agentic engineers who accelerate away with the entire world. It’s not even a dystopian fantasy, for the tech industry’s success the past three decades provides ample proof of this trajectory.

The usual points were raised in various discussions, from universal basic income to some form of techno-communism, ideas that have already failed concretely or intellectually. What’s not even being discussed is what happens when bots have their own wallets and the very notion of property starts to break down. Who counts as a person? What is their (or its) relationship to property and contracts? Can an AI legally possess its own compute? Why not?

I was thinking about that on the walk back to the hotel. Answers didn’t flow, and I understand the deeply unsettling world we are all entering when very important and fundamental questions on the arrangement of society seem to be openly ambiguous and widely unanswered. Then again, as I walked by one decrepit motorhome after another, maybe unsettling some of our society’s (or at least, California’s) social failures is the disruptive breath we need to inhale. Our hearts might beat fast, but they will beat strong.

Don’t miss

Water services in the United States are decentralized by design, but the introduction of digital technology into water utilities makes these systems vulnerable to cyber attack. Small-town water suppliers are especially squeezed between affordability for customers and infrastructure resilience, but AI data center investments in rural communities may offer a way to lift small communities above the “security poverty line” and harden local water services against cyber attacks.

Local water infrastructure shocks can become a national strategic interest. The second-order effects can reach far beyond drinking water.

For example, agriculture needs clean water for irrigation; without it, crops run the risk of spreading deadly diseases or withering in summer heat. The agricultural sector particularly relies on water service providers that are most vulnerable to cyber attacks. Further, the energy sector needs water for power generation and turbine operations. Water services doubly impact AI and data center operations, meanwhile, because they require energy and water for cooling. A 5,000-person town becomes more of a strategic target if they have a critical asset, for example, when that water supply serves a critical data center or agricultural irrigation, the calculus changes.

In modern life, access to clean water is so ubiquitous that few consider what is required to sustain and secure it. Yet we understand that access to water is essential; in fact, when asked about their top environmental concerns, most Americans list access to water.

The United States’ complex system of watersheds and aquifers means there is no single point of attack that could affect the water supply for the entire United States. Localism in our water system means systemic resilience for the “national water supply.” But distributed water services must now reckon with cyber threats that could disrupt a critical municipality. The vast capabilities of a hostile nation-state could be levied against a single American township. What now exists as a localized network could have second order effects that require a national strategic response.

Before water services relied on digital technology, water security was primarily physical. Now utilities have adopted operational technology (OT) into their workflows to assess and control their systems in real time, but the benefits have come with the side effect of introducing a cybersecurity vulnerability. Now, perpetrators do not need to be physically present to attack the water supply.

Chilling reports detail the lack of cybersecurity in water treatment facilities. Because municipal governments orchestrate water management, the federal cybersecurity guidelines are just that: guidelines. More than 70% of facilities fail to meet basic cybersecurity requirements, such as changing default passwords (wait, 1234 isn’t a secure password?). The information technology systems that are vulnerable to hacking directly connect to the OT systems controlling chemical treatments. Simply investing in air gapping would dramatically reduce this threat. Perhaps the simplest cybersecurity measure Americans could take today is for residents to pester their city council to ensure the water department has changed its OT password from the system default. This simple fix could produce a disproportionate security boost from something requiring relatively minimal effort.

Centralizing water management is another obvious temptation for policymakers and strategists alike, but distributing national resources to ensure cybersecurity at the local level would maintain the built-in national resilience of a distributed system. Finding solutions to proliferate cyber resilience for our water supply should be preferred over centralized management. Chesterton’s fence suggests that we should not overturn localized water management practices until we can account for the financial, geographic, and political reasons municipalities were tapped to manage water in the first place. These reasons do not change because of cyber threats. Instead, local water services should locally execute robust cybersecurity standards.

Despite the benefits of localized water management, many small water services face economic pressure because cybersecurity is often a fixed cost shared by fewer customers. Roughly 91% of the water service systems in the United States serve just 17% of the population. These utility providers are desperately trying to ensure affordability, and every dollar spent on cybersecurity increases consumer water bills. Duke University’s Nicholas Institute outlines the trade-offs between infrastructure investment, affordability, and fiscal stability for the service provider. Although there are some creative solutions to mitigate the intensity of these trade-offs, cybersecurity concerns can create a labor arbitrage problem. While the water services provider attempts to keep costs low, it relies more on technology to automate their operations. But this technology comes with cybersecurity risks that are expensive to counter.

For municipalities that lack the funds to harden their cyber infrastructure to protect their water supply from cyber threats, there are even some charities where cyber professionals do work pro bono to help undersourced municipalities. Wendy Nather describes this phenomenon as living below the “security poverty line.”

Another way to secure funds is AI data centers. For now, many rural communities have mounted well-founded pushback on proposals to build AI data centers in their area. But the conversation remains fixated on energy prices. Howell Township, referenced in the WSJ’s recent piece on this topic, chronicles a community of 8,000 residents rejecting a proposal for a Meta data center. While some may caricature their campaign as Ludditism, the reality is rural communities have a lot to gain and a lot to lose on a deal like this, particularly when it comes to utilities.

Commitments from hyperscalers to invest in infrastructure resilience could dramatically change these negotiations. Structuring a deal to be a net benefit for utilities in a rural community would eliminate much of the rural hesitance to AI data center buildout while simultaneously being one of the few ways to revitalize aging, vulnerable infrastructure and ensure continued local control of water infrastructure.

Cybersecurity for water services directly benefits the data center, as such centers also rely on water for cooling, and the energy they use requires water services. As data centers come to rural communities, they often have significantly more leverage than local leaders in negotiations and typically offer only a modest form of infrastructure investment. If municipalities and hyperscalers approach infrastructure revitalization commitments as mutually beneficial, it would promote shared, long-term thinking that will ultimately lead to security for both parties. Many rural communities face an apparently impossible tension between affordability and water service resilience. Hyperscalers can invest in cybersecurity for water utilities out of self-interest while maintaining affordability and local control for residents.

A cyberattack won’t eliminate my ability to boil snow in the mountains, but it can certainly shut off the water supply to a data center. Small communities must incorporate cybersecurity for water services into the infrastructure proposals they bring to hyperscaling businesses that build in their towns. The right deal with a hyperscaler can lift a small community above the security poverty line and protect their water supply from cyber attacks.

The analysis, opinions, and recommendations contained in this memorandum are solely those of the author in a personal capacity. They do not reflect the official policy, position, or endorsement of the Department of the Air Force, the Department of Defense, or any other U.S. government entity.

Don’t miss

In 2025, SpaceX launched 170 times. Europe, whose Ariane 1 rocket pioneered commercial space launch in 1979, launched just eight times.

This is the negative compound interest on a decision made 12 years ago, when European officials rejected designing their new rocket, the Ariane 6, to be reusable. The next year, SpaceX’s Falcon 9 made its first vertical landing. France’s former economy minister Bruno Le Maire would later admit: “there was a fork in the road, and we didn’t take the right path.”

Since then, Europe has wallowed in the doldrums of a launcher crisis, struggling to retain sovereign access to space. From the final flight of the Ariane 5 in July 2023, the two-year grounding of Vega-C, and the much-delayed lift-off of Ariane 6 in July 2024, Europe finally began its sputtering recovery last year. Although it has restored the bare minimum of access, Europe is only beginning to claw its way toward reusability. As the stunners-wearing French President Emmanuel Macron recognized at Davos: Europe is “sometimes too slow. Foshur.”

Europe now finds itself in a tough space trilemma. It wants strategic autonomy, a world-leading space capability, and to protect the industrial status quo—that is, jobs. It can’t optimize for all three. If you think these sound like classic Riskgaming tradeoffs… you would be right, since that’s what our next game in development, Dead Reckoning, will be about.

In its current approach, Europe loudly proclaims its need for space autonomy while refusing to disrupt the industrial status quo. The inevitable result will be middling capabilities.

Consider ArianeGroup, the manufacturer of Europe’s only operational heavy-lift launcher, the Ariane 6. The company, itself a 50/50 joint venture between Airbus and Safran, employs 7,600 people across France and Germany and is dependent on up to €340 million in subsidies per year. Ariane 6 is scheduled to launch just 6–8 times in 2026 as it ramps up operations, reaching a tempo of ten launches per year by 2027. ArianeGroup’s subsidiary responsible for launch, Arianespace, will increase its launch rate further only if demand increases. As the company’s chief executive, Alain Charmeau, framed it in 2018, a rocket that can be reused ten times, scheduled for ten launches, would only be built once a year: “I cannot tell my teams: ‘Goodbye, see you next year!’”

Fortunately for Arianespace, 18 launches have been ordered for Jeff Bezos’ Starlink competitor, Amazon Leo (formerly Project Kuiper). Still, relying on a different American company to sustain launch demand is hardly a sovereignty-forward approach for Europe. Moreover, a triangular strategy of playing Amazon off SpaceX will cease to work once Blue Origin‘s reusable rockets come online. As it stands, there is not enough organic demand in Europe to incentivize ArianeGroup to expand their operations, let alone pursue rapid progress toward reusability.

Arianespace’s call for a mandated European launcher preference makes sense from their perspective, but criticism of the company’s track record abounds. As John Holst of Ill-Defined Space observed, “the company appears to be behaving in a manner that is daring the [European Space Agency] to cut off its antics. It’s not behaving as if it’s serious about the launch business. Instead, it’s behaving as an entitled monopoly.”

Europe’s intergovernmental space organization, the ESA, is correct in diversifying from ArianeGroup and Avio, the Italian maker of Vega-C — the under-used medium-lift launcher also reliant on €21 million in subsidies per year. ESA is now pursuing reusability with the European Launcher Challenge, first established in 2023. Critically, the ESA is implementing a significant reform to the ELC, softening the organization’s adherence to the geographical return principle, or geo-return, which requires every euro contributed to the ESA by a member state return a euro to domestic industries.

Such forced fairness has predictably resulted in market distortions and lowered competitiveness. At a 2023 conference, the chief executive of ArianeGroup Germany, Pierre Godart, complained about geo-return: “I cannot choose my supplier… Even if I have suppliers who are not performing, I cannot change them.” Geo-return’s critics are legion: the Draghi report on European economic competitiveness advocated scrapping the policy and the head of France’s space agency called it “poison.”

With its launch challenge, the ESA is finally listening to the critics. Rather than rely on geo-return, the ESA is experimenting with a concept called “fair contribution,” with five competition finalists selected first for their technological and business maturity, with member states then funding their desired champions. The program allocates €900 million across five reusable launcher startups, with the ESA acting as an anchor customer for future missions to stoke continued private investment.

The idea is to mirror NASA’s success with the Commercial Orbital Transportation Services (COTS) program, which provided SpaceX $396 million in milestone-based funding to seed the development of Falcon 9 in the early 2010s. Since then, SpaceX has raised a publicly reported $11.9 billion of venture capital, and is now preparing for a blockbuster IPO that could value the company well above $1 trillion.

Competition is useful to speed innovation, but that competition assumes a prize on the other side of the race. Europe is offering no prize today, neither for ArianeGroup or Avio as incumbents, nor for the five new ELC entrants. By comparison, after SpaceX proved its worth with NASA, the agency contracted $1.6 billion for 12 operational cargo missions to the International Space Station. Europe lacks such a unified prize. Without one, Europe will spend more to launch less, ultimately developing middling capabilities that pale to SpaceX.

While public revenues were valuable, SpaceX’s true genius was harnessing reusability to unlock an extraordinarily lucrative prize of its own: a virtuous cycle with their proliferated Low Earth Orbit constellation, Starlink. Starlink operates roughly 9,000 satellites in LEO today, accounting for two-thirds of all satellites in orbit, and contributing up to 80% of SpaceX’s $8 billion in profit ​last year.

The European Union’s flagship space program, IRIS², is an attempt to create a similar incentive, namely a sovereign, multi-orbital constellation of 290 satellites providing a “Starlink alternative” for Europe by 2027. Unfortunately, the EU’s procurement process for the now €10.6 billion project (up from €6 billion), relying on a single non-competitive consortium of primes given a 12-year contract, all but guarantees that the project will be doomed to mediocrity.

Sven Meyer-Brunswick, a principal at Alpine Space Ventures, made this comment on IRIS² prospects: “I think IRIS² is dead in the water, frankly... It’s not a competitive program. It should have been started completely differently.” IRIS² won’t reach full operational status until 2030 at the earliest, for nearly twice the original cost, and with serious uncertainty around its private financing. This risks a repeat of the full-blown governmental rescue of Galileo, the EU’s sovereign GPS-alternative and previous attempt at a public-private partnership in the space sector. At yet another fork in the road, this time around procurement, Europe chose yesterday’s approach.

Europe is indeed focused on sovereignty, as diminished as it may be, and it is getting mediocre capabilities, but the upshot of course is the allocation of jobs through procurement. Europe’s space manufacturing sector employs over 66,000 people, with four large industrial groups (Airbus, Thales, Safran and Leonardo) directly responsible for more than half of total space industry employment in the EU. Unfortunately, job allocation has taken precedence over entrepreneurial merit. When Airbus’ space division announced 2,500 job cuts in October 2024, citing $1 billion in losses over the preceding year, the cuts had to be politically apportioned across Germany, France, Britain and Spain. Moreover, they would not be implemented until mid-2026, taking a full two years to “right-size” the company.

Meanwhile, the primes will justify consolidation under the guise of efficiency, framing their efforts as necessary to become “European champions.” The proposed merger of Airbus, Thales and Leonardo’s space divisions would create a 25,000-person behemoth. With €6.5 billion in annual revenue versus SpaceX’s $16 billion, the primes may be correct in their desire to benefit from scale. But it will prove far too tempting for European leaders to direct their spending to this mega-prime thanks to its convenient industrial spread, with jobs across Belgium, France, Germany, Italy, Poland, Spain and the United Kingdom.

If they do, it will exacerbate Europe’s worst tendencies against competitiveness; as noted by Raycho Raychev, CEO of Bulgaria-based Lux portco EnduroSat: “The European agencies are biased towards a few very big players… tenders typically almost exactly describe the technical solution offered by a certain company instead of describing the challenge and choosing a winner who is best able to solve it.”

To compete properly with SpaceX, Europe may eventually need an excellent, near-monopoly provider to emerge, but only through disruptive competition that is allowed to take market share from incumbents — regardless of where the jobs are. Until that happens, politically-driven mega-primes would simply ensure even less competition and innovation. Europe would be left as vulnerable in satellite manufacturing as it was in the doldrums of the launch crisis, reliant as they were on Arianespace.

In sum, Europe doesn’t have reusable launch capability because it doesn’t have enough satellites worth launching. Europe can’t afford to launch a proliferated LEO constellation because it doesn’t have affordable launchers. Europe’s launchers won’t be affordable until they are reusable and launching at capacity and at cadence. Shot through this Catch-22 are Europe’s issues with geo-return and non-competitive primes — the industrial status quo.

The only path to cutting-edge capability without accepting industry disruption is to partner with leading American companies. If Starship is able to bring launch costs even remotely close to the $10 per kilogram Musk mused about, all manner of orbital economics become tenable. If AI data centers can become cost-effective with Starship, so too can European satellite constellations. With Starship, Europe could launch an expanded and viable IRIS²; continue evolving Galileo’s position, navigation, and timing services; and establish its new remote sensing capability, called the Earth Observation Government Service. By hitching a ride with SpaceX, the European Space Shield could become a reality.

European concerns about Elon Musk are valid, but here’s a counterintuitive strategy: post-IPO, SpaceX will need to be even more responsive to growth-demanding shareholders. European institutional spending represents the vast majority of the available space market in Europe. Instead of framing the dependency of Europe on SpaceX, the frame should really be that SpaceX is dependent on European expansion for revenue and profit growth. SpaceX shareholders will not tolerate cutting off Europe in Muskian shenanigans if it means a slash in market cap. After all, the oft-told story of Musk’s disabling of Starlink access for Ukraine misses a key detail: he had been providing Starlink effectively for free. Once the Pentagon began officially footing the bill, the disruptions ceased.

Europe has two valid options: rely on SpaceX’s launch capacity for cost-effective launch or avoid SpaceX completely by pursuing painful procurement reforms. The middle is what we dub the “dependency trap,” with Europe reliant on SpaceX for critical European launch but not a large enough customer for shareholders to care if access is shut off. When SpaceX launches 170 rockets in a year, half of those should be European funded. If Europe treats Starship not strictly as a threat, but as an opportunity, the continent can show Musk that dependency cuts both ways.

Mandatory Disclaimer: The views expressed are my own and do not constitute endorsement by the Department of Defense, Department of the Air Force, or the U.S. Government. All references, including external hyperlinks, to non-federal entities do not constitute or imply Department of Defense, Department of the Air Force or U.S. Government endorsement of any company or organization.

Americans pride ourselves on resilience, yet our critical infrastructure struggles to withstand both natural and man-made shocks. Each year, there is a natural disaster that “no one saw coming.” Whether policy decision or inaction is to blame (Texas had zero snowplows in North Texas until 2011, and power companies around the country have chosen not to improve their local grids) our society veers toward absolute efficiency at the cost of protections from tail risks.

Winter Storm Fern is no different, which raises an important question: What would happen if the damage came not from a natural disaster but from a hostile state actor intentionally targeting transportation and power systems?

Mother Nature may dump inches of slush across America, tripping up society for a day or two. But a clever adversary can use hybrid warfare to seize asymmetric advantage. Fragile things can be destroyed at a significantly lower cost than it takes to produce or replace them. For instance, it would take minutes to destroy one of the transformers that enables our substations. By contrast, manufacturing another one to replace it has a 120-week lead time. It does not take a skilled aggressor (or even a smart one) to sabotage the right substation and cause mass disruption. Mother Nature might get lucky; an adversary can optimize.

Consider Asheville, North Carolina, and the time required to clean up the aftermath of Hurricane Helene. During a road trip to Tennessee this October, I was shocked to be diverted around Asheville due to a partial closure of I-40 still causing delays. The Interstate Highway System was originally designed to make America ready for a nuclear holocaust; now it can barely survive the wild.

Statistician Nassim Taleb’s book Antifragile offers a classic risk framework for understanding why our infrastructure is so intolerant of disorder. He characterizes systems within three categories. Some systems are fragile: they fall apart with small amounts of disorder. Other systems are durable and can withstand some disorder, yet they are still worse off afterward. The most resilient systems — which he calls antifragile — improve with disorder. Ultimately, inanimate objects are typically categorized as durable or fragile, but systems composed of humans can become antifragile.

The web of bureaucratic requirements around critical recovery work often gets in the way of the workers tasked with those efforts. The question is what we can do to turn this morass into an antifragile system that recovers quickly and improves its recovery speed with each disruption. This gets the power on faster, and the increased production should improve the quality of the components being repaired.

In Breakneck (and on the Riskgaming podcast), Dan Wang poignantly observes that the United States is a lawyerly society compared to China’s engineering society. Each type has its benefits and drawbacks, but the way he has described China’s fast-paced infrastructure improvements is compelling. Even more striking, Wang says that the America of the early 20th century was an engineering society. This was a time when Americans built great things, including much of our current infrastructure.

The tension lies between the extent to which we pursue antifragility versus durability. An optimist could argue that infrastructure fragility is simply an engineering problem, but the reality is that if Wang’s observation is true (and I think it is), calcified legal hurdles will continue to be a barrier to expanding and improving infrastructure — a particular challenge as the United States contemplates reindustrialization.

As Stewart Brand argues in his new book Maintenance, just using and possessing infrastructure is not where excellence lies. Instead, and especially without supply chains and labor forces to quickly build anew, we must be able to maintain what we have already built. Indeed, maintenance is what will make infrastructure resilient — or in Taleb’s terms, antifragile.

The exact mix between systemic redesign in pursuit of antifragility and simply maintaining what we have will be hard to find. Thankfully, the goal of each is the same: to create reliable infrastructure that serves Americans. The enemies of both are indecision and apathy.

For the United States to develop power, water, telecommunications and transportation systems that are truly antifragile, we will need to prioritize repairs. This would look like both improving our existing infrastructure and increasing the production of parts needed to repair it, a trend that Lux has been looking at called “fixware.” While continuing to emphasize safety, this process should focus on production speed and on developing the supply chain to maintain the newly reinvigorated system.

An individual substation cannot become antifragile, but a company that can build or repair substations rapidly is part of system-wide antifragility. A single transformer may always be vulnerable to attack or natural disasters, yet a diversified supply chain that continually improves transformers can ensure the power grid remains resilient in the face of disorder.

Acts of nature and hostility will persistently threaten our critical infrastructure, but an antifragile system that rapidly heals and improves will minimize the damage they cause. Whether America has the capacity to build such a system is a question Winter Storm Fern once again brings to the fore.

The analysis, opinions, and recommendations contained in this memorandum are solely those of the author in a personal capacity. They do not reflect the official policy, position, or endorsement of the Department of the Air Force, the Department of Defense, or any other U.S. government entity.

Don’t miss

The complexity comes from how these options interact. Easy exit undermines voice — why organize if you can just leave? Strong loyalty can enable voice — you fight harder for something you’re committed to. And the threat of exit amplifies voice. “Give me a raise or I’m gone” makes both more powerful than either alone.

Yesterday at Davos, Canadian Prime Minister Mark Carney raised his voice louder than any other middle power in my lifetime.

Canada cannot fully exit the American-led system — geography and economics make that impossible. But loyalty has become untenable now that Trump’s America refuses to even pretend to play by the rules and continually threatens Canada’s sovereignty.

So Carney has built the infrastructure for partial exit: renegotiating trade with China, India, Qatar, ASEAN and the EU; doubling defense spending and securing European procurement arrangements; signing strategic partnerships across four continents. And then he used his voice.

Carney used Václav Havel‘s 1978 essay “The Power of the Powerless” to frame his remarks. Havel, a dissident Czech poet who later served as his country’s president, described a greengrocer in communist Czechoslovakia who places a sign in his window each morning: “Workers of the world, unite!” He doesn’t buy the slogan — truly, no one does. But he displays it to avoid trouble. And because every shopkeeper does the same, the whole system persists, not by force, but because of the fear-making power of collective performance.

For decades, Carney says, countries like Canada “placed the sign in the window.” They praised the rules-based international order, benefited from its predictability, and avoided calling out “the gaps between rhetoric and reality” — that the strongest exempted themselves when convenient, that international law applied with “varied rigor, depending on the identity of the accused or the victim.”

Carney acknowledged what you’re not supposed to: the rules-based order was American hegemony, and for most middle countries most of the time that was a good thing. “This fiction was useful” he said.

But now, Carney argued, “this bargain no longer works.” And then he said the thing no Western leader ever, ever says: “You cannot live within the lie of mutual benefit through integration when integration becomes the source of your subordination. Friends, it is time for companies and countries to take their signs down.”

In other words, he names the noble fiction — and says it is no longer noble.

There’s a reason no one talks about the noble fiction: to name it is to shatter it. When I was at the State Department, someone somewhere decided we needed to write a big keynote speech or a splashy essay in Foreign Affairs defining and defending the rules-based international order. The job landed on my desk.

I was relatively new to foreign policy and had always been confused about what the “international rules-based order” was exactly and why it had to be such a mouthful. So I set about trying to figure it out. I asked policy advisors in the U.S. mission to the United Nations, who pointed me to State’s Policy Planning office. They told me to talk to the National Security Council. The NSC sent me to some lawyers. The lawyers told me there was no official definition besides “an international order based on rules.” I felt like I was taking crazy pills.

The speech/essay never happened. I suspected, at the time, that perhaps the phrase was designed to resist definition, so as not to expose the occasional gaps between what we said and what we did. Reinhold Niebuhr diagnosed this problem when he argued that nations dress up their self-interest as universal principle, and then they come to believe their own propaganda.

I’m not sure if we had just come to believe our own propaganda — I do believe the rules-based order was the closest the world’s powers have ever come to binding their own might, and that the UN was an incredible force for peace — but Trump shattered any pretensions or illusions about the nobility of American hegemony. So all that was left was the fiction. Carney was just the first leader honest enough to dispel the fiction, too.

I’m a U.S.-Canadian dual citizen, and I travel to Canada several times a year. But when we went last spring to Toronto to run a Riskgaming scenario on AI and national security, right after Trump had started making hoopla about turning Canada into the 51st state, I was shocked.

Unlike the United States, one never saw Canadian flags in Canada. It just wasn’t a thing. Suddenly I saw the maple leaf flying everywhere — apartment windows, office parks, construction sites. When I went to grab a coffee, the café had renamed the Americano the “Canadiano.” American liquor was pulled from shelves, apps like Maple Scan helped shoppers avoid U.S. products, and border crossings plummeted.

The political shift was just as dramatic. In January last year, the Liberals were polling at 20 percent. By April, they’d won. Carney was propelled into office by pure Canadian defiance in the face of American imperialism. So when he speaks his truth, he’s also speaking to his base.

A former leader of Canada’s Liberal Party, Michael Ignatieff, quoted Freud in an excellent recent essay on the narcissism of small differences — Canadians defining themselves against America precisely because the countries are so similar. But truth be told, what I saw wasn’t narcissism. I saw fear sharpening into resolve. Carney is channeling that energy at the precise moment it gives him leverage.

What Carney proposed with his leverage is “variable geometry” — essentially, different coalitions for different issues, middle powers coordinating for collective bargaining power. So for Ukraine, that’s a coalition of the willing. For Arctic sovereignty, solidarity with Greenland and Denmark. And on AI, that means cooperation among democracies to avoid choosing “between hegemons and hyperscalers” — a striking phrase that names tech companies as somewhat-sovereign actors.

Unfortunately, variable geometry doesn’t provide a forum for resolution when coalitions harden or disputes arise. And while Carney named the UN, WTO, and COP as institutions “under threat,” he offered no vision or prescription for preserving them. The threat of World War III looms large.

And the AI math is brutal. According to the Federal Reserve, the United States controls 74 percent of global high-end AI compute; China holds 14 percent; the EU has less than 5 percent. Canada doesn’t register. U.S. hyperscalers are spending over $400 billion annually on AI infrastructure — more than five times China’s investment. The logic of concentration is relentless: compute, data, and talent flow toward whoever already has the most.

Canada does have real leverage as the only Western nation with serious reserves of cobalt, graphite, lithium, and nickel, all of which are essential for the chips and batteries that power AI. But critical minerals are inputs, not capabilities. Unless open source wins, it’s unclear how even the strongest collective bargaining group could have real leverage over the AI race’s winners.

But I think Carney knows this. As he conceded, “A world of fortresses will be poorer, more fragile, and less sustainable.” His gambit isn’t that variable geometry is a great idea or solves all these problems, but rather, that it’s the least-bad option available to middle powers like Canada.

Havel spent his life arguing that truth-telling was the most radical act available to the powerless. But once he became president, he discovered that the dissident’s clarity doesn’t easily translate into statecraft. A literal poet, he never quite figured out how to go from campaigning in poetry to governing in prose.

Carney is attempting to succeed where Havel failed, to be both the greengrocer and the prime minister all at once. “We are taking the sign out of the window,” he declared. “We know the old order is not coming back. We shouldn’t mourn it. Nostalgia is not a strategy…”

Nostalgia is not a strategy, but the greengrocer’s gambit is. Carney has broken the first rule of the rules-based order: don’t talk about the noble fiction. He’s chosen to raise his voice right at the moment of maximum pressure, and bet his country’s future on the power of ending a pretense.

The sign is down, the noble fiction is dead. Now we find out if the truth has any leverage, or if Canada’s threat of exit is real.

Don’t miss

But flip it around: Why bother to write if you don’t have to? What if the LLM writes better and faster than you? Why is it worth your time if you can have the result in seconds?

For most of history, you couldn’t separate the process of writing from having the writing exist. The artifact and the practice were fused. Unless, of course, you were rich enough to pay someone to separate them for you — or, now, unless you have a chatbot.

Before AI, I was that someone. I signed NDAs preventing me from sharing who I ghostwrote for, precisely because those people were afraid they’d get accused of not having spent the time to do the thinking and writing. Even people who constantly used speechwriters, like President Barack Obama, undermined his wife by saying her memoir was ghostwritten. (Guess what: all of President Obama’s books besides Dreams of My Father were ghostwritten. Of course they were!)

When ghostwriting is done well, you don’t know it’s been done at all. But when it’s done poorly, people notice. Politicians sound wooden. CEOs sound like their lawyers, and celebrities like their publicists. And we’ve seen plenty of backlash to what’s perceived as a lack of authenticity. Notice that both President Donald Trump and Representative AOC eschew speechwriters and teleprompters.

Still, there are times when we’re more forgiving, when we understand that having the thing written matters more than who wrote it. I wrote 250 speeches a year as a speechwriter for the U.S. ambassador to the UN, who was also a member of President Joe Biden’s cabinet. I did it because it would have been a colossal waste of her time to decide whether we should say we were happy that the rules-based international order existed or glad that it was around (RIP). No one cared that I wrote her speeches, so long as they truly represented her thoughts and views and the stance of the United States government.

But speechwriters are rare. A privilege of the powerful. For everyone else, as my writing teacher Tony Tulathimutte liked to say: No one is going to do your writing for you.

Well, now someone will. Or something will. Instantly, and for peanuts.

Except, not really. Because I don’t think AI has democratized writing. I think it’s democratized editing.

That’s because writing is really two activities bundled into one: the generation of new text and ideas, and the refinement of existing text into something better. Creation and curation, drafting and revising.

Most people, when they imagine “writing,” picture the first thing — the terror of the blank page, the taunting of the blinking cursor. But what AI offers us is the opportunity to start from the second. From now on, you never have to start from scratch. You can always have writing and ideas to react to.

Forget founder mode. We are all in editor mode now.

You’ll see this everywhere once you start looking. Vibe coders and Claude coders ‘edit’ apps into existence. RLHF is just: Did you like this, or did you like that? Musicians using Suno or Udio aren’t playing instruments, they’re selecting generated tracks. Same with the movie-makers using Runway or Midjourney. Lawyers review AI-drafted contracts instead of crafting them from scratch. Your emails are pre-written by AI, and it’s your job to edit them.

Even the phenomenon of calling out AI writing, as Meservey did and as so many others tagging Pangram on X have been doing, is a way of exercising editorial judgment. These slop-callers are performing taste, announcing they can discern the difference and that it matters to them. Pangram itself, a machine that somehow does a better job of judging the Turing test than a human, is just the editorial function made algorithmic. It’s not writing anything new. It’s judging what has already been written.

There’s something freeing about ubiquitous editor mode. It’s democratizing. It helps, for example, dyslexic people (who are significantly overrepresented among entrepreneurs) get their ideas written out. So too with the person whose first language isn’t English, or whose background is primarily in STEM. The blank page, for centuries the great equalizer and the great barrier, is now optional.

But we should also be concerned. Because this is all so new, I worry we’re still figuring out which parts of the writing process are valuable to us and why.

The linguist Emily M. Bender has a nice line: “Teachers don’t ask students to write essays because the world needs more student essays.” The science fiction writer Ted Chiang extended the metaphor. “Using ChatGPT to complete assignments,” he wrote, “is like bringing a forklift into the weight room; you will never improve your cognitive fitness that way.” The lesson is, you don’t go to the gym because the weights need to be lifted. The point is for you to lift the weights.

So the case for the student is obvious. They’re learning how to write and think. But what about us adults in the real world, those of us who presumably already know how to write and think? Are we lifting weights, or do we just want the weights to have been lifted? Are we writing to improve our cognitive fitness or to have people read our ideas? Are they even “our” ideas if the LLM came up with them?

Danny has already argued that we’ll stop caring about who “owns” ideas anyway. He might be right. But there’s a difference between generally caring where other ideas come from and caring whether your own ideas were given to you or worked out by you. Did you encounter the idea or did you earn it? Do you just agree with it, or did you arrive at it? And is the difference important?

I learned to write using what my mentor Vinca LaFleur calls the “Frankenstein” method. You barf onto the page everything you know — research, stray lines, half thoughts — and then you tinker and shape. The fragments cohere into paragraphs, paragraphs into arguments. You discover what you think through the struggle of making the mess cohere.

That’s editing, of course. But the difference is whose mess you’re editing. In the Frankenstein method, the fragments are yours. They’re incoherent because you haven’t figured it out yet. The editing is the thinking.

With AI, the coherence comes pre-installed. If you use the tool with nuance and clarity of thought, like Yoni Rechtman’s proposed “bicycle method,” then you might just wrestle your own confusion into clarity. But if you are essentially evaluating a stranger’s draft and saying “Sure, I guess” then the mess truly happened elsewhere, distributed across a trillion tokens of other people’s prose.

This is what the slop-detectors suspect: that there’s no mess, that no one is home.

So we’re all in editor mode now, and there’s no going back. The question is whether you’re editing yourself, the AI, or a mix of the two. The first is thinking. The second is outsourcing. And the third is the new thing we’re all going to have to figure out.

The weights don’t care if they get lifted. But you might.

Don’t miss

Prediction markets may be focused on conquering traditional sports betting powerhouses like DraftKings and FanDuel, but their viral hits so far are political and cultural. For instance, will Jesus Christ return before the end of 2025? As Bloomberg reported, that Polymarket bet paid out at a rate of 5.5% — better than a high-yield savings account.

Yet defining events and whether they take place or not is a complicated art. Controversy brewed at Polymarket this week after it announced it would not pay out bets on a U.S. invasion of Venezuela, contending that only control of territory would trigger the payout. Thus, Nicolas Maduro’s abduction is not classified as an invasion.

Here’s the challenge: the squishiness around words like “invasion” or “war” is intentional. Adversarial ambiguity is used by nation-states, corporations and even individuals to take action while avoiding some or all of their consequences.

When did Russia invade Ukraine? Was it when the so-called little green men started maneuvering in Crimea and the Donbas in 2014? For Moscow, these unidentified Russian soldiers masquerading as an insurgency force offered plausible deniability in its diplomatic negotiations. For Washington, war is technically only official when Congress passes a declaration, which hasn’t happened since World War II.

Given the high potential costs of war — up to and including nuclear strikes — nearly all of today’s great power conflicts are hybrid or use some form of gray-zone tactics. For instance, in 2017, Russia was widely reported to be seizing Georgian territory by simply moving the border sign and fence between the two countries a few inches to a few feet at a time, day after day. Does that inch count as an “invasion” since it involves control of territory? It’s precisely the ambiguity that allows Russia to avoid a wider conflict.

Over in the private sector, we see adversarial ambiguity in the insurance industry. Homeowners who survive hurricanes quickly learn that there can be a massive disparity in benefits between wind damage and water damage, the latter covered by the federal flood insurance program while the former is often covered under a private policy. What caused the damage to a home?

Insurance attorneys can spend years arguing over definitions and evidence in an attempt to transfer future payouts to the government or another carrier. We’ve seen a similar pattern in the aftermath of last year’s wildfires in Los Angeles, where lawyers continue to argue over what caused the damage to specific properties, just one of many reasons rebuilding efforts in the region have stalled.

Insurance companies repeatedly fight each other from different sides of the table, so there is an economic incentive to resolve adversarial ambiguity and make the industry more efficient and therefore lucrative. In response to high legal costs and legions of frustrated policyholders who may wait years to get their lives back on track, insurers have become increasingly receptive to the idea of parametric insurance. Instead of arguing endlessly over the cause of particular damage, parametric insurance pays out based on objective factors. For instance, a policyholder could buy a policy for wind speeds exceeding 120 miles per hour for over 15 minutes and would get paid even if their property had no damage. On the other hand, they would also not be paid if their house was destroyed but wind speeds were 119 miles per hour.

Such objectivity is elusive in politics, nor are there incentives for nations to cooperatively resolve ambiguity. Take “Will Trump acquire Greenland before 2027?,” which is currently trading at a 15% chance with a market volume of $3.3 million and is one of the bets the Polymarket homepage is pushing to me. There are four paragraphs of rules attached to the bet. The first is relatively simple:

This market will resolve to “Yes” if the United States officially announces that Greenland will come under US sovereignty by December 31, 2026, 11:59 PM ET. Otherwise this market will resolve to “No”.

The second paragraph is much more complicated:

Sovereignty is defined as the transfer of the majority of the territory of Greenland from its current status as an autonomous territory within the Kingdom of Denmark to being under the formal governance or jurisdiction of the United States, either as a state, territory, or other classification within the US system.

Territory here is interesting. Is it the physical land of Greenland, or does territory include Greenland’s territorial sea and exclusive economic zone? Since the bet pays out if President Donald Trump acquires “the majority of the territory,” figuring out the denominator is crucial. Jurisdiction is another open-ended and ambiguous component, but let’s move on to the third paragraph:

An official announcement made by the United States and Denmark that Greenland will come under US sovereignty will qualify, even if the actual transfer of sovereignty is yet to occur. Only announcements of official agreements or actions (e.g. executive order, signed legislation, etc.) will count - mere posts on Social Media will not.

Did you notice what just happened? We started the rules with “if the United States officially announces” and now we have “an official announcement made by the United States and Denmark” (emphasis added). Which is it? With adversarial ambiguity, it’s entirely possible for the United States to take Greenland and announce it while Denmark never acknowledges the loss but rather acquiesces to the reality of American military might in order to partially assuage popular anger. Practically every diplomatic negotiation will include such ambiguity as a way to usher parties toward an agreement. Let’s move on the fourth and final paragraph of rules:

The resolution source for this market will be official information from the governments of the US, Greenland, and Denmark, however a consensus of credible reporting confirming that Greenland has come under U.S. sovereignty will also qualify.

Oh my. “A consensus of credible reporting” is one of those beautiful phrases that harkens back to a halcyon era when adversarial ambiguity and weaponized truth were mere figments in the imaginations of the most inventive sci-fi authors. It reminds me of my article “Consensus Functions” from a few years ago, in which I describe why it’s become impossible to reach any form of democratic consensus in any context, and what we should do to push forward institutions like science.

Once again, we see a contradiction in the rules. “Has come under U.S. sovereignty” is not the same as “under the formal governance or jurisdiction of the United States.” Greenland could be under U.S. sovereignty but not formally under the jurisdiction of the United States.

I can point out half a dozen other asinine, pedantic problems. I realize I am closely reading what is otherwise a gamified gambling experience that’s meant more as entertainment than professional investing. Ultimately, Polymarket can cancel a bet and just refund every investor’s wager, and outside some aggrieved would-be winners, everyone can move on. Indeed, as a market, prices can include our relative beliefs about Polymarket’s judgment in resolving these bets. The premium for Jesus’s arrival is not a commentary on religious belief, but rather the market pricing in the complication of determining that bet’s outcome and Polymarket’s fidelity in doing so. The potential ambiguity of resolution raises the cost.

Prediction markets and sports gambling have massive negative repercussions for society as Charles Fain Lehman noted in The Atlantic and I pointed out in “America’s Gambling Fetish.” But one upshot is that more people will be exposed to the elite battles over words, terms and definitions that have made adversarial ambiguity a hallmark of great power politics. It’s not just about reading the rules, it’s about realizing that the whole point of many actions is to be able to ignore the rules entirely.

Don’t miss

South Korean President Lee Jae-myung has the toughest job on the world stage. He has to placate U.S. President Donald Trump and protect his country’s massive and growing trade surplus with the United States. Then he has to rebuild long-frayed ties with China while pressing the country to help him improve relations with always-be-firing-missiles North Korea (the latest salvo came Monday, complete with alleged hypersonics). Plus, Lee must balance the Korean public’s historical antipathy toward Japan while forging a relationship with Japan’s most nationalist leader in years in hopes of buttressing Northeast Asia’s axis of democracy.

But wait, that’s not all! Lee must also orchestrate a complex arrangement of deals with the Middle East to promote South Korea’s massive nuclear power buildout, including what could be a blockbuster bid in Saudi Arabia. That buildout is part of the region’s sovereign AI initiatives, a challenge Korea is facing as well as it competes against heavyweights from both West and East. Finally, there’s the ongoing awkwardness of South Korea’s defense exports to Ukraine, with the Korean defense industry growing to record highs even as it bans exports to active conflict zones.

This is the endless struggle of a straddle power without the ability to drop special forces on a presidential palace and ransom a country’s oil supplies. Such a position might be extraordinarily challenging, but it also presents equally extraordinary opportunities that can only be seized through a dauntless and ever-evolving strategy.

During the Cold War, political scientist Robert O. Keohane wrote a review article titled “Lilliputians’ Dilemmas,” which first appeared in International Organization before coalescing into a Foreign Policy essay. He discussed how small powers balance relations with big ones, arguing that the first and most important criterion for a small power to navigate the whitewaters of great power politics is that it “…must have a high degree of maneuverability within limits set by its own domestic politics.”

What does that maneuverability look like? Well, it’s constantly inventing new reasons to attract big powers to you on favorable ground. In one of his most profound lines, Keohane wrote “American policymakers are too busy to think up grievances for states that cannot manufacture their own.” In other words, straddle powers need to assiduously create complex conditions where they can secure favorable tradeoffs.

For instance, while Korea is at the pinnacle of a number of industries, AI memory is its most strategic asset right now. SK Hynix is one of the world’s most important manufacturers of high-bandwidth memory, which is required for the most complex AI computations.

The dizzying growth of the AI sector is pushing SK Hynix’s stock and that of competitors like Samsung Electronics to record highs. Over the past month, the company’s stock has jumped more than 35%, and it’s up more than 160% over the past 6 months. Roughly 80% of its employees are domestic, leading to a memory boom in Korean employment and bonus wages that amounted to $80,000 per employee this year (Korea’s GDP per capita is about $36,000). Samsung is right behind with a gain of 130% over the past six months.

Korean memory exports may not be a rainbow unicorn like Nvidia’s next-generation Rubin chips, but they are indeed supply-constrained, which is leading to record global price increases. That gives the Korean government leverage in negotiations with the United States and China, since neither country wants to surrender an edge to the other by losing priority access to this essential ingredient in the race for AI supremacy.

A single dependency is still a linear negotiation, though, and that doesn’t offer strategic maneuverability (or frankly, a good Riskgaming experience). So let’s remember Keohane’s diktat about manufacturing grievances. Trump has manufactured myriad complaints about Korea, including its yawning trade surplus and its lack of American auto imports. Korea got a significant if temporary reprieve in the aftermath of Trump’s crackdown on a Hyundai and LG Energy Solutions battery factory in Georgia back in September, which saw hundreds of Korean workers deported as they built up American manufacturing.

Yet that factory is precisely the right model. Korea’s most potent strategy is to position itself as the sole enabler of American reindustrialization. While I was off on holiday in mid-December, a major deal was announced that has Korea Zinc upgrading an existing zinc refinery in Tennessee through a $7.4 billion joint venture funded by J.P. Morgan and the U.S. government directly. In the final cap table, the Defense Department would own a 40% stake, Korea Zinc would hold less than 10%, and other investors would own the balance.

It’s a massive project. The Tennessee site is the sole zinc refinery in America, and no domestic facilities have opened for decades. Korea Zinc will take over the existing facility from Singaporean commodities conglomerate Trafigura and upgrade and scale it up to global competitive standards. If the plan is successful, it should shore up America’s zinc supply chain in 5–10 years, improving the country’s resilience in a range of products from metal alloys, like galvanized steel, to chips, batteries and weapon systems.

Since the announcement three weeks ago, Korea Zinc’s stock has gyrated. Initial euphoria sent it zooming to its highest levels in a year, but the stock has since lost much of those gains. It dropped nearly 9% earlier this week when ETF provider MSCI announced the company didn’t qualify for inclusion in its indices due to the share sale required to fund the joint venture with the American government.

Deepening co-dependency isn’t just happening in critical minerals, but also shipbuilding. Korean defense prime Hanhwa closed its $100 million purchase of the Philly Shipyard a year ago, announcing a massive expansion and reinvestment plan that would see volumes increase 20-fold as part of the official — and I kid you not — “ Make American Shipbuilding Great Again” plan (we are definitely better marketers than those Soviet apparatchiks and their “fifth five-year plan” snooze-fests). Hanwha has highlighted that it will train American workers at its Korean facilities, with the goal of improving sovereign shipbuilding capabilities.

Hanwha’s investment complements other announced projects. HD Hyundai Heavy Industries (HII) signed an agreement back in October with Huntington Ingalls Industries (also HII, so obviously a match made in heaven and a nightmare for every journalist and analyst going forward) to rebuild American shipbuilding and invest in “distributed shipbuilding.” HII (the American one) is the shipbuilding company at the center of our first Riskgaming scenario, Hampton at the Crossroads, and it’s the company best known for constructing Gerald R. Ford-class aircraft carriers.

Last month, a three-way agreement brought together American defense prime General Dynamics alongside Korea’s Samsung Heavy Industries and DSEC to focus on next-generation ship design and manufacturing, particularly on the U.S. Navy’s Next Generation Logistics Ship program.

I can keep on going in this Nor’easter flurry of announcements. How many of them will come to fruition much less lead to positive economic outcomes? My cynicism says that there is little chance that more than a fraction of them will ever ultimately be realized. However, I admit I have recently been more impressed at the level of focus on reindustrialization from leaders like Pennsylvania senator John Fetterman, who stopped by last month.

Back to straddling. Korea is attempting to enmesh itself in a complex weave of co-dependent relationships, where it always has another lever to bolster its relative negotiating position against larger superpowers. That’s the right approach for a straddle power: you want to be differentially and strategically valuable enough to raise the cost of intentional or unintentional adverse actions.

That’s obviously the opposite direction of America’s current positioning. Trump’s unilateral cuts to organizations like USAID mean that relatively cheap and featherlight threads connecting the United States to the rest of the world have been burned to smoke. Our trade autarky and increasingly venomous perspective on high-talent visas mean that even the social and business bonds that accrue as American power are weakening.

This approach does give a superpower like the United States more raw power. As I wrote yesterday, we can indeed “just take it” and there’s not much most countries can do in response. Where that strategy goes wrong, though, isn’t when other countries acquiesce, but when they don’t. If and when a country refuses, Keohane’s manufactured grievance toolbox will be empty, and America will be left with only much more expensive strategic actions available.

It’s perhaps the most important lesson I’ve observed across dozens of Riskgaming sessions, and particularly the ones we have conducted using No Man’s Land, our scenario focused on AI capabilities in commercial and defense procurement. Blunt strategies are satisfying to players since they are psychologically thrilling and so much easier to think through. It makes us feel powerful to stab another player in the back, particularly if they can’t immediately respond. Long term though, there are high costs to these fleeting moments of dominant power. Brewing vengeance, for one. But more importantly, the increasing dearth of co-dependencies means that bad outcomes aren’t as easy to prevent in the first place.

The players that succeed in Riskgaming — no different than straddle powers like Korea and others — understand that it is precisely the dense layers of value that have to be unwound that stop undesired decisions by adversaries. Blunt decisions don’t increase strategic maneuverability but instead do the complete opposite, transforming a complex game of negotiations into a linear tug-of-war. It assumes that other players will tug back on that rope rather than ignore it in favor of new strategies and tactics (think hybrid war or gray-zone activities). Brash audacity makes for viral social media posts and the occasional fun selfie, but it’s ultimately deleterious to long-term strategic power.

Don’t miss

Even the sales-and-deals turn of Christmas has its origins in reciprocity. Exchanging gifts is about deepening the ties that bind us all, looking back to pay it forward. It’s the same foundational cultural trait that powers up the startup world too, the reason Silicon Valley can’t be replicated. Nowhere is there a higher density of richly talented and richly wealthy people who are willing to take bets on projects with nothing but trust that it will one day all work out.

That spirit of Christmas is dying. The world is shrinking, resources are getting scarcer and it’s harder than ever to get what you want.

Apparently, what we want is oil, and it’s not for anointing the messiah. This weekend, the United States scarfed up Venezuela’s Nicolás Maduro in a surgical military operation that must be hailed as one of the most extraordinary successes in special forces history. Thanks to the wonders of this strange cosmic universe, Maduro and I now share the same gentrified borough.

The Venezuelan people have suffered mightily under Maduro and his predecessor Hugo Chavez. For almost three decades, the two have impoverished one of the most entrepreneurial countries in the world — a country that also happens to sit on an extraordinary endowment of natural resources that would be the envy of all but a handful of nations. Upward of a million refugees — from a population of about 30 million — have fled the miseries and trepidations of an autocratic government and a shrinking economy. And that’s just to the United States — an estimated 7 million more have fled to neighboring countries.

That’s not why we captured Maduro, though, for that would be far too humanitarian and charitable for the new Christmas spirit. It was oil, spoken plainly by President Donald Trump. This is no longer the Bush administration and its soaring (if at times sputtering) rhetoric of freedom and democracy. You have what we want, and we are going to just take it.

It’s the international relations equivalent of California’s $950 theft limit, a marker that divides a misdemeanor from a felony and which led the City by the Bay to become the City that Locks Up the Toothpaste. There’s no loss prevention strategy in Venezuela, though, since there’s not much to lose, empty shelves being Bolivarian socialism’s lasting legacy.

Just take it. I wish this wasn’t a pattern in American life, but it’s commonplace. Ro Khanna, Silicon Valley’s left-progressive congressman and a man who could really use that locked-up San Francisco toothpaste, certainly understands the need to just take it. Over the holidays, in spat after spat after spat with venture capitalists, he vociferously defended a proposed wealth tax meant to fund whatever it is that California considers a functional government.

Apparently, after years of walking the Fisher Price-colored corridors of the startup wealthy with hat in hand asking for coins every biennial (a hat that was ultimately filled to the brim, I might add), he’s finally had enough with it all and has decided to just take it. One time only, of course! I guess he hopes we can all feel the warmth of collectivism during the holidays. Call it the Wrath of Khanna, the superior man who lost his mind to aggrieved vengeance and is now biting the hand that feeds him.

Biting the hand that feeds you, or just pocketing it. I grew up in Minnesota, a land noted for its Nice people. It’s really hard for coastals to understand what it’s like to live in a society where your neighbors will always do the right thing, where charity — religious or otherwise — is paramount in the lives of most, and where calls for food donations or blood drive flyers have to include a proviso apologizing in advance that they won’t be able to accept everything that will be donated. That was just Minnesota Nice culture, which continued long after I transplanted to the golden foothills of South Bay.

That culture of egalitarian benevolence is now in serious jeopardy. Feeding Our Future was the name of the non-profit that’s at the heart of the multi-billion-dollar Somali fraud story that has spread like wildfire since my colleagues at City Journal published their deep dive right before Thanksgiving (an irony missed by most). Tim Walz, Minnesota’s governor and former vice presidential candidate, announced yesterday that he wouldn’t seek a third term because of the growing scandal, putting an ignominious exclamation mark on an otherwise decent man’s tenure.

The response from my coastal friends has been sheer bewilderment at the stupidity of the Minnesota officials overseeing welfare distribution. Take it from a still-prideful home state lover — naivety is justly earned. Leaving your doors unlocked is only foolish if you believe it’s impossible to live in civilization. Civilization isn’t the police and courts, it’s a glass house that’s wide open with a sign that says “no one is home” and nothing happens.

Venezuela, California, Minnesota: these three stories are merely the leading headlines of a repeated pattern in our culture today. America’s enlightened ethos of can-do “just do it”-ism has become today’s darker “just take it” psychology of the Beltway robber, the Barbary petrol-man and the Bitcoin fraudster. Positive-sum economics has been slaughtered by the atavistic dominance of zero-sum or even negative-sum thinking. It’s every man, woman and child for themselves.

I just watched Park Chan-wook’s new film No Other Choice, a story about a down-on-his-luck blue collar worker who is laid off after 25 years at a paper company. He realizes that the perfect job still exists out there, but there’s only one position and a lot of highly-qualified candidates waiting to be interviewed. How far must he go to eliminate the competition, save himself and secure his livelihood?

It’s a thrilling, terrifying and at times downright hilarious romp. It’s also part of an increasing oeuvre of films that critique competition in modernity like Josh Safdie’s Marty Supreme, Bong Joon Ho’s Parasite and Galder Gaztelu-Urrutia’s The Platform. These films observe this zero-to-negative-sum culture, but they either direct the blame for it to structural factors beyond individual control or naturally assume this culture is part and parcel with modernity. Absolutes have become relatives. Rather than trying to outrace others in the great game of life, we at some point realized we can just take a hammer to the engines of our competitors and watch as they sputter while we drive at whatever speed we want.

Competition is inevitable in any resource-constrained environment. Earth is no exception, and even becoming a multi-planetary species won’t release us from this basic law of economics. It’s how we address competition and its implications that determine the tenor of civilization. We can cultivate competition and the human desire for victory to frenetically recombinate brilliance and push forward the frontiers of technology and productivity. Alternatively, competition can be extraordinarily corrosive, undermining the will to work or invest and dragging everyone down together (see Venezuela, Country of).

America’s most important strength was — and always will be — its precarious meritocracy of the market. Good products generally beat bad products because consumers can suss out a better deal. Good ideas overcome incumbents because Americans are optimistic and willing to embrace better futures. Good institutions supplant bad ones because efficiency and performance matters to generous but impatient funders.

All those safeguards have washed away. Consumers are cynical and desperately fighting capitalism’s dark patterns and enshittification. Pessimism has replaced optimism, entrenching narrow ideas over the potential of tomorrow. We no longer expect our government, courts, non-profit organizations and civil society to be fair but rather capricious, something best avoided rather than engaged.

“Just take it” has a double meaning, one that extends back to The Melian Dialogue of Thucydides. It’s the claim of the victor to what’s not rightfully his, and the submission of the loser acquiescing to a tragic fate. Two sides of the same coin, a penny that is fought over even as it is taken out of circulation.

Our energy to fight for a prosperous system is drained as we battle each other. Just-take-it culture is an economic game of musical chairs, but the chairs are made of wood and the participants have flamethrowers. What looks like an ample supply of seating quickly burns in an inferno.

Our present moment reminds me of Koushun Takami’s Battle Royale, which I read during my very first Lux partner offsite, sitting in earthquake-prone California below a hotel ceiling with a very concerning crack. The novel, which became a movie and an entire genre of multiplayer video games, is about a group of school kids sent to an island in which only the last remaining survivor is allowed to walk away and live.

The most poignant moment comes during a scene at a lighthouse in which a group of students desperately tries to create a small island of civilization against the murderous impulses of their classmates outside. Their small society works for a time, yet zero-sum thinking will come for them, too, even as they rebel against it. You can’t be good in a bad system.

It is hard not to feel anguish as we gaze at the grand expanse of history; millennia of evidence accumulated across innumerable societies and civilizations proves that the simplest yet most potent answer to what makes a good life is that positive sums not only exist, they are almost omnipotent in their ability to ameliorate the worst aspects of human existence and give our lives meaning. We still have that culture in parts of Silicon Valley, and it’s still the default mode for most Americans, even as it’s getting harder to sustain.

Why do we give Christmas gifts? Often the monetary value of these offerings are normalized within families, and so it can feel circular on that great economic spreadsheet of existence. But even if the dollar values are the same, those gifts come attached with emotions and ultimately memories. No net value was exchanged, yet incalculable value was created and offered. The magic of Christmas is the magic of civilization and what’s made humanity such a force for good: a gift economy means always being surprised that something better is in store. If I have faith in anything, it’s that we can, even now, still just do it.

Don’t miss

For those counting, we published 104 newsletters, 47 podcasts and four new scenarios. We also hosted about 24 events including runthroughs of Riskgaming scenarios, community meetups and geopolitical dinners. In total, about 500 people got to join us live for an experience — definitely an upgrade thanks to Laurence Pevsner joining as my partner at the tail end of 2024.

As always, thanks for reading, listening and attending — your commitment has allowed Riskgaming to turn into a powerful institution for profound thought on some of the most complex issues facing the world today.

Highlights from 2025

• We hosted 60 senior leaders from the United States and the United Kingdom for a biotech summit at the U.K.’s embassy in Washington DC as part of Prime Minister Keir Starmer’s trade mission to the United States.

• I pissed off some Pennsylvania politicians with an op-ed in The New York Post on President Trump’s horrifying deal approving U.S. Steel’s acquisition by Nippon Steel.

• I pissed off some public sector labor unions and mayors with an aggressive essay in City Journal on the need to use AI to automate more of city government.

• Dozens of people in New York City and in Los Angeles played our biotech game Experimental Automata during Upfront Summit.

• We were profiled by Business Insider, Payload Space and The Wall Street Journal, complementing past coverage in NBC News and The Information.

• I learned that Senator John Fetterman is a complete giant compared to me:

• We hosted a senior delegation of Canadian political and business leaders in Toronto to play No Man’s Land.

• Laurence launched Gray Matter with sellout events in New York City and in California, doubling down on biotech gaming with our first live trading card game. Negotiations were fierce, and you know you have a winning game when it is impossible to bring the room back to order so you can give the final results. We’ll publish the game in early 2026.

• We hosted a prominent former U.K. defense (I’m sorry, defence) minister for a Riskgaming session (who is excellent at negotiation, I have to say).

• We had a great time launching Ian Curtiss’s Southwest Silicon to a distinguished group in London from the U.K.’s Foreign and Commonwealth Office, the Department of Business and Trade and other ministries as well as in DC with Eric Schmidt’s Special Competitive Studies Project.

• We hosted our first comedy revue with Truth Santa, which will always have a special place burned into my heart (note to self: never write a comedy show under deadline).

New Riskgaming Scenarios for 2025

• Experimental Automata: A Global Investigation of the Future of Digital Health and Biotechnology — a salon experience for 60 that educates on the progress coming in human wellness and agriculture.

• No Man’s Land: The AI Singularity and the Future of American National Security — the most in-depth full-immersion game we’ve ever made. This game takes eight players through the challenging tradeoffs of balancing commercial and defense security in the AI industry alongside a zany storyline that has now come to fruition in the real world.

• Southwest Silicon: Water, Semiconductors and the Future of American Economic Security — a full-immersion game that simulates the strict tradeoffs facing water-parched homeowners, farmers and the emerging chip fab industry in Arizona’s Phoenix Valley.

• Truth Santa: A Lux Riskgaming Holiday Experience — our holiday comedy revue that we hosted in SF and NYC. You had to be there, but you can see some notes and photos on the game’s unlisted page.

Best Newsletters of 2025

We switched to Substack this year, which was a huge help in terms of user experience and that all-important growth metric. Here were my favorite pieces of writing this year:

• Make gray-zone war expensive again — Unusually for me, defense was not one of my main topics this year, but gray-zone warfare is a perennial interest of mine. The costs of hybrid war have declined dramatically, encouraging more state and non-state actors to use these tactics to tremendous effect. We need to raise the costs again.

• The machine burns down — A meditation on a massive data center fire in South Korea and E.M. Forster’s The Machine Stops. Our lives are increasingly handled by machines that are now central points of failure for our society. Catastrophe awaits without better resilience and chaos engineering.

• Open secrets harm society far more than AI or disinformation — It’s the age-old story of the emperor’s new clothes: Everyone knows and agrees on facts that can’t be said out loud. It’s a pattern of behavior profoundly corrosive to democracy.

• Should AI recommend God? — AI chatbots like ChatGPT respond from an agnostic/atheistic point of view that reflects Silicon Valley’s mainstream values. To what degree should popular religions be brought up as paths in these chats? Is proselytizing through technology immoral?

• Why is economic data disappearing? — The declining quality and transparency of American and Chinese economic data mean that the twin superpowers are increasingly weaponizing perceptions of their economies at the expense of reality.

• The taps are dry and the rivers are flooded — A meditation on rivers inspired by James C. Scott’s In Praise of Floods. We need to pay more attention to water security.

• We’re finally predicting the weather accurately — Small progress made annually can add up to massive change over time. Weather forecasting improves by about a day every decade, and that is radically enhancing industries from logistics and agriculture to air travel.

• Uncle Sam is a terrible board member — My screed against the U.S. government taking board seats at U.S. Steel. The government can’t be a robust board member when it doesn’t have economic alignment and fails in its fiduciary duties.

• Tech is turning the Middle East into a pivotal region — Our image of the Middle East needs a dramatic renovation. It’s now a bountiful land of opportunity, dynamism and surprise — and it’s becoming a powerful economic fulcrum in innovation.

• Riskgaming against a world on fire: A manifesto, of sorts — This is the longest post of the year and a summary of my goals with Riskgaming as we moved to Substack. We constantly strive to avoid easy solutions to complex problems while uncovering the deep interconnections that constrain and direct today’s events.

• Is Plagiarism Dead? — AI is radically changing the definition of creativity and ownership. What is the future prognosis of the norm against plagiarism as AI tools reshape what we think of as originality?

• Asteroids, xeno-kidneys, planes and H5N1 — A deep dive into different types of risks and why it is still so hard to mitigate long tail risks even when they are predictable.

• Should you wait to have kids? — New biological technologies under development will radically improve the health of children right from conception. In the interim, that dramatically raises the relative risk of having children the next few years.

Subscribe now

The Best Podcasts of 2025

For your listening pleasure over the holidays, we’ve posted a special podcast episode recapping the year titled “11 Clips That Defined 2025.”

• The inside story of the billionaires fighting for space — The Washington Post’s Christian Davenport on Elon Musk / SpaceX, Jeff Bezos / Blue Origin and the race to colonize the atmosphere, the moon and ultimately Mars.

• Why AI safety is like a bolt in a croissant — Former Popular Science editor Jacob Ward on the ethical dilemmas facing the big general-purpose LLMs from OpenAI and Anthropic.

• America’s degrowth lawyers need to learn from China — Dan Wang talks about his everyone-is-reading-it book Breakneck and what we can learn from China’s engineering state and America’s lawyer society.

• The CIA in the 21st Century — The epic espionage chronicler Tim Weiner talks about his new book about the CIA over the past two decades, from the challenges of the Global War on Terror to its renewed war against superpower adversaries Russia and China.

• Intel, chips and America’s future — Dylan Patel of Semianalysis and I debate Trump’s partial bailout of Intel and what it means for the future of industrial policy and chips.

• What’s next for European defense autonomy — With Russia continuing to wage a brutal war on Ukraine, I talk with venture capitalist Eric Slesinger on the prognosis for defense innovation on the Old Continent.

• How Jane Jacobs got Americans stuck — The Atlantic editor Yoni Appelbaum discusses his new book Stuck on how declining geographic mobility is limiting opportunity for millions of Americans — and how to get people moving again.

• Can we ever defend against agricultural warfare? — Alicia Ellis, Arizona farmer and global security professor at ASU, talks about a new domain for conflict: agricultural security and the food supply chains that keep people nourished.

• Europe needs national champions, now — An in-depth discussion of Europe’s present place in the world from BCA Research’s chief strategist Marko Papic.

• How can we make the internet fun again? — Renée DiResta is one of the original researchers of the internet, and she’s made it her mission to find the joy again in this once optimistic medium.

• How Russia is bringing the cost of global sabotage to zero — My favorite episode of 2025 was with Daniela Richterova of King’s College London. We discuss how Russia has reduced the cost of sabotage roughly to zero by taking advantage of the gig economy, cryptocurrency and increasing polarization in the West to find solitary allies on the cheap.

That’s the end of 2025 — as I joked on the podcast, I am sure 2026 will be so quiet that we will have to shut down for a lack of content. Okay I’m kidding, but let’s just hope there isn’t a nuclear war before we all return.

Earlier this year, I talked with researcher Daniela Richterova on the Riskgaming podcast about this very subject in an episode titled, “How Russia is bringing the cost of global sabotage to zero.” In her research, she describes a Snow Crash–esque cyberpunk gig economy of would-be saboteurs and bored teenagers who pick up quick murders and propaganda work from listings posted online by terrorism franchises that pay out in cryptocurrency. The terrorists walk away all but anonymous.

At the same time, The Economist declared “the end of the rip-off economy” this week, arguing that the internet had accelerated the spread of quality and pricing information to consumers, allowing them to increase their economic surplus through better knowledge and negotiation in the marketplace.

Left unsaid in such analysis is that intelligence agencies such as Russia’s FSB and China’s Ministry of State Security have similarly been avoiding rip-offs in the freelance saboteur marketplace. It used to be extraordinarily difficult and risky to recruit an asset overseas. The process required prodigious training and careful psychological profiling of case officers, who were then sent behind enemy lines to identify leads, cultivate them over time, suss out their reserve prices for activities ranging from mundane surveillance to brazen homicide, and do it all without getting caught. Tradecraft was incredibly important, and it’s one reason why critics like grand-strategy expert Edward Luttwak are so dismissive of the CIA’s efforts (read Santi Ruiz’s excellent interview with him in Statecraft and also listen to our podcast episode with Zach Dorfman, who is writing a book on the CIA’s non-official cover clandestine officers).

Now, however, these same spy agencies can post a gig on the dark web (or apparently Snapchat), offer $50, and get a random citizen to conduct propaganda work — or worse.

It’s not just the ease with which these transactions are being conducted through social media and consummated with cryptocurrency. It’s also how fast the price has fallen. Data on costs is obviously hard to come by, but Richterova’s research and analysis of Russia’s more open recent tactics indicate that even serious sabotage or arson might cost only hundreds to low thousands of dollars. Since these gigs are handed out through what amounts to a curated auction, prices have dropped precipitously from the heights of the Cold War. There are clearly a lot of would-be saboteurs, and Russia and others have only so many targets they can blow up simultaneously (they can’t trigger all-out war, after all).

These activities are intentional. Arson can undermine the defense industrial base powering Ukraine’s army. Sabotage can induce fear among citizens, changing their behavior. Propaganda efforts can swing elections, or simply force politicians to respond to new crises and remove their gaze from overseas.

The price of construction is always legions higher than the price of destruction. Over decades, Jamaica grew into a challenged but growing economy, transitioning from a wrenching $423.70 per capita GDP in 1960 to over $7,000 last year (both present-day U.S. dollars). Six decades of work, 16x improvement. Hurricane Melissa, which swirled at an Earth-record peak speed of 185 miles per hour at landfall earlier this week, will have torn much of that growth asunder in just a few hours.

A catastrophic storm is no different than a Category 5 autocratic “chaos engineer” bent on undermining democratic polities, to use Giuliano da Empoli’s phrase. The cost of tearing apart a France or a Belgium — or even an America — is not Nvidia-scale in the trillions of dollars. The cost is cheaper than a fraudulent SPAC. Gig skullduggery is cheap, fast, efficient, flexible and globally available.

Technological innovations have ushered in efficiency around hybrid war tactics against which we at present have no defenses. People are willing to burn down factories for roughly the cost of an economy seat from New York to San Francisco. Despite the fears of privacy advocates, we don’t have a Minority Report–style crime prediction engine, even if we were to throw every dollar we have at Palantir.

The social psychologist Adam Mastroianni decried “the decline of deviancy” this week, asking “Where has all the weirdness gone?” Well, it seems to me that at least some of the weirdness is flowing right here, into scams, cryptocrimes, internet-mediated independent spywars and all the rest.

Two weeks ago, the United States seized $15 billion in cryptocurrency from a “pig butchering” scam network originating in Cambodia with deep ties to China. These organizations abduct tourists and recruit overseas workers with the promise of riches, and then have them dial for dollars across an ever-evolving series of scams targeting countries all over the world. This is a form of transnational crime that was impossible even a few years ago, one that undermines political legitimacy while also conveniently siphoning dollars from the West to the rest.

What can democratic nations do to fight back? It starts with building up immunity. That means getting a grip on inequality and a tattered social safety net that pushes desperate citizens to seek out any form of lucre, legal or illegal. It also means strengthening the rule of law, police forces and criminal prosecutions to ensure culprits are brought to justice, and swiftly. We also need to improve our surveillance systems, from funding data collection and analysis tools like Palantir to building new platforms like recently de-stealthed Valthos in biodefense to foil newly-discovered threats.

Even more importantly, though, we must respond with much better and wider education on how to be vigilant in this new era. It’s not enough to blast a 15-second pre-recorded warning to tired riders of mass transit (“See it. Say it. Sort it.”). Very few understand the complexities and tactics of state-to-state technology-enhanced terrorism and influence operations and the signs to look for. That’s one reason why most of the gig workers who sign up for these activities don’t even realize they are pawns on the global chessboard of power. That’s a critical societal weakness, but one that can be plugged with a much more robust program of education.

These immunizations need to be coupled with more direct outbound actions. Democratic countries can’t keep taking blow after blow from adversaries with nary an attempt at a hybrid-war response. That’s a breakdown of even the simplest game theory model of international relations. Instead, it’s time to augment older playbooks on undermining the authority of autocrats everywhere with a much more improvisational toolkit, one that is centered around new technologies that offer new capabilities.

We won’t undo the innovations that have spread over the past decade, but we can build new institutions to protect ourselves from their worst effects. We live in rich nations with deep resources and citizens who still — mostly — believe in the future of their countries. It’s time to take those resources and kick the vandals out.

Don’t miss

Well, here we are again.

This morning, Amazon Web Services suffered a catastrophic outage in its US-EAST-1 data center complex in Northern Virginia starting at around 3am Eastern and continuing into the morning rush. As of writing at 10am, 71 services at the data center remain offline, while 37 have been brought back online. Amazon’s US-EAST-1 is one of the most important data centers in the world, and its failure has led to outages of dozens of well-known applications according to DownDetector and social media reports.

It’s obvious that we are building our critical applications on foundations of sand; just as with South Korea, an outage in US-EAST-1 is hardly unprecedented. As The Verge notes, “AWS outages in the US-East-1 region have created widespread disruptions in 2023, 2021, and 2020, forcing multiple websites and platforms offline for several hours before regular service was restored.” Yet, developers and IT architects keep on returning to that same data center, while failing to mitigate against its predictable outages.

If you are thinking “WTF,” I don’t blame you. This is far beyond technical acumen or engineering prowess, and far more about organizational design around risks and rewards. It reminds me of Dan Davies’ book The Unaccountability Machine and his notion of “accountability sinks.” The premise is that organizations adapt themselves to ensure no one working at them can be held accountable for what the organization does. Stafford Beer’s koan that “the purpose of a system is what it does” needs a Davies corollary, namely that “no one here chose that purpose, sorry.”

In other words, rather than taking a resilient and risk-mitigating approach to engineering these systems to ensure their robustness under even complicated and unique outages, organizations have chosen cover-your-ass-at-all-costs by flocking to the same US-EAST-1 data center in mutually-assured defensibility. Don’t blame us, it’s Amazon!

In fact, that’s basically a perfect sales pitch for Amazon’s services. No one can be fired by selecting Amazon, since the alternative would be an equally incompetent employee at another organization making the exact same decision. This herd mentality induces catastrophic and correlated societal risks. As the world has become more deeply integrated, these outages and failures metastasize farther.

As the fire in South Korea and this morning’s AWS outage show, essential technical infrastructure has a habit of going offline. Cloudflare pushed out a bad patch and computers worldwide were bricked. Jeep pushed out an over-the-air update last weekend and summarily froze the Wrangler 4xe for many owners. Texas lost all power in 2021, the Suez Canal was blocked by a cargo ship, and the Panama Canal sometimes lacks sufficient water to allow vessels to pass.

These events can feel random when reading headlines, but they aren’t emergent behaviors from complex systems that are impossible to foresee. This is not the butterfly effect causing a hurricane with a flap of the wings. Every one of these crises was predictable, and therefore they are all possible to mitigate.

There is stochastic degradation of any real-world system, which means we must diligently conduct chaos maintenance to ensure stability. Yet, such maintenance is increasingly not budgeted or worse, per Davies, is just actively ignored in the hopes that no one can be held accountable for any failures.

I’m a strong nuclear energy advocate, but if you are curious why so many people don’t believe humanity could ever safely run a complex technical system that could melt down, this is why. Even the most robust organization will struggle to maintain the highest incentives for safety and risk mitigation. I reference Charles Perrow and his book Normal Accidents perhaps too much on Riskgaming, but his central thesis bears repeating. Highly-coupled and highly-complex systems will inevitably fail due to the interlocking and emerging interactions of the underlying assemblage of components. We can reduce such failures, but we can’t eliminate them.

A single computing service having an outage shouldn’t lead to push notifications from every major news service and a major concern that stock markets will crash today (that’s President Trump’s job!) We can’t allow accountability to be shorn from the critical nature of these applications to society. Hospital systems fail when servers go down, and there are patients relying on the outcome of those computations. With power comes responsibility they say, or used to. I don’t know, my social media feed is blank.

Don’t miss