Neal Stephenson on History, Spycraft, and American-Soviet Parallels

Novelist Neal Stephenson returns to Conversations with Tyler to discuss Polostan, his new historical novel set in the Soviet Union and America of the 1930s, ranging from the psychology of spies and the leak of atomic secrets through to augmented reality’s stalled economics and what current AI still cannot do.

Key ideas

1. History has moved from formal schooling to leisure-time autodidacts. Stephenson thinks formal historical education is at a low point even as podcasts and YouTube let anyone become a self-taught expert in a chosen era on their own time — part of why he returned to historical fiction with Polostan after years of science-fiction-flavoured work.
2. The best spies are psychologically unexceptional, not James Bond. Recruiters want people who disappear into the background rather than stand out; maintaining multiple identities takes a toll over time regardless of who starts the job. He doubts the leak of atomic-bomb secrets to the Soviet Union was highly contingent — the incentive to build a bomb was too strong, and the gun-type design was simple enough that duplication was likely regardless of any one spy’s success — a parallel he draws to today’s AI model weights.
3. 1930s America and the Soviet Union resembled each other more than at any other point. Both had a densely industrialised core — Chicago to the Atlantic in America, the Soviet west — set against a vast, sparsely populated, resource-extraction hinterland with a poorly treated indigenous population, alongside a shared idealisation of the frontier and the drive to industrialise.
4. Augmented reality’s problem is content economics, not hardware. Stephenson argues AR/VR investors are comfortable funding hardware but not the years and millions of dollars content creation actually requires, and that even mature LLM-driven applications will be dominated by the vastly larger market of ordinary phones and screens rather than niche headsets.
5. Art’s value comes from human connection, not just quality. Stephenson’s reason for being unmoved by AI-authored fiction is economic as much as aesthetic: an algorithm can generate infinite variations at zero marginal cost, and once supply is infinite, value collapses regardless of how well any single output reads.

Content

From speculative futures to historical fiction — why Polostan

Cowen opens by asking why Stephenson periodically swaps illustrating the future for illustrating the past. Stephenson calls it a continuum rather than a break: his science-fiction work has always carried historical callbacks, and returning to a historical setting after some years away felt ‘like going home.’ He notes a paradox of the present moment — formal historical education is probably at an all-time low, even as podcasts and YouTube videos give autodidacts unprecedented tools to become deep experts in a chosen era on their own time, outside school. Polostan’s own title, he explains, was chosen partly for search-engine practicality (publishers now reject titles that will not auto-complete cleanly) and partly because polo, the sport, recurs through the story; the joke behind the word is explained only near the book’s end.

The psychology of espionage and the atomic-secrets leak

Asked whether spies are psychologically broken people, Stephenson argues the job breaks them over time rather than selecting for brokenness from the start: maintaining multiple identities while worrying constantly about exposure is ‘not a natural way for people to live.’ The best operatives, historically, were unexceptional and unmemorable — not the James Bond archetype but people who disappear into the background. He recounts a story, of uncertain accuracy, in which American intelligence services once tried recruiting deliberately blemish-free ‘straight arrow’ types to reduce blackmail risk, only for the British to grow frustrated dealing with such naive, small-town characters — one reason, Stephenson jokes, the CIA drew disproportionately on Mormons. On whether the leak of atomic-bomb secrets to the Soviets was highly contingent, Stephenson is sceptical: the incentive to build a bomb was overwhelming, and the simpler gun-type design (as opposed to an implosion device) would likely have been reproduced with or without any particular spy’s success. Cowen extends the question to AI: Stephenson observes that whatever counts as today’s equivalent secret — trade secrets, or the weights produced by training runs costing billions of dollars — is something labs will fight hard to keep, precisely because it was so expensive to produce.

Faking death, and writing by hand

A lighter detour on staging one’s own death turns serious quickly: Stephenson notes that ubiquitous cameras and DNA testing have made total physical disappearance far harder than a century ago, leaving something like a small-plane crash with a mid-air parachute exit as one of the few remaining routes — and even then, the kind of underwater search effort mounted for the Titanic shows how hard it now is to guarantee a body is never found. On craft, Stephenson confirms he still writes novels in longhand on fancy imported Italian paper, not because it is economical (it isn’t) but because typing lets ideas hit the screen almost as fast as they occur, denying a sentence the chance to sit ‘in the buffer’ where it is easy to improve or discard while still ‘between your ears.’ Once a sentence is on the screen it acquires a kind of permanence that takes more effort to undo; editing on paper — a box around a paragraph, an X through a line — is also, he says, simply faster than editing with a mouse.

Soviet communism, Western infatuation, and the Chicago–Magnitogorsk parallel

Cowen presses on why so many Western intellectuals became enthusiastic, not merely tolerant, about Stalinism, often even after visiting the USSR themselves. Polostan’s protagonist is the daughter of a Russian father and an American mother, both committed to leftist politics in the 1920s — a generation of Westerners who saw the early Soviet project as a chance to build something great, before Stalinism’s full excesses were visible, even as the Cheka (the KGB’s precursor) was already committing abuses visible to anyone willing to look. Stephenson cites Malcolm Muggeridge’s 1930s roman-à-clef Winter in Moscow as a portrait of Western expatriates who knew about the Ukrainian famine and refused to see any downside to the revolution regardless. Asked what he most admires about the USSR, Stephenson’s answer is narrowly technical, not moral: excellent rocket engines, and the sheer speed with which Stalin’s industrial drive of the 1930s dragged a nineteenth-century country into the twentieth by building an enormous industrial base — ‘I’m not taking into account the body count.’ He also credits the Soviet Union with genuine, if instrumentalised, progress on women’s rights (education, medical careers, military roles, easier access to abortion and divorce), which he thinks was itself a draw for many Western leftists. Asked about a parallel between 1930s America and the USSR, Stephenson agrees the two countries resembled each other unusually closely at that moment: each had one densely industrialised, urbanised zone (Chicago to the Atlantic; the Soviet west) set against a vast, largely empty hinterland with a poorly treated indigenous population, functioning as a resource-extraction zone feeding the industrial core — alongside a shared idealisation of the frontier and the wide-open wilderness even while both raced to industrialise and urbanise.

Augmented reality’s stalled content economics

Turning to current technology, Stephenson argues AR and VR — Meta’s Orion, Apple’s Vision Pro — have not moved meaningfully past what Magic Leap was already attempting years ago, because nobody has solved the content problem. Building for AR is a much harder problem than ordinary game development, itself no small undertaking, because AR content must account for far more real-world variables than VR. Hardware-focused investors, in his account, are comfortable funding devices but not the years and many millions of dollars content creation actually costs, operating on a ‘build it and they will come’ assumption that creators will simply appear once hardware exists. Pressed by Cowen on whether large language models finally supply the missing organic connection — an AI companion that recognises people at a party, or flags sale items at the supermarket — Stephenson agrees LLMs work and are promising (he is using one in a current project) but sees no obvious structural link between LLMs and AR specifically: any application built for a market of a few thousand headset owners will always be economically dominated by the same application built for the billions of people who already carry a phone. On group chats replacing open social media, he thinks anything that removes the ‘glare’ of a big public platform and lets people speak among people they know, without constantly fending off bots, is an improvement over Twitter. On rising domestic surveillance, he is struck less by its scale than by how little concern it seems to generate, and by the historical accident that the open internet had a head start of some decades before authoritarian states — China, Russia, North Korea — learned to wall it off and turn the same surveillance tools naively provided by Western technology against their own populations.

AI, art, and the primer from The Diamond Age

On his own influence over technologists, Stephenson’s approach to morally fraught subjects — solar geoengineering in Termination Shock, most recently — is to present arguments on both sides so no reader can accuse him of one-sided propaganda, and then let people draw their own conclusions. Asked when a human-AI co-authored novel might match a talented human’s best work, he reframes the question: something close probably already exists, but the real reason people undervalue AI-generated art is not quality but economics and connection. Art matters because it exposes readers to ‘a very dense package of micro decisions’ made by a specific person, living or dead — a human-to-human communion. Because a script can generate a million more variations at essentially zero cost, ordinary supply and demand collapses the value of any one of them to near zero, regardless of whether it reads well; concealing AI authorship to manufacture a fake human connection strikes him as possible but uninteresting, given how many human-written books he has not yet read. On current LLMs’ biggest weakness, Stephenson points to their lack of spatial and basic logical understanding of the physical world, illustrated by their tendency to answer arithmetic confidently and wrongly rather than admitting the limits of the architecture. He is doubtful that reinforcement learning on robot-generated or synthetic driving-simulation data will fix this within an LLM specifically — that looks to him like a different kind of AI system entirely, even if it can be trained cheaply and effectively for tasks like self-driving. Asked whether The Diamond Age’s ‘Primer’ — a device that teaches a child almost anything — already exists via LLMs, Stephenson credits language models with solving the knowledge-delivery half of the problem, but not the harder half: the special relationship between learner and object that makes a physical, tailored book compelling in a way that querying a chatbot on the internet still is not. Closing on AGI and IQ, Stephenson is unconvinced abundant machine intelligence changes much on its own — plenty of high-IQ people fell for ‘ridiculous bullshit’ political movements in the twentieth century — and when pressed on what is actually scarce, half-jokingly nominates kindness, before noting that the same self-deceiving belief in one’s own kindness is exactly what let earlier generations excuse the worst political movements they supported. He confirms Polostan opens a planned trilogy, ‘Bomb Light,’ that will follow different characters (including one barely glimpsed in this book) through the 1930s physics discoveries — Fermi, the Curies, Compton, Millikan, Rutherford, Chadwick, working out the universe’s mysteries by deduction and comparatively simple experiments — that fed into the Manhattan Project, converging with the Nazi rise, Stalinism, and the Spanish Civil War by the 1940s.

See also

• Neal Stephenson — speaker
• Tyler Cowen — host
• Polostan — the 2024 novel and ‘Bomb Light’ trilogy discussed throughout