He Bought an Android, Two SIMs, and Left Claude Watching His Inbox

Cognitive Revolution host Nathan Labenz is back after two weeks in China, the longest gap in the feed since he started the show almost three and a half years ago. He had meant to record from the road, but a business visa — not a journalism visa — and a packed calendar pushed the telling home. This is the first of three episodes: tech and everyday AI, then Chinese AI safety and governance, then the U.S.–China relationship as capabilities accelerate. Across all three he is observing Chatham House rules, not because every interlocutor demanded it, but so no one is put in the position of speaking for an institution.

He is candid about the sample. Almost everyone he met spoke excellent English, which marks a particular class, and everything still runs through his own lens. He cites an old foreign-press-corps joke: stay a week and you think you know a lot; stay a month and you know a little; stay a year and you know nothing. He paid his own way; the only gifts were a handful of meals he sometimes could not pick up the check for. Hospitality, he says, was excellent — which itself could be a bias.

Search and chatbots were unusually unhelpful on the practical questions. Friends were worse in a different way: they contradicted one another. A colleague at a major international firm said he goes constantly, the company has an R&D center in China, and nobody takes precautions. Asked about IP, the colleague flipped the usual Western anxiety: Chinese competitors are formidable, and "if anything, maybe we should be trying to steal their IP." National-security friends, including people who had held senior U.S. government jobs, demanded the opposite. Burner phone, all-new accounts, never connect the device to home Wi-Fi, and assume anything that enters China is compromised forever.

Labenz did not have trade secrets, and he talks about most of his work in public. The national-security argument that moved him was not that he was a target. It was that he is connected to people at frontier labs, and a compromised device could become an attack vector. He never felt the precautions were necessary on the ground. He still kept the promise and built a parallel stack.

Entry looked simpler online than it felt. Stays of ten days or less can now be visa-free, but he did not want to fly around the world and then be asked if he was a journalist with a podcast. He applied for a business visa with an invitation letter from a host he will not name. A local expediter filled the forms, hand-carried the passport to the Chicago consulate, and returned it in eight to ten days. The result was a ten-year visa allowing stays of up to ninety days. One appointment, half an hour of paperwork.

He also bought a one-way ticket from Detroit through Washington to Beijing. After his son's cancer treatment, he did not want a return date locking him overseas if bad news arrived. Internet lore said a one-way ticket can cause trouble at immigration; with the business visa, it did not. He figured out the itinerary on the fly once he was there.

He bought an Android, mostly to try one after years on iPhone. In hindsight that may not have been optimal. Apple operates in China: Apple Maps is well regarded, and the App Store works. Google Pixels are rare. The larger surprise was that a foreign cellular plan made the Great Firewall almost irrelevant. Labenz's model — he is not an infrastructure expert — is that the phone attaches to a local tower, but requests go home to the U.S. carrier, out to the open internet, and back along the same path. Filters designed for domestic traffic never see it.

He ran two numbers on dual eSIMs: a one-month Google Fi line with roaming, plus his ordinary T-Mobile number moved onto the burner, also with a month of international roaming. He could send and receive on both. A China-oriented VPN with its own eSIM sat in reserve and went almost unused. He created a travel Gmail, signed up for Chinese services on that identity, and still found Google Play available over roaming — contrary to the advice that you must download everything before you land. He still recommends finishing the stack at home so you are not fumbling in an airport at 3 a.m.

The must-have apps were few. WeChat is the super app for social, business, and payments; Alipay is the other payments network. Both now accept international credit cards. Didi is Uber with a twist: on smaller fares you get a quote, ride, then pay after, so the next trip sometimes waits on the last bill. Trip.com is the tourist platform for hotels, trains, flights, and attractions. Meituan covers food and fast delivery. WeChat, Didi, and Trip.com have English interfaces; Alipay offered a floating translate button; Meituan did not, so he screenshot heavily into Google Lens.

Cards loaded into WeChat and Alipay, but person-to-person transfers and other in-app money moves were blocked for international cards. He could not test a live payment before takeoff. Landing in Beijing at 3 a.m., his first errand was a self-serve kiosk: prove the stack works before leaving the airport. China, he found, is effectively cashless. He never saw a bill in ordinary life. When he wanted souvenir notes for his kids, a lunch companion had to raid an office desk drawer.

The juice was under two dollars for something that would be six to eight at a U.S. airport. There was no clerk. He scanned the bottle, stalled, then realized he had to present his payment QR code to the machine's camera. A beep, "processing," and two seconds later it was done. Online he had read that many shops refuse cash; on the ground he was told the law still requires merchants to take it. Using cash would still be a different, bumpier experience, especially at automated gates.

QR codes were everywhere else too. Restaurants put them on the table: scan for the menu, order in the app, pay in WeChat or Alipay, never speak to a server. Adding a WeChat contact is a mutual scan so fast the code need not even fill the frame. Trip.com tickets live in the app; you flash a code at the gate. When heavy rain closed the Great Wall, Trip.com pushed a notice and refunded the ticket automatically. The economy felt formal: posted prices, almost no haggling, almost no hard sell, even at tourist souvenir stalls.

He left his main Claude and ChatGPT accounts at home as part of the alter-ego plan, and left the Mac Mini running. Claude already had Gmail access — a setup he has described in earlier episodes, including one with Daniel Meisler. Before departure he told it to monitor communications, send a digest a couple of times a day, flag anything urgent, and take instructions from the travel inbox. He also stood up Outlook, because Microsoft products are not blocked the way Gmail is, and never needed it. Roaming kept Gmail open the entire trip.

Claude wanted a password in every inbound email. Labenz refused: if the burner were compromised, the password would be sitting in the same mailbox. He told the model to apply a sensible filter — reasonable, low-risk requests that sound like him can run; otherwise issue a challenge only he would know. The challenge never fired. He had Claude reply from his real address so messages would not fall out of priority inboxes, but introduce itself as Claude. Later he told it to skip the "I am an AI assistant" preamble; in his circles, people already know.

Gmail auth expired twice. Claude would not refresh the token through the browser on its own, so he asked his wife to open a new terminal session on the desk machine, reauthorize, and click through — about two minutes. For AI-industry travelers he now thinks the isolated profile is cheap insurance: a burner, two lines, a VPN backup, and an agent on a machine that stays home. For his retired parents he would skip the theater. Second number for redundancy, WeChat, Alipay, Trip.com, Didi, enough roaming data. Do not lose the phone.

Crosstown Didi rides of thirty or forty minutes often landed in the six-to-eight-dollar range — in many U.S. cities, Labenz says, close to ten times that. Pickup in Beijing and Shanghai felt as liquid as Uber. The detail that stayed with him was the traffic light. When the assigned car sat at a red, the app showed the countdown to green. The signal is on a central system Didi can query. In one hotel, inserting the key card woke the TV, which displayed the live status of free washers and dryers downstairs. The property had no laundry service; it had wired the machines into the room.

The most practical demo was medical. After a sudden allergic reaction in the night, a companion ordered Claritin on Meituan to an address twenty minutes ahead on their bus route. The box was waiting when they arrived. Outlets in hotels took U.S. two- and three-prong plugs; he never packed an adapter. Combine the prices, the plumbing, and the lack of friction, and he offers a litmus test. Hide the history and the price tags, park someone in New York or San Francisco for a week and Beijing or Shanghai for a week, and ask which country is richer. He thinks an overwhelming majority would pick China. He is talking about the major cities he saw, though he was told the stack has spread farther than outsiders assume.

As a tourist city he puts Beijing with Paris: enough deep history and architecture for an easy week. He went to the Great Wall and still met almost no aggressive selling. Translation is a solved social protocol. Drivers, hotel desks, and shop staff record a sentence and flip the phone around; he answers the same way. Twenty or thirty years ago, he says, you would have been immersed and constantly negotiating in broken language. That world is gone. The remaining failure mode is simple: lose the phone.

He spent the first days sightseeing before WAIC, mostly without a human guide, on the free tiers of DeepSeek, Kimi, and MiniMax. Timing mattered. President Xi opened the conference, and model companies clustered releases around it; Kimi K3 landed near the event, so the tourism days ran on Kimi 2.6. For explaining a temple, he argues, you do not need frontier intelligence, long-horizon agents, or cyber capability. You need a fast, decent explainer. At the Lama Temple — a Tibetan Buddhist complex in central Beijing that surprised him — the apps supplied history he did not have. At the Temple of Confucius and the Imperial College he asked why a philosopher who is not a god has a temple. The answer, about reverence without worship and ritual as an imperial duty, underlined how poorly Western church-and-state categories fit.

Refusals still showed up. He was not probing the "three T's" — Tibet, Taiwan, Tiananmen — for sport; earnest questions about a Tibetan temple or the Manchu Qing dynasty were enough. Sometimes a chain of thought and a partial answer appeared, then the text vanished into an apology: the Sydney-Bing pullback, he says, in a Chinese product. Not every app censored the same way. That matches reporting he has seen that the base models are less censored than the deployed apps, which are trying to stay on the right side of regulators. Over-refusal, he notes, is not only an Anthropic problem.

Quirks differed. DeepSeek expert mode, on the best model, blocked file attachments and search; instant mode did not. Kimi kept search on the strongest model. MiniMax felt the most "down the fairway" and, in his use, refused the least. For taste he preferred Kimi, the most Claude-like of the three — which raises distillation questions he will save for a later episode. He thinks Moonshot's aesthetic is more than cloning: the name nods to Pink Floyd's *Dark Side of the Moon*, and the company takes the look of its models seriously. When the hives hit at 3 a.m., though, he still downloaded ChatGPT without logging into his home account. He wanted the health product he trusts from covering OpenAI's work, including a conversation with Karen, who leads health there. The reaction passed while he chatted; the tell is that in a possible emergency he reached for the Western app.

People were, in his accounting, extraordinarily kind. He had essentially zero ugly interactions and only a couple of mildly pushy bike-taxi pitches. It is not a tipping culture: WeChat and Didi checkouts never ask if you want to add ten percent, and service at hotels was still high. Transaction fees help explain the hospitality hacks. He was told Chinese payment costs run around a quarter of a percent, versus two to three percent on U.S. cards — no rewards, and WeChat Pay is a debit against a known balance, not a monthly tab. Hotel staff could order Meituan, buy soccer tickets, or hire a Didi by the hour on their own apps, ping him an invoice, and lose almost nothing on the pass-through. The English Didi app lacked hourly hire; a concierge booked a nine-hour driver for about $120 to $125, created a WeChat group with the driver, and kept translating after Labenz had already checked out.

Huawei and Xiaomi stores were uncannily Apple: the layout, the devices, even the lighting — plus cars on the floor, as if the Apple car had shipped. Founders get a mythic treatment that reminded him of Bezos, Zuckerberg, and especially Elon: Huawei's founder, DeepSeek's, Zhipu's. Confucius, 79 generations of named descendants later, is the original version; Mao is still on the cash. WAIC at the Shanghai Expo Center felt, in scale and booth aesthetic, like CES or NVIDIA GTC. If you opened your eyes, he says, you might think you were in Las Vegas. The floor was remarkable mostly for being unremarkable: the companies you expect, showing the things you expect.

Robots were more visible in daily life than in the United States — meal-delivery units in a hotel, R2-D2-like sweepers in parks — but the tasks looked simple. At WAIC, Tencent offered a one-to-two-minute pressure-point massage, the first time a robot had really touched him; it had real force, and he would still take a human. Many cooking and magic demos were paused. A magician who had spent six weeks teaching a robot a trick said he was not worried about his job, and the trick was scheduled once. A humanoid walking down a hall was teleoperated; the operator shook her head when Labenz offered a handshake. Hardware and government priority are real, he thinks. General-purpose autonomy still looked like an American lead, and a possible complement rather than a mirror.

A LinkedIn connection at a Chinese hyperscaler invited him to lunch. The cafeteria — guest badge, overloaded food stations, long tables — felt exactly like Google, which he does not think is an accident. That evening the same host walked him into a startup mixer the company was throwing: early founders pitching, free space and food, maybe cloud credits, almost no alcohol. Clean living, he notes, now rhymes with Silicon Valley. A tech-transfer office described itself as sitting between universities (zero to one) and companies (ten to a hundred), doing the one-to-ten check that a technology can scale and that a market wants it — language that may nod at Peter Thiel, and a structure that felt familiar.

The political difference that struck him is the "mayor economy." Local officials are not, in general, competing for local votes. They compete with peer cities to please party superiors by delivering growth. That, he argues, makes them more reliably pro-business when Beijing names a priority such as EVs or humanoid robots. Entrepreneurs still start the firms, but mayors want *their* local maker to be one of the survivors. He contrasts that with U.S. mayors who can win by attacking a president or a celebrity founder, and with New York rhetoric around finance that is hardly a love letter to the city's biggest taxpayers. China is hyper-capitalist and hyper-competitive in the new industries; the officials know most of the new firms will die, so they try to keep theirs alive.

Daily life had other tells. Beijing, especially, is dense with clean, free public toilets — a national priority at some point — in a way New York and San Francisco are not. Cameras are constant, indoors and out; as a visitor you should assume you are on camera anywhere but a private room. High-speed rail and the Forbidden City required passport checks; some parks charged a token two yuan, perhaps two dollars a year for an annual pass, which he wondered might be another logging function. The upside is a public attitude that there is basically no crime. He saw children out at 11 p.m. Women and families, he was told, move at night with a comfort he would not assume in New York. Lose a phone in a cab, go to one of the many police posts, and the story he heard repeatedly is: pull the camera, read the plate, call the driver, phone back in an hour or two. He never heard of a failure.

On a recent *Hard Fork* discussion, Labenz recalls the worry that Western law assumes not every violation is caught; AI reviewing everything would suddenly tighten a loose system. China is already tight. People expect to be watched and not to have digital privacy. Crime is low. The abuse potential, he thinks, is high, and much of the risk may be stored in the future. Asked what happens when President Xi retires or dies, nobody he met could name a successor. A power struggle plus this apparatus could be ugly. For now, few people step out of line, so little force is required, and the lived experience is comfortable. Brazil, he notes, is adding similar network surveillance against phone theft, and the statistics appear to be moving. Every society, with AI watching tape at scale, will have to pick a point on the spectrum. China has picked.

A separate observation, one a counterpart called astute: Chinese society seemed more comfortable with contradiction than the Western habit of interrogating inconsistency until it resolves. People might participate in Christian and Buddhist practice and still call themselves atheist. He points back to a ChinaTalk episode with David Kang: if the Westphalian binary of sovereign-or-not had not been imposed, a China–Taiwan arrangement of "independent, and also part of one China" might have been livable. VPNs are illegal, and everyone he met uses one; Chinese AI researchers are all over Twitter. Some companies issue an official corporate VPN. Getting in trouble for it, someone told him, would be like getting struck by lightning — possible if you are already a problem, not a worry for ordinary use.

Censorship drew a similar double move: complaint, circumvention, then a paternalistic defense. Worldly people can handle unfiltered information; maybe the broader public cannot, and panic is worth preventing. Drones are a national "low-altitude economy" priority and banned in Beijing; a professor at lunch freely called the permitting catch-22 ridiculous, then added that he understood why the capital might not want drones overhead. At the Imperial College museum, the civil-service exam exhibit showed centuries of extreme fairness — three-day stints in a cell, nine days of testing, scribes recopying papers so handwriting could not identify a candidate — beside a long history of corruption once the post was won. Leaving tensions unresolved, he argues, is not unique to China, but the comfort with it may create wiggle room in future U.S. deals that a demand for full coherence would close.

The one human guide he hired, at the Forbidden City, lit up when she understood he worked in AI: "It's a good job." The product she actually used was Doubao, ByteDance's assistant — by several accounts the most broadly diffused AI in Chinese society, on the order of 150 million users, voice-native, companion-like. DeepSeek excitement, he was told, is more of a Western story; Qwen models are for Western startups fine-tuning on price, not a retail habit at home. Parents use Doubao. Two women in their late twenties or thirties reported mothers asking why they could not be as sweet as the bot. The guide's line was that some things you cannot tell anyone, but you can tell Doubao and nobody will judge you — not the model, and not society for using it. Her example was wanting to leave a boyfriend. Before Doubao? "Just keep it to yourself." He could not download Doubao or its international twin DOLA from a U.S.-registered store; one of the few times he turned on the VPN was to look like he was in Singapore.

Headline surveys that China is more optimistic about AI than the West may overstate the mood. Several people said the economy feels shaky and jobs feel contestable. Some of the enthusiasm is the old line that AI will not replace you — someone using AI will — so you adopt in self-defense. He heard, from a seemingly credible English-speaking source, of a Didi driver who used to be a software engineer and lost the job "for AI reasons." He cannot vouch for the anecdote; he treats it as a weather vane for anxiety. Nobody is starving. Nobody wants to be the one who did not get on the tool.

Diffusion, as he saw it, is happening through platforms more than through a viral OpenClaw-style agent. WeChat is building a native agent now in late-stage beta. ByteDance tried a super-agent that would drive other super apps, including WeChat; WeChat blocked it. Because appointments, payments, and social life already live in one place, a WeChat agent would have context and spend authority no Western app can match on day one. Labenz calls the forthcoming launch potentially one of the most important in agent history — if Tencent has the compute to serve a billion-plus users, which he heard conflicting reports about. Ant Group, Alipay's parent, is putting payment profits into an AI doctor, with the same free, non-threatening, don't-replace-your-physician posture as ChatGPT or Claude. The aging math is brutal: two generations of only children means one young adult facing two parents and four grandparents. People talk about robots for elder care in family conversations, not only on expo floors. When those firms decide a product is ready, they can push it to phones that almost everyone already has. That lever, he thinks, will be hard for the West to copy.