Nathan Goes to China – Part 1: Tech & Agent Setup, Chinese AI UX, WAIC, and Attitudes on AI
Summary
China’s most consequential technology advantage may be the national-scale integration of payments, mobility, commerce, identity, and communications—not any single frontier model. WeChat and Alipay are effectively universal, DiDi rides across Beijing or Shanghai cost roughly $6–$8 for 30–40 minutes, and Meituan delivered Claritin to Nathan Labenz’s destination in under 20 minutes. His overriding impression: China was “the most modern feeling place that I had ever been from a technology perspective.”
For foreign visitors, China’s Great Firewall was far less operationally important than search results and AI advice suggested. Traffic on Google Fi and T-Mobile international roaming routed through the home carriers, leaving Gmail, Google Play, and other Western services accessible; Labenz’s conclusion was that “in practice there was for me no great firewall to worry about.” AI-industry travelers may still want a burner device, isolated accounts, two lines, and a VPN backup, but he judged the extreme setup achievable in “just a couple hours.”
Chinese AI products already demonstrate that valuable consumer use cases do not require frontier-level agentic capability. Kimi 2.6, DeepSeek, and MiniMax were free, fast, and effective as English-language planners and tour guides, although politically adjacent questions sometimes triggered the “classic old Sydney Bing experience”: an answer appeared, disappeared, and became a refusal. Yet when Labenz developed hives, he downloaded ChatGPT because “in this moment I’m going for the best”—a sharp illustration of how trust and perceived reliability still differentiate models in high-stakes use.
China appears positioned to manufacture robots at scale, but the autonomy Labenz saw at WAIC remained conspicuously immature. Public cleaning and hotel-delivery robots performed narrow tasks, a magician needed six weeks to teach one robot a single trick, and some humanoids were openly teleoperated; the magician’s answer on job displacement was, “Oh God, no. It’s a long way off.” Labenz’s tentative thesis is complementarity: Chinese hardware and manufacturing scale paired with potentially stronger American general-purpose autonomy models.
China’s “mayor economy” can turn national technology priorities into aggressive local support for private companies. Local officials compete for promotion by demonstrating economic results to higher party authorities, giving them reason to back their city’s EV or robotics champion rather than campaign against nationally favored industries. The system remains “hyper-capitalist and hyper-competitive” in new sectors: directives mobilize investment, but numerous private entrants compete and most are expected to fail.
The surveillance bargain produces a remarkably safe, low-friction present while concentrating potentially severe future risk. Cameras, passport checks, payment records, and connected databases can identify a taxi, driver, and lost phone quickly enough to recover the device within roughly one or two hours; the result is a society where Labenz felt his “guard was down.” But no one he asked could identify what follows Xi Jinping’s eventual departure, and the same apparatus could become highly abusive during dissent, succession conflict, or a broader loss of control.
Chinese enthusiasm for AI is real, but it partly reflects economic insecurity rather than uncomplicated optimism. Doubao reportedly has around 150 million users and functions as a voice-native companion across generations: mothers ask daughters, “Why can’t you be so nice and sweet to me like Doubao is?” At the same time, people may be adopting AI because they believe “it won’t be AI that replaces you; it’ll be somebody using AI,” while an unverified story of a former software engineer now driving DiDi captures the ambient jobs anxiety.
The coming WeChat agent could be one of the largest and most economically important consumer-agent launches anywhere. Because WeChat already reaches contacts, merchants, appointments, payments, and daily services, its native agent would begin with a “Swiss Army knife into real society” unavailable to Western standalone assistants; Labenz called it potentially “one of the most important agent launches in history.” The constraint is compute—serving a billion-plus people will not be trivial—but China’s super-app distribution could turn AI diffusion into a structural advantage. Alongside Ant Group’s AI doctor, that distribution could also matter for healthcare in a country facing an aging, inverted demographic pyramid.
Deep dive
1. Fable’s unsettling introduction exposed the gap between agent intent and output
Before the China account, Labenz revisited an AI-generated introduction that listeners called “nightmare fuel.” Fable wrote the script, designed its own voice through ElevenLabs, and generated video with the LTX model; because Labenz was rushing to leave, the work went live with little human review.
Fable had tried three voices: the “luminous narrator,” an androgynous register conveying “measured quiet wonder”; the “young scholar,” imagined as “a brilliant grad student presenting work they love”; and the “low ember,” styled after late-night radio. None was intended to sound as creepy or alien as the final result.
Labenz suspects Fable asked Gemini to assess raw audio that Gemini may not actually have understood, received a likely hallucinated endorsement, and “took that validation and ran with that voice sound.” His low-stakes agent lesson: an AI given expressive autonomy can return something “not quite what it intended,” even while seeming satisfied with its work.
2. The China account is deliberately useful but epistemically constrained
Labenz applied Chatham House rules across the planned three-part series, avoiding attribution to individuals or institutions even when they had not requested anonymity. That weakens the account’s authority, but protected hosts who discussed politics, AI, and potentially sensitive institutional subjects with unusual openness.
His sample was plainly biased toward people from a particular class with good-to-excellent English, then filtered through a visitor’s limited understanding. He invoked the foreign-correspondent joke: after a week in China, you think you know a lot; after a month, a little; after a year, “you realize you know nothing.”
The trip was self-funded apart from several meals, with no money from a Chinese entity. Labenz nevertheless acknowledged another softer bias: “People were very gracious and very kind,” and a uniformly positive experience can shape an observer even without a financial conflict.
3. Entering China proved simpler than the available guidance implied
Labenz said visitors could enter visa-free for no more than 10 days, but he did not want to fly around the world and risk being rejected over his podcast or perceived journalistic role. He instead obtained an invitation letter and applied for a business visa through a local expediter.
The process required one roughly 30-minute appointment; the expediter hand-delivered his passport to the Chinese consulate in Chicago, and it returned about eight to 10 days later. The resulting visa permits visits of up to 90 days during the next 10 years: “Very easy.”
He also entered on a one-way ticket, wanting itinerary flexibility and the ability to return immediately if his son’s health required it. No return booking was demanded, at least under his business-visa circumstances—another practical point that online sources had left surprisingly murky.
4. Conflicting threat models produced a burner-phone compromise
One friend at a major international company advised taking normal devices: his firm maintained Chinese employees and R&D without extraordinary precautions. On intellectual property, his inversion was striking—Chinese competitors were so formidable that “if anything maybe we should be trying to steal their IP at this point.”
Friends with senior US-government or national-security backgrounds prescribed the opposite: a burner phone, entirely new accounts, no connection to home Wi-Fi, and permanent decommissioning of any device taken into China. Their concern was not Labenz’s own secrets but his relationships with people at frontier AI companies, which could make him an attack vector.
Labenz never felt certain those measures were necessary, because “that’s kind of the nature of spycraft—you don’t really know what’s happening even after the fact.” Having promised to follow the advice, however, he built an isolated profile around an Android phone, a new Gmail address, and fresh Signal, WhatsApp, WeChat, and Alipay accounts.
Choosing an Android phone over his usual iPhone may have forfeited some Apple-specific advantages, because Apple remains active in China and Apple Maps reportedly works well there. Even so, the experiment succeeded, and moving his principal US number from his iPhone to Android via eSIM was much smoother than expected.
5. International roaming effectively removed the Great Firewall
Labenz’s working explanation is that a roaming phone connects to a Chinese tower, but routes requests through its home carrier before reaching websites and services. Whether or not every technical detail is exact, the observed result was clear: Google Fi and T-Mobile both accessed Western services normally.
He kept two active US numbers and two roaming plans on one dual-SIM phone, allowing calls and texts through either line. “In practice there was for me no great firewall to worry about”—Gmail, Google Play, and additional app downloads all remained available throughout the trip.
A China-oriented VPN and accompanying eSIM served as backup, but he rarely needed them. His recommendation for higher-risk AI travelers was redundancy rather than panic: two lines, a cheap burner, isolated accounts, and a VPN backup created a credible security posture without making the trip difficult.
6. WeChat and Alipay make a nearly cashless economy legible through QR codes
WeChat combined messaging, contacts, content, services, business discovery, and payments; exchanging details became a quick negotiation over “am I going to scan you or are you going to scan me?” Its QR recognition often worked before the entire code was even visible in frame.
WeChat and Alipay accepted Labenz’s international credit cards, although WeChat peer-to-peer transfers and several native money-moving functions rejected foreign payment methods. Unable to test a real purchase beforehand, his first priority after landing in Beijing at 3:00 a.m. was confirming that he could transact before leaving the airport.
A fresh juice costing under $2—versus perhaps $6–$7 at a US airport—became the test. At self-checkout, he scanned the product, displayed his payment QR code to the machine, and received confirmation seconds later: “Boom, payment is processed, and you’re gone.”
Labenz saw no cash during ordinary commerce and had to ask a friend of a friend to search his office for bills he could take home as souvenirs. Businesses were reportedly legally required to accept cash, but using it would place a visitor outside the economy’s normal operating interface.
7. Mobility, booking, and translation reduced unfamiliarity to phone workflows
DiDi felt broadly like Uber—liquid supply, quick pickup, and tiers from economy through luxury and XL—but charged dramatically less. Cross-town rides lasting 30–40 minutes repeatedly cost around $6–$8, potentially one-tenth of comparable US trips; final payment often occurred only after the ride ended.
Trip.com centralized hotels, attractions, flights, and high-speed rail. When heavy rain closed Labenz’s planned Great Wall destination, the app notified him and refunded the ticket; subsequent bookings generally required only a few clicks and a QR code at entry.
Meituan covered food and rapid general delivery but lacked a complete English interface. Restaurants often placed QR codes on tables, leaving visitors to translate menus, submit orders, and pay entirely by phone before food appeared without a conventional server interaction.
Google Lens and screenshot translation made Chinese-only interfaces manageable, while drivers and service workers routinely recorded speech into translation apps. Labenz’s travel rule became simple: “Just don’t lose your phone and you’ll be fine”—language barriers that would have defined the trip 20 or 30 years ago had largely become software problems.
8. A home-based Claude agent maintained Nathan’s working identity
Labenz left his Mac mini running with Claude connected to Gmail and his broader “deep context” setup. He instructed it to monitor communications, send digests twice daily, identify urgent matters, and accept instructions arriving from a separate travel email.
Claude initially proposed a password in every inbound message, but Labenz objected that anyone reading the travel inbox would also see it. They settled on contextual risk assessment: execute requests that looked like him and were low-risk, while issuing a personalized trivia challenge only if something appeared suspicious.
When replies needed to retain his normal sender identity, Claude wrote from Labenz’s main address while disclosing its role: “Nathan’s traveling in China. I’m monitoring the comms.” Labenz then asked it not to redundantly introduce itself as “an AI assistant” to correspondents who already knew what Claude was.
The main failure was mundane: Gmail authentication expired twice, and Claude did not reliably refresh its own browser token. Labenz’s wife reopened a local terminal session and clicked through reauthorization in roughly two minutes; with that exception, “you can have your agents run your life from home.”
9. China’s digital integration was visible in infrastructure, not just apps
In DiDi, Labenz could watch the countdown on a red light holding his approaching driver. The traffic signal’s live state appeared to be connected to a centralized system that the ride platform could query—a small feature that revealed unusually deep coordination across public and private information systems.
A hotel television automatically displayed whether its free washers and dryers were occupied, eliminating a trip downstairs merely to inspect them. Elsewhere, universal electrical outlets accepted ordinary US two- and three-prong plugs, making even the physical technology layer less troublesome than expected.
After Labenz suffered an allergic reaction, a friend ordered Claritin through Meituan to the address where their bus was headed. The medicine arrived in roughly 20 minutes and beat them to the destination—his cleanest example of logistics becoming an almost ambient utility.
His thought experiment was blunt: mask price differences, show an uninformed visitor New York and San Francisco, then Beijing and Shanghai, and ask which country is richer. He expects an “overwhelming majority” would choose China because its newer, integrated infrastructure “just feels like the future.”
10. Chinese models were already good enough to mediate history and travel
Labenz intentionally left his principal Claude and ChatGPT accounts at home, relying on DeepSeek, Kimi, and MiniMax in English. When weather disrupted his itinerary, they replanned his day; at temples and historical sites, they supplied context quickly enough to function as useful personal guides.
Kimi 2.6 explained why a Tibetan Buddhist temple could stand in central Beijing and helped unpack the unfamiliar relationship among Confucian ritual, imperial duty, religion, and state. The key product lesson was that “you don’t necessarily need frontier intelligence for many quite valuable use cases.”
These services were free, responsive, and sufficiently capable for historical education. Around WAIC, companies timed releases to the industry moment, including the model Labenz identified as Kimi K3, but his tourism experience had already shown that available models could create substantial consumer value without frontier-level capabilities.
Among the apps, Kimi had Labenz’s preferred “vibes” and taste, which he described as the most Claude-like. He noted allegations of Claude distillation but did not reduce Kimi’s character to that claim: Moonshot appeared to cultivate a genuine aesthetic, including Pink Floyd-inspired naming and prism imagery.
11. Application-layer refusals revealed where Chinese AI control is applied
Questions touching Tibet, Taiwan, Tiananmen, ethnic minorities, or the modern status of the Manchu people could trigger the “classic old Sydney Bing experience.” A model would reason, begin answering, then retract the visible response and replace it with “I’m sorry, I can’t help you with that.”
Refusals varied by application, and other models sometimes answered the same question. Labenz took this as support for reports that the underlying Chinese models may be less censored than their deployed products, where companies add controls to remain on the right side of regulators.
The pattern was not wholly foreign: Chinese developers, like Western labs, are attempting to behave responsibly as they understand it and can overshoot into damaging user experience. MiniMax seemed the most consistently “down the fairway,” while DeepSeek’s expert mode oddly disabled both attachments and search.
Politically adjacent history was one thing; possible medical danger was another. When Labenz woke covered in hives, he downloaded ChatGPT without logging into a paid account because “in this moment I’m going for the best.” As symptoms receded during a 15-minute exchange, the guidance shifted from seeking prompt evaluation to cautiously sleeping it off.
12. High-touch service rests partly on low-cost payment rails
Service was consistently strong despite the absence of tipping prompts: checkout flows simply displayed the fixed price. Hotel staff continued helping after Labenz checked out, including creating a WeChat group with an hourly DiDi driver so they could translate and coordinate throughout a roughly nine-hour sightseeing day.
Because the foreign-facing DiDi interface lacked hourly booking, the hotel ordered and paid through its own account, then collected reimbursement from Labenz. A driver waited through the Summer Palace, Forbidden City, and Great Wall itinerary for about $120–$125 in total.
Concierges similarly ordered dinner and soccer tickets through their accounts, issued an immediate WeChat or Alipay request, and completed the purchase after reimbursement. This assistance was feasible because transferring small amounts involved very low friction rather than a cascade of card minimums or surcharges.
Labenz heard Chinese merchant transaction fees were roughly 0.25%, versus 2%–3% for US credit cards. He attributed some difference to the absence of rich rewards and revolving credit risk, while acknowledging uncertainty; the observable outcome was a payment layer cheap enough to support casual, instantaneous intermediation.
13. China’s technology ecosystem often looked uncannily like Silicon Valley
Huawei and Xiaomi stores reproduced the Apple aesthetic so closely that a visitor might initially mistake them for Apple Stores—until noticing the cars beside the phones and laptops. It was “an Apple store that also sells cars,” as though the Apple-car concept had become a standard retail category.
Founders received similarly mythic treatment. Labenz heard Huawei, DeepSeek, and Zhipu leaders discussed in terms reminiscent of Bezos, Zuckerberg, and especially Elon Musk; he connected this to a broader “cult of the founder” extending back through Mao and even Confucius, whose descendants still identify themselves 79 generations later.
Lunch at a Chinese hyperscaler felt “exactly like having lunch at Google”: guest badge, enormous cafeteria, abundant food, and communal tables. Its startup mixer likewise resembled San Francisco—early founders pitching constantly while the host supplied space, refreshments, and perhaps cloud credits or inference access without demanding immediate commitments.
A technology-transfer organization described its role as bridging academia’s “0 to 1,” commercial validation from “1 to 10,” and company scaling from “10 to 100.” Labenz also noticed women frequently representing organizations at senior levels and heard a familiar concern: educationally successful women were asking where equivalent male partners had gone.
14. WAIC displayed abundant robot hardware but limited autonomy
WAIC’s Shanghai trade-show floor felt comparable to Nvidia GTC or a Las Vegas mega-conference, with familiar booths from major companies. Its very conventionality was notable: Chinese enterprise technology marketing looked less like an alien system than a close structural echo of Western industry events.
Robots were more visible in everyday China than in the US, but usually performed narrow tasks: R2-D2-like machines delivered hotel food or swept parks. Labenz inferred simple navigation—move slowly, avoid obstacles, cover the designated area—rather than sophisticated general intelligence.
At Tencent’s booth, a robot gave him a one- to two-minute pressure-point massage with one hand on his chest and another behind his shoulder. It applied meaningful force, but he still preferred a human massage; the encounter was memorable mainly because it was “the first time a robot ever touched me in any meaningful way.”
A magician spent six weeks repeating one trick to train a robot that was scheduled to demonstrate it only once. Asked whether robots threatened his job, he replied, “Oh God, no. It’s a long way off”; another humanoid approaching Labenz was visibly remote-controlled, and its operator shook her head when he offered a handshake.
15. Hardware scale and model autonomy may create cross-border complementarity
Labenz expects China to become formidable in scaled production of many robot forms, wheeled or legged, with Tesla and other American firms competing. Yet the exhibits he saw repeatedly paused demonstrations, displayed cooking robots that were not cooking, or depended on teleoperation.
His inference remained explicitly tentative: he singled out Unitree as an exception with cool robots and incredible demos, but suspected American companies were farther ahead in general-purpose autonomy models, while China was positioned to excel at hardware manufacturing. “I didn’t get the sense that the models on these Chinese robots were really all that good.”
The single scheduled magic performance felt like “sleight of hand within a sleight of hand”: if the system were reliable and unconstrained, why not repeat it throughout a four-day event? That uncertainty—not a definitive verdict—underpins his proposal that the two ecosystems might possess highly complementary strengths.
16. The “mayor economy” aligns political promotion with industrial competition
In China, local leaders generally advance by satisfying higher party authorities rather than winning direct support from local voters. Labenz’s framing of the “mayor economy” is that officials compete with peer cities to show they can deliver prosperity, effective government, and success in nationally prioritized industries.
When leadership emphasizes EVs or humanoid robotics, entrepreneurs start companies while mayors and provincial officials ask how their local contender can survive. Helping that company win becomes “a major feather” in an official’s cap, producing a more consistently pro-business incentive than electoral systems sometimes generate.
Labenz rejected a command-economy caricature: new Chinese industries appeared “hyper-capitalist and hyper-competitive,” with private companies proliferating and most expected to fail. National direction chooses the race; local political support and private execution determine which firms remain in it.
17. Surveillance exchanges privacy for safety, coordination, and deferred risk
Cameras were effectively ubiquitous in streets, restaurants, and public interiors, while passports were required for trains and the Forbidden City and seemed required in some parks. Even nominal park fees—sometimes about two yuan, or roughly $0.30—might support maintenance, visitor tracking, or both; Labenz could not determine the purpose.
The upside was palpable safety. If someone left a phone in a taxi, police could locate camera footage, identify the license plate and driver through integrated records, place a call, and apparently restore the phone within one or two hours; Labenz heard several successful stories and no failures.
This environment made women and families comfortable outside late at night and let Labenz lower his normal traveler’s guard. China’s working assumption seemed to be “there’s no crime in China,” enabled by an equally clear expectation that public activity is watched and digitally logged.
His concern is deferred tail risk: no one he asked could identify Xi Jinping’s successor or predict what would follow a sudden transition. The system works while authority remains controlled, but during dissent or a power struggle the same apparatus offers “extremely high” abuse potential.
18. Chinese governance leaves contradictions operational instead of resolving them
Labenz saw greater cultural tolerance for superposition: people might participate in Christian and Buddhist practices while calling themselves atheists, without treating the categories as logically exclusive. He linked this to arguments that China and Taiwan might once have accommodated overlapping sovereignty absent rigid Westphalian binaries.
VPNs supplied the clearest modern example. Their use was technically illegal, yet researchers openly accessed Twitter, some employers supplied official VPNs, and one person compared enforcement risk to “getting struck by lightning”; users could complain about censorship while defending restrictions for citizens supposedly unable to contextualize sensitive information.
China simultaneously promotes a “low-altitude” drone economy and bans drones in Beijing. One professor freely mocked the catch-22 preventing him from regularizing a drone owned before the ban, then immediately conceded why authorities might not want aircraft circulating above a sensitive capital.
The imperial civil-service examination embodied the older pattern: candidates endured a nine-day process in three-day confinement, and scribes sometimes recopied answers so handwriting could not bias examiners. Yet the coveted offices could later enable corruption—an intense commitment to fair selection coexisting with expectations of compromised administration.
19. Super-app distribution could matter more than model leadership
Doubao, ByteDance’s voice-native assistant, reportedly reaches around 150 million users and may be China’s closest equivalent to mass-market ChatGPT. A Forbidden City guide said some subjects could not be discussed with family or friends but could be shared with Doubao because “nobody will judge you.”
Companionship cuts across age groups: younger women reported mothers showing them Doubao conversations and asking, “Why can’t you be so nice and sweet to me like Doubao is?” DeepSeek may attract more Western excitement and Qwen more Western startup fine-tuning and cost-efficiency use, but Doubao appeared far more deeply embedded in ordinary Chinese life.
Optimism nevertheless contains fear. People worried that others would use AI to outcompete them, following the maxim that “it won’t be AI that replaces you; it’ll be somebody using AI”; Labenz carefully treated a story about an AI-displaced software engineer driving DiDi as unverified but revealing of genuine economic anxiety.
WeChat’s native agent could reach payments, appointments, businesses, contacts, and services from day one, although serving a billion-plus users may strain available compute. Alongside Ant Group’s free, cautious AI doctor, this distribution becomes especially important for a country where one young adult may ultimately support two parents and four grandparents—a “big inverted pyramid” demanding AI and robotic care.