AI’s Uneven Arrival, TikTok’s Potential Departure, Xiaohongshu and the Delights of Cultural Exchange
Summary
- AI’s near-term economic impact will be uneven: in 2025, individual power users and AI-native startups should capture more value than large incumbents buying copilots. Employees with agency can become “tremendously more productive” or “tremendously more lazy” while maintaining output, leaving the gains with workers rather than employers. Thompson’s long-term winners are entities that “start without” human-centered processes, not enterprises trying to retrofit them overnight.
- The unit of enterprise software could shift from seats and salaries to completed jobs priced by value, accuracy, and compute. Humans today are proxies for output — “half of them work, and I don’t know which half” — much as ad impressions once proxied purchases. That creates two openings: AI-native challengers replacing incumbent workflows and measurement companies that help enterprises determine which automated jobs worked.
- Thompson defines the AI progression operationally: assistants answer, AGI executes assigned tasks, and ASI decides which tasks matter. AGI resembles “a very conscientious but fairly dumb employee” capable of multi-step work without perfect reliability; ASI flips control so “the AI is starting to tell humans what to do.” That distinction makes capability milestones more testable than a vague godlike-intelligence standard.
- Per-seat SaaS is structurally exposed, but the transition could take years because every incumbent process assumes humans are the unit of work. Thompson compares the lag to consumer-goods companies staying with television long after digital advertising was clearly superior: “This whole business model is kind of screwed up,” yet institutional inertia can defer the reckoning. By 2035, workers raised with ChatGPT may turn AI from individual advantage into table stakes whose gains re-accrue to employers.
- Cheap, fast base LLMs retain substantial value even as expensive reasoning and agent systems emerge. Better models can generate synthetic training data, while low-cost models handle high-volume, low-consequence tasks such as recognizing products across Meta’s feed — potentially making “every single item on Facebook” an ad. The new capability layer therefore adds to aggregation rather than automatically killing it.
- Thompson supports restricting TikTok on a narrow national-security case grounded in demonstrated algorithmic influence, not data collection or hypothetical future abuse alone. An adversarial state receives a targeted, opaque channel into American “hearts and minds”; his concrete evidence was TikTok returning highlights for every NBA team except the Houston Rockets during the China–NBA dispute. “We wouldn’t have let the Soviet Union control a television network,” and TikTok is more precisely targetable.
- For Thompson, the TikTok decision is a “51/49” trade-off because a shutdown could destroy real creator value, weaken competition for Meta, and conflict with free-market and free-speech principles. Thompson conceded he was “hoisting myself on my own petard,” subjected the case to “strict scrutiny,” and only narrowly let national security prevail. Waiting since 2020 made the disruption vastly more painful, while the statute’s actual targets — app stores and Oracle hosting — left room for tactical maneuvering short of TikTok voluntarily shutting down.
- The migration of TikTok users to Xiaohongshu exposed both China’s censorship problem and America’s unusually effective soft power. Chinese state media interpreted Americans attacking their government as repudiation of the US system; Thompson called that freedom-based propaganda “in the water,” because dissent without punishment demonstrates the system itself. He expected Xiaohongshu’s overwhelmed moderation and inability to induce American self-censorship to end the exchange quickly, possibly through a Chinese-phone-number requirement.
Deep dive
1. AI adoption rewards employees before it rewards enterprises
Thompson is not skeptical that AI helps individual workers now; he is skeptical of the top-down instruction that “everyone has an assistant now. Go and use it.” Like newspaper companies enjoying supposedly free internet customers in the 1990s, employees see only upside before the surrounding economics adjust.
The immediate arbitrage belongs to workers with enough agency to experiment. They can become “tremendously more productive” or “tremendously more lazy while doing the same amount of work,” while corporations struggle to capture those dispersed gains at an enterprise level.
Google’s decision to bundle Gemini into Workspace and raise the overall price struck Thompson as excellent monetization: locked-in customers pay more whether or not productivity transforms. Usage might “trickle in,” but access alone will not make every company “a gazillion times more productive next year.”
Adoption may enter through attrition rather than transformation programs: a ten-person team becomes eight, responsibilities remain unchanged, and the survivors reach for AI. Wholesale replacement resembling the mainframe’s elimination of back-office work would require difficult top-down integration that Thompson doubts most enterprises can execute quickly.
2. AGI executes the agenda; ASI sets it
Thompson’s assistant stage describes current LLMs: humans ask, models answer, and the interaction remains directly reactive. Even diagnosing a patient from symptoms is still assistance, however impressive the information processing, because the model is returning an answer to a human-defined question.
AGI begins when a model can accept a task, access several tools, and complete the necessary steps. Thompson’s analogy is “a very conscientious but fairly dumb employee”: it will not decide what the organization needs, but it can perform assigned work at a good-enough, not perfect, reliability level.
ASI is the control flip. It might inspect patient data, identify an emerging problem, schedule the appointment, direct humans to perform a scan, and write the prescription: “The AI is starting to tell humans what to do instead of humans telling the AI what to do.”
3. AI exposes the employee as a proxy for economic output
Companies organize around humans as units of work, but Thompson argues that this has always been a proxy for completing tasks. Generative art revealed the same hidden separation: ideation and manifestation seemed inseparable until a person could express an idea in a prompt and let a model implement it.
The commercial endpoint is payment per completed job rather than per seat or employee. Freelance marketplaces already approximate this when a company pays for a logo rather than the designer’s labor, but AI could make job completion “the defining concept” for buying work.
Google’s advertising breakthrough supplies the analogy. Newspapers priced exposure because circulation was measurable, even though the advertiser wanted a purchase; performance advertising moved payment closer to that actual outcome. AI similarly replaces a fuzzy labor proxy with a priced result.
Thompson’s formulation: determine what a correct job is worth, the accuracy required, the compute needed, and its cost. “Humans are the old advertising. Half of them work, and I don’t know which half”; conspicuously busy employees may contribute little, while apparently idle ones may hold the company together.
4. CPG’s long retreat from television previews enterprise inertia
Consumer-goods companies illustrate how an obviously superior technology can arrive slowly. Axe and Dove sit inside the same company, yet branding differentiates them so thoroughly that CPG product managers are called brand managers; the organization exists to manufacture habitual, often subconscious selection.
Early Facebook targeting could reach a precisely defined customer, but doing so was expensive and poorly matched businesses built around “scale, scale, scale, scale.” Broad television or ESPN advertising remained cheaper per consumer, while the actual purchase decision occurred later in the supermarket aisle.
Digital finally won as television audiences shrank, Facebook improved customer-finding, and COVID pushed purchases online, tightening conversion measurement. The lesson is not that digital failed P&G or Unilever; it is that adoption “took a lot longer than you might have thought.”
SaaS faces the same temporal distinction. Thompson rejected the idea that he was optimistic about it — “This whole business model is kind of screwed up” — but compared it with television circa 2015: correctly doomed over the long term, yet capable of surviving years beyond bearish expectations.
5. AI-native entrants will create the market that incumbents resist
Facebook did not merely convert old advertisers; it enabled new e-commerce, app, direct-response, and Shopify businesses. When large CPG companies boycotted the platform, their withdrawal lowered ad prices for challengers capable of taking their market share — an “anti-fragility” that let Facebook win either way.
Apple’s ATT changes initially devastated Facebook but ultimately reinforced its competitive position. Thompson expects a similar dynamic from AI: new companies built around the technology will form its native customer base and attack incumbents from below before established businesses fully adapt.
This is “not a forecast about the next 50 years”; the narrower call is that wholesale enterprise transformation is “not gonna happen in 2025.” AI-native firms will chip away first, eventually forcing incumbents to reorganize once the tools improve and competitive pressure becomes unavoidable.
The generational mechanism matters. SaaS benefited from millennials already comfortable working in browsers and Google Docs; by 2035, more workers will have grown up assuming ChatGPT exists. AI use then shifts from individual advantage to table stakes, allowing productivity gains to re-accrue to corporations.
6. Better reasoning models add a layer without erasing aggregation
Sharp connected the incumbent-versus-blank-slate question to Rich Sutton’s “bitter lesson,” but Thompson narrowed the analogy. Computation-first methods may govern fundamental capability development; productization is a separate discussion requiring iteration and other product decisions, so there is no equivalent bitter lesson eliminating product work.
New models can generate useful synthetic data that feeds back into model training, while base LLMs remain cheaper and faster — characteristics that continue to matter for aggregation and work performed at enormous scale.
Meta could recognize a bag in a photo and turn it into a commerce link, moving toward “every single item on Facebook becoming an ad.” Mislabeling one bag carries little cost, so an inexpensive model that is usually right can be more valuable than a costly system optimized for near-perfect answers.
Measurement remains the gating layer. ATT hurt Facebook because advertisers could no longer know which ads worked, not because performance vanished; enterprises likewise need credible attribution before buying automated work confidently. Existing companies have no native process for pricing a job with that precision.
7. TikTok’s law leaves tactical uncertainty around a strategic dispute
At recording, the Supreme Court ruling, a January 19 shutdown, possible involvement from Elon Musk, and a Trump executive action were unresolved. Thompson doubted an executive order could directly undo Congress, though litigation timing, an injunction, or enforcement discretion could still shape what happened.
His legal clarification: the statute did not itself switch off TikTok. It barred app stores from distributing it and Oracle from hosting relevant data, meaning existing installations might continue working; TikTok’s threatened full shutdown could be an attempt to “force the issue.”
Thompson regarded personal data as secondary to giving an economically, militarily, and ideologically adversarial power “a direct sort of pass through the hearts and minds of the American people.” Targeting makes that channel more powerful and much less observable than a conventional broadcaster.
His concrete exhibit came from the Hong Kong and Daryl Morey controversy: TikTok searches returned highlights for every NBA team except the Houston Rockets. Later studies, he said, also found China-related terms receiving unequal treatment, making the risk an observed “thumb on the scale,” not merely a hypothetical weapon.
8. A justified restriction can still destroy real economic value
Telling creators to move elsewhere understates what disappears with a TikTok account. A following of 10,000 or 100,000 people creates leverage that can become income or opportunity; because no market price sits beside it, the loss is underrated despite functioning as “an economic taking.”
TikTok also forced Meta to improve. A prohibition in 2020 would already have hurt, but waiting five years enlarged the creator economy being disrupted and left Facebook better positioned to capitalize on the competition. Thompson was openly sympathetic to the users, businesses, and competition sacrificed.
China’s refusal to permit a sale might demonstrate TikTok’s strategic importance, but Thompson would not claim that categorically: ByteDance’s rational negotiating strategy was to refuse until the last possible moment. Evidence that Beijing officials, rather than executives, controlled any deal would strengthen the concern.
The control concern is concrete: Sharp pointed to the CCP’s 1% stake in a ByteDance subsidiary, its golden share on ByteDance’s board, and Chinese legal obligations to comply with state intelligence work; Thompson said these facts were already known in 2020.
Pressed on free speech, Thompson called his position “51/49” and admitted to “hoisting myself on my own petard.” After “strict scrutiny,” he narrowly let national security override free-market and speech principles while accepting the inconsistency. Sharp separately framed such freedom-versus-security conflicts as case-by-case.
9. Xiaohongshu turned American dissent into unintended soft power
TikTok refugees choosing Xiaohongshu — a literal mainland Chinese app — struck Thompson as hilarious. Chinese state media treated their migration and anger as proof Americans reject their government, missing that publicly giving the government “the middle finger” without punishment is itself a defining American freedom.
Thompson described the cultural gap through Chinese four-character sayings: omitting one character can turn an apparent compliment into a devastating accusation about the missing trait. Chinese discourse often relies on implication, while Americans take words literally; Chinese officials, conversely, assume stated US ideals conceal some unstated lever.
That mismatch makes the Xiaohongshu episode unusually effective propaganda precisely because it is not presented as propaganda. Americans attacking their own system are demonstrating its tolerance: “It’s not a billboard. It’s in the water.” Sharp called the juxtaposition a striking testament to the American system.
The exchange was also genuinely two-way, exposing Americans to Chinese people, clean streets, and functioning infrastructure. Yet Xiaohongshu faced a “Sword of Damocles”: Americans would not self-censor, moderation could be overwhelmed, and Thompson predicted the episode might end within a week through app-store withdrawal or Chinese-number registration.
10. The Great Firewall created a rival ecosystem but not a universal model
Thompson’s reciprocity argument is blunt: China blocks American consumer-internet companies, so the US can block China’s. A free-trade regime without credible “tit for tat retaliation” lets one side violate its commitments until corrective action becomes far more disruptive than it needed to be.
Setting morality aside for analysis, he called the Great Firewall “one of the smartest things that any country did.” It enabled censorship and self-censorship, maintained political control, and protected a domestic software ecosystem until China became the only plausible technological rival to the United States.
Bill Clinton’s image of China trying to “nail Jello to a wall” proved “totally wrong,” including to Thompson’s earlier expectations. Other countries nevertheless missed their window: they are unlikely to block US platforms now and reproduce China’s protected ecosystem because they cannot “out-China China.”
The chip analogy marks his limit. He opposed newly announced controls that shifted from constraining China toward a “permission structure on all of tech,” treating every country as an enemy by default. Yet on unavoidable alignment he was candid: even if Washington controls US companies too, “I’m a US citizen, so I’d rather we be in charge than them” — eventually, “you have to pick sides.”