Vol.66 We’re Not Just AI Content Movers—A Conversation with 36氪, 特工宇宙, and Z Potentials
Summary
- Two AI-native content teams are rewriting how institutional media is organized. Z Potentials publishes daily with 5 part-time partners, several interns, AI, and Feishu; it takes on about 5 sponsored assignments a month, turns away or refers roughly half due to limited capacity, and handles RMB200K–300K in monthly deal volume, according to Yuca. 特工宇宙 has 6 people offline, with only about 2 actually focused on writing; the rest mainly develop Agents.
- The key to scaling is not getting a model to write one article, but breaking a column into SOPs that can run in parallel. Z Potentials has standardized sourcing, topic selection, transcription, translation, proofreading, editing, and layout; 20 topics can run simultaneously, hundreds of drafts sit in the queue, and the system runs 24/7. 特工宇宙 pushes the same logic into an Agent-native MCN: humans define account positioning and information sources, while Agents handle topic selection, generation, and publishing.
- The scarce asset is moving upstream from wording to exclusive sources, editorial taste, and creators’ private memory. 庄明浩 can have AI search Himalaya’s financing history and SEC filings and ask follow-up questions about liquidation preferences, but it cannot recover old industry stories or emotions “sealed in my head.” 仲泰 also notes that without the unique sources held by outlets such as 36氪, AI cannot produce comparable original content; product reviews require firsthand experience.
- AI will first erode journalists’ processing work while strengthening founders’ ability to manage their own narratives. Ryan sees more founders building in public from day one, using AI for cross-language posts, user emails, founder stories, and brand narratives. 杨轩 cut interview整理 from about 3 business days to 1, lifting efficiency by at least 50%, though hallucinations still require review. Writing is therefore becoming a core founder skill.
- There was no consensus on whether style matters, but everyone placed human value upstream. 仲泰 believes users may care only about the result and that most content will eventually be AI-generated; Yuca says an account is defined by its “animating intent” and vision, while 庄明浩 and Ryan emphasize taste and the original “Soul.” What scales is codifying human positioning, selection, and judgment into a process—not simply reproducing a writing style.
- By around Q2 2025, AI product launches began showing a Web3-style template—and the risk of backlash: China’s first, benchmark SOTA, invite codes, and KOL promotion. 杨轩 believes the playbook may have originated with Manus’s accidental breakout, while 仲泰 says some partners now ask only to “copy a few products that are already hot.” Invite codes raise expectations and delay actual use; if the product fails to deliver, distribution turns from a growth engine into a backlash engine.
- The potential business-model shift is from software pricing to outcome-based pricing, but subscriptions and advertising still dominate. AI may eventually price directly for outcomes as vertical applications reduce hallucinations; Token costs also make it more likely that AI startups monetize from day one and validate PMF faster. “Fake it until you make it” works only if value is ultimately delivered.
- The next wave of demand will come from AI as a buyer on one side, and as a companion and labor force on the other. Ryan sees startups buying SaaS, infrastructure, and other services for their own AI—“make something AI want.” Yuca is bullish on long-term companionship in the mold of Character.AI and LoveDovey, while 仲泰 expects e-commerce, digital-human livestreaming, and intelligent customer service to land first because they are closer to revenue. The more distant vision is an IT department transformed into an Agent recruiting department, with 特工宇宙 becoming “the Boss Zhipin of the next era.”
Deep dive
1. AI-Native Media First Breaks Fixed Costs
杨轩 framed the discussion as existential for traditional media. He moved from print media to 36氪 and once demonstrated that traditional reporting methods could still work in the mobile-internet era; now that AI can produce content directly, he worries that the previous generation of media professionals could once again “die on the beach.”
Z Potentials’ organizational experiment is more radical than its content: 5 part-time partners who all have other jobs, several interns, AI, and Feishu have already achieved daily publishing at the output level of an institutional media outlet, with “zero full-time employees.” The team receives roughly 5 sponsored assignments a month, but limited time makes most difficult to complete; about half are rejected or referred to other partners. Yuca puts monthly deal volume at RMB200K–300K.
特工宇宙 has only 6 people offline, plus some additional partners online. Most work on Agent development; roughly 2 people actually write. 仲泰 did not disclose revenue from 特工宇宙’s media business.
2. A Column Is Not a Tag, but a Parallelizable SOP
Yuca explained that every column name appearing before a Z Potentials article maps to a human–machine collaboration SOP. For an overseas translation piece, the workflow includes scraping and filtering content; if the source is video, the team chooses Tongyi, Claude, ChatGPT, or another tool to produce a first draft, then proofreads, edits, lays out, and embeds the process in Feishu.
Traditional media measures output by how many articles one person can write in a month. Yuca instead puts 20 topics on a board and runs them in parallel. As long as the front end has enough topics, “the supply from the back end should theoretically be infinite.” Hundreds of unpublished drafts are currently in the queue, and the full workflow runs continuously 24/7.
杨轩 asked why Z Potentials often follows quickly when 36氪’s “智能涌现” publishes an exclusive. Yuca said the person responsible for the news-flash column maintains a media-monitoring database and uses AI to summarize content quickly. The exchange exposed the tension between institutional media’s proprietary information and AI-powered rapid aggregation.
3. Agent-Native MCNs Turn Editors into Account Architects
仲泰 has followed Agents since 2023 and deliberately adopted the term “特工.” Agents may eventually outperform ordinary humans in one or more capabilities, with content accounts serving as an early experiment in a symbiotic human–Agent universe.
特工宇宙 is testing an Agent-native MCN. Humans spend substantial effort defining an account’s positioning, name, logo, overall tone, separators, and layout, then connect RSS, API, RPA, and other information sources. Once the positioning is set, AI handles topic selection, generation, and publishing wherever possible.
The stack is mainly prompts, plugins, and RAG. The team generally uses overseas models because different models have different strengths in memory, instruction following, and “human feel.” Business analysis and inspirational content already perform well, but product reviews do not work because they require firsthand experience. An event may invite its account matrix in H2; the organizers apparently believe the accounts are operated by real people.
4. Founders Are Bypassing Media to Run Their Own Narratives
Ryan believes content production has become a core founder skill. 10 or 20 years ago, founders did not necessarily need to put themselves forward to sell their brand, product, or personal image; now more teams are building in public from day one and publicly building their brands on social media.
The simplest use case is converting a founder’s technical content directly into English and publishing it on overseas accounts. More sophisticated SaaS tools help founders write user emails, social media posts, startup experiences, and complete narratives. Vivid stories—living with one’s parents or meeting an investor for the first time—can also be generated by AI.
杨轩 asked whether some reporters and profile writers would therefore lose their jobs. Journalists can hand interview notes to AI, while founders can process their own material directly. Media professionals and founders can use the same class of tools to produce narratives, creating a new challenge for the media industry.
5. AI Can Establish Facts but Cannot Reconstruct a Creator’s Private History
After news broke that Tencent Music was acquiring Himalaya, 庄明浩 used AI to look up the company’s financing history, its final valuation before going public, and its investors, while searching SEC filings directly. He also asked whether the final-round investors had liquidation preferences, because that would affect how the acquisition proceeds were distributed.
He draws a clear line: AI can find the materials, but it cannot complete the episode’s logic, content planning, memes, or jokes, much less retrieve the “old, stale industry stories” remembered only by people who lived through them. “There are many moments of emotion sealed only in my head,” so AI provides foundational assistance, not the work itself.
The same gap appears in PPTs. Explaining AI to fifth-grade students can be handled with a single sentence that prompts a tool to generate the presentation; producing an H1 2025 review of the AI industry requires designing the framework and jokes oneself. As the task moves from general knowledge to dense personal judgment, automation drops sharply.
杨轩’s own test was closer to news production. Organizing an already transcribed Q&A with a Chinese-American Silicon Valley billionaire took about 3 business days using traditional methods; AI assistance cut that to 1 day, improving efficiency by at least 50%. The model still hallucinates, so supervision and verification remain necessary.
6. The Style Debate Ends with a Choice Between “Results” and “Soul”
仲泰 says personal style is “not that important.” Most content will eventually be generated by AI; users may not be able to tell whether AI wrote it and may not care. They will care about the result. Human creativity should go into the account’s initial positioning and final oversight; how it grows can be left to AI.
His longer-term view is that generation and recommendation will jointly rebuild “the next Toutiao.” The same introductory economics material could be filled with citations for university students, illustrated for kindergarten children, or delivered in the voice of their favorite cartoon character. Content will no longer have a single fixed version.
Yuca calls the source of style an account’s “animating intent.” Z Potentials’ vision is to focus on young people and globalization, with AI using those keywords to select and express material. The expression follows the intent; whether the wording comes from a person, AI, or both matters less than what gets selected and why.
庄明浩 and Ryan add taste and “Soul.” Early 36氪 operated as a Chinese-language TechCrunch site, using Reader to retrieve TechCrunch articles and compile and adapt them; its intent, taste, and selection principles shaped how readers understood the world. Ryan compares the division of labor to a magazine editor-in-chief and an editor: AI can take over downstream prose processing, but “if that original Soul was generated by AI,” the content loses its roots.
7. Exclusive Sources and Basic Skills Define the Automation Boundary
仲泰 believes the biggest limitation is not writing style but insufficient context and information sources. Articles from 36氪 and other established media outlets often rely on exclusive sources that AI neither possesses nor can recreate. When a model cannot access first-hand information, even strong generative capabilities can only recombine existing material.
特工宇宙 therefore amplifies what models do well: style imitation, routine business analysis, and inspirational content. In other fields, it fills the information gap through human input or RSS subscriptions. Product reviews perform poorly because they require real experience, not simply language ability.
Ryan points to a subtler limitation: tools are so easy to use that they create the illusion of having mastered writing. “If you never write it yourself, you will lose the ability.” If someone does not know how to control AI, they should first train the most basic skills.
8. AI Product Launches Are Converging on a Web3-Style Template
杨轩 observed that by around Q2 2025, the first-wave marketing SOP for domestic AI products appeared to have formed. He believes the template may have originated with Manus’s accidental breakout; once success is visible, deliberate or involuntary imitation spreads quickly.
仲泰 describes the standard moves more specifically: first claim a “China’s first” position, then achieve SOTA on a benchmark, deploy an invite-code strategy, and hire KOLs for promotion. Some partners have no clearly defined product requirements and simply ask to “copy a few products that are already very hot.”
Ryan agrees that the surface resembles Web3 but rejects equating the two. Web3’s commercial nature makes it more dependent on marketing; AI ultimately still has to prove usage and payment. Concentrated KOL distribution can also create reverse bias: since 2024, some people have assumed a product is “buying rankings” as soon as they see promotion. DeepSeek faced similar skepticism when it first appeared.
Yuca believes viral marketing also faces pressure from technology cycles. User numbers for products such as ChatGPT and DeepSeek can surge, while the next technical inflection point may replace the incumbent at any time. Rapid acquisition, iteration, and fundraising become a “conspiracy” produced by multiple forces acting together.
9. Hype Works Only If Expectations Are Eventually Fulfilled
Ryan says the issue is not whether invite codes or KOLs should be banned, but whether expectations are controlled. Invite codes raise anticipation while delaying real product experience; if delivery falls short of expectations, the momentum reverses. The safer approach is still to be factual and let users vote with usage and payment.
When preparing a marketing plan, 仲泰 first asks whether the product is strong enough. If its capabilities are solid, moderate distribution and amplification are acceptable; if the product is mediocre but adopts a hit-product template, the backlash is often larger. The team also built a platform called “码多来” to ease user frustration with invite-code mechanisms.
Yuca factors delivery time into the judgment. Cursor may have lacked many capabilities before the new Claude model appeared, yet its value rose rapidly after the model upgrade; Builder.ai, by contrast, may struggle to deliver value over the long term. A short-term gap does not necessarily equal deception. The key question is whether the product ultimately delivers value.
庄明浩 summarizes the two-way wager as “fake it until you make it.” If a founder believes the product will become “the first in the world,” creating attention in advance may help it reach that goal; if it ultimately fails, the same traffic becomes the cost. He also joked that an AI product may not have made money yet, while the friends marketing it with AI already have.
10. Outcome-Based Pricing Could Rewrite SaaS, While Product Iteration Moves Backward from General to Specialized
庄明浩 believes “pay by outcome” is one of the more widely shared views this year. Traditional SaaS charges for software; if AI can reduce hallucinations in vertical scenarios or narrow functions, it may charge directly for broadly defined business outcomes, creating a different business model.
He also keeps the current constraints in view. The AI industry still mainly relies on internet-era subscription and advertising models and has not completed a fundamental shift. New pricing models can work only when outcomes are definable, verifiable, and consistently deliverable. He offered no concrete example, using a meme instead to explain the product path.
Agile development starts by building the wheels, frame, and seats, then assembling a car; traditional PMF moves from bicycle to motorcycle to tricycle and finally to car. AI may first build a heavily armed, winged vehicle equipped with rocket launchers, then keep stripping it down until it moves from general-purpose to specialized.
11. Token Costs Push Monetization to Day One—and Create AI as a Buyer
Ryan believes AI startups follow a different logic from traditional internet startups: they incur Token costs from day one, which may force earlier monetization and faster PMF validation. That is a major difference in building an AI company.
After validating PMF and becoming profitable, some startups may no longer need further financing after raising a Series A because they already have relatively healthy cash flow. Monetization in AI startups may therefore happen much earlier.
Ryan also sees a new purchasing relationship. Some founders are not buying ordinary products for their teams, but SaaS, infrastructure, and other fast-moving services that help their AI complete tasks more effectively. As long as the economics work, founders will decide from the perspective of “what products does my AI need?”—the logic behind “make something AI want.”
12. Companionship and Agent Recruitment Form Two Endgame Visions for To C and To B
Yuca moved from One Hundred Years of Solitude to the To C opportunity: it is difficult to find someone who understands you, stays with you forever, never leaves, and never betrays you. Products such as Character.AI and LoveDovey are already building early versions of that persistent companionship. In a few years, she believes this class of AI may genuinely solve part of the problem of long-term loneliness.
仲泰 takes the To B view. Over the past year, every vertical industry has experimented with LLMs and Agents to improve productivity, but deployment speeds vary: “the closer they are to money, the faster they will land.” E-commerce, digital-human livestreaming, and intelligent customer service are the early scenarios he sees.
He cites 黄仁勋’s vision that IT departments may eventually disappear and become Agent recruiting departments, with companies simply looking for Agents capable of completing tasks. If 特工宇宙 can bring together its own Agents and those of its partners, it could become “the Boss Zhipin of the next era.”