中文精读
从信息到行动，从观看到参与：曲凯在2025年初押注的两种变化
42章经单期精读 · 年度复盘与产品交付
主讲：曲凯，42章经创始人（单口）
发布：2025年1月4日｜时长：31分34秒｜整理：2026年10月18日
官方单集与节目介绍
乐观是一组待验证的条件
这期标题很确定，论证却包含许多“如果”：如果模型跨过可用临界点，如果Agent真能交付结果，如果组织愿意围绕新模式变化，应用才可能迎来不同的增长。Agent在这里是能使用工具、推进任务的模型系统，而非所有聊天窗口的统称。
曲凯把2023年的热情与2024年的失望看成两次摆动，认为市场低估了能力、创始人经验和应用模式的积累。阅读价值在于还原他2025年初怎样理解机会；节目没有提供后来一年的兑现检验，本文也不以事后结果倒填预测。
1. 融资回暖与产品变好，是两条不同证据链
线索：00:35–10:13。 他回顾2023年资金追逐大模型、2024年转向具身智能和硬件。应用端许多团队仍在打磨或寻找PMF，即产品与市场需求的契合；估值出现，并不等于用户已经稳定付费。
他提到覆盖范围内约二三十家应用公司估值超过五千万美元，又分述零收入、百万与千万美元年化收入的不同层次。节目没有名单、统计口径或财务核验，不宜把这些数拼成全市场分布。对2024年融资“最低点”的判断也属于观察，不能当完整行业统计。
乐观的第二条线来自产品条件：性能提升、调用成本下降、多模态进步，以及有AI业务实战的新一代创业者。视频、图像与音乐的进步可能打开过去不够用的任务。他用“超过人类”“接近GPT某代”作概括，未定义比较任务和测试标准，因此不能据此认定某模态全面可靠。
编辑推论：融资改善可以延长探索时间，却不能证明应用有用。立项时分别记录可完成的任务、质量门槛、单次成本与团队实战，再看资金是否足以支撑验证；新背景带来不同理解，也不保证交付或融资成功。
2. Prosumer的机会，在传播与付费交界处
线索：10:45–15:22。 曲凯预期市场继续分化，资本更集中于准备充分的团队。对尚无AI经验的人，他建议先进入现有团队做出成绩，再考虑创业。这是当时的一种学习路径，不是对所有职业选择的统一答案。
Prosumer在本期指有专业任务、又以个人方式选择工具的人，例如视频作者、歌曲作者、保险经纪人和工程师。他们可能既有个人传播渠道，又愿意为工作价值付费，所以被称为“小B大C”。它不是单一人口画像，更不是天然有预算的标签。
他的推导从国内企业采购成本、渠道分成和个人付费能力出发，但都是宽泛市场判断。不同国家、行业、客单价和交付要求可能完全不同。AI产品上线即产生调用成本，也只是提醒免费增长需要算账，不足以断言广告模式一律不可行。
编辑推论：应问具体专业任务是否高频、节省的成本由谁获得、购买者能否自行决定、能否通过作品传播。创作者、经纪人与程序员的验收标准不同；把三者都写成Prosumer，不能替代逐类需求研究。
3. Agent按结果收费，首先改变的是责任与组织
线索：15:22–22:35。 曲凯把传统软件、SaaS和Agent串成一条交付变化。SaaS是持续提供软件服务；Agent的设想则把软件和执行工作打包。市场所说“价值十倍SaaS”，在这里是叙事，没有可复核市场测算。
最有力的例子是销售提效：卖一套软件，让一百位员工使用，与直接承诺完成工作、按节省金额分成，是不同的合同。后一种若可实现，供应方就要承担失败、验证与成本波动，不能只按开通席位确认交付。
他用传统软件转云订阅作类比，认为新商业模式会牵动开发、销售、人员管理与组织结构。新公司可从第一天围绕结果设计，旧公司可能经历犹豫、内部试验和换人，给新团队学习窗口。不过旧公司也有渠道和客户，可能更早卖出AI产品；渠道收入与新模式成立仍须分开。
节目进一步推演结果收费可能弱化销售：客户既然成功才付钱，何不都试用？这有重要前提。企业还需要数据接入、权限、采购、信任和结果归因，试用也占员工时间。编辑推论：按结果计费不会自动消除销售；它把“能不能交付、怎样计算结果”的解释责任放得更重。新公司机会取决于适应速度，而非公司年龄。
4. 搜索之后的那一步，决定是否真正完成任务
线索：23:04–26:10。 曲凯以Perplexity解释从答案到行动的想象：买鞋的人要结合季节和温度找款式、看评价、跨平台比价，再下单。若这些环节在同一环境连接，搜索入口就可能变成任务入口。
这不是说找到链接已等于买到鞋。推荐必须满足条件，价格要可用，下单要经用户确认，履约还在后面。节目把购物搜索作为当时观察，将未来平台地位作为推演；没有实测证明所有环节已经可靠完成，也不能由估值反推能力成立。
另一个例子是搜集视频素材。若用户最终为了做视频，产品可以追问下一步是否需要组织素材、生成成片，而非停留在信息列表。编辑推论：扩大交付范围也会增加版权、授权、质量和人工返工；并非每个信息工具都应一口气承包整条工作。应找用户愿意交出的步骤，再定义人在哪一步确认。
5. 多模态的新意，是用户能改变内容的进程
线索：26:41–31:03。 曲凯认为，给既有界面加总结或聊天按钮，不足以解释AI原生产品。这里“原生”指能力改变组织内容和使用方式。他以NotebookLM式体验为例：把文章变成两位AI的对话，用户在不理解处插话，让讨论转向解释。
这个组合不仅有声音，还包括材料理解、角色、对话组织与用户干预。他再设想，把一个待学概念变成辩论视频，用户作为辩手加入。这是产品构想，不是当期已上线案例。关键问题从“能否生成更多内容”变成“用户的输入是否真的改变后续”。
视频会不会成为最大输出形式、视频与游戏会不会融合，他都以开放问题提出。更多模态未必更适合学习或娱乐：有些任务文字更清楚，有些用户只想安静听完，互动也会增加延迟和认知负担。
编辑推论：参与式消费的检验应包含主动加入、理解或满足的变化、再次返回，以及系统回应是否切中插话。片尾“早入场者得到回报”来自身边观察，有幸存者偏差，不能成为要求人人立即创业的证据。
把乐观写成两张验收表
以下是编辑练习。选择一个专业任务和一个内容体验，各写一张小表：
| 实验 | 先定义什么 | 观察什么会推翻假设 |
|---|---|---|
| 从信息到行动 | 用户最终结果、授权范围、谁确认、失败由谁承担 | 仍需大量人工收尾；节省额无法归因；成本高于可收费价值 |
| 从观看到参与 | 用户何时插话、改变哪段进程、系统如何回应 | 互动只是按钮；打断后理解更差；没人再次主动加入 |
先用少量、可回读任务观察，再讨论扩张和收费。融资、调用能力、一次演示与用户重复选择，分别回答不同问题。这期的乐观可以转为行动动力，前提是每个“如果”都有具体的验证位置。
回听导航与关联阅读
- 05:17：应用估值与收入的距离。
- 06:35：能力、资本和创业者变化。
- 13:06：Prosumer的交界人群。
- 15:30：结果交付为何可能改变商业模式。
- 22:57：搜索之后的任务。
- 26:25：多模态与AI原生交互。
- 30:28：乐观与早参与者的观察边界。
相关主题：结果交付、专业个人工具、参与式内容。可对照澜码谈企业任务落地和RC谈人与Agent协作，比较愿景进入企业流程后需要哪些机制。
资料边界：完整阅读所供转写、旧精读与官方介绍，未逐句复听或实测产品。融资、收入、模型比较和2025预测保留2025年初归因；未把转写中未经核验的收购关系和人物身份误称写为事实。本文未检验预测后来是否兑现。
English edition
From Information to Action, From Watching to Participating: Qu Kai’s Early-2025 Thesis
42 Zhangjing episode analysis · Annual review and product delivery
Speaker: Qu Kai, founder of 42章经, in a solo episode
Published: January 4, 2025 | Duration: 31 minutes 34 seconds | Prepared: October 18, 2026
Official episode and introduction
Optimism consists of conditions still needing evidence
The title sounds certain, but the argument contains many conditions. Applications might grow differently if models cross a usability threshold, if agents can deliver results and if organizations adapt to a new business model. An agent here is a model system that uses tools and advances tasks, rather than a name for every chat window.
Qu interprets 2023 enthusiasm and 2024 disappointment as successive swings. He believes the market underestimates accumulated capability, founders’ experience and application models. The value is reconstructing his reasoning in early 2025. The episode does not test how that year eventually developed, and this analysis does not retrofit later outcomes into its predictions.
1. Better funding and better products require different evidence
Episode reference: 00:35–10:13. Qu recalls capital chasing large models in 2023 and shifting toward embodied intelligence and hardware in 2024. Many application teams were still building or seeking product-market fit: alignment between a product and market demand. A valuation does not establish sustained user payment.
He reports roughly twenty to thirty application companies in his field of observation valued above US$50 million, then distinguishes zero revenue, million-dollar and ten-million-dollar annualized revenue levels. Without a list, sampling definition or financial verification, these figures cannot establish a distribution for the entire market. Calling 2024 a financing bottom is likewise an observation rather than a complete industry statistic.
His second reason for optimism concerns product conditions: improved performance, cheaper calls, multimodal progress and founders with practical AI business experience. Video, images and music might open tasks previously beyond usable quality. Phrases such as “beyond humans” or comparisons with a GPT generation lack defined tasks and test standards. They do not establish general reliability across a modality.
Editorial interpretation: Better financing can extend exploration without proving usefulness. Record achievable tasks, quality requirements, unit cost and practical team experience separately, then assess whether funding supports validation. A different background can bring insight without guaranteeing delivery or financing.
2. The prosumer opportunity lies between distribution and payment
Episode reference: 10:45–15:22. Qu expects further polarization, with capital concentrating on well-prepared teams. He suggests people without AI experience first join an existing team and produce results before founding a company. That is one contemporary learning path rather than a universal career prescription.
Prosumer here describes people with professional tasks who select tools individually: video creators, songwriters, insurance brokers and engineers. They may combine personal distribution with willingness to pay for work value, hence the Chinese shorthand for small-business and advanced-consumer characteristics. This is neither one demographic nor a label guaranteeing budget.
His explanation draws on enterprise purchasing costs, channel margins and personal payment capacity in China, but these remain broad observations. Country, industry, price and delivery requirements can produce different conditions. AI usage costs from launch mean free growth needs an economic model; they do not prove advertising is impossible everywhere.
Editorial interpretation: Ask whether a particular professional task recurs, who captures savings, whether the buyer can decide independently and whether work output distributes the product. Creators, brokers and engineers use different acceptance standards. Grouping them as prosumers cannot replace research into each task.
3. Charging for agent results first changes responsibility and organization
Episode reference: 15:22–22:35. Qu connects traditional software, SaaS and agents through changing delivery. SaaS supplies an ongoing software service; an agent might bundle software with execution. The suggestion that agents represent ten times SaaS’s value is a market narrative here, without a reproducible calculation.
Sales productivity is the clearest example. Selling software for a hundred employees differs from promising completed work and taking a share of savings. If the latter becomes feasible, the supplier assumes failure, verification and variable-cost responsibilities. Opening seats cannot alone establish delivery.
Qu compares the shift to moving from traditional software to cloud subscriptions. A business model changes development, sales, management and structure. A new company can organize around results immediately; an incumbent might hesitate, experiment internally and replace personnel, giving newcomers time to learn. Incumbents also have channels and customers and may sell AI products sooner. Channel revenue and validation of a new delivery model remain separate.
He further imagines result pricing reducing sales needs: if customers pay only for success, why not try everything? This requires qualifications. Enterprises still need data access, permission, procurement, trust and outcome attribution. Trials consume staff time too. Editorial interpretation: Result pricing does not automatically eliminate selling; it increases the responsibility to explain delivery and measurement. The opportunity depends on organizational adaptation rather than company age.
4. What follows a search determines whether a task is complete
Episode reference: 23:04–26:10. Qu uses Perplexity to illustrate moving from answers to action. A shoe buyer considers season and temperature, finds styles, reads reviews, compares platforms and orders. Connecting those steps could turn a search entry point into a task entry point.
Finding a link does not mean buying suitable shoes. Recommendations must satisfy requirements, prices must be available, orders need confirmation and fulfillment follows. Qu presents shopping search as a contemporary observation and platform leadership as a future possibility. No product test establishes reliable completion of every step, and valuation cannot prove that capability.
Collecting video material provides another example. If the eventual goal is making a video, a product can ask whether the user also needs organization and a finished edit rather than stopping at an information list. Editorial interpretation: Expanding delivery adds copyright, authorization, quality and rework requirements. Every information tool need not immediately own the entire workflow. Identify which steps people want to delegate and where they should confirm decisions.
5. Multimodal novelty lies in users changing how content proceeds
Episode reference: 26:41–31:03. Qu argues that adding summaries or chat buttons to existing interfaces does not explain an AI-native product. Native here means capability changing content organization and use. His NotebookLM-style example turns an article into a conversation between two AI speakers, with a listener interrupting an unclear point and redirecting the explanation.
The combination involves more than sound: understanding source material, roles, conversational structure and user intervention. He extends it to a debate video explaining a concept, with the user joining as a debater. This is a proposed experience rather than an already launched example in the episode. The question changes from generating more content to whether input actually alters what follows.
He leaves open whether video becomes the dominant output and whether video and games converge. More modalities are not invariably better for learning or entertainment. Text can be clearer, some listeners want an uninterrupted experience and interaction adds latency and cognitive effort.
Editorial interpretation: Test voluntary participation, changes in understanding or satisfaction, return visits and whether responses address the interruption. The closing observation that early participants received rewards comes from people around Qu and is vulnerable to survivorship bias. It cannot establish that everyone should start a company immediately.
Write optimism as two acceptance tables
As an editorial exercise, choose one professional task and one content experience:
| Experiment | Define first | Observations that would challenge the hypothesis |
|---|---|---|
| Information to action | Final result, permitted actions, confirmation and failure responsibility | Extensive manual completion; unattributable savings; costs above payable value |
| Watching to participation | When users interrupt, what changes and how the system responds | Interaction is merely a button; interruptions reduce understanding; nobody voluntarily returns |
Begin with a few tasks and inspectable results before expansion or pricing. Funding, available capability, a demonstration and repeated user choice answer different questions. Optimism can motivate action when each conditional claim has a specific place to be tested.
Listening guide and related reading
- 05:17: Application valuations versus revenue.
- 06:35: Capability, capital and changing founders.
- 13:06: The prosumer intersection.
- 15:30: Why results could change business models.
- 22:57: Tasks beyond search.
- 26:25: Multimodality and AI-native interaction.
- 30:28: Limits to observations about optimism and early entry.
Related themes: result delivery, professional individual tools and participatory content. Compare 澜码 on enterprise tasks and RC on human-agent collaboration to see which mechanisms visions require inside organizational workflows.
Source boundary: The complete supplied transcript, earlier note and official introduction were read, without sentence-by-sentence audio checking or product tests. Funding, revenue, model comparisons and forecasts retain their early-2025 attribution. Unverified acquisition relationships and incorrectly transcribed executive identities are not asserted as facts. This analysis does not verify whether the predictions later materialized.