中文精读
等待模型成熟时，产品仍要交付：曲凯在2024年初提出的三条路
42章经单期精读 · AI产品与商业交付
讲述者：曲凯，42章经创始人；本期为单口，无独立嘉宾
节目发布：2024年1月13日｜时长：40分44秒｜整理：2026年10月18日
原节目与官方介绍
大方向乐观，为什么当期仍要缩小产品承诺
这期标题押注2024、2025年是关键两年，正文却用了很大篇幅谈不够成熟的技术。两者构成它的核心问题：怎样相信长期机会，同时用当下能做到的事情获得客户、收入和继续学习的时间？
曲凯自述2023年服务十多家AI公司融资、参与四家投资，结尾也推广自己的融资顾问和社群业务。因此市场观察与项目标准带有他的业务视角。本文把“今年”放回2024年，把“未来两年”限定为2024—2025年，不用后见之明补写预测是否兑现，更不当作2026年的创业或投资指令。
1. 技术大方向的共识，不等于商业落地的共识
节目线索：00:37–10:47。 曲凯回忆移动互联网：后来觉得进入晚的人，在当年仍面对争议。AI也可能同时有广泛乐观与巨大分歧——大家相信技术重要，却不确定何时、用什么产品、由什么公司兑现。
他描述2023年机构先看模型、再看应用，八月后渐冷，估计约200家应用拿到资金。这是个人所见范围的判断，未给完整融资数据库；不能当作全市场统计。关于一轮模型融资可支持许多应用的比较，也是资本配置观点，不能推出投资回报。
“AI还不是赛道”在节目中的意思，是应用没有形成稳定可比的细分类别，机构有时更多在押创始人，而非验证过的品类。曲凯因此希望创业者把注意力从融资叙事移到收入利润。编辑判断：可操作的追问是：谁愿意为何种结果付钱？这不意味着所有长期研发应停止，也不能用资金环境替代逐个需求判断。
2. 可控性不足，有三种不同的交付设计
节目线索：11:21–16:06。 第一种是卖结果：AI与人工在服务商内部协作，客户收到满意成果，不必面对每次失败。节目用FancyTech举例。模型问题被转成内部审阅、返工与管理成本；技术进步可能降低这些成本，但“有人兜底”不自动带来好毛利。
第二种是分开生产与消费：在内部或创作者端使用AI，让最终用户仍以熟悉方式消费成品。它可以是持续的产品，而非一次物料。转写有一句前后矛盾的表达，结合上下文应理解为少把不稳定生成直接暴露给消费者，而不是完全不让生产使用AI。
第三种是受限场景、有限使用：让AI处理它擅长的一段，其余交给传统自动化、知识结构或人工。曲凯认为分析已有材料可能更可控，举10%到20%的AI占比说明不必追求纯度。分析仍会出错，比例也不是设计标准。
编辑判断：三条路实际在安排谁承担不确定性：服务团队、创作者，还是流程约束。选择时要同时检查错误在哪里被发现、由谁修、最终如何验收。产品经理的工作，是找到技术边界与用户需求的交点；模型更强后，这个交点还要重新画。
3. 形态会变，团队要积累的不能只有一层界面
节目线索：16:22–23:40、37:45–40:26。 曲凯关注三个变量：上一年融资团队的产品开始上线、模型能力变化、资本环境。他预计2024年落地、2025年爆发，也说当期产品大多会被淘汰。具体模型发布时间和“炮灰”措辞，属于当时的预测与强调，不是已验证的规律。
头条到抖音、团购到外卖等例子，被用来解释团队可能在新形态中延续。类比不能保证任何公司转型成功。更稳妥的理解是，短期收入可以支持组织、客户理解与技术学习，而不是把当前产品质量视作不重要。
节目最后采用两类尺子：一类看能否交付规模收入利润，不要求AI占比；另一类看是否巧妙结合技术特点与新需求，容忍短期不成熟，但要求明确边界与学习能力。编辑判断：可以把探索与经营写成不同假设，分别保留证据；“有长期想象力”不能免去短期的失败记录，“现在能赚钱”也不能独自证明长期优势。
4. 生成可能改变供给，但社区和执行责任没有消失
节目线索：24:11–33:06。 曲凯先设想外包行业：代码、设计、广告、财税等由AI降低执行成本，再结合行业知识交付结果。他认为自营服务可能比向碎片化服务商卖软件更有价值。五分之一、十分之一价格是战略想象，节目没有给成本账，不能成为降价承诺。
另一组推演是内容平台。用户画像和搜索意图，可以组合成生成条件；如果所需内容能即时产生，旧库存和推荐匹配的价值会怎样变化？“生成打败推荐”是极限推演，曲凯明确暂不展开社区与创作者角色，也承认质量、算力和速度限制。省略的因素仍然重要，不能把它总结成现有平台已无价值。
线下服务则是另一种重组：赶飞机的人关心半小时内到达，系统可以考虑骑车、地铁、打车的组合；家务服务也可根据具体房屋和任务，而非只选标准时长。编辑判断：表达灵活需求只是第一步，供给是否存在、时间是否准确、责任如何分配、实际完成怎样确认，决定它能否成交。内容无限的设想，也不能让信任和注意力无限。
5. 聊天入口和生成渠道，分别改变哪一层
节目线索：33:38–37:04。 曲凯并未主张所有软件都变聊天。简单点选可能比打字更快；多参数、复杂操作则可能用语言降低门槛。外卖商家在后厨口述修改配送时间和地址，是目标表达替代复杂后台操作的例子。
关键价值在理解后完成操作。编辑判断：需要核对订单身份、修改条件和生效结果；对话界面不能替代执行系统。一个界面更自然，不等于权限、确认与失败恢复自动解决。
“内容生成即渠道”进一步讨论议价权：假设一个工具生成各平台20%的视频，它能否影响内容转场和品牌展示？这是示意，不是市场份额或广告方案。猫项圈带品牌的例子，用来解释供给入口可能影响分发。是否展示商业内容仍需尊重用户预期和授权，不能把隐性植入当既定设计。
两组设想共同提醒，AI的机会未必只在模型性能，也在谁定义目标、组织供给并对结果负责。节目提出的是研究方向，价值要靠具体任务而非宏观叙事验证。
练习：把同一项需求放进三种交付方式
以下为编辑练习。选一项明确任务，例如将客户资料整理为可发布的产品介绍，分别设计服务商交结果、作者审阅再发布、受限自动流程三种方案。
| 方案 | 应记录的成本与证据 |
|---|---|
| 交付结果 | 谁审阅、返工几次、客户何时接受 |
| 生产消费分离 | 作者能改什么、消费者最终看到什么 |
| 受限流程 | AI处理哪一段、规则怎样查错、失败交给谁 |
用同一批材料比较合格率、总用时与总成本。若人工耗时下降却返工变多，就不能只报模型提效。再写一个远期假设，标出它需要哪项尚未具备的能力，以及怎样观察其出现。把今天可交付与未来值得研究放在同一张路线图里，但分别判断。
回听导航与关联阅读
- 07:51：为何广义AI标签不足以定义品类。
- 09:16：从融资叙事到交付收入。
- 11:18：三条可控性路径。
- 19:11：形态变化与团队积累。
- 26:17：生成供给对平台的极限推演。
- 33:33：复杂任务与聊天入口。
- 35:56：生成工具的渠道议价想象。
相关主题：可控交付、商业假设、内容供给。可对照Monica的场景入口和FlowGPT的创作者生态，补看本期暂未展开的用户与供给关系。
来源边界：依据完整提供的转写、官方介绍与既有精读，未逐句复听。曲凯的融资顾问、投资与社群业务立场已保留；未经公开确认的竞品“内部渠道”转述未采纳为事实。融资数量、行业状态、模型发布与市场预测均限定2024年初，未以当前信息反向补齐。
English edition
Deliver While Waiting for Better Models: Three Routes Proposed at the Start of 2024
42 Zhangjing episode analysis · AI products and business delivery
Speaker: Qu Kai, founder of 42章经; a solo episode with no separate guest
Published: January 13, 2024 | Duration: 40 minutes 44 seconds | Prepared: October 18, 2026
Original episode and official introduction
Why does a long-term optimist narrow the immediate product promise?
The title identifies 2024 and 2025 as crucial years, while much of the episode addresses immature technology. Its central question is how to believe in a long-term opportunity while obtaining customers, revenue and learning time through work deliverable now.
Qu reports helping more than ten AI companies raise capital and investing in four during 2023. He also promotes advisory and community services. His market observations therefore carry a business perspective. “This year” means 2024 and the next two years mean 2024–2025. This analysis does not retrospectively declare forecasts fulfilled or turn them into 2026 startup or investment instructions.
1. Agreement about technology does not mean agreement about commercial adoption
Episode reference: 00:37–10:47. Qu recalls that people who later felt late to mobile internet still faced controversy at the time. AI can similarly combine broad optimism with disagreement about when, how and through which companies value appears.
He describes investors moving from models to applications in 2023, cooling after August, and estimates around two hundred funded applications. This is an observation without a complete funding database. Comparing a model round with funding many applications expresses a capital-allocation view, not investment returns.
His statement that AI is not yet a sector means applications lack stable, comparable categories and investors sometimes back founders rather than validated markets. He wants attention shifted toward revenue and profit. Editorial interpretation: Ask who pays for what result. This does not require abandoning all long-term research, and funding conditions cannot replace assessment of individual needs.
2. Weak controllability permits three different delivery designs
Episode reference: 11:21–16:06. First, sell results. AI and people work inside a service provider while customers receive accepted output without confronting every failure. FancyTech is the contemporary example. Model uncertainty becomes internal review, rework and management cost. Human backup does not automatically produce healthy margins.
Second, separate production from consumption. AI supports internal teams or creators while consumers receive finished work through a familiar experience. This can remain a product rather than a one-off asset. One contradictory transcribed sentence is interpreted through its surrounding argument: reduce direct exposure to unstable output, rather than prohibit AI in production.
Third, constrain tasks and use AI selectively. Let it handle a suitable portion alongside existing automation, structured knowledge or people. Qu considers analysis potentially more controllable than free generation and uses a 10–20% share illustratively. Analysis can still fail, and the percentages are not design standards.
Editorial interpretation: Each route assigns uncertainty to a service team, creator or constrained process. Check where errors are detected, who repairs them and how output is accepted. Product management finds the intersection of technical boundaries and user needs, then redraws it as capability changes.
3. Product forms change; teams need more than an interface layer
Episode reference: 16:22–23:40 and 37:45–40:26. Qu watches product launches from recently funded teams, model progress and capital conditions. He predicts implementation in 2024 and expansion in 2025, while saying contemporary products will largely disappear. Specific release predictions and dramatic language are historical forecasts, not established laws.
Examples from content feeds to short video and group buying to delivery illustrate teams continuing through changing products. Analogies cannot guarantee successful pivots. A more cautious reading is that revenue can sustain organizations, customer understanding and learning, without making current quality unimportant.
His closing framework uses two standards: deliverable revenue and profit without demanding a high AI share, and inventive combinations of technology with needs that may not mature immediately but require clear boundaries and learning. Editorial interpretation: Record exploration and operations as distinct hypotheses. Long-term imagination does not eliminate near-term failures; present profit does not independently establish durability.
4. Generation may alter supply without eliminating community or execution responsibility
Episode reference: 24:11–33:06. Qu imagines AI reducing execution cost in coding, design, advertising and accounting services, combined with domain expertise. Integrated service businesses might capture more value than selling software to fragmented providers. Prices at one fifth or one tenth are strategic scenarios without cost evidence, not promises.
For content platforms, user attributes and search intent can become generation constraints. If desired content appears instantly, how does stored supply and recommendation change? “Generation beats recommendation” is an extreme long-term scenario. Qu explicitly leaves community and creators unresolved and acknowledges quality, compute and speed limits. Those omitted factors cannot be treated as irrelevant.
Offline services provide another recombination. Someone catching a flight may prioritize arrival within thirty minutes, allowing a bicycle, subway and taxi combination. Housework could reflect a particular home and requested tasks rather than standard time blocks. Editorial interpretation: Flexible expression begins the process. Supply, timing, responsibility and completion determine delivery. Imagined unlimited content also does not make trust or attention unlimited.
5. Conversational interfaces and generation channels change different layers
Episode reference: 33:38–37:04. Qu does not propose replacing every interface with chat. Simple selection can beat typing, while complex parameter-heavy tasks may benefit from language. A restaurant operator verbally changing an order’s delivery time and address illustrates replacing a difficult administration screen with a goal.
The value requires action after understanding. Editorial interpretation: Verify the order, allowed conditions and effective result. Conversation cannot replace execution, permissions, confirmation and recovery merely by feeling natural.
“Generation as distribution” asks about bargaining power. If a tool hypothetically generated 20% of videos across platforms, could it influence transitions and brand appearances? This is a scenario, not market share. A branded cat collar illustrates influence at the supply entry. Commercial content still requires appropriate user expectations and authorization, not hidden placement as an assumed feature.
Both scenarios locate opportunity in defining goals, organizing supply and owning outcomes alongside model performance. Concrete tasks must validate these research directions.
Exercise: Deliver the same need in three ways
Choose a bounded task, such as turning customer materials into publishable product copy. This editorial exercise designs a managed service, creator-reviewed production and a constrained automated process.
| Route | Cost and evidence |
|---|---|
| Results as a service | Reviewer, rework count and customer acceptance |
| Separate production and consumption | Creator control and final consumer experience |
| Constrained process | AI’s portion, rule-based checks and failure owner |
Compare acceptance, total time and total cost on the same material. Less manual time with more rework is not simply higher efficiency. Add a long-term hypothesis specifying missing capabilities and observable signs of their arrival. Deliverable work and research can share a roadmap while retaining separate judgments.
Listening guide and related reading
- 07:51: Why the AI label does not define a product category.
- 09:16: Revenue rather than financing narratives.
- 11:18: Three controllability routes.
- 19:11: Changing forms and accumulated team capability.
- 26:17: The extreme scenario for generated supply.
- 33:33: Complex tasks and conversational entry.
- 35:56: Generation tools and channel bargaining power.
Related themes: controlled delivery, business hypotheses and content supply. Compare Monica’s contextual entry points and FlowGPT’s creator ecosystem for user and supply relationships left open here.
Source boundary: Based on the complete supplied transcript, official introduction and earlier close reading; audio was not checked sentence by sentence. Qu’s advisory, investment and community interests remain explicit. Unconfirmed competitor claims from unnamed internal channels were not adopted as facts. Funding counts, industry conditions, releases and forecasts retain their early-2024 context without retrospective completion.