中文精读
模型一直变，应用该坚持的是问题，还是今天的解法？
主讲：曲凯，42章经创始人｜本期为单口，无受访嘉宾
原发布：2026 年 7 月 19 日｜时长：26 分 36 秒
整理：2026 年 10 月 18 日｜官方单集
“应用已死”是这期最刺耳的说法，也是曲凯从部分投资机构那里听到的市场情绪。它不等于用户已经不再使用应用，更不是一份证明应用消失的统计。模型势能、融资偏好与用户价值，是三个可能同时朝不同方向移动的变量。
这次年中复盘有不少判断，也明确留下许多没有答案的问题。它适合拿来检查创业动作与叙事有没有脱节：我们究竟在服务一群具体用户，还是在等市场下一次偏爱自己会讲的故事？
一、资本对应用转冷，不直接裁决应用有没有用
回听：01:03—05:34。 曲凯回顾几轮摆动：2023 年仍在摸索模型与应用的关系，2024 年先冷后暖，2025 年推理与 Agent 拉动预期，2026 年在他接触的圈子里，模型强而应用融资冷。他甚至说这是从业多年第一次听到机构直接宣告应用赛道死亡。
这些观察来自创投、融资顾问与创业者交流，并没有完整机构样本或投资数据库。它可以说明主讲接收到什么信号，不能替全市场下结论。标题的夸张，是把矛盾拉到面前，不是定义产业事实。
同样，模型公司被资本奖励，也不意味着所有应用都没有空间。不同融资策略有不同约束：大基金可能跨行业、跨阶段覆盖，小基金未必能复制；创业团队更不能把投资机构的组合打法当成自身经营计划。
编辑推论： 复盘时至少留两栏：客户是否继续获得价值；外部资金是否愿意在当期承担风险。融资转冷会影响生存时间，但不能代替需求研究。反过来，拿到钱也不能代替交付和付费。
二、说“面向海外”，需要有真的海外执行
回听：05:39—09:07。 曲凯批评一种错位：团队融资时说做海外，英文宣传却主要在中文网络流转，最终用户仍集中于原来的圈子。他强调，如果原本就没有海外目标另当别论；问题是口头目标与实际动作不一致。
他举 Dify、Workmagic、ACE Studio、Meshy 等持续投入海外市场的案例，经营数字与比较仍是当期观察，未经本篇审计。其海外付费能力较强的概括，也不能替代具体国家、品类、渠道与成本验证。
这里的机制不是“换一个地方就成功”，而是将获取反馈的环境真正移到目标用户所在处。语言只是一个表层；渠道、沟通方式、购买理由、支持时区与付款体验，都可能影响能否形成持续使用。
编辑推论： 若产品声称服务某个市场，就检查真实触达从哪里来、有效用户在哪里、谁付了钱、反馈怎样进入下一版。一个英文演示不构成国际化，一个海外注册也不证明付费价值。让地图、渠道和收入记录回答问题，比反复强调标签更有效。
三、技术路线变化时，旧排名不能替新任务验收
回听：09:34—16:47。 曲凯认为值得研究的不只是一直被关注的领先者，也包括评价短期内反复变化的公司。他谈 Google、智谱等例子，试图说明模型竞争可能在聊天、代码、长程自治任务之间重排。
关于具体公司的领先名次、人才流动与路线选择，属于 2026 年 7 月主讲的判断和转述，节目没有提供统一评测。引用公开信也不等于本篇已经复核原文。因此这里不列当前模型榜单，而提炼一个方法：新任务出现后，重测原有选择。
模型知识的信息差还可能迅速缩短。人人讲 Agent、主动 Agent、Harness 时，讲明白用户要完成什么，反而更稀缺。Harness 在此可理解为组织模型、工具和执行约束的外围系统；懂这个名词，并不能证明某个业务已经需要这套方案。
编辑推论： 模型升级应带来具体任务比较：成功率、失败类型、接手成本和完成确认如何变化？旧模型某项限制消失后，产品还剩哪些价值？这些问题比永久拥护某家公司，或每半年跟随一次热词，更接近经营。
四、随技术改解法，不等于随资本换服务对象
回听：17:13—21:24。 曲凯说要“随模型的波，逐资本的流”，但紧接着限定：持续尝试更好的方式，服务同一批用户、解决同一类问题。稳定的是需求与组织积累，可变的是技术实现。这与不停换热点有本质不同。
他用 Ambi 创始人莫子皓的故事自我反省：面对一个正在做软硬结合体验的团队，他下意识建议融资时讲模型。对方反问，既然当下工程已能支持体验，未来有需要再做模型，为什么今天必须讲？故事里的产品效果是主讲转述，未独立验收；真正值得保留的是他意识到建议只是迎合资本口味。
如果公司把未来计划提前包装成当下能力，融资叙事就可能反过来支配资源。团队会为了维护说法投入并非当前最必要的研究，而不是先完成用户问题。讲清今天做了什么、未来依赖什么，能减少这种偏移。
“产品和创始人是模型的期权”是节目转述的比喻，不是已核实的具名引语。它表达今天积累可能随模型成熟而增值；曲凯强调要保持学习与现金流，别提前归零。但不是所有积累都会兑现：需求不存在、权限拿不到、经济性不成立，未必由更强模型解决。
五、历史周期帮助提问，不能给应用复兴定日期
回听：21:24—25:23。 曲凯把 GPU、token 收费与历史算力、运营商短信流量收费类比，讨论上游收费以后，是否可能出现更低成本和新的应用模式。token 是模型处理内容的计量单位，它的价格只是产品成本的一部分。
这些类比可以提醒产业利润位置会变，但 IBM 具体收费与运营商历史周期未在本篇查证，不能把故事当精确时钟。即使模型单价下降，使用频次、上下文、工具与人工成本也可能继续增长。“免费后可以用广告”需要广告收入、用户规模和服务成本共同成立，并不自动可行。
他引用《技术革命与金融资本》的导入、转折、展开框架，最后以问题收束：现在是爆发期还是狂热期？转折还有多远？这是一种观察镜头，不是节目已经找到的答案，也不能从别的技术革命平均出某个日期。
关于不会一家长期独大、模型不会吃掉一切，是主讲信念，不能升级为经济学定理。“世界模型之后还有什么故事”的疑问，也不意味着技术路线已经走到终点。
编辑推论： 历史最有用的作用，是帮助列出不同情境：资本更冷怎么办、成本继续降怎么办、供应商进入应用怎么办。每种情境都要看用户需求与自己的可迁移能力，而不是只等待某个宏观拐点。
应用练习：为下一次转向保留一条连续线
选一个最近想跟进的新技术或新名词，写四段短说明：
- 保持什么： 仍服务哪类用户、哪种任务？若对象都变了，承认这是新探索，不用“升级”掩盖。
- 改变什么： 新模型具体替换哪一步？用同一任务集比较结果、成本和人工接手。
- 积累什么： 即使这版界面被替代，哪些客户理解、合法反馈与交付能力还能用？
- 能等多久： 当前现金、承诺和团队能力支持多久验证？哪种失败会迫使缩小范围？
再把融资演示里每个能力句标成“已做到”“在测试”“未来计划”。若最吸引人的一句都落在未来，先检查今天的价值，而不是补一个更热的标签。
来源边界
依据官方介绍与完整逐字稿整理。市场冷暖、公司评价、收入与竞争路径均保留 2026 年 7 月归因，未独立审计或测试。损坏的品牌、发布时间和估值不扩成最新事实；Ambi 故事与期权比喻为主讲转述。片尾音乐不作观点。编辑推论和练习另行标识。
回听导航
- 01:03—05:34：模型与应用的融资情绪摆动。
- 05:39—09:07：出海话术与真实市场行动。
- 09:34—13:21：技术阶段与领先者变化。
- 13:57—16:47：人、模型知识与真实需求。
- 17:13—21:24：连续服务用户、融资故事与存活。
- 21:24—25:23：历史周期与尚未回答的转折。
相关主题：应用战略、需求连续、模型依赖。延伸收听：曲凯在 2024 年初提出的交付路径、朱宁谈必要不充分与判断偏差。
English edition
As models change, should applications preserve the problem or today's solution?
Speaker: Qu Kai, founder of 42章经 | Solo episode with no interview guest
Originally published: July 19, 2026 | Duration: 26 minutes 36 seconds
Edited: October 18, 2026 | Official episode
“Applications are dead” is the episode's sharpest phrase, reported from some investment firms. It does not mean users have stopped using applications or constitute statistics proving disappearance. Model momentum, financing preferences, and customer value can move differently at the same time.
This midyear review contains strong judgments and unanswered questions. It helps examine whether startup actions have separated from their narrative: are we serving particular users, or waiting for capital to prefer our story again?
1. Colder financing does not directly establish whether applications are useful
Listen: 01:03–05:34. Qu recalls successive swings: uncertainty about model–application relationships in 2023, cooling and recovery in 2024, reasoning and agents sustaining expectations in 2025, and stronger models alongside colder application funding in his circles during 2026. He says this is the first time he has heard firms declare the category dead.
These observations come from venture investment, financing advice, and founder conversations, without a complete institutional sample or investment database. They describe signals he encounters rather than the entire market. The provocative title highlights a contradiction without defining an industrial fact.
Capital rewarding models also does not eliminate every application opportunity. Funding strategies face different constraints: large funds may cover multiple sectors and stages, while small funds cannot necessarily copy them. Startup operators should not mistake a fund's portfolio strategy for their own operating plan.
Editorial inference: Keep separate records of customer value and external capital's willingness to take risk. Financing affects survival time without replacing demand research. Funding success likewise cannot replace delivery and payment.
2. “Serving overseas” needs actual overseas execution
Listen: 05:39–09:07. Qu criticizes teams claiming an international focus while distributing English promotional material mainly through Chinese networks and attracting their existing circles. Choosing a domestic focus is another matter; the issue is a mismatch between declared objectives and actions.
He cites Dify, Workmagic, ACE Studio, and Meshy as persistent international examples. Business figures and comparisons remain contemporary observations without audit here. Broad statements about stronger willingness to pay overseas cannot replace validation by country, category, channel, and cost.
The mechanism is not that relocation guarantees success. Feedback collection must actually occur among intended users. Language is one surface; channels, communication, purchasing motives, support hours, and payment experiences affect continuing use.
Editorial inference: Examine where reach originates, where effective users are, who pays, and how their feedback changes the next release. One English demonstration is not internationalization, and an overseas registration does not establish paid value. Geographic, channel, and revenue records answer more than repeated labels.
3. When technical routes change, old rankings cannot evaluate new tasks
Listen: 09:34–16:47. Qu suggests studying companies whose reputations change quickly, not only familiar leaders. Google and 智谱 illustrate his view that competition may reorder around conversation, coding, and extended autonomous tasks.
Specific rankings, personnel movements, and strategic choices are July 2026 judgments and reports without standardized evaluation in the episode. Referring to a public letter does not mean the original was independently checked here. The useful method is to retest existing choices when tasks change, rather than publish a current leaderboard.
Model knowledge advantages may also narrow quickly. When everyone discusses agents, proactive agents, and harnesses, explaining the user's actual job can become scarcer. A harness here means the surrounding system organizing models, tools, and execution constraints. Understanding the word does not establish a business need for that system.
Editorial inference: Compare concrete tasks after an upgrade: success, failure types, takeover costs, and confirmation of completion. If an old model limitation disappears, what product value remains? These questions are more operational than permanent allegiance to a company or regularly following new vocabulary.
4. Changing solutions with technology differs from changing users with capital
Listen: 17:13–21:24. Qu advocates following model and capital waves, immediately qualifying the idea: keep finding better ways to serve the same users and solve the same class of problems. Demand and organizational learning may remain stable while implementation changes. This differs from continually switching fashionable categories.
He describes reconsidering advice to Ambi founder Mo Zihao. Qu instinctively suggested emphasizing models while fundraising for a product combining software and hardware. The founder asked why that was necessary now if current engineering supported the experience and models could come later. Product effectiveness remains Qu's report, not independently validated here. The useful moment is recognizing advice driven by investor taste.
Presenting future plans as current capabilities can redirect resources. Teams may fund unnecessary research to maintain a narrative rather than solve the user's immediate problem. Explicitly distinguishing present delivery and future dependencies reduces this drift.
The idea that products and founders are options on models is a metaphor relayed in the episode, not a verified named quotation. It suggests accumulated work might gain value as capabilities mature. Qu stresses learning and cash flow to avoid disappearing first. Yet not every asset will pay off: absent demand, missing permissions, and broken economics may survive stronger models.
5. Historical cycles offer questions rather than a date for application recovery
Listen: 21:24–25:23. Qu compares GPU and token charges with older computing and telecommunications charges, asking whether cheaper inputs could support new application models. Tokens measure processed content, and their price is only part of product costs.
The comparisons suggest that profit locations change, but the specific IBM charging history and telecommunications periods were not checked here. They are not precise clocks. Lower unit prices can coexist with more usage, longer context, tool costs, and human work. Advertising requires adequate revenue, audience, and service economics rather than becoming viable automatically after cheap inference.
Qu invokes the installation, turning point, and deployment framework in Technological Revolutions and Financial Capital, ending with questions: eruption or frenzy, and how far from a turning point? This is an observational lens, not an answer found in the episode or a date averaged from other revolutions.
His belief that one company will not dominate permanently and models will not absorb everything is not an economic theorem. Asking what story follows world models does not establish that technical development has reached an endpoint.
Editorial inference: History helps define scenarios: colder capital, falling costs, or a supplier entering applications. Each requires examining demand and transferable capabilities rather than merely waiting for a macroeconomic reversal.
Exercise: preserve continuity through the next change
For a new technology or term your team wants to adopt, write four short accounts:
- Preserve: Which users and tasks remain? If both change, acknowledge a new exploration rather than calling it an upgrade.
- Change: Which step does the new model replace? Compare results, costs, and human takeover on the same tasks.
- Accumulate: Which customer understanding, lawful feedback, and delivery skills survive replacement of this interface?
- Survive: How long can cash, commitments, and skills support validation? Which failure requires narrower scope?
Mark each capability in the fundraising presentation as delivered, under test, or planned. If the most attractive sentence is entirely future tense, examine current value before adding another fashionable label.
Source boundary
This reading uses the official introduction and complete transcript. Market conditions, company assessments, revenue, and competitive routes retain July 2026 attribution without independent audit or testing. Damaged brands, release dates, and valuations are not expanded into current facts. The Ambi story and options metaphor are the speaker's reports. Closing music is excluded. Editorial inferences and the exercise are identified separately.
Listening navigation
- 01:03–05:34: Financing sentiment around models and applications.
- 05:39–09:07: International narratives versus market execution.
- 09:34–13:21: Technical stages and changing leaders.
- 13:57–16:47: People, model knowledge, and demand.
- 17:13–21:24: Continuing user service, funding stories, and survival.
- 21:24–25:23: Historical cycles and unanswered turning points.
Related themes: application strategy, continuous demand, model dependence. Further listening: Qu's early-2024 delivery routes and Zhu Ning on insufficient conditions and judgment biases.