中文精读
三种AI产品的PMF：让不会的人完成、让专业的人改完、让喜欢的人回来
42章经单期精读 · AI产品与市场匹配
嘉宾：赵充（AiPPT创始人）、Joe（ACE Studio创始人）、胡修涵（捏Ta创始人），身份依节目当期介绍｜主持：曲凯
发布：2024年5月11日｜时长：79分21秒｜整理：2026年10月18日
官方节目与收听入口
先分清产品需要证明什么
PMF是产品与市场匹配：有一群人持续觉得产品值得使用或付钱。本期把三家公司放在一起，恰好说明它没有一个适用于所有业务的数字。低频办公工具可以先证明每次需求值得付费；专业创作软件要进入真实交付；角色内容产品则要解释人为什么不断回来。
主持人介绍的收入、正ROI与留存，是2024年节目的阶段性判断。ROI指投入带来的回报，但本期未给统一计算方式；“421”即次日40%、一周20%、月度10%的历史留存经验线，也缺少完整队列与分母说明。因此，三家的数字不能直接排成优劣榜，更不能外推为今天的经营情况。
1. AiPPT的入口：替低频用户跨过空白页
节目线索：05:28–07:25、11:17–16:37。 熟练的投行或广告公司员工，可能对每一页排版都有要求。赵充描述的另一群人，一年只做两三次PPT，甚至第一次做就是用AiPPT；具体例子是家长帮孩子准备作业。对他们，学习传统软件的大量按钮，本身就是完成任务前最大的成本。
AiPPT把入口改成一句主题或导入既有大纲，再选风格生成演示文档。当时AI主要负责大纲，后续版式匹配依赖已有产品。赵充估计AI只占实现的10%–20%，用户却强烈感到“有人替我做”。这里可迁移的机制是重排工作顺序：先表达意图，再检查结果，减少从零搭建。
这并不证明质量无需提升，也不说明模型不重要。它说明产品价值可以集中在某个阻碍最大的步骤。低频也留下经营问题：团队计划用配图、文档、数字人和协作扩大使用场景，但没有展示这些新增功能已经改善留存的结果。“多做产品留存自然提高”仍要用共同用户的真实需求验证。
2. AiPPT的放量：先知道买来的用户能否养活业务
节目线索：04:44–05:10、17:27–30:17。 赵充称，团队在做产品时已经研究流量分布、获客成本和付费率，能赚钱才持续放量。他把月收入达到100万元视为内部加码节点。开场还提到数百万元月收入，与中段说法的时点和口径没有完全交代，不能替节目补成一条精确增长曲线。
这条路径背后有十多年经营经验、已有赚钱公司、招聘与渠道资源。赵充通过访谈目标行业多家公司不同岗位的人，判断竞争、内容、增长和技术，再找有经验的操盘手共创。CEO负责组织和关键资源，是这种组织条件下的分工；主持人也追问，小团队可能没有钱招到同样的人。
版权同样是交付的一环：PPT需要配图，向大客户输出API与素材时，对方会关心授权风险。节目提及视觉中国的战略资源，价值在于支持内容供给及合作信任。不能据此推导素材库永远安全、生成图天然没有版权问题，或资源型创业适合所有团队。
3. ACE Studio的付费点：最后一段修改决定能否交付
节目线索：32:45–40:37、50:37–51:33。 Joe把歌声选为切入点，因为编曲的许多环节可在电脑里完成，真人演唱却常要约人进录音棚。他希望先用AI歌声补上昂贵断点，再扩成完整音乐工作站。
他用“85%满意”的生成结果说明专业工具的难题：歌词或唱法还有一点不对，整首歌就难交付；重新生成又可能丢掉满意部分。ACE Studio强调逐音符、逐词和唱法修改，使创作者保留成果、继续打磨。85%是解释用的例示，并非公开质量测试。
另一个具体发现是“转谱”：把哼唱转成可编辑的音符与歌词。音乐人原本就会哼唱作曲，功能把这种自然表达接入软件，省去逐个输入音符。团队称它让用户翻倍，但节目没有给观察窗口、用户定义与对照组，适合视为需求线索，不能当独立因果结论。
Joe还判断工作流未来能整合模型，理由之一是音乐模型门槛可能有限。相关训练成本与行业规模都是当时估计。产品经理可以据此追问模型更换后用户为何仍留下，但不能把“工具将吞并模型”写成已经实现的产业结局。
4. ACE Studio的出海：信任在行业网络里形成
节目线索：41:29–50:02、55:22–56:26。 团队最初按替代歌手成本定价：一年约千元，使用两次似乎回本。真实用户却可能担心学不会，因此经济上划算不足以自动促成购买。Joe称国内新增付费率约1%–1.5%；这属于团队口径，未说明完整统计范围。
冷邮件联系百余位YouTuber无人回应后，团队赴NAMM音乐展现场演示。观众带朋友再来，部分行业人士留下联系方式，随后推广者开始回应。值得学的是从陌生推销变成同行看过、愿意转述的产品；展会并非一个可以机械复制的获客按钮。
Joe称少量合作视频推动收入增长，美元收入占比从约10%升到90%，海外新增付费率约15%。这里同时发生了市场、渠道与用户构成变化，不能推出所有海外用户都更愿付费。团队还发现音乐人用Instagram联络、在YouTube学习；搜索广告表现很差。同一产品需要找到真正的行业触点。所谓ROI从0到100缺少算法、成本范围和周期，不宜拿来算别人的投放预算。
5. 捏Ta的转向：喜欢角色，不等于想编漫画
节目线索：58:32–01:10:25。 捏Ta最初把用户生成角色壁纸的热情，推成了创作漫画的意愿。实际碰到两端：专业用户要精确控制，普通用户不想构思剧情。团队继续观察留存者，发现他们在“养”自己的角色，反复为同一角色增加形象与经历。
产品于是把创作拆开。专业创作者提供剧情骨架，用户选择或塑造角色进入剧情；AI结合人设与过去事件生成反应，并在关键点生成照片，新经历回到角色记忆。主线相对固定，角色反应个性化。用户获得“我的角色经历过这件事”的结果，少了承担整部故事编剧的负担。
胡修涵把AI定位为组合者：组合预制资源、角色与个人线索。固定剧情是维持质量的一种产品选择，不能把它说成所有用户都不想自由创作，也不能把2024年的原创能力评价当作AI的永久上限。关键问题是哪些部分需要稳定，哪些部分变化才能让人感到属于自己。
6. 捏Ta的留存：消耗增加之后，还要证明付费成立
节目线索：01:10:59–01:19:07。 团队描述旧版本有少数长期用户，但次日和周留存弱。补上剧情线索后，生成次数、分享和留存提高，投诉也从创作体验、角色功能、社交转向“体力不够”。体力会随时间恢复，用户不愿等，提示当下存在继续玩的需求。
但抱怨体力不足不是商业模式完成。团队当时仍在通过激励广告、小范围购买测试付费，提到部分产品方案已有续费复购，尚未大规模商业化。过早向创作收费也可能压低内容供给，应同时观察消费意愿与创作者活跃。
团队用社媒内容提前约三个月试探普通人是否愿看，避免把科技爱好者称赞技术当成内容需求；又让一两名核心成员连接用户观察与模型优化。机制是缩短“看见问题到动手调整”的距离，不是人数越少必然越强。
用这期做一次证据练习
以下是编辑建议。为自己的产品写三句话，每句配一条可观察证据：
- 用户原本卡在哪一步？参照AiPPT的空白页，记录一个不会使用传统工具的人如何完成真实任务。
- 用户得到初稿后还要改什么？参照ACE Studio，保留一段满意结果，只修改出错局部，观察是否仍需重做全部。
- 用户为何下次还回来？参照捏Ta，把角色延续、奖励驱动、生成消耗、付费复购分别记录，避免用一种信号代替另一种。
三种路径可以共存，但每一条都需要明确客户、场景与统计周期。经营数字是验证入口，解释数字如何产生才帮助下一次决策。
回听导航与关联阅读
- 11:39：AiPPT究竟服务谁。
- 22:01：增长、产品与内容版权。
- 34:36：生成与可控修改的差别。
- 44:47：NAMM现场如何形成信任。
- 58:31：捏Ta怎样纠正漫画创作假设。
- 01:09:06：AI作为组合者。
- 01:15:07：内容探路与短迭代闭环。
相关主题：产品与市场匹配、专业创作工作流、角色关系与内容消费。可结合Albert谈寻找PMF与陶芳波谈个人模型继续比较。
来源边界：完整阅读所提供转写、旧精读与官方当期介绍后重新组织，未逐句复听。名称按官方介绍校正；多人转写的说话者标记不完整，按官方三段访谈归纳，未把不确定句子作逐字引语。收入、留存、付费、技术和竞争判断均保留2024年访谈口径，未经经营后台或独立实验验证。
English edition
Three Routes to AI Product Fit: Help Beginners Finish, Professionals Refine, and Fans Return
42 Zhangjing episode analysis · AI product–market fit
Guests: Zhao Chong, founder of AiPPT; Joe, founder of ACE Studio; Hu Xiuhan, founder of 捏Ta. Roles follow the episode’s contemporary introduction. | Host: Qu Kai
Published: May 11, 2024 | Duration: 79 minutes 21 seconds | Prepared: October 18, 2026
Official episode and listening link
Decide what the product needs to prove
Product–market fit, or PMF, means that a group of people repeatedly finds a product worth using or paying for. These three interviews show why no single number proves it for every business. An occasional office tool can first demonstrate that individual tasks merit payment. Professional software needs to participate in real delivery. A character entertainment product must explain why people return.
Revenue, positive ROI and retention are the program’s assessments of progress in 2024. ROI describes returns on an investment, but the episode provides no consistent calculation. Its historical “421” retention reference means 40% next-day, 20% weekly and 10% monthly retention, without complete cohort or denominator definitions. These numbers cannot rank the companies or establish their current performance.
1. AiPPT: Help occasional users get past the blank page
Episode reference: 05:28–07:25 and 11:17–16:37. Experienced investment bankers or advertising staff may have exact requirements for every slide. Zhao describes another audience: people making presentations two or three times a year, sometimes for the first time. One example is parents helping children with homework. Learning numerous software controls can be their largest obstacle before the actual task begins.
AiPPT changes the starting point to a topic sentence or an imported outline, followed by a style selection and generated presentation. At the time, AI mainly produced the outline; existing product systems handled subsequent layout matching. Zhao estimated AI represented only 10–20% of implementation, while users strongly perceived work being done for them. The transferable mechanism is a change in sequence: express intent, then inspect a result, with less construction from scratch.
This does not establish that quality needs no improvement or that models do not matter. Value can concentrate in the step that previously blocked completion. Occasional use also creates a business question. The team planned illustration, documents, digital presenters and collaboration to expand use, but supplied no results showing those features had improved retention. The belief that a broader portfolio naturally raises retention still requires evidence of shared customer demand.
2. AiPPT: Establish whether acquired customers can support the business
Episode reference: 04:44–05:10 and 17:27–30:17. Zhao says the team investigated traffic distribution, acquisition costs and payment conversion while developing the product, then scaled while it remained profitable. Monthly revenue of RMB1 million served as an internal point for increasing investment. The introduction also mentions several million renminbi per month; differences in timing and accounting are insufficiently explained to reconstruct a precise growth curve.
This approach relied on more than a decade of operating experience, an existing profitable company, recruitment and distribution resources. Zhao describes interviewing people across several functions at competing businesses, assessing content, growth and technology, then recruiting experienced operators. A CEO focused on organization and strategic resources makes sense within those conditions. The host explicitly asks whether a small startup can afford equivalent hires.
Copyright also belongs in the delivery process. Slides need images, and large customers receiving APIs and content care about authorization. The interview describes Visual China’s strategic resources as supporting content supply and trust in partnerships. It does not establish permanent protection from a content library, copyright immunity for generated images, or the universal suitability of resource-driven entrepreneurship.
3. ACE Studio: The final edit determines whether work can ship
Episode reference: 32:45–40:37 and 50:37–51:33. Joe chose singing as an entry point because many production tasks already happen on a computer, while a human vocal may require booking a singer and studio. His proposed progression was to fill this expensive gap with AI vocals, then build a complete music workstation.
An output that is “85% satisfactory” illustrates the professional problem. A small lyrical or vocal defect can make a song unusable, while regenerating may destroy the parts the creator likes. ACE Studio emphasizes editing individual notes, words and singing expression so creators can preserve progress and refine it. The percentage is an explanatory example, not a published quality benchmark.
Another discovery was transcription: turning humming into editable notes and lyrics. Musicians already compose by humming; this feature connects a familiar expression to the software and avoids entering notes one at a time. The team says it doubled users, but gives no observation period, user definition or comparison group. It is a demand signal rather than an independently established causal effect.
Joe also argues that workflows may eventually integrate models, partly because he believes music models have limited technical barriers. Training costs and market sizes mentioned here were contemporary estimates. Product managers can ask why customers would stay after a model change, but should not turn the idea that tools will absorb models into an established industry outcome.
4. ACE Studio: Build trust within the industry’s actual networks
Episode reference: 41:29–50:02 and 55:22–56:26. Initial pricing used the cost of replacing a singer: roughly RMB1,000 annually appeared to pay for itself after two uses. Customers could still worry about being unable to learn the software. Financial value alone therefore did not complete the purchase decision. Joe reports domestic new-user payment conversion of approximately 1–1.5%, without a full statistical definition.
After cold emails to more than a hundred YouTubers went unanswered, the team demonstrated at the NAMM music trade show. Visitors brought friends, industry participants left contact details, and promoters subsequently began replying. The lesson is the move from an unfamiliar sales pitch to a product peers have seen and can describe. A trade show is not a mechanically repeatable acquisition button.
Joe attributes revenue growth to a small number of sponsored videos and reports dollar revenue increasing from roughly 10% to 90% of the total, with overseas new-user payment conversion around 15%. Market, channel and audience composition changed together. This does not show that every overseas audience is more willing to pay. The team also found that musicians networked on Instagram and learned through YouTube, while search advertising performed poorly. Identifying the industry’s real meeting places mattered. Reported ROI moving from zero to 100 lacks cost boundaries, formula and period, so it cannot support someone else’s advertising budget.
5. 捏Ta: Caring about a character does not mean wanting to write a comic
Episode reference: 58:32–01:10:25. The team initially extrapolated from enthusiasm for character wallpapers to a desire to create comics. It encountered two groups: professionals requiring precise control and ordinary users unwilling to invent plots. Continued observation of retained users showed people repeatedly developing the appearance and experiences of their own characters.
The product then separated the creative work. Professional creators supplied plot structures. Users selected or shaped characters to enter them. AI combined personality and past events to generate reactions and pictures at key moments, with new experiences added to character memory. The main plot remained relatively fixed while character reactions varied. Users could obtain the experience of “my character went through this” without taking responsibility for writing an entire story.
Hu describes AI as a composer of existing resources, characters and personal cues. Fixed plots are a product choice for maintaining quality. They do not show that all users reject free creation or make a 2024 assessment of originality a permanent ceiling on AI. The practical question is which parts should remain stable and which variations make the result feel personally owned.
6. 捏Ta: More consumption still needs a payment test
Episode reference: 01:10:59–01:19:07. The team describes an older version with a small group of persistent users but weak next-day and weekly retention. Adding plot cues reportedly increased generation, sharing and retention. Complaints moved from creation problems to character features, social functions and finally insufficient energy. Because energy replenishes over time, unwillingness to wait suggests immediate demand to continue.
An energy complaint does not complete a business model. The team was still testing rewarded advertising and small purchase offers, mentioning renewals and repeat purchases for some offers without large-scale commercialization. Charging creation too early might also suppress content supply. Willingness to consume and creator activity both need observation.
The team tested content on social media roughly three months ahead of the product, asking whether ordinary viewers wanted it, rather than treating technical enthusiasts’ praise as content demand. It also connected user observation and model optimization through one or two core people. The mechanism is a shorter distance between seeing a problem and changing the product; fewer employees do not automatically produce better decisions.
An evidence exercise
These are editorial suggestions. Write three statements about your own product, each paired with an observable example:
- Where was the customer previously stuck? Following AiPPT’s blank-page example, record how someone unfamiliar with traditional software completes a real task.
- What must change after the first output? Following ACE Studio, preserve a satisfactory section and edit only a defect, observing whether the whole output must be redone.
- Why does the user return? Following 捏Ta, record character continuity, reward incentives, generation consumption and repeat payment separately, so one signal does not substitute for another.
The three approaches can coexist, but each requires a defined customer, situation and measurement period. Business numbers begin verification; explaining how they arise helps the next decision.
Listening guide and related reading
- 11:39: Who actually uses AiPPT?
- 22:01: Growth, product and content rights.
- 34:36: Generation versus controlled editing.
- 44:47: Trust emerging at NAMM.
- 58:31: Correcting the comic-creation assumption.
- 01:09:06: AI as a composer.
- 01:15:07: Testing content and shortening iteration.
Related themes: product–market fit, professional creative workflows, character relationships and content consumption. Compare Albert on finding PMF and Tao Fangbo on personal models.
Source boundary: Reorganized after reading the complete supplied transcript, prior analysis and official contemporary introduction; audio was not checked sentence by sentence. Names follow the official introduction. Speaker labels are incomplete, so attribution follows the three official interview segments and uncertain statements are not presented as verbatim quotations. Revenue, retention, payment, technical and competitive claims remain attributed to the 2024 interview and have not been verified against company records or independent experiments.