Short-term personal agent competition depends on models; long-term winners compete on supply and security trust
Cai Wei believes model capability will eventually plateau with minimal provider differentiation; what truly separates winners is who controls more offline supply and whose security compliance work runs deeper.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Open-source projects proved models crossed the threshold for long-horizon task execution
Cai Wei notes that over this past Lunar New Year, an open-source agent project (called Open Cloud in the transcript) went viral. Though ordinary users spend two or three hours installing and configuring it—the experience is highly technical—it proved that models had crossed a threshold for long-horizon task execution: properly configured personas and skills can then orchestrate and execute numerous tasks automatically. This signal prompted nearly every major player to start personal agent projects after the new year. A hidden secondary theme is security, privacy and compliance. Muse also invested considerable time here, because for personal agents to be truly useful, users must willingly give up context to them—trust is a prerequisite, not a bonus feature.
— Cai WeiMuse notifications feel like WeChat, not marketing spam
Cai Wei sees several personal agent products with similar feature lists but divergent experiences: the gap is primarily in what he calls 'boundary sense.' He observes that every Muse notification is carefully curated with sufficient information gain, creating a feedback loop where users think 'this message is worth opening'; but if you send morning and evening digests daily with low information density, users next time treat them as marketing messages and ignore them. This explains why personal agent products generally have higher retention—push has stopped being an engagement tool and become genuine feedback from the real world.
— Cai WeiThree generations of AI products map to service industry worth 10 trillion yuan
Cai Wei maps the three generations of AI products unlocked by large models directly to market scale: Chatbots solve information retrieval, corresponding to the information sector at around 100 billion yuan; coding agents like Claude Code that complete tasks upend the software industry at the trillion-yuan scale; Personal Agents point to service industry restructuring—nutritionists, weight loss coaches, tutors and similar services now delivered by humans stand at 10 trillion yuan. He emphasizes the logic: estimate based on what currently solves this problem and how big that market is—the level of problems models solve maps directly to market scale.
— Cai WeiPersonal agents reshape supply, they don't just shift where scenarios happen
Cai Wei distinguishes two classes of innovation: mobile internet's transformation was fundamentally demand-side—smartphones unlocked location, letting users do what they couldn't before. Personal agents from this wave of large models are fundamentally supply-side restructuring: models can write answers on the fly, can recombine services scattered across different places into entirely new experiences. This distinction matters because a decade of O2O was actually laying the foundation for today's personal agents, letting them directly access offline supply that's now been moved online.
— Cai WeiKPIs as targets inevitably distort execution
Cai Wei's product management philosophy: compress a direction into a specific metric—5 million DAU, 40% retention—and you necessarily lose massive amounts of high-dimensional information; teams instinctively pick the path easiest to hit the metric, but that's usually not the right path. He prefers explaining the product's principle and value clearly, letting the team understand the 'highest-level core,' so when external conditions shift the team can adjust without worrying whether work done against an old goal counts—especially in large companies where year-end performance reviews happen and constant KPI changes confuse everyone.
— Cai WeiOne error in ten steps means the whole sequence fails
Cai Wei explains with a mathematical intuition why personal agent experience remains far from ready: even if you improve single-step accuracy from 70% to 90%, a task split into ten steps means any failure breaks the whole chain; final success rate is one minus the product of all error rates. Even if each step has only a 3% error rate, multiply that over ten steps and overall success drops to 60-70%. This is why there's a canyon between 'can call tools' and 'calls them well and satisfies users'—leaderboard models only care about the former.
— Cai WeiIntelligence will plateau; security and supply are the durable moat
Cai Wei breaks down personal agent competitive barriers by time horizon: in the short term, intelligence matters because model capability is still rapidly differentiating; but he judges intelligence will eventually plateau like the smartphone industry after iPhone 4—base models will all be much of a muchness, and what pulls ahead will be supply and trustworthy security. He sees this not as an indefinite hypothesis but as objective industry law, though we can't yet see when it will happen in the near term.
— Cai WeiPersonal agents compete on productivity replacement, not on traffic tolls
Cai Wei identifies two business models for personal agents: productivity logic, measuring whether the AI assistant's value creation exceeds its cost—if it creates 100 yuan monthly value and costs only 10 yuan, users have no reason not to use it, closer to hiring hourly labor; and internet-style traffic platform/toll logic, dependent on centralized traffic aggregation. He judges that if the personal agent ecosystem fragments, the first logic is likelier to hold because it doesn't depend on winner-take-all platform effects, just on individual unit value being high enough.
— Cai WeiIn their own words · checked verbatim
I might think of it as a marketing text rather than a real message, but when Muse sends you something, it feels like a WeChat message. With WeChat's style, nobody treats it as marketing spam.
我可能就会觉得我就会当成一条营销的短信 而不会当成一个真正的一条信息 但比如说Muse给你发消息 他可能感觉就像微信一个消息 微信谈的扑式 大家是没有人会把它当成一个营销短信的
Cai Wei37:31
Personal agent experience changes stem from model capability changes, so it's fundamentally a supply-side shift. Whether it's personal agents, the task completion we discussed, or chatbots, the essence is restructuring of intelligent supply.
PA的体验的变化 还是由模型能力的变化带来的 所以它本质上是一个供给侧的变化 所以不管是PA还是说前面讲的这个任务的完成 还是像这个Chatbot 它的本质其实是一个智能的供给重构
Cai Wei49:38
A principle I've always believed in: direction matters more than targets. Not that targets don't matter, but when we discuss a target, we usually like to make it concrete.
我一直相信的一个 理念是 我觉得方向比目标更重要 不是说目标不重要 而是当我们讨论一个目标 我们通常喜欢把它具象化
Cai Wei1:19:55
But if it's a ten-step task, any single error is an error. Your final success rate is one minus the product of all error rates. So the error compounds. Even if each step has only a 3% error rate, multiply that across a few steps and your overall success rate is just 60-70%.
但你如果一个十步的事情 你每错一步就是错 所以你最后的成功率是所有的错误率乘起来 再用一减去那个数 所以他其实错误率会非常高 你每一步哪怕只有3%的错误率 然后你多乘下几个 你最后整体的成功率只有6-70%
Cai Wei1:57:19
Without completion, you're at the ninety-mile mark of a hundred-mile journey. No matter how many steps you've completed, without finishing, you've completed nothing.
你在没有完成的情况下 你就是行百里者半九十 你没有完成 就是你不管之前你完成了多少步 你最后其实都是没有完成的
Cai Wei1:59:22
This really is something you have to fight for. Business competition is brutal. When we talk about AGI, there's a lot of idealism; when we build technology or products, there's inevitably lots of idealism. But ultimately whether you can hold ground, whether you can stand—that's all something you have to fight for.
这真的是打出来的 商业竞争还是很残酷的 我们聊AGI的时候 会有很多的一些理想的成分 包括说我们在做技术 做产品 不可避免 其实都会有很多理想的成分 但是最终能不能够守住 能不能够立住 其实最终都是打出来的
Cai Wei2:13:33
In China's fiercely competitive environment, even Sam Altman would have to hand out red envelopes.
他说在中国这样的充分激烈竞争的环境 如果是Sam Ottoman来 他也要发红包
Cai Wei2:19:38
Figures
| Three generations of AI products and China's market scale | Chatbot (information sector) 100 billion yuan; Coding Agent (software sector) 1 trillion yuan; Personal Agent (service sector) 10 trillion yuan | 44:37 |
| Ten-step task success rate | Single-step error rate 3%; ten steps compounded yields 60-70% overall success | 1:57:19 |
| Lingguang's DAU target | 10 million scale | 1:24:58 |
| Personal agent productivity model calculation example | AI assistant creates 100 yuan monthly value at 10 yuan cost; user has no reason not to adopt | 2:28:41 |
Glossary
- harness
- Task orchestration framework; the outer layer handling task decomposition and tool scheduling that determines how models are guided in specific scenarios.
- MCP
- Model Context Protocol; interface standard enabling AI to connect to external services and data under unified criteria.
- A2A
- Agent to Agent; connection protocol enabling different agents to interface with and invoke each other's services.
How to listen
Founders, product managers and investors focused on major tech AI strategy, consumer AI product design methodology and AI business model analysis.
Around 1:16:55, chat about the origin of Lingguang's name—pure trivia, can skip.