Dai Yusen: Revisiting AI Investment Judgments, Harnesses, and Startup Opportunities

Zhang Xiaojun invites ZhenFund managing partner Dai Yusen to revisit his judgments about AI and markets six months earlier and discuss investing, products and entrepreneurship. They state that personal investing experience is shared for discussion, not as investment advice.

In the previous episode, Dai was cautious about AI returns and a market correction. Facing subsequent market and product developments, he says early investors must judge with incomplete information and often be contradicted by new facts; continual review matters.

He recalls attention shifting rapidly among model companies over recent months. In AI, a product that looks ahead at one moment may soon face new competition; these comparisons reflect the interview’s timing, not a permanent ranking.

Dai invokes “strong opinions, weakly held”: form a clear view, understand its premises and adjust when those premises change. Public statements can make it harder to change one’s mind, he acknowledges, and he is working against that tendency.

He believes he underestimated the speed of growth in coding agents and enterprise demand. As models and tools made longer coding tasks feasible, he adjusted his view of the timeline for AI’s commercial returns.

Dai says his own public-market style is trading rather than long-term holding. After seeing coding-tool usage grow he rebuilt some positions related to hardware bottlenecks; he notes his prior public views affected his decisions and does not recommend others follow them.

The more strongly one expresses a view publicly, he says, the harder it can be to acknowledge change. Friends encouraged him to keep discussing and learning; he treats market feedback and listener criticism as tests of his thinking.

Reviewing his earlier view, he still sees merit in concerns about raising consumer subscription prices and monetizing advertising and commerce, but did not anticipate the speed of coding and enterprise usage changes. His growth and revenue assessments are observations.

Dai thinks AI capability can cross a threshold where a task suddenly becomes useful, making timing hard to predict. He uses heating water as an analogy for nonlinearity but acknowledges that hindsight does not make a turning point predictable in advance.

He thinks high-quality coding data and agent training may have contributed to longer task loops, but does not treat model naming or any one factor as sufficient evidence. In retrospect he sees the product experience change as a clearer turning point.

Asked when he understood Anthropic, Dai first stresses that much of his information is secondhand; he can offer observations but cannot claim a complete internal view.

He believes a decentralized, bottom-up organization may suit exploration, while aligned goals and concentrated execution may help when the main direction becomes clearer. He compares Anthropic and OpenAI as an outside observer.

Dai discusses hiring, values and internal communication, suggesting alignment may support rapid iteration while a more divergent organization may preserve seeds of innovation. A new paradigm could change which form is suited, he says.

Coding is not only an isolated vertical, he argues: tool capability can support research, healthcare and other work. This became clearer after products appeared, but was not universally foreseen.

On model-company business models, Dai says selling APIs alone may face pricing and substitution pressure; a user-facing product affects experience and may, under appropriate conditions, provide data for improvement. Actual feedback depends on the product and permissions.

Zhang asks whether only model companies can build good harnesses. Dai says no: model companies know their models, but independent developers can innovate in product design and workflows, with OpenClaw among his examples.

Discussing valuations, Dai separates short-term market sentiment, medium-term returns and long-term potential. Different horizons can yield opposite judgments; his valuation, listing and return projections are predictions from the interview, not facts or investment advice.

He says the gap between leading coding tools can change, while habits, configuration, price and distribution affect choice. His price and performance comparison reflects that moment rather than a lasting relationship.

On returns, Dai argues model-company revenue is also customer expenditure. Customers must turn tokens into software or services and ultimately increase revenue or reduce costs; he frames this as input, output and outcome.

He questions whether more coding capacity automatically yields proportional revenue: some companies may lack worthwhile product ideas rather than programmers. Zhang suggests the results may also take more time to appear.

Dai sees a potential timing mismatch: token costs occur now, while new product discovery and revenue may emerge gradually. That does not make all usage worthless; concrete value needs testing.

As coding costs fall, he observes some companies entering one another’s product areas and copying existing features more quickly. Finding genuinely new demand remains a product and innovation problem, not just a coding problem.

On cost reduction, he distinguishes layoffs from automation: engineers also communicate, coordinate, build trust and take responsibility. Automating some tasks does not mean a role can simply be removed; he compares this to edge cases in autonomous driving.

Fast model-company revenue growth alone does not prove end customers have sufficient returns, he says. New users may be trying tools while others leave for insufficient value, so retention and actual outcomes need observation.

For small firms or individuals who had ideas but lacked engineering resources, coding tools may be especially useful, Dai says. Larger organizations may also be constrained by product direction and coordination.

Zhang notes that coding capability can affect knowledge work more broadly. In investment research, Dai says, AI can produce more reports, but people still check information and own decisions, so increased output does not necessarily yield proportional outcomes.

He admits it may be early to judge returns, but argues infrastructure spending is large enough that the question cannot be deferred indefinitely. His expenditure and revenue figures are estimates from the time of recording.

On 2026’s main theme, Dai says he has long followed AI productivity, especially agents that can perform tasks with some independence. He discloses that his fund invested in related companies, providing context for his observations.

When agents take on longer tasks, people may turn some attention elsewhere, he says; actual time savings still depend on the task, supervision and result quality.

Explaining a harness, he compares it to a racing team: the model supplies capability, while context, tools, task loops and a runtime help it perform a specific task.

Context may include live information or proprietary organizational knowledge; tools and loops determine how an agent acts. Dai cites scheduled checks, file operations and sandboxes as engineering combinations that can extend tasks.

He discusses persistent runtimes: an agent may need an environment that retains state, accesses software and continues tasks. Model companies or third parties can supply these layers around the model.

Zhang asks whether this conflicts with “the model is the product.” Dai thinks both model capability and product packaging matter: people normally interact through a product, and interface and workflow affect usability even when capability comes from a model.

Advanced models remain essential for tasks, he says, while the product layer can supply distinctive value and feedback. He sees model and application improvements as potentially reinforcing rather than mutually exclusive.

Dai suggests real-task traces and feedback may improve models, citing coding tools and training. Such a flywheel depends on actual permissions, data quality and training practices; one cannot assume all user data is used for training.

As an early investor, he argues that even if a product category is eventually consolidated, it can still create user value or investment returns in the meantime. This reflects his investment perspective.

He observes that users may stay with a harness to retain workflows, memory and configuration while switching underlying models to compare performance and cost. He infers the product layer may develop some stickiness.

On researchers’ differing views of frameworks and models, Dai says they need not conflict: both strong models and good tools affect performance. Model providers know their models, while application teams may focus more on specific user experiences.

He uses presentations as an example: a researcher may consider appearance secondary, whereas a user needs an attractive result. Model capability and product design can reinforce each other, and third-party applications can also use stronger models.

Model vendors will make important applications, Dai says, but need not cover every user need. Operating systems host both platform-owned and third-party software, illustrating a possible ecosystem with many participants.

He reviews sandboxes, browser operations and parallel research in products such as Manus as application-team innovations around models and task loops. He notes his perspective as an investor in related companies.

On command-line interfaces and Claude Code, he sees tool interfaces, skills and task loops helping models work. Product advantages come from interaction and engineering design as well as the underlying model.

Dai says OpenClaw’s local access, scheduled checks, memory and integration with familiar chat tools make it feel closer to users’ work. He acknowledges stability, context-confusion and permission-security concerns.

He likens a harness to an operating system and a model to a processor, with applications or skills on the runtime. The analogy explains division of labor, not an assurance that models are interchangeable without cost.

In his vision, application developers can use existing tool layers for loops, memory and model communication while focusing on their own features and experience. That may reduce some development barriers.

On entrepreneurship, he cautions against following trends for safety when technology changes quickly. He argues general-purpose products can discover needs across settings and that vertical niches are not the only opportunity.

Asked whether coding agents have become too homogeneous, Dai sees that risk and says public standards for AGI keep changing. Creating genuinely new knowledge and value remains an open question.

Could stronger models absorb the application layer? Possibly, Dai says, so products need advantages in experience, workflows, data or other areas; application teams may also develop their own models.

He imagines agents exchanging value because different users’ agents accumulate different knowledge and context, producing different results for the same task. This network effect is only a possibility he proposes.

For example, a host’s agent might acquire distinctive interviewing experience and others might pay for help; the price could include model costs and proprietary expertise. This is a thought experiment, not a validated market.

Asked whether stronger models reduce application opportunities, Dai says he has observed more funded projects and higher valuations and thinks stronger models and infrastructure enable more use cases. It is not a comprehensive market survey.

As memory, sandboxes and other components mature, application teams may experiment faster, he says, while brand, user relationships, data and concrete product experience can still affect competition.

He advises entrepreneurs to notice opportunities large companies temporarily overlook: a direction may appear too small, early or unfashionable while its demand and product value are underestimated. His examples and categories are retrospective.

Open source may help products attract users and later explore commercialization, he says, but sustainable revenue must be demonstrated. He acknowledges practical permission and stability risks in local agents.

At ZhenFund, Dai says selecting projects still starts with people. He cites young robotics researchers whose technical judgment and entrepreneurial intent the team followed for years, rather than simply pursuing a hot topic.

Learning, leadership, creativity and determination remain important across technology cycles, he says; founders at the frontier also need independent understanding of problems not yet widely recognized.

He mentions teams working on dexterous hands and sees potential interaction among hardware, costs and training data. His claims about company rankings, valuations and prospects are investor observations.

For a particular founder, Dai values sustained commitment, independent technical understanding and motivation. He admits wondering if the direction was too early; later progress does not erase the original risk.

He describes the fund’s founder framework—learning, creativity, leadership and determination, plus trust, complementary partners and team experience. These are judgment tools, not a formula guaranteeing success.

Asked whether founders need doctorates, Dai says deep technical understanding matters but a PhD is not mandatory. Research-driven fields and application startups may require different experience.

On AI product managers, he says leaders need their own judgment about users and the future and must organize resources to execute. Even as model training becomes systematic, individual product judgment can matter.

He cites a developer who connected a computer and messaging tool to keep using a coding agent from a phone, then expanded the product. Hands-on use and technical understanding can reveal new needs, Dai argues.

Zhang asks which lessons from the mobile internet era may no longer apply. Dai says attention and time spent were important measures, but if agents can spend long periods completing valuable tasks, keeping users glued to a screen may not be the best goal.

He suggests that closed platforms’ former advantages could become constraints under new forms of interaction: if a user-authorized agent cannot access information somewhere, users may turn to tools that collaborate more easily. This is his observation about potential product changes.

Mature products must balance safety, stability, and the needs of a broad user base, while new products may initially serve people willing to tolerate complexity. Dai says the tradeoffs differ; innovation does not automatically remove permission and security risks.

He believes subsidies can attract trials, but sustained use requires clear value. He cites voice experiences as an example of potential word-of-mouth adoption, without denying the role of marketing and distribution.

On daily active users, Dai says the metric is useful, but productivity tools should also be assessed on whether they complete valuable tasks. Requiring frequent visits to increase activity can conflict with the goal of saving users’ attention.

For early products, he looks for users who truly love them rather than only a grand narrative. Using a product himself, he says, helps him judge whether it solves a specific problem.

Dai compares agents entering the digital world to new participants entering human society: first more people acquire agents, then agents use existing software and infrastructure, and eventually tools better suited to machine-to-machine collaboration may appear.

He discusses possible needs around payments, software interfaces, and sandboxes: machine tasks may involve frequent small transactions, while existing interfaces are mostly designed for people. These are ideas for product opportunities; deployment still requires solutions for authorization, security, and accountability.

If programming becomes cheaper, he imagines existing software could improve and more personalized or one-off software could emerge. People must still discover these needs and platforms must provide suitable interfaces, so change will not happen overnight.

On startup organization, he believes AI could make product teams smaller and more cross-functional, while sales, business development, and offline work will still need people. Different tasks may change at different speeds.

New companies that organize knowledge and processes for agents from the start may find AI easier to use, he says, while established companies must deal with scattered data and existing departmental boundaries.

Dai uses the historical transition from steam power to electricity in factories as an analogy: new technology does not automatically produce productivity gains; layouts and work organization also had to change. He believes adopting AI likewise requires changes to processes and organizations.

Zhang asks what he means by saying the next company on ByteDance’s scale may not resemble ByteDance. Dai says important companies in each era may solve different problems and use different interfaces; copying the previous generation’s feeds and marketing model could mean competing directly with established leaders.

He says entertainment products compete directly with existing games, short videos, and other entertainment: users must first find them fun, rather than accept them simply because they use AI. He is comparatively more optimistic about productivity applications that solve specific work problems.

Founders with ByteDance backgrounds may be highly capable, he says, but may sometimes need to question or even overturn what they have learned instead of applying familiar product formats directly to new technology.

Dai compares this with the internet and mobile internet: early applications of a new technology often solve old problems, while wider adoption and mature infrastructure can enable previously unviable products and business models. Collaboration among agents could offer similar opportunities, he believes.

He imagines communication, advertising, data exchange, and organization changing as more people and services use agents. Machines might exchange data through interfaces, for example, rather than always needing forms designed for humans.

He stresses that these scenarios remain early: people and multiple agents collaborating in a shared workspace and agents trading capabilities are emerging ideas. He discloses an investment in a related collaboration product.

Comparing startup themes in China and Silicon Valley, Dai says Silicon Valley has both research labs and vertical application companies, while China has visible activity in robotics, deep tech, productivity, and consumer applications. This is his observation, not a comprehensive market survey.

His areas of interest in Silicon Valley include coding, world models, and automated research. Teams define world models differently and use different data and technical approaches; their specific capabilities remain under research and validation.

In robotics, he says teams are exploring how data can improve manipulation, stability, and generalization, and experimenting with methods borrowed from model training. He notes that “world model” remains a term without a settled meaning.

Dai is cautious about consumer AI hardware: adding chat to existing devices may not create enough new value, he worries, while hardware also carries supply-chain, manufacturing, and sales costs.

For wearable recording devices, he says users must form habits around wearing and maintaining them, so products need to deliver substantial ongoing value. He emphasizes that this reflects his investment preference, not a verdict on any particular company.

His fund has invested in robotics components and model-related projects, but does not cover every area. He believes humanoid robots remain early and useful manipulation, mechanical design, and models all need progress; specific investment judgments are uncertain.

Dai says he often passes on projects that simply follow a trend. He prefers founders who pose independent questions and offer different product ideas rather than build similar versions merely because a sector is popular.

On gathering information, he says he usually prefers reading podcast transcripts and uses a small tool he built to check for updates and organize them into notes. For him, reading generally saves time compared with listening in full.

He describes using coding agents to build personal tools that turn podcasts, newsletters, meeting notes, and writing material into a daily digest and maintain an archive around people and companies. He believes this has made his personal information processing more efficient.

But he worries that accepting AI answers directly and skipping one’s own reasoning could affect understanding and independent thought. He compares deliberate thinking to exercise and cites preparing arguments for interviews as an example of practice.

He believes innovation and using AI also require practice: begin with concrete problems such as schedules, meetings, or small web tools, then gradually learn what is possible and where judgment is still needed, rather than expecting complex results immediately.

He and friends have also tried organizing data from family health devices and building a simple health dashboard together. These projects may not directly generate commercial value, he says, but helped him experience the difficulties of human-agent collaboration.

On employment disruption, Dai believes technology may spread faster than some people can adapt, making short-term unemployment and redistribution important concerns. He is optimistic about new jobs emerging over the long term, while acknowledging potentially turbulent change.

He uses historical stories about industrial technology spreading to other industries to illustrate how new opportunities may emerge through multiple stages. He cautions against applying history directly to AI, especially because the current pace may differ.

He observes that prolonged collaboration with agents may produce stronger trust, dependence, and emotional projection. Many questions remain about how society will adapt to these relationships.

Asked what abilities children should develop, Dai highlights asking questions proactively, choosing problems to solve, and remaining curious, while admitting he has no definitive answer.

He distinguishes familiar tasks from genuinely novel problems and says creativity, learning, and autonomous action deserve further study. His assessment of current models reflects his views at the time of the interview, not a definitive statement of their limits.

On whether to continue recording the program, Dai says expressing views publicly exposes them to feedback and reveals where his thinking is incomplete. He wants to keep revisiting his judgments and regards correction by new facts as part of learning.