Xiao Hong on Building from Monica to Manus Amid Changing Technology

Zhang Xiaojun explains that this episode combines two interviews, one in autumn 2024 and another after Lunar New Year in 2025. Xiao Hong discusses entrepreneurship, Monica, and an agent product not yet launched; “this year” in the first interview means 2024.

Xiao introduces himself and Butterfly Effect. He previously built enterprise products in the WeChat ecosystem and started a new company in 2022 to explore generative-AI applications. Funding and product status reflect the interviews’ timing.

His first company built an editor for WeChat public accounts and later a WeCom CRM. Xiao says the two products were sold together in 2020; his second venture developed Monica for overseas users.

Why stay in Wuhan? He says the team began near its university, and familiarity, costs, and speed of building influenced the choice. For a global product, starting in a familiar place could allow faster execution, he thought.

Distance from Beijing and Shanghai meant fewer local industry connections, but the team also learned from overseas independent developers, social platforms, and technology videos. He says that exposed them to different product ideas.

He also identifies talent and fundraising access in Wuhan as challenges. Looking back at the first company, he says the team grew well but lagged some competitors in financing scale and access to capital.

In the first year after graduation, attempts including anonymous social networking and a secondhand marketplace failed to take off. Positive feedback at university did not translate automatically into startup success, and repeated failures shook the team’s confidence.

At university, he combined a technical club with work on the school’s media account. Knowing some code and account operations gave him an advantage at their intersection. Working with stronger technical peers also helped him recognize his own product strengths.

Writing a software blog trained his product judgment: explaining why a product is good requires understanding it and comparing alternatives. He says articulating value clearly became an important foundation for product work.

By 2016, savings were nearly gone and the team considered taking jobs in Beijing. A hackathon let them prototype a tool for public accounts and meet an investor willing to support them in continuing.

Xiao describes the shift from preparing to find jobs to receiving an investment agreement as surreal. The support gave the team time to refine its product, he says, rather than instantly solving every business problem.

How should a public-account editor charge? He recalls choices including advertising matchmaking and software subscriptions. The team had to consider not only users but also how product revenue could support continued improvement.

After changes to features, growth, and pricing, product revenue changed markedly. Looking back, he says a business model must match value users will actually pay for rather than merely fit a fundraising narrative.

On fundraising, he calls venture capital an expensive way to raise money: the cost may be most visible when a company succeeds and the equity given up becomes valuable. Founders need to understand capital’s long-term effect on decisions and ownership.

Xiao suggests distinguishing an investor’s business opinions from a founder’s own judgment. Investors can provide information and an outside view, but founders bear responsibility for decisions that affect the team and substantial resources.

He advocates serious discussion of consequential disagreements, recording views and reasons when needed. If a decision later proves wrong, the team can examine the information and judgment available then rather than remember only the outcome.

In competition around WeCom CRM, he recalls the team trying to anticipate actions by platforms and larger companies. A startup must consider not just user needs but also how a platform it relies on might change its rules.

He says revenue from the existing editor gave the team some room to wait and explore new directions. A product that sustains operations is itself a resource in uncertain competition.

When a platform change may happen at an unknown time, teams must balance short-term tactics against long-term dependence. Xiao says no single trick can eliminate the risks of relying on an external platform.

Capital competition changed his perspective: when rivals can mobilize far more resources, product skill alone does not automatically erase differences in funding and team size. Founders must learn the capital dimension they may not initially understand.

He recalls high expectations around enterprise software and capital markets, saying market enthusiasm affects company choices. Bubbles look easier to identify in hindsight than to time while inside them.

After selling the company, he reflects on lifestyle and mindset: continually increasing spending and expectations makes future choices more expensive. He says he still lives in the home rented during his first startup.

At the start of his second venture, he wanted to build products used by many people. Generative AI suggested new possibilities, and the team began with browser extensions and familiar search experiences.

Xiao observed that generative AI could help people understand information, write, and work, not only create entertainment. He used concrete use cases as market signals rather than discussing model parameters alone.

The early team tried placing model capabilities in search and browser workflows. Product value depended on whether people would actually use it in existing work, not merely whether a demo felt novel.

He recalls using an existing extension and distribution channels to reach users while developing Monica. Early growth requires product and distribution together; a launch cannot substitute for ongoing improvement.

Reviewing key decisions in 2023, he says the team chose to focus on overseas markets. For a resource-constrained team, serving multiple regions’ product, operations, and distribution needs at once could dilute attention.

He stresses coordination among product, technology, and growth. When a window opens, teams need both a usable product and ways to reach and support users; excellence in only one area is insufficient.

Comparing proprietary models with external ones, Xiao sees tradeoffs: model companies control core capabilities, while application companies can choose and integrate different models and adapt the experience to users.

When technology changes quickly, he says application teams can gain opportunities through flexible model selection, but model providers may also enter applications. Outcomes depend on execution, product, and timing, not an inevitable victory by one organizational form.

On whether models democratize technology, he says open models and APIs make powerful capabilities more accessible to small teams. Access alone does not solve product definition, reliability, or distribution.

He distinguishes applications that supplement existing workflows, new experiences enabled by model capability, and capabilities extending into particular industries. Each requires matching user problems with what models can do.

The second interview takes place after Lunar New Year in 2025. Xiao revisits applications such as Cursor: better models can suddenly make a previously limited product useful, and its users may extend beyond professional engineers.

He observes non-engineers using coding tools to analyze data or complete everyday tasks. By lowering interaction barriers, products can bring capabilities once confined to developers to a broader audience.

Xiao compares AI industry division of labor to chip design and manufacturing: model and application companies can focus on different problems. He respects foundational model research while arguing that additional work is needed from a user and product perspective.

Monica integrates multiple models so users need not separately register and manage an experience for each new one. Xiao says integration can reduce switching friction, while model choices change with capabilities.

In his observation, model capability often appears first and applications turn it into experiences users can perceive. Product teams need to understand capability limits while recognizing that providers may not have anticipated every use.

He proposes a startup approach: anticipate the next model capability, build a product around it, and wait for it to mature. The approach also carries risk because neither timing nor the eventual capability is guaranteed.

Zhang asks how applications can compete if model providers follow. Xiao sees opportunities in vertical workflows, demanding engineering work, and temporary windows, but says there is no fixed answer about whether a provider will enter.

He cautions that competition changes with entrepreneurs’ own actions: if a team already serves a need well, a model provider’s later choice may differ from the initial expectation. Other participants’ strategies should not be treated as fixed.

On DeepSeek’s spread, he says showing a reasoning process was a user-visible change in experience. His explanation of why other companies chose different presentation is outside speculation.

He says model capability and product presentation jointly shape user experience: technical improvements need accessible interfaces and pricing to be felt. The episode’s note corrects a few spoken product-name slips to DeepSeek.

Xiao says the team is building an agent that looks like a chat interface but can perform multi-step work. The product had not launched at the time; he thinks the area is worth exploring even though model providers may also participate.

He revisits early agent concepts: when models were weaker, application teams often had to integrate search, knowledge bases, and other APIs one by one. He compares the accumulation of features to a feature phone.

His goal is a system that can use code, tools, and a browser for long-tail needs without a developer hand-building each feature in advance. Actual reliability still depends on the model, tools, and environment.

While using a coding tool, he saw it retrieve code and attempt to resolve runtime errors, suggesting tool calls could form longer task chains. The experience prompted him to think further about an agent for ordinary users.

He imagines giving an agent a cloud virtual machine and browser so it can run code and use tools in an isolated environment, then release resources after the task. Isolation still requires appropriate permissions and security design.

The product is not solely a coding tool for programmers: Xiao wants ordinary users to state goals in a familiar chat interface and the system to determine the steps. Some multi-step tasks were possible to try, he says, though models were still unreliable.

He thinks agent interaction should support asynchronous work: after a task is assigned, the system shows a plan and progress, accepts changes, and delivers the result when done. That differs from chat requiring an immediate answer each turn.

He gives an internal-test example in which the system opened a YouTube video, located a frame, and answered a question about animals. He presents it as a surprising observation, not a guarantee for every video task.

Zhang asks what ordinary users might assign it. Xiao mentions information retrieval and data analysis and sees agents as extending chat’s boundaries rather than eliminating existing question-answering needs.

The team chose a separate product because Monica had established user habits, while a new product could fit asynchronous work more freely. Xiao admits he did not yet know how frequently it would be used.

He wants the product aimed at ordinary users rather than only engineers; pricing should match that audience while accounting for the cost of heavy usage. Specific plans remained ideas at the time.

Xiao says the team’s tests at the time found certain models better suited to planning and stepwise problem-solving. He limits the observation to their tests and speculates that agent training may have different requirements from single-turn chat.

He describes the surprise of seeing new capabilities in testing, while stressing the need to manage user expectations. The team planned limited testing and concrete examples to help users understand what the product could do.

In an unstable industry, Xiao recommends keeping one’s own pace, remaining optimistic about technical change, and focusing on problems one wants to pursue. He admits he cannot know the company’s position from an all-seeing perspective.

He compares the Lunar New Year periods of 2023, 2024, and 2025: unexpected technologies appeared each year, requiring the team to reassess opportunities. Current setbacks or advantages should not simply be projected forward linearly, he says.

Looking back at Monica, he thinks that if a team is confident it leads, it might need to expand its team and users faster. This is his retrospective judgment; resources and risk tolerance differ across companies.

He advises founders to think about the “age of the era”: consider not only personal age but also the stage of the technology cycle. He values learning and exploration with a team and acknowledges that business outcomes depend on many factors outside its control.

Why not build a foundation model initially? Xiao says the team lacked resources at the start, so it used existing models to build products. He does not rule out change, arguing that capability, user needs, and partnerships should guide the decision over time.

Asked how far models might advance, he says he does not know. Existing capabilities can be extrapolated to a degree, but previously unimagined dimensions may be what truly change product direction.

He predicts that agents may change some office work while acknowledging that the specific changes have not happened and are hard to predict. He sees physical health and intellectual life as aspects worth reconsidering in the future.

On startup lows, he says setbacks can tempt people to extend the present indefinitely into the future; a new technological dimension may change the situation. Continuing to work and adapt is more realistic to him than pretending to predict everything.

He reiterates that models and applications are not simply opposed. Application teams rely on model research while creating value in user workflows and products; open models may expand choice but cannot eliminate competition.

Xiao says founders should think strategically: other people’s choices respond to your actions, and the environment is not a fixed set of conditions. Instead of assuming a large company must or must not act, consider how participants change one another’s choices.

He sees a founder as a variable that can change the situation rather than someone merely deducing a fixed outcome from current resources. That requires observing the outside world and understanding what one can do.

The interview closes with a discussion of a founder’s life, worldview, and personal habits. Xiao emphasizes continued learning and curiosity before a quick-fire exchange ends the episode.