Wang Xiaochuan on Model Technology, Applications and Starting Again

In 2024, Zhang Xiaojun interviewed Baichuan AI founder Wang Xiaochuan about his technical choices after starting another company and how he hoped to combine technological ambition with commercial applications.

Wang responds to an earlier debate over AI startup paths. He says technological ambition and commercial reality each illuminate part of the problem and companies must consider both.

He summarizes his position as moving a step more slowly on ambition but three steps faster on application: follow frontier models while seeking settings where technology creates practical value. This was his strategic framing at the time.

He proposes matching technology to problems: understand both what a tool can do and what an industry actually needs, rather than seeking uses for a new technology in isolation.

He recalls the thinking before Baichuan was founded, connecting his long-standing interest in life sciences and public health with new language-model capabilities.

He hopes future technology can help study complex questions about life. His description of modeling life is a research vision, not evidence that existing systems can reliably understand or treat disease.

He believes language capabilities may help systems process literature, communicate and integrate information, but model outputs still require expert validation and cannot replace medical judgment.

He describes product directions the team considered, including improved conversation, search for external information and further work in fields such as health care.

In his view, a good application should be useful with technology available then and continue to gain value as it improves. Health care is one area he considers worth long-term exploration.

On entertainment, he distinguishes written narrative, film and interactive games, envisioning a combination of story generation, multimodal content and user interaction.

Responding to investor Zhu Xiaohu’s criticism, he says he understands different investment strategies and believes participants see different parts of the market.

He recalls that conversations abroad after a model release helped him articulate the technology and the company’s position more clearly.

Reflecting on Sogou, he says experience with input methods, search and natural language technology was valuable; he also recalls the difficulty of building a team and products against established competitors.

He discusses whether to pursue the capabilities of larger models and stresses balancing long-term technical goals, resources and practical demand when choosing applications.

He says prior collaboration helped Baichuan’s early team get started quickly, while it also needed new researchers with skills suited to later stages.

Considering business customers and individual users, he discusses their different needs and how the company chooses applications it can continue to support.

On health care, he argues that companies need sustained investment and support from leadership. His comparisons among companies are personal judgments, not evidence of medical effectiveness.

He says Sogou’s experience with products, distribution and commercialization helps him think about a model company, although developing a new generation of models requires different technical talent.

He also recalls constraints in search competition, discussing how to find opportunities against a market leader and regret over missed directions.

He cautions against both underestimating technology and treating it as capable of anything; in his view, practice is needed to determine its abilities and limits.

Looking ahead from 2024, he discusses model capabilities and application bottlenecks. His timelines are expectations from the interview, not accounts of what later happened.

He says colleagues proposed investing in video generation, but he chose to focus on other priorities, emphasizing the need for clear tradeoffs when resources are limited.

He recalls rethinking the relationship between products and models over the holiday, shifting from simply following existing models toward a closer integration of his own applications and model capabilities.