Qi Fanchao on Moving from AI Models to Information Products

Zhang Xiaojun interviews Shenyan Technology founder Qi Fanchao about how an AI startup that raised funding chooses a direction, builds products and organizes its team.

Qi recalls studying electronics and computer science at Tsinghua, moving from computer vision to natural language processing and participating in large-model research during his doctorate.

He began exploring entrepreneurship during his doctoral studies, motivated both by technological change and by the prospect of turning research into products more people could use.

He says products can use existing models, but particular user needs may also require additional technology or targeted engineering.

He recalls meeting investors before ChatGPT appeared. Investors understood the technology and its applications differently, requiring founders to keep explaining the problem they wanted to solve.

After ChatGPT’s release, funding conditions and market attention changed. He describes both opportunities for startups and rapidly changing competition.

He recalls that by the second half of 2023 it was becoming harder to differentiate solely through model training, prompting the company to rethink strategy and product direction.

He does not equate a model with a complete product, arguing that companies must combine technology, interaction and real user needs. His view of the industry’s future is an opinion.

He describes using multiple models together to manage uncertainty in individual outputs. Combining models does not automatically guarantee accuracy; evaluation and verification remain necessary.

His product idea came from information overload: as his research field expanded, he found more papers to read while his time for filtering and understanding them remained limited.

He explores news reading and information assistance, focusing on how users find relevant material, understand key points and follow questions they care about.

He discusses changes in content creation, consumption and distribution, arguing that digital tools keep changing the volume of information and how people access it.

In his framework, content carries information; analyzing smaller units of information might help a system filter material for users more effectively.

He describes a product concept that breaks articles, video and audio into information points, identifies overlap and unique details, then reorganizes them around user interests. Accuracy, sources and context still require checking.

He says users cannot always formulate a clear question easily, so an information product should reduce interaction costs rather than merely provide an input box.

He observes a shift toward consumer needs in the content chain and considers how automation might change creators’ and readers’ work.

Introducing the recently launched Yu Jing product, he describes reading features including hierarchical outlines for papers and long articles. The feature description reflects the product at the time.

In a smart outline, users can expand an article’s structure level by level, moving from its overall theme to sections and paragraphs before choosing where to read deeply.

Discussing products such as Cursor, NotebookLM and Perplexity, they consider how model capabilities can fit concrete workflows and provide users with perceptible value.

He compares entrepreneurship to sailing without a complete chart: a team needs a direction but must keep adjusting to feedback and changing conditions.

He compares structured reading to recovering hierarchy from linear text and imagines eventually connecting related ideas across different pieces of content.

He recalls introducing excessive switching costs in pursuit of organizational neatness and later paying more attention to whether process benefits justified the burden on the team.

Discussing management as a young CEO, he reflects on information sharing, communication and delegation, saying managers must learn amid change rather than rely on their title.

He says uncertainty in model outputs affects product development and collaboration: teams need experiments, evaluation and feedback to determine whether changes actually improve the experience.

As user feedback grows, he wants to turn scattered requests into actionable priorities and assess whether the team can deliver effectively.

He believes technology evolves quickly, so product teams and managers need to understand the fundamentals and their changes while continuing to listen to users.