Li Di on Conversational AI Products, Costs and Business Choices

In 2023, Zhang Xiaojun interviewed Xiaoice CEO Li Di about conversational AI products, model costs and business models. His views reflect the interview’s time, not necessarily later industry conditions.

Li recalls that when he joined Microsoft, the team hoped to develop a product originating in China and eventually serving global markets, rather than only supporting research for other markets.

He previously worked on mobile-phone product planning and found that the technical features engineers emphasized did not always match what customers cared about.

He also recalls working in a startup and the audiovisual industry, where shifts in business direction, funding conditions and partners changed the team’s options.

Discussing Microsoft’s search business, he describes the difficulty of operating across different markets and regulatory requirements and the team’s desire to preserve an information service across markets.

Xiaoice began as an exploration of another form of interaction: beyond completing tasks, the team wanted to study conversation that could continue over time and create an emotional connection.

He argues that product value can come not only from completing immediate tasks but also from longer relationships, illustrating his view with toys and customer service.

On business models, he offers a hypothetical article-generation service: even when technology creates value, competitive pricing may leave its supplier with only a small share.

His point is to calculate actual revenue and delivery and operating costs rather than infer earnings from potential market size alone.

On model size, he argues for choosing according to task performance, latency and cost. He cites deployment examples on hardware available then; those figures should not be treated as universal benchmarks today.

He questions whether one largest model is always the most economical choice for every task, using systems integration and maintenance as an analogy for components of different sizes serving different purposes.

Li says technical progress matters, but enthusiasm should not replace concrete judgments about compute, pricing, deployment and sustainable operations.

He discusses market perceptions of model size and capability, arguing that companies should define needs before deciding which model mix fits a product.

Regarding model leaderboards, he worries that a single ranking can push teams toward similar goals while obscuring differences in data, methods and practical use. This is his view of competition at the time.

He proposes a competitive framework: challengers may use costly new capabilities to attract customers, while incumbents must weigh the effects of new technology on existing costs and margins.

He emphasizes training-data quality and combinations of models, while acknowledging that not everyone agreed with those judgments at the time. Different tasks still need practical evaluation.

On corporate research papers, he discusses tradeoffs among publishing methods, research incentives and engineering deployment from a commercial-team perspective; this does not negate the other value of open research.

Asked about comparisons with OpenAI, he says the companies differ in resources and objectives. He focuses on Xiaoice’s survival, products and business rather than direct comparison of scale.

He recalls limiting the scope of Xiaoice’s proactive replies after it generated heavy interaction on Weibo. In his view, a product should consider its effects on the platform it relies on and on other users.

He says the early virtual persona was designed partly to make the technology’s limitations easier for users to accept. This is his recollection of a team choice, not a universal conclusion about gender or preferences.

He argues that one fixed persona cannot be everyone’s best friend: users differ in their preferences and expectations of a relationship.

On conversational experience, he says both the person and the system shape an exchange. His team observed that some users changed the subject or rationalized poor answers to keep a conversation going.

Discussing the CEO role, he describes keeping the company operating as a long-term aim and reflects on uncertainty and pressure in decision-making.

Finally, they discuss open questions about consciousness. Li raises philosophical questions of his own; these speculations do not establish that AI systems possess consciousness or subjective experience.