Li Zhifei on Listing, Model Competition and the Cost of Frontier-Tech Entrepreneurship

Zhang Xiaojun interviews Mobvoi founder Li Zhifei about the company’s listing, commercial pressures in frontier technology and multimodal model developments in 2024.

Li says listing preparations involve extensive disclosure, regulatory communication and intermediary work. A company must explain its business and risks to public investors in a way consistent with reality.

He recalls the fundraising environment and challenges facing AI companies: investors must assess losses, valuations, competition and business models. These are his observations at the time, not investment advice.

He argues frontier technology often leads to project work: when technology is immature and customer needs unclear, suppliers and clients may need extensive customization for each use case.

Looking back on twelve years of entrepreneurship, he felt more than joy at the listing bell: he also had complicated feelings about the relationship between the effort, hardship and outcome.

He argues that a broad vision of the future is insufficient: founders also need a path and timing for reaching goals, making choices under limited resources.

First-mover advantage is no guarantee either. He says a startup needs to use its head start to build capabilities or find a distinct position, while funding and luck can also affect outcomes.

Turning to GPT-4o, he focuses on combining modalities such as text, audio and images and sees unified models as a direction worth exploring.

He hypothesizes that language models can provide a foundation for other modalities and shares his team’s speech-generation experience. His claims about particular systems’ internal architecture are speculation.

He favors autoregressive methods, arguing that prompting and in-context learning can make interaction flexible. This is a research judgment and does not rule out progress with other architectures.

He notes that audio and video involve more information than text, making input length, inference cost and real-time performance engineering concerns. Actual resource use depends on the model and implementation.

He compares end-to-end voice interaction with staged processing, where some audio information can be lost during conversion. Whether a system retains and uses that information correctly still depends on its design and validation.

He sees multimodal capability as important for applications while acknowledging that combining capabilities and keeping them reliable is difficult.

He discusses choices for technical leaders: teams must decide among possible paths while balancing resources, timing and uncertain results.

On robots and agents, he sees perception, adaptation and action in physical environments as major challenges; substantial work remains between a model demonstration and a reliable, usable system.

As model-service prices fell at the time, he argued that companies relying only on API margins needed to reassess their competitive advantage and long-term business.

He suggests founders imagine their position after mature competition: why customers would choose them, whether the business is sustainable, and whether the team has the capabilities to deliver.

He draws on strategy lessons from established industries, distinguishing relatively knowable business questions from technological uncertainty before defining customer value, positioning, goals and a path.

Looking back at his company, he says it once invested across too many devices and use cases at once. A large company may be able to support such exploration, but a resource-constrained startup must be more selective.

He roughly groups model-related companies into frontier-model developers, companies building capabilities on existing models and users of external services, each with different resource needs and competitive approaches.

He recalls an early vision of a cross-device virtual assistant serving users through watches, headphones and other devices. In hindsight, he saw a gap between the ambition and the company’s resources.

He says the company’s difficulties brought significant psychological pressure and forced layoffs and business contraction. He describes the period as seeing the possibility that the startup could die.

He recalls controlling headcount and calculating profit and loss for individual product lines. More detailed cost allocation helped business leaders see resource constraints.

He says maturing generative technology opened new possibilities for tools such as dubbing that had previously been difficult to deliver well, while attracting more competitors.

To pursue his interest in technology without sending the company after every new direction, he chose to research and write about models. Organizational investment, he says, needs stricter judgment.

He worries that fundraising and expansion can create a cycle of losses and further fundraising. He prefers starting with a real need, building a sustainable small business and expanding gradually.