Li Guangmi on Perplexity, AI Search, and Model Applications in 2024

In June 2024, Zhang Xiaojun and Li Guangmi review large models and AI applications in the first half of the year. Guangmi says independent AI applications have not broadly taken off beyond model companies’ own products, and focuses on Perplexity and search.

He reviews Perplexity’s team and financing, saying its growth attracted Silicon Valley researchers and investors. Valuation, user, and revenue figures in the episode are descriptions or forecasts from that time, not current data.

Guangmi argues that information retrieval fits model capabilities at the time: a system can search, combine information, and answer complex questions. He sees exploratory research and scientific discovery as harder tasks.

On whether to start with product or model, he says there is no standard answer. An application can first validate demand with an external model, then optimize cost and quality, potentially using smaller models for specific tasks.

Zhang asks how an independent product can build a moat if model providers also improve search. Guangmi acknowledges the risk and suggests deeper model optimization or partnerships; any proposed future deal is hypothetical.

On Perplexity’s valuation, he starts with growth, retention, and the search market, while acknowledging competition from Google, model companies, and other information gateways. These assessments reflect his personal view in mid-2024.

He says AI search can improve information access by answering directly and organizing sources, but products still need users to return and must address sources, cost, and accuracy.

He summarizes Perplexity’s advantages at the time as a fit between product and model capability and some established user habits; frequency, monetization, and differentiated competition with giants still need exploration.

On Chinese AI search, Guangmi mentions Metaso and says differences in content-ecosystem openness affect retrieval. Competition involves not just models but reliable access to and organization of information.

The host asks about Perplexity’s moat and whether it might remain independent or be acquired. Guangmi says its relationship with models, user habits, and competitors’ actions could all affect the outcome.

He distinguishes foundation-model companies from applications: training models needs a different resource structure, while applications must demonstrate demand, distribution, and retention. Valuation differences alone do not establish which is more valuable.

He compares overseas model-company financing and cloud relationships, saying funding, compute, and partners all affect competition. His views on teams and financing prospects reflect mid-2024 observations.

Guangmi believes AI-native applications had not taken off systematically in the first half of 2024, possibly because of model capability, inference cost, and latency. He emphasizes this is his assessment of the ecosystem at that time.

Looking ahead, he watches smaller models, on-device computing, and multimodality. Running models on devices could change costs and privacy experiences, though actual effects depend on hardware, models, and implementation.

Making models smaller is not just reducing parameters: training data, retained capabilities, latency, and device memory also matter. Guangmi says these engineering tradeoffs affect on-device usability.

He thinks reasoning and multimodality may be priorities for the next stage and discusses preserving useful capability when compressing models. These directions and timelines are forward-looking research judgments.

They ask listeners to consider which currently impractical applications might work if model-call costs fell sharply. Guangmi connects price, speed, and user experience without assuming that lower prices automatically guarantee success.

On next-generation model release timing, he says training, infrastructure, and productization need coordination. His comments about internal progress and release dates are outside speculation, not confirmed plans.

He compares model research and infrastructure to a train and the track being laid ahead of it: researchers push capability forward while engineers keep compute, data, and training systems aligned. The analogy expresses his view of the coordination challenge.

He cautions that a capability leap as visible as GPT-4 may not occur every year. As expectations rise, incremental improvements can look less dramatic while still changing application costs and experience.

On business models, he compares subscriptions with established advertising platforms and says model investment, returns, and defensibility still need testing. His revenue estimates and market comparisons reflect assumptions from the interview.

Beyond scaling, he focuses on data and compute efficiency: can useful capability be obtained with fewer examples and less compute? He treats this as a research question rather than a settled answer.

Guangmi sees human and AI abilities as potentially complementary: models can process large amounts of information, while people can set goals and judge results. His description of future work is a projection, and actual capability varies by task.

On robotics, he says translating planning into reliable physical control remains difficult and useful applications must be discovered. He predicts autonomous driving may deploy sooner than general household robots, though timing is highly uncertain.

They discuss the positions of OpenAI, Apple, Google, NVIDIA, and other technology companies. Guangmi’s assessments of staffing, model capabilities, and strategy are outside observations from 2024.

He says Apple needs to see whether AI features change device experience and upgrade decisions; for Google, he notes both model competition and existing distribution advantages. These are not guarantees of future company performance.

He says relationships among chips, cloud platforms, and enterprise customers will affect how AI industry returns are distributed, while acknowledging that some cloud-revenue assumptions lack sufficient evidence.

He says many startups compete with large companies while depending on their compute, cloud, and distribution. For investors, assessing research teams and long-term resource support becomes harder.

On retrieval-augmented generation, he sees enterprise information retrieval as a real need, but says effective implementation requires coordinating data, retrieval, ranking, and answers. Organizations must also address access permissions.

Comparing innovation ecosystems in China and the US, Guangmi says long-term capital and tolerance for failure support early exploration. He hopes for more technically distinctive startups, a generalization rather than an assessment of every company.

Looking to the second half of 2024, he focuses on applications that might work as costs fall, on-device experiences such as phones, and longer-term general robotics. The episode closes with questions that remained open at the time.