Wang Mengqiu on ChatGPT, Large Models, and Investment Opportunities in 2023

Zhang Xiaojun discusses ChatGPT, large models, and investment with Wang Mengqiu, founding partner at ClearVue Partners and former Baidu technology vice president. The video labels the episode April 8, 2023; its judgments belong in that technological and market context.

Wang says her team had begun studying energy, materials, and semiconductors; after seeing images from Midjourney and Stable Diffusion, she turned to AI. She values whether ordinary users can directly perceive a technology’s usefulness.

She likens ChatGPT’s conversational product to an important user-experience moment: underlying model capability needs an appropriate interface before more people will try, understand, and reuse it. Her account of internal release discussions is hearsay.

She uses the iPhone to illustrate product definition: combining existing components into an accessible experience can open new application space. In Wang’s view, both model progress and the product vehicle matter.

In her view, demonstrated capabilities offer research and startup directions, but the routes to better results remain unclear. She expects AI to be a long-term topic while continuing to study the underlying technology.

She jokes that a day without AI news brings relief: the pace of technology and application updates creates constant pressure to learn. She says the feeling is stronger than during the early internet and phone-hardware cycles.

Why does change appear so fast? She offers two possibilities: long-accumulated technology becoming visible to users at once, and existing software rapidly incorporating AI. She acknowledges that she does not fully understand every cause.

Image, video, and 3D tools are also evolving. She cannot yet determine how general multimodal models and independent generation tools will ultimately compete, and says their data, users, and product advantages need observation.

Comparing the previous AI wave, she says research directions, compute, funding, and long-term investment all influence development. She does not present herself as a model researcher; her technical explanations are an investor’s understanding.

She uses DeepMind’s history to discuss commercialization of fundamental research: exploring general capabilities needs long-term resources, while customer-facing products must answer more immediate payment questions. Different research teams face different constraints.

Turning to private markets in 2023, she says investors are watching energy, materials, consumer businesses, and AI, but studying projects is different from investing immediately. Market excitement and the number of suitable investments can diverge.

She observes some funds looking across models, applications, and infrastructure while others wait. Because capabilities and competition are changing quickly, she wants to understand the technical foundations before deciding.

Zhang asks when AI will have its “iPhone 4 moment.” Wang says timing is hard to predict: a major application needs not only users but also a sustainable business model.

She thinks generative AI and search can complement one another: users sometimes seek verifiable information and sometimes want ideas or generated content. She also warns that chat models can produce confident-sounding errors.

One possibility is that AI appears inside existing large applications and workflows rather than as a single independent entry point. That could still have a large impact, she says, though she is unsure whether a new consumer killer app will emerge.

On people and AI, she treats models as tools and asks how education can develop question formation, judgment, and original ideas. Her assessment of job replacement is a prediction from that time.

On enterprise software, she says slow adoption is not solely a software problem: customer data, processes, and willingness to pay matter too. She calls uneven customer readiness an “uneven field.”

She also mentions robotics research interacting with physical environments and sees a gap between laboratory progress and products ordinary users can experience. Real applications must address hardware, compute, and deployment conditions together.

Comparing investment in China and the US, she admits she had not spoken directly with every US investor and cannot judge how cautious each decision was. She observes active study by Chinese funds but comparatively restrained investing at the time.

ClearVue had not invested in general foundation models at the time. Wang says training frontier models requires substantial capital and talent; a small early-stage fund may find it hard to pick a winner or match the financing need.

She says ruling out some investments is easier than identifying what to fund. For a single feature easily covered by a general model, she worries that product differentiation may be hard to sustain.

Vertical domains may differ: specialized knowledge, lawfully obtained data, and real workflows can create distinct requirements. She names materials and medicine as examples for exploration, not as permission to use sensitive data.

Foundation-model startups also face requirements for talent density, compute, engineering, and a potentially long commercialization cycle. Her funding and cost figures are industry estimates from the time, not universal thresholds for every team.

She calls her stance cautiously positive: excited by progress and willing to study it, but not rushing to invest just because the market is hot. Different funds can move at different speeds.

Comparing the history of search, she identifies user-visible technical value and a business model as separate milestones. Search shows that they can be years apart, although AI’s timing cannot simply be copied from that history.

The mobile operating-system example leads her to consider relationships between foundational platforms and applications. Platform stability, developer access, and user accessibility all shape an application ecosystem.

For robotics and autonomous driving, she expects deployment may begin in settings with defined tasks; hardware, sensors, data, and on-device compute are challenges. This is her judgment in 2023.

On major Chinese companies, she sees potential advantages in existing technical experience, funding, and user workflows, but teams and strategies differ. Her knowledge of some companies is limited, so she avoids a definitive conclusion.

She considers foundation-model development better suited to large institutions, but that does not mean startups have no opportunity. Competition also depends on research, execution, resources, and product choices.

Discussing generative models, she stresses that generated content is not necessarily reliable and important facts need checking. Her analogy about randomness and search is a personal simplification, not a strict definition of either system.

She observes both excitement and anxiety about AI: some expect new productivity, while others worry about their jobs. The feelings of early adopters do not represent everyone.

She again notes that frontier models can require substantial capital, whereas application-layer economics differ. Existing companies with customers and distribution can also adopt new capabilities, so startups need specific advantages.

On model “emergence,” Wang uses learning and feedback as analogies and acknowledges substantial uncertainty about both models and brains. Her analogy is an explanatory perspective, not a scientific claim that they are equivalent.

She thinks prompt engineering may be a transitional response to technology that is not yet easy to use, comparing it with learning search-query techniques in the early web. Whether that happens depends on future technology and user needs.

She sums up her investment stance as cautiously optimistic: the technology matters, but at the time she had not found a particularly suitable project for the fund. Her understanding will continue to change with new evidence, she says.

She recommends reading AI interpretations from product and human perspectives while warning that commentary is not necessarily accurate on every technical detail. She is still watching whether AI creates an independent new application or mainly reshapes existing workflows.