Baidu Ventures' Gao Xue: Nothing Stays the Same — That's the Destiny of AI Investing

By Lili Yu

The OpenAI drama has reached a tentative resolution. Whatever the final outcome, one undeniable fact remains: large language models led by GPT have evolved to a point where humanity can no longer afford to ignore them.

LLMs are opening a Pandora's box of unknowns. The world with GPT is clearly entering a different temporal narrative.

In fact, the first developer conference — dubbed the "Spring Festival Gala for Developers" — was enough to make everyone scream, and many developers shudder.

A boundless behemoth is expanding its territory. The debut of GPTs will likely swallow up those building AI widgets and GPT assistants. The new GPT's support for images, voice, and more will likely impact those working on multimodal models, not to mention those developing AI agents or building knowledge bases.

For China's AI entrepreneurs, the path forward is no longer as traceable as it was during the internet and mobile internet eras. They must rethink: What are the parts that large models cannot erode? What is their unique weapon?

In the face of this reckoning, Sam Gao offers his answer: entrepreneurs should seek out native AI applications that satisfy previously unmet long-tail demands, as well as vertical models that can form data flywheels and possess deep industry data, information, and knowhow.

Sam Gao is CEO and managing partner of Baidu Ventures. A Beihang University materials science PhD who entered the internet industry straight after graduation, after witnessing the full cycle of the mobile internet and Xiaomi's ecosystem buildout, he chose to start anew at Baidu Ventures in 2020.

Over the past three years, Baidu Ventures has consistently focused on artificial intelligence, investing in a large portfolio of AI startups including Agibot, Xihuan Xinchen, Shengshu Technology, Molecule Mind, and JiuShi Intelligence. More than 90% of these investments were made before Series A.

Whether native AI applications can serve as effective countermeasures against large models remains hotly debated among investors. What has achieved broader consensus is that enterprise service companies with deep vertical data, scenarios, and accumulated customer resources will have distinctive advantages. For the latter, though, it also depends on the speed of technology penetration.

In the face of new technology's flood, imagination and resources are undoubtedly solid fortresses. In this tug-of-war between new technology and resources, who will be the winners — the new generation of startups or existing enterprise service companies? How do this generation of AI entrepreneurs differ from the mobile internet era? And what kinds of application companies does Baidu Ventures favor?

On these topics, we spoke with Sam Gao.

He seems to offer something of a template for doing AI investment in this era of constant disruption.

Amid chaotic change, he constantly reminds himself: "Don't preset too early," because "technology is changing every day. The only thing we can do is grasp the trend of technological change to drive our investment strategy, and thereby avoid being swept along by the market."

In fact, the arrival of this AI transformation is itself quite subtle. An internet banquet had already dispersed, when suddenly from some corner came a clamor of voices, followed by wave after wave of smaller climaxes.

Perhaps this is what the era truly wants to show us: you never know what tomorrow will bring.

The following is the full interview:

How Startups Can Avoid Being Swallowed

"Dark Surge": After the OpenAI developer conference, many developers were chilled to the bone. OpenAI's seeming omnipotence could be a fatal blow to many application startups.

Sam Gao: I'd hardly call it fatal. On the contrary, it can help entrepreneurs see the long-term path more clearly. We always believed this industry has only just begun.

ChatGPT didn't appear out of nowhere. It emerged when parameter scale reached a certain threshold, triggering an intelligence explosion. After that, everyone started building around it, but no person or rule could tell you what was right.

The brutal reality is that model capabilities will only keep getting stronger. That means it's basically meaningless for startups to compete on technology — don't try to benchmark against the big players' models.

Also, companies closer to core model capabilities will face greater challenges. There's an American company called Jasper that helps you write documents, but this capability sits very close to the model's native text generation. If your prompts are good enough today, you can replicate its product. Products with prompts as their sole moat abstract single-point business knowledge into tool products without forming a complete feedback loop, which is why they later saw user decline and operational issues.

"Dark Surge": As of now, which segment is most severely impacted?

Sam Gao: From the current stage, middleware software between large models and AI native applications — especially demo-level AI tool platforms for C-end users built entirely on GPT — may face the most direct competition. The emergence of GPTs and the Assistant API hasn't fully covered all capabilities for building upper-layer applications, but the trend is moving in that direction, so these companies face relatively greater immediate impact.

Of course, if middleware tools can deeply root themselves in industries, they will still hold value in the future.

"Dark Surge": Which types of companies might escape the devouring spell of large models?

Sam Gao: Currently, we see two types of companies as having relatively stronger moats. One type can form their own data flywheels when building applications. The other possesses deep industry data, information, and knowhow — such vertical models hold greater uniqueness in our value system.

Why has Midjourney developed so well? Because it adjusts its model based on user feedback data, especially from professional designers and illustrators. As a vertical model, after being iterated through professional feedback, its results keep improving.

So for startups, forming closed-loop data flywheels, collecting user experiences, and iterating on technology and products — all of this is crucial.

"Dark Surge": Which vertical models are more likely to form moats?

Sam Gao: Mastering industry information is key. Take healthcare — ChatGPT needs to learn from many actual cases before it can examine scans and analyze lab reports for individual patients like a doctor. But such data isn't publicly available online; it's accumulated through industry practice.

Or consider legal models. To build vertical judgment document models, there's no public data online either. But if you can collaborate with relevant authorities, you can form your own industry moat.

"Dark Surge": What's the current progress of application startups in China?

Sam Gao: Only when model capabilities are strong enough and licenses are obtained will many people seriously start building. We expect many good product managers will emerge next year with ideas for native applications.

"Dark Surge": How do you understand what constitutes an AI native application?

Sam Gao: An AI native application must be architected on top of large models. Currently, ChatGPT is the world's largest native AI application.

How user demands emerge is fundamentally different from before. In a sense, these are long-tail needs you previously couldn't imagine or that didn't exist — needs that can now be satisfied with a prompt after large model capabilities emerged, with entirely unprecedented usage processes and experiences.

"Dark Surge": For example?

Sam Gao: Like asking a program to write you a poem. In the past, no one developed for such niche demands because traditional algorithms couldn't support them — they only supported mass-market needs, not personalized ones.

Have you seen Guo Degang speaking English? It's not simple translation; it understands tone, intonation, and semantics, then generates output with matching lip movements and Guo Degang's characteristic delivery.

A popular American example is Character.AI, where users set a prompt and it generates a chatbot. But some users don't want conventional chat interactions — they want to chat with Li Bai or Confucius. There are even more personalized demands here, many of what we call long-tail needs.

"Dark Surge": During the mobile internet era, various applications evolved through a process. Will something similar happen with this wave of native applications?

Sam Gao: When people are accepting something new, especially when an industry holds vast unknowns, don't preset too much.

When entrepreneurs understand model capabilities and have unique ideas, they can all solve previously unsolvable problems.

I experienced the full mobile internet era. Today is quite different. People say the mobile internet era had gold everywhere. Early on, you just needed to solve penetration — moving various industries online, hailing cabs, getting around, reading books aloud, etc. The second half solved growth.

Now the bar is higher for entrepreneurs because it's more about competing on creativity. You need strong imagination to discover unmet niche demands, what we've been calling "long-tail demands." At the same time, you need to understand model capabilities and know how data is generated.

But the upside is that in the past, startups had to invest heavily in algorithms. Now algorithm investments are somewhat lighter because you can use mature models from big players. The key remains whether the team has the ability to create and implement AI native applications — to think about how to apply this model to their desired scenarios.

"Dark Surge": Will next year be a big year for AI?

Sam Gao: Large models are currently just language models. To apply them across industries, multimodal capabilities still need time. Now that everyone can access models, once various multimodal capabilities are added, there will be some new applications next year. Of course, starting a company now takes considerable courage.

"Dark Surge": Do you think the capital environment will improve?

Sam Gao: Sharp question. We always believe markets are cyclical and industries spiral upward, so it will definitely improve.

Technology-Driven Investment Strategy Is the Only Way to Avoid Being Swept Along by the Market

"Dark Surge": For a long time after ChatGPT 3.5's debut, AI investment in the primary market was quite bleak. Did Baidu Ventures experience fluctuations?

Sam Gao: Frankly, no. Well before ChatGPT became hot, we were already looking at AIGC and its path to implementation, making a series of investments in text-to-image, text-to-video, and text-to-multimodal companies.

"Dark Surge": We didn't invest in some of the so-called representative companies in large models?

Sam Gao: We ran many simulations of what large models would become. Our conclusion was that a closed-source ecosystem and an open-source ecosystem would emerge, like iOS and Android. They would roughly split the market. In this context, we wouldn't follow the hype or make bets.

What we seek are companies with underlying technology, and those that deeply immerse themselves in vertical applications across law, healthcare, finance, and other domains. Technology will eventually be commoditized, but industry knowhow is what large models struggle to penetrate. That's what we truly want to do.

"Dark Surge": Though we did invest in some large model companies like Xihuan Xinchen.

Sam Gao: Right, Xihuan Xinchen was invested before ChatGPT became hot. Its founder Zhenzhong Lan previously worked at Google on many lightweight large models. Lightweight large models solve specific problems — for example, chatting with depression patients. We believe not all problems will require large models in the future; in certain scenarios, lightweight large models can solve them better.

Our investment in Xihuan Xinchen, plus several multimodal models like Shengshu Technology and Morph, was all based on valuing capabilities beyond text models.

"Dark Surge": How do you view the evolution direction of the major large model companies today?

Sam Gao: After the model catch-up phase, they need to find their own vertical industry applications and go deep in one domain. This is like the previous generation of AI companies — everyone talked about the "AI Four Little Dragons," but each was actually different: some did transportation, some finance, some mobile customers. Technical capabilities ultimately must land somewhere.

The previous generation of AI technology had poor generalization capabilities, while large models are generalizable. They may solve the problems of poor generalization, high costs, and lack of universality that plagued AI in many scenarios back then.

"Dark Surge": How would you summarize your three years at Baidu Ventures?

Sam Gao: We're an AI-themed fund. We're actually quite pragmatic. Our core belief is that technology can change industries — either reducing costs, improving efficiency, or creating new opportunities.

We don't just have AIGC. We've made investments across industry, agriculture, energy, transportation, healthcare, and more. From how machine tools polish, to how autonomous vehicles operate in mines, to how drug manufacturing is solved — these are all things we follow.

"Dark Surge": Business is also a world of signals. What important signals have you captured in recent years?

Sam Gao: Looking back, we made good investments in the second half of autonomous driving. The earliest stage of autonomous driving was primarily about technological breakthroughs. As technology gradually matured, we found that general-purpose autonomous driving wasn't so applicable when deployed in scenarios — it needed to do specific things for specific scenarios. So we thought through each vertical industry: what technology, team, and deployment scenario does this actually need?

In this process we basically covered all application scenarios — for mining, we invested in Main Function Intelligence; for logistics, JiuShi Intelligence; for sanitation, Yunchuang Zhixing; and so on.

"Dark Surge": Have large models brought new variables to autonomous driving?

Sam Gao: Yes, everyone is working on end-to-end autonomous driving. For example, Tesla demoed its end-to-end AI autonomous driving system FSD Beta V12 in August this year, relying entirely on onboard cameras and neural networks to recognize road and traffic conditions and make corresponding decisions. We think the results are quite good.

End-to-end autonomous driving takes raw sensor data as input — camera and sensor data — skipping intermediate modular processing to directly output steering wheel angles and throttle/brake controls. If this can be achieved, true autonomous driving becomes possible at very low cost, with just a few cameras.

Of course, this will still take time to develop. First, it requires data training. Second, data collection costs are currently high — big companies spent considerable time collecting and training this data. We've also invested in newer startups trying to do end-to-end autonomous driving.

"Dark Surge": Beyond autonomous driving, large models have also brought variables to robotics. How far do you think embodied intelligence is from explosion?

Sam Gao: Embodied intelligence research and applications focus mainly on physical agents like robots and autonomous vehicles, and its applications are continuously expanding and deepening. Influenced by large language models, embodied AI will bring a major breakthrough — from simple capabilities like image recognition, to learning how to execute complex human-like tasks through multiple steps.

Simply put, embodied intelligence is the vehicle for AI to interact with the physical world. The first step is helping robots reach everywhere humans can reach and complete all tasks humans can complete. The second step is helping robots reach places humans cannot reach and complete tasks humans cannot complete. Currently, only a handful of companies worldwide have the capability to achieve this technology.

Additionally, robot intelligence brains won't generalize at first either — they'll start by helping solve specific problems in certain operating environments, like tightening screws in auto factories. Only after collecting sufficient data can these two parts be optimized, so it's still a very long process.

"Dark Surge": What dimensions do you consider when investing in humanoid robot companies?

Sam Gao: We invested in Agibot because Zhihui Jun has strong hands-on ability and problem-solving skills, and we also value industry charisma. This industry needs to gather many hands-on capable, quick-thinking geeks and enthusiasts, so founders need charisma. Also, fundraising ability must be strong — after all, embodied intelligence implementation may be a very long road, and these are all strengths of the Agibot team.

In robotics, if a founder has strong hands-on ability and strong electromechanical integration capability, they can build products. Like Unitree in the previous era of robot dogs — they could attract similar enthusiasts in their circle to build products together. But in this wave of large model entrepreneurship, talent is extremely critical and scarce, especially in embodied intelligence. Beyond Agibot, we've also invested in several "spark" companies. To achieve embodied intelligence, you still need to continuously collect various data through scenarios.

"Dark Surge": When entrepreneurial teams come out to start companies now, what advice would you give on fundraising?

Sam Gao: Be pragmatic. Today's startups — without clear cash flow planning and fundraising rhythm — many will die. When the economy slows, founders need to know even more how to survive.

We have a complete post-investment team that regularly follows our portfolio companies' operating conditions and communicates on whether things are healthy. On fundraising, many people harbor fantasies and inaccurately estimate cash flow. We constantly communicate with founders about these matters.

---

Industries and Companies Always Have Ups and Downs

"Dark Surge": What attracted you to join Baidu Ventures in 2020?

Sam Gao: Previously I did investment in Xiaomi's industrial ecosystem. Later Xiaomi went public, and the brother companies I invested in also went public one after another. I was 40 then, so I was thinking whether to take another step and try something different.

"Dark Surge": When you arrived at Baidu Ventures, what was the first problem you wanted to solve?

Sam Gao: First, building the team. It was a period of extreme VC prosperity — projects and capital were abundant, but few people professionally understood AI.

My thinking then was: since we're a technology fund, we still need people who understand technology. So we found many people from within the industry.

"Dark Surge": When building a new team, beyond selecting people, what unique mechanisms did you implement?

Sam Gao: Our investment committee is open and transparent. Whether you're in software, hardware, ToB, or ToC, everyone can discuss at the same level — how much was invested in which project, what the thinking was at the time, and so on.

Today's new technology is the intersection of many disciplines. Understanding it requires stimulation from different people. If the company atmosphere is technology-oriented, everyone can collide on where the opportunities lie. People investing in applications need to know how far models have developed, and also how chips constrain technology development.

"Dark Surge": Among the entrepreneurs you've invested in, who impressed you most?

Sam Gao: Changjing of Roborock. He's very simple — even now, he maintains a very pure, focused state when discussing products. If you talk products with him, you can talk for hours.

Another is He Qing, founder of domestic EDA company Xingxin. When he first returned to China to start his company, we visited him frequently for a period and found he'd set up a camp bed in his office. This camp bed eventually collapsed from him sleeping on it. At critical moments, he went three months without leaving his office. Like us, they are firm believers in technology — believing technology can change life and push society in a positive direction.

"Dark Surge": In the zone where Baidu Ventures can freely explore, what companies have you invested in?

Sam Gao: We invested in Moting Medical, a company focused on ophthalmic diagnosis, ophthalmic surgery, and refractive testing equipment. Through independent R&D, Moting achieved domestic substitution of OCT, biometers, and other ophthalmic equipment, breaking import monopolies and giving Chinese medical institutions more choices. Meanwhile, it continuously tries to apply AI technology to ophthalmic medical equipment. I deeply admire such founders — they all have extreme pursuit of certain technologies.

"Dark Surge": Over these years, what's the most important thing you've learned about investing?

Sam Gao: Industries and companies always have ups and downs. Often, a company's essence hasn't changed, but there will always be people saying what's cold and what's hot — this is the greatest noise. Every industry has its own cycle. I always believe technology moves forward.

"Dark Surge": Doing AI investment, have you summarized its essence, or how do you remind yourself?

Sam Gao: Investment always needs its own logic, but the key thing about AI investment is that technology isn't 100% settled. Today's projects could be replaced by technological iteration tomorrow.

So to do AI investment well, you must constantly ask yourself: when will this be replaced? You need psychological preparation — nothing is unchanging.

Image Source | IC photo

Layout | Yunxiao Guo