Yunqi's Chen Yu: The "80/20 Rule" in the LLM Era | Tencent Tech × Tidal Wave AGI+
Getting a model to 95% accuracy is definitely a moat.
Source | Tencent Tech. Reporter: Xiaoyan Zhou.
➤➤➤ The AGI+ wave is already here — ChatGPT became a "mass-market app" in just six months. MiniMax, an early angel-stage project we invested in 18 months ago, has rapidly deployed its self-developed multimodal large model across multiple industries, tangibly improving productivity and quality of life. At the AGI+ infrastructure and application layers, several Yunqi Capital portfolio companies are also actively embracing new technologies (click to see the full Yunqi AGI+ portfolio), exploring new possibilities.
In this edition of Riding the AGI+ Wave, we've partnered with Tencent Tech's "AI Future Compass" series to bring you the latest thinking and insights on large models and their applications — covering how new technologies are reshaping industry landscapes, differences between domestic and overseas development, and future regulatory measures.
The key takeaways of this piece include:
1
In the large model era, a "Pareto principle" will likely apply to future applications: roughly 80% will be "upgraded versions" of existing apps integrated with large models, while 20% will be entirely new applications built on the novel capabilities of large models.
2
The overall structure of most industries won't fundamentally change because of large models. Even if large models shift application entry points, this won't substantially alter the underlying business landscape — it's ultimately just an experience upgrade. Opportunities for AI-native applications may lie in: cross-border e-commerce, AI for Developers, private deployment, and AI security.
3
Many believe there are limited entrepreneurial opportunities at the model layer, with more potential at the data layer. But the market isn't actually short on data — public data remains underutilized. ChatGPT's training data only goes up to September 2021, so considerations around proprietary data can wait; focus on making good use of public data first. Moreover, data uniqueness matters more than data volume — companies that truly possess distinctive proprietary data and can convert it into commercial value are rare to nonexistent.
4
A model scoring 80 points may not constitute a moat, but one reaching 95 points certainly does. Despite numerous domestic companies building models, none has yet comprehensively surpassed GPT-4, because its core moat lies in engineering optimization — the "secret sauce" that OpenAI hasn't disclosed.
5
Differences in large model investment strategies between China and abroad stem from exit pathways. Overseas, investors are willing to back projects with single-point技术创新 because such projects can exit through M&A. In China, where M&A cases are scarce, investment institutions prefer to bet on big ideas that can eventually go public.
Full interview below:
1 What new application opportunities will large models create?
① Tencent Tech: ChatGPT has made the large model赛道 quite buzzy for entrepreneurship and investment. What projects has Yunqi Capital focused on, and what were the key considerations behind those investment decisions?
Chen Yu: ChatGPT entered public consciousness six months ago; we invested in MiniMax, a company focused on multimodal large model R&D, about 18 months prior. They have three foundational modalities: text to visual, text to audio, and text to text. Their main product based on large models is the intelligent conversational robot Glow, which currently has five million users. They've also opened API interfaces and partnered with numerous enterprises including WPS.
Our decision logic was primarily betting on the team and the general technology direction. As early as January 2021, we began engaging with MiniMax's founding team — at that point, they hadn't formally decided to start a company. GPT-2 and GPT-3 had already emerged, though neither was as polished as ChatGPT would later become. By late 2021, the team felt the timing was right, MiniMax was formally established, and we led their angel round.
MiniMax's starting point was to build products along the direction of artificial general intelligence, with multimodal large models at the foundation. This aligned with our own thinking — we had been contemplating general-purpose large models ourselves. To put it concretely, AGI is equivalent to replicating a digital human brain. Humans have free thought and can infer new knowledge through learning; AGI should be able to think like humans, complete multi-task operations, and provide answers even when encountering things it has never seen before.
We've already invested in MiniMax and spoken with most model startups on the market. After this wave heated up, we don't currently need to invest in a second foundational model company at the infrastructure layer. Our investment focus has shifted to the application layer, specifically projects that can leverage the characteristics of large models to build applications. We'll also keep an eye on what new technologies might emerge beyond OpenAI, hoping to invest in things with more genuine innovation.
② Tencent Tech: There's a saying that all industries are worth rebuilding with large models, and many sectors may use them for upgrades — could businesses like Meituan and Taobao be completely redone with large models? Do startups have opportunities to help enable these transformations at large companies?
Chen Yu: In the large model era, a "Pareto principle" will likely apply to future applications: roughly 80% will be upgraded versions of existing apps integrated with large models, while 20% will be new applications built on the novel capabilities of large models.
Whether large models are suitable for all industries remains undetermined. The first challenge is that China doesn't yet have a mature Chinese-language large model available — optimizing just this step will take many years. We also can't yet determine which domestic model offers a superior user experience, because the feeling of use is highly subjective, and each model has its own strengths. In the end, it comes down to comprehensive user experience.
If Chinese-language large models become mature, this wave of industrial upgrading will certainly begin with the internet industry, then gradually transition to traditional industries — a process that may take 5-8 years.
Large models may become the entry point for all applications in the future. For example, right now to buy something you open the Meituan app or mini-program; in the future you might simply tell a large model "help me buy this thing" and complete the entire purchase operation. Entry points like Meituan's app or mini-program would become secondary rather than primary entry points. As human-computer interaction methods evolve, these kinds of changes are quite likely.
But similar changes have been attempted before — Apple Siri, various smart speakers — and many attempts became punchlines, probably because the language models weren't good enough, affecting user experience. If it takes the same amount of time but delivers inferior experience, users won't stick with it.
As for startup opportunities in this industrial upgrading, I believe the overall structure of most industries won't fundamentally change because of large models. Take food delivery: there were only Meituan and Ele.me to begin with; startups weren't really a factor. Even if large models change application entry points, this won't substantially alter the underlying business landscape — it's just an experience upgrade.
③ Tencent Tech: If 80% of existing applications may upgrade by integrating large models, then the 20% of new applications you mentioned represent incremental market opportunities. In which scenarios do you think these AI-native applications will emerge first? Looking at Yunqi Capital's portfolio, which investments fall into this 20% incremental category?
Chen Yu: Let me address the second question first. There's a common misconception that you invest today and see results tomorrow — that's impossible. For a startup to actually build something and show results takes at least 18-24 months. Something producible in a week has no moat or imagination; we focus on more substantive things.
From our investment cases, because we generally invest at very early stages, the projects themselves are still in early development, making it difficult to judge whether these 20% native applications have already materialized. Even with MiniMax, which has shown clear potential, looking back at our investment process, we actually observed them for nearly a year before investing.
Returning to predictions about AI-native application opportunities, I think the process of discovering and挖掘 these opportunities resembles the淘金历程 of Western cowboys — it requires a spirit of exploration. The opportunities may lie in these areas:
First, cross-border e-commerce, because it involves multilingual communication;
Second, AI for Developers, helping programmers write code or check for errors and security vulnerabilities;
Third, private deployment, as enterprise clients will have needs around personalized models and data security;
Fourth, AI security, as widespread large model deployment will bring a series of security issues, such as the spread of false information from model-generated content, and prompt injection attacks targeting models that could lead to internal enterprise information leaks.
2 How can startups build core moats?
④ Tencent Tech: Among the opportunities created by large models, is the ability to efficiently utilize limited data — such as data annotation, data refinement — as well as using proprietary data for vertical models, a potential growth point for startups? Where do the opportunities lie for investors?
Chen Yu: A document was leaked from Google internally some time ago, in which a Google researcher argued that OpenAI had no moat. It mentioned that you could fine-tune an open-source model to get an 80-point model, but there's still a large gap between that and a 95-point model like GPT-4.
While entrepreneurs can drastically reduce model-building costs through open source, that doesn't mean they can build a 95-point model, and 95-point models will capture most of the market. This is analogous to Google in history: initially the search engine market was flourishing with many players, but eventually over 90% market share consolidated on Google. When something is pushed to极致, few end consumers will want to use non-mainstream products.
In this context, many believe model-layer entrepreneurial opportunities are limited, with more potential at the data layer. But I think the market isn't actually short on data — public data remains underutilized. ChatGPT and GPT-4's data only goes up to September 2021, so considerations around proprietary data can wait; focus on making good use of public data first.
As for how to make good use of this data, that's where teams must demonstrate their capabilities — requiring continuous improvement of algorithms and engineering capabilities, which fundamentally comes down to people. In software technology innovation, we've always emphasized investing in "smart," "ambitious" "young people," because technology iterates every year, and young people keep pace fastest.
We're reserved about investing in vertical models based on supposedly unique proprietary data. From studying machine learning, we understand that data uniqueness matters more than data volume, but I believe companies that truly possess distinctive proprietary data are rare to nonexistent. And when general-purpose large models have sufficient domain-specific data in their training, vertical models may not necessarily outperform them.
For example, a certain education-listed company disclosed having lots of proprietary exam data, but in reality you can train decent results using publicly available exam data online — the listed company's proprietary data isn't needed. Much of what companies claim as proprietary data isn't as important as imagined. Truly valuable proprietary data is unique data, and doesn't require large volumes to achieve good results. But such unique data is extremely scarce — you might not find more than a handful of companies with proprietary unique data in the entire market. Even with such data, possession doesn't guarantee the ability to leverage it or generate commercial value. Once you apply these filters, you find very few investable projects.
Moreover, even without large models, traditional industries like intelligent manufacturing have their own data. But if they weren't utilizing this data effectively before large models emerged, having large models won't necessarily significantly improve data utilization efficiency. If they truly had investment value, they should have been recognized before large models appeared.
⑤ Tencent Tech: On one hand, according to your view, the emergence of models won't enhance the moats of companies with proprietary data. On the other hand, building large models doesn't seem that difficult — when ChatGPT first became hot, the industry默认 large models were the exclusive domain of major tech companies, but in the past three months, an increasing number of non-big-tech models have emerged. In this context, can we say that the barriers to large model entrepreneurship are themselves unstable? Do investors need to reassess opportunities?
Chen Yu: Regarding model moats, we need to look at this in layers. A model scoring 80 points may not constitute a moat, but one reaching 95 points certainly does.
Although various domestic companies are building models, none has yet comprehensively surpassed GPT-4, because its core moat lies in engineering optimization — the team's repeated refinement of engineering details, and not just optimization in one or two areas, but potentially involving 50 or even 100 sub-engineering optimizations. This is OpenAI's undisclosed "secret," requiring technical teams to repeatedly test and tune during their own innovation process. It's difficult to generate methodologies that domestic companies can借鉴 — it's an engineering innovation, not a技术创新.
Before GPT-3.5, OpenAI was relatively open, carefully describing their work in papers. At the GPT-4 stage, they became "deliberately vague." The outside world can only speculate or chat with OpenAI insiders to learn bits and pieces, but employees can't leak company secrets. For outsiders, it's like an inscrutable black box.
I previously worked as an engineer at Google. Google's search engine quality is excellent, and its moat also lies in engineering optimization. Fifteen years ago, Google already had over 200 engineers working full-time on search quality improvements. Each engineer would optimize the search engine in different ways, using A/B testing to determine whether an improvement could ultimately be launched. The result was quantitative change leading to qualitative change — hundreds of improvements annually greatly enhanced search quality, but no single person knew the full picture.
In this context, if I were to summarize investment criteria for such projects, I think there are several key points:
First, team members must have long-term深耕 in this domain; second, they need strong engineering innovation and execution capabilities; third, this kind of entrepreneurship burns through GPUs and compute power, requiring strong fundraising ability — otherwise you can't even get to the table.
How do domestic and international AI ecosystems differ?
⑥ Tencent Tech: Domestic large models are emerging like bamboo shoots after rain. Comparing domestic and international large models, we can see that the representative overseas model ChatGPT currently follows a to-C subscription model, while many domestic models are integrated at the industry level. From an investment perspective, what are the differences in large model investment methods and styles between China and abroad? Overseas AI development experience is ahead of ours — is there anything that can be "copy to China"?
Chen Yu: Looking at differences between domestic and international models, the core difference is still technical. The first-tier OpenAI GPT-4, second-tier Google PaLM 2, and Anthropic (founded by former OpenAI employees) are all excellent. If you run evaluations, you'll find the differences between Chinese and foreign models are primarily technical.
Additionally, although the mainstream overseas large model ChatGPT currently uses a to-C subscription model, this doesn't mean overseas isn't doing to-B. Their model applications are also flourishing in diverse ways. But one advantage of overseas to-C is stronger user willingness to pay — for example, Midjourney has only 11 people but $100 million in revenue, which is unimaginable in China.
From an investment perspective:
First, overseas we're more willing to invest in single-point技术创新 projects, while domestically many investment institutions evaluate whether a project idea can reach IPO;
Second, from investors' experience, China and abroad each have their strengths — overseas software is generally better than domestic, while domestic hardware is better than overseas. So domestic investment opportunities lean more toward hardware出海 like robots and electric vehicles, because they leverage China's supply chain advantages to reduce production costs. If you achieve 80 points at the AI layer and combine it with hardware, the product competitiveness becomes very strong. We're optimistic about general-purpose humanoid robots — they'll be the combination of large models and robotics.
⑦ Tencent Tech: Why do these differences in large model investment strategies emerge between China and abroad?
Chen Yu: Differences in large model investment strategies between China and abroad stem from project exit pathways. Overseas, investment in single-point技术创新 projects can exit through M&A. In China, where M&A cases are scarce, investment institutions prefer to bet on big ideas that can eventually go public through IPO.
Different investors may have different styles. For early-stage investors, investment is about betting on "variables" — if a project lacks future imagination space, we won't invest; if it only stays at the imagination level without team execution capability and firm conviction, we also won't invest. Large models have just begun; reaching IPO stage will take several years at minimum. If you want to invest in "certainty," you might wait until one or two years before IPO to invest — that's probably not early-stage fund style.
On the other hand, people do consider macro environment. For instance, hard tech projects are relatively more popular now, while model and software projects are somewhat less favored. But the underlying thinking hasn't changed — still following variables, investing in things with large imagination space, and continuously paying attention to technology-driven industrial transformation.
Some people have been quite anxious about exit issues these years due to macro environment. I don't think we need to worry excessively — for example, USD funds can invest through Sino-foreign joint ventures or QFLP, and exit pathways are relatively diverse, with Hong Kong stocks being one option.
We're a relatively early-stage fund; many of our projects go through multiple funding rounds, during which we can also achieve moderate exits, so our current investments maintain their own rhythm.
The critical next step for generative AI
⑧ Tencent Tech: OpenAI's ChatGPT counts as a representative case of overseas single-point技术创新 in recent years. Can you predict OpenAI's next critical inflection point?
Chen Yu: OpenAI likely does have "moves in reserve":
First, large language models will develop toward multimodality. "Multimodal" applies not just to input — output can also be text, images, or video;
Second, future generative AI output may evolve from 2D images to 3D models. OpenAI recently released the conditional generative model Shap-E, which can generate 3D assets. Shap-E can generate parameters for implicit functions that can be rendered as textured meshes or neural radiance fields (NeRF), enabling the generation of versatile, realistic 3D assets.
⑨ Tencent Tech: Finally, returning to domestic AI governance — recently the Cyberspace Administration issued data management measures related to large models. How should China's generative AI properly respond to regulatory requirements?
Chen Yu: Previously regulation relied purely on manual oversight; now we likely need to use models to regulate models, because model-generated content is voluminous and increasingly difficult to distinguish from real content. Regulatory departments need to consider how to better leverage technology to standardize the industry. Additionally, models themselves embody values — when enterprises train models, they must carefully screen and clean training data, aligning it with correct values. OpenAI has invested considerable effort in this area, which domestic enterprises should learn from.