BlueRun Ventures in Conversation with Moonshot AI and Muyan Zhiyu: At the Crossing of the AI Industry | The Road to AGI
China's AI Can Be Different
After two years of breakneck growth, the AI industry has reached a crossroads.
Is the Scaling Law still holding up? Has the competition between tech giants and startups begun to show a clear winner? What paths will the domestic and overseas AI markets take? The answers to these questions will determine the direction of the entire battle.
At BlueRun Ventures' 2024 Annual Fund Partners Meeting, Terry Zhu, Managing Partner at BlueRun Ventures, sat down with Yueguang Zhang, founder of Muyan Zhiyu, and Dikang Du, Head of Strategy and Investment at Moonshot AI. Standing at this crossroads of technology and industry, the conversation took on an undefinable quality — from model evolution to frameworks for product definition, from market selection to the dialectic of comparative advantage. Both of these highly representative AI star companies shared their contrarian thinking.
We hope these reflections will inspire the industry. Below is an edited transcript.
Terry Zhu: Let's start with a rapid-fire round. Over the past couple of years, AI products have been emerging nonstop. What's a recent product you've most admired?
Yueguang Zhang: What impressed me most was the savage AI built on Wordware. You just feed it a Twitter link, and it automatically roasts the tweet. Its takedown of Elon Musk went viral, and it will almost certainly lock in Product Hunt's #1 spot for the year. It validates a possibility: AI understanding people — their personalities, social behaviors, social information.
Terry Zhu: It leverages a great data source — Twitter. AI's summarization capabilities analyze it, presented in a pretty entertaining way. Dikang?
Dikang Du: The product I've admired most this year is Cursor, an AI coding tool that helps programmers with code completion and bulk code generation. This AI-native IDE has built a devoted following. Many former Microsoft Copilot users have gradually shifted to Cursor and become paying customers. I used to think there was no opportunity for startups in AI coding — after all, Microsoft had GitHub's data, VS Code's development environment, and its partnership with OpenAI. But Cursor broke through in a domain where the giant seemed impregnable. I think the founders' rapid iteration and their understanding of AI models — better defining context across various code generation tasks — made the difference.
Terry Zhu: There's something undefinable about that. What we love most is seeing startups punch above their weight and break through in the gaps between giants — that's where VC gets its sense of accomplishment. There's been a lot of technical discussion lately: Is the Scaling Law still effective? Or is it slowing down? Can Transformer go all the way? What do you think?
Yueguang Zhang: Generative AI model technology has likely hit some bottlenecks at this stage. We've also seen that progress in industry models has entered a phase of relatively minor optimization. But I believe you should never bet against "technology won't advance further." I started working on computer vision in 2016 and have seen more than one wave of AI. Each time, after climbing a step, it enters a relatively flat curve and gets slightly stuck. But there will always be smart people who find the next thing that can scale up, and the gap between these moments is getting shorter and shorter. If we extend our horizon to a 5-to-10-year cycle, I'm optimistic.
Dikang Du: Right. ChatGPT's success harvested the technical fruits of tens of trillions of tokens on the internet, unleashing the Transformer architecture's capabilities while receiving strong compute support. But today, high-quality internet data is relatively limited. So how to help models find high-quality data during pre-training has become an urgent problem.
The recently buzzy OpenAI o1 model explores using real synthetic data to give models internalized reasoning capabilities, reflecting during inference to find more reasonable chains of thought. This is inseparable from OpenAI's high standards in the data domain. So the path of using model intelligence as leverage to produce more high-quality data has begun to emerge. There's also research on solving math problems through self-play, converting inference compute into intelligence. So we shouldn't overestimate technology's short-term development, nor underestimate its long-term development.
Terry Zhu: So will the industry consensus on Transformer and Scaling Law continue?
Yueguang Zhang: I don't think there's any non-consensus on Scaling Law in the industry. Regardless of what architecture or data paradigm subsequent technologies adopt, the next-generation technical paradigm will definitely be something that can scale up. Audio, image, and video modalities are all trying to unify through Transformer methods, feeding more types of data in to scale up. I think Transformer's own technical value hasn't been fully mined yet — there will be a lot of multimodal technology emerging this year.
Dikang Du: Many people previously believed that Transformer combined with next-token prediction faced challenges in solving arithmetic problems, because humans do arithmetic from the last digit while models need to generate the first digit first, then subsequent digits. But we found that even so, Transformer plus next-token prediction can solve this problem, and without particularly large models, it can achieve very good results in high-digit operations and complex arithmetic. We've also tried combining traditional diffusion models with auto-regressive models to simultaneously solve some problems considered challenging for next-token prediction, and achieved very good results.
Terry Zhu: Both of you are at application startups building from models. How do you think about more efficient trial and error? This might be a black box of product development — it's rare to hear both of you share on this.
Dikang Du: The exploration of combining model capability boundaries with PMF is very important to us. To extend the base model's capability boundaries, you need continuous investment in model training and data quality. Second, you need to define your differentiated positioning in the market. Large language models are a consensus opportunity with many competitors. We want to establish a "useful to users" mindset. What's useful to users? Last October, we focused on long context representing long memory. This October, we released the Exploration version, hoping to explore long-inference products.
Terry Zhu: Why choose long-context and long-inference products? What's the trade-off logic?
Dikang Du: Initially, we divided model capabilities into three layers: First, base capabilities: scaling law plus next-token prediction. Second, data-related: how to find unified data representations? How to solve data bottlenecks language models encounter during pre-training? The top layer is user perception: including instruction following and long context.
When we released our model, GPT-4 already existed, and Claude had models too. We hoped to leverage our understanding of the Chinese market, using a tokenizer better suited for Chinese to improve the model's Chinese performance, and through synthetic data, help the model learn to achieve better results on long-context tasks. At the same time, we believe long context actually represents the model's memory. While everyone is expanding CPU and compute resources, if we can improve the model's memory, it can solve more user problems.
Terry Zhu: Excellent. So you can see that insight into the fundamentals of model training is indeed needed to perceive its differentiated characteristics. We often say what's most lacking now are product-manager-type entrepreneurs who understand AI. How are such entrepreneurs made? Having gone through 0-to-1 on products is very important. Yueguang, what do you think? How can AI product entrepreneurs trial and error more efficiently?
Yueguang Zhang: Let's push back to the most fundamental dimension. Industry booms are always about changes in key production factors. For current AI, it's the equation of energy and compute being converted into intelligence. From a technical perspective, development mainly moves in two directions: First, improving intelligence levels, from low-level intelligence gradually advancing to high-level intelligence; second, improving conversion efficiency, getting larger-scale intelligence at the same cost — this will be the most important change to watch in the industry's development.
Then several questions follow: First, who is intelligence provided to? Ordinary users? Or professional users or enterprises? This corresponds to what we often call ToC (consumer-facing), ToP (professional-facing), and ToB (enterprise-facing). But ToC doesn't necessarily mean delivering capabilities to users to realize market value. Like recommendation algorithms — they first empowered enterprises, which then used that capability to empower every user.
Second, is the capability you're providing more of a repetitive instruction, or something only very few people can do? The former is like AI search, the latter like AI for Science.
Then there's technical maturity. When building an AI product, users don't particularly care whether this is currently the best technology — they only care whether the problem is solved. So it's best to choose a track where "70 points is already solved, but with some effort and definition, you can get to 100" — that's the low-hanging fruit.
There are many more dimensions. For example, does the ultimate commercial space align with your vision? In competitive commercial factors, is there an opponent you simply can't beat? I'd say there are at least a dozen decision spaces. After all the elimination, the options left for you must be very few.
Terry Zhu: If you strictly eliminate by these standards, you might not have any choices left. So there must be certain standards where entrepreneurs are willing to take some risks. On which boundary conditions are you willing to make adventurous trade-offs?
Yueguang Zhang: First and foremost, what you should take risks on is still technology. If technology is evolving, if technical boundaries are changing, then even if something can only do 30-60 points now, but getting to 100 points would be especially valuable — that's worth trying.
Terry Zhu: So many choices are based on technical boundaries. This brings up the next question: Whether the model evolution curve is in a steep or flat phase, what does that mean for how you define product boundaries and pace?
Yueguang Zhang: Different companies' technical visions determine a lot. If my company's vision is to drive AGI realization, then we might choose domains with the largest technical fluctuation space. Even if we don't develop products that all users can fully utilize, as long as we maintain global or local leadership on technical boundaries, there's very clear value from both a capital and internal team perspective.
If the company vision is to provide users with some category of experience improvement, then you might focus more on technologies that can solve problems at the current stage. Here's a counter-question many people ask: Those technology-oriented companies keep pushing technical development forward, so how do application companies find their place in that? Will the product you develop get iterated away by OpenAI tomorrow? Actually, don't worry about this. If you can solve 100% of the problem, even if new technology comes later, it can at best only match what you've done. There might be some improvements, but the user experience will be relatively similar. For application companies, as long as you hold your competitive moat, it's not that dangerous.
Dikang Du: Our model team is divided into pre-training and post-training teams. For domains with large technical variables, the pre-training team covers more — base model iteration is a 0-or-1 variable. Either you can achieve exploratory reasoning, or you can't; there's no middle ground. For domains with relatively stable technical iteration, we have the post-training team collaborate with the product team to improve user experience.
Terry Zhu: Let's talk some big-picture topics. Actually, in this wave of AI, although China's large model research started slightly later than the United States, both countries are very active in application development. China's AI researchers and entrepreneurs are extremely abundant, and information exchange between China and the US is also very active. For both of you, how would you approach globalization? Born global? Or develop stably in the domestic market first and then go overseas in the future?
Dikang Du: Kimi is currently still focused on the China market from a business perspective. Because Kimi's base model has done quite a bit of optimization for Chinese users, it has relatively stronger advantages in Chinese, English, and some smaller languages. And the United States does have阶段性 advantages in base model training, compute resources, and talent. We need to accumulate more momentum on the model side before we can go overseas to compete with giants like OpenAI, Claude, and Google.
This also involves which technology curve we should bet on in the next stage. For example, OpenAI is currently betting on model reasoning capabilities, hoping to find higher intelligence through this. Claude is betting on coding, on AI using computers. Last October, Kimi focused on long context, which had certain leadership both domestically and overseas. For the next generation, we still hope to define our core competitiveness before moving toward the global market.
Terry Zhu: So had you thought about this before? Was it determined from the start to do domestic first? Or was there actually deliberation and choice along the way?
Dikang Du: There were definitely many discussions. First, we believe China is a large enough single market that it can grow a large enough C-end enterprise, and Kimi has taken the Super App path. We have indeed achieved certain results, but market competition is far from over. We hope to first establish the mindset of "China's most leading AGI company" and get the ticket to AGI that leads to the world. When both conditions are ready, we'll have more energy and model competitive advantages.
Terry Zhu: I've also discussed this with Zhilin Yang. His view is there are no shortcuts — first build internal strength, strive to catch up, then talk about the rest. I very much agree with this. Yueguang, what do you think?
Yueguang Zhang: We look at both domestic and overseas markets simultaneously. AI-native-generation entrepreneurs must have a born-global perspective. There are also two relatively objective technical issues:
First, AI applications with relatively high market-recognized value share an important characteristic: they don't directly generate content but process raw information and deliver it to users — like chatbots, search. This is a common paradigm. I believe the products that ultimately deliver the greatest value will emerge here. Not relying on AI's internal self-circulation within the model, but using the intelligence that humans produce.
But the information silo effect in China's business environment is greater than in the United States. The information silos between several major social platforms make it difficult for AI to fully leverage its information processing capabilities — whether in extracting information, processing information, or providing information to users — there will be certain limitations. This is a relatively practical problem.
Another point: the domestic competitive ecosystem also means startups can't abandon observation of overseas markets. Domestic giants are investing heavily in AI, iterating very fast, and competition is more intense.
Terry Zhu: So is observation of overseas markets just a strategic perspective? Or is it really worth cultivating overseas markets, with real harvest to be had?
Yueguang Zhang: Chinese companies going global have a clear advantage: a complete product R&D system. Not just engineers, but also product managers, designers, operations, and other talent — the entire industry has a complete talent system supporting it. Whether from a cost or efficiency perspective, there's an absolute advantage. This is essentially our advantage in the software supply chain.
Terry Zhu: You just mentioned that domestic and overseas markets present different elements and characteristics, involving talent supply, data, technology, and demand. The bargaining power of relevant stakeholders also differs. On this foundation, how can China's AI industry carve out its own development path? In the past, Chinese industry was often in a Tier 2 or Tier 3 position. But now, we need to start competing globally as a Tier 1 competitor. As entrepreneurs, how do we walk our own China path?
Yueguang Zhang: Beyond the complete software supply chain I just mentioned, Chinese enterprises have also accumulated a lot of experience in user demand insight and experience detail polishing from the mobile internet era. For example, abroad there isn't really the concept of content operations or community operations — these positions don't exist. But in China, not only do these concepts exist, they're quite important. Our work is very meticulous, placing user experience at a very high position. So in the ToC domain, Chinese enterprises — not considering geopolitical issues — have certain competitive advantages.
Dikang Du: We're also very optimistic about Chinese enterprises ultimately going global. We previously tried video music products. With almost no internal traffic support or R&D resources, we were still able to attract loyal overseas users and achieve significant user payments, even reaching unit-level breakeven. I think this validates Chinese startups' accumulated advantages in product talent and operations/growth talent.
Moreover, China not only has software supply chain advantages. Our hardware supply chain advantages will also gradually play out in the AI domain. China's embodied intelligence enterprises and overseas enterprises more dominated by science have different underlying cognitions and starting points regarding AI software-hardware integration. I think China has already established very complete channels in robot supply chains, future consumer electronics supply chains, and overseas channels. These can all help carry more Chinese AI companies' overseas ambitions. This is also a direction we're observing.
UniUni Completes $30 Million C2 Round | BlueRun Ventures Family
A VC Leads the Way Out of the Quadrant | BlueRun Ventures Perspective
