ZhenFund's Yusen Dai: A Deep Dive on AI Agents — Every Industry Will Face Its "Lee Sedol Moment" (Part 2)

Last month, ZhenFund managing partner Yusen Dai sat down with LatePost for an extended conversation on AI and Agents. We've transcribed the full interview, releasing it in two parts.

In "ZhenFund's Yusen Dai: A Long Conversation on AI Agents — Every Industry Will Face Its 'Lee Sedol Moment' (Part 1)," Dai analyzed the breakthroughs of the o1 and R1 models, noting that "in the Agent era, attention is not all you need." In 2025, the phenomenon of AI surpassing humans in specific domains will become increasingly common. How will this transformation reshape the future, and what opportunities and challenges will it bring?

Q: Another development that's had a huge impact on the current landscape is DeepSeek going viral. The impact itself is significant, and it's adopted an extremely thorough and consistent open-source approach. I think we can break this down into several layers. First, the major tech companies — many previously closed-source giants have made moves. Tencent and Baidu, for instance, have both integrated DeepSeek. Tencent integrated it across numerous products: its flagship AI product Yuanbao, its largest national-level app WeChat, and roughly a dozen other products. Baidu integrated it into Ernie. Alibaba and ByteDance, however, have not.

When do you think Doubao will integrate DeepSeek?

Yusen Dai: I'd be quite surprised if Doubao chose to integrate DeepSeek. In my view, ByteDance is particularly eager to explore the frontier of intelligence and places heavy emphasis on developing its own foundation models. Integrating DeepSeek would represent a significant shift, both externally in terms of image and internally in terms of morale.

That said, from another angle, if Doubao users find DeepSeek more useful, then from a user-value perspective for Doubao, it would make sense. But I don't think this aligns with ByteDance's original intentions for AI. From what I understand, they still want to achieve comprehensive leadership in AI, and they have abundant talent and resources.

Q: What about Tencent?

Yusen Dai: What we're discussing here is hearsay — as angel investors, we don't have access to their decision-makers' thinking. People used to say Tencent's approach to video was "late to start, first to finish," letting others run ahead for three years, confident they could always mobilize their massive WeChat user base. I'd heard that Tencent took a similar "latecomer advantage" approach to models, figuring that with its user relationships and data, and with everyone dependent on WeChat, it could wait until model technology converged or matured before integrating. Moreover, WeChat is user-facing infrastructure — you can't make drastic changes without significantly impacting users. So I actually think Tencent deserves credit for integrating DeepSeek. I'd heard AI search was being pushed since last year, but the decision to integrate DeepSeek was certainly made at the highest levels.

I think this is good for Tencent's users. I've heard that after integrating DeepSeek, many Tencent products have seen solid data growth — possibly double-digit. From a DAU perspective, when people tap WeChat search now, a prompt appears to download "Yuanbao powered by DeepSeek-R1." That traffic-driving capability is simply unmatched. So Yuanbao is now number two on the App Store; I wouldn't be surprised if it's number one tomorrow.

Q: So you see this as Tencent's choice? That it's less aggressive on self-developed large models, slightly behind, knowing someone will build a better model, and then leveraging WeChat as its killer weapon to actively integrate. Do you think this was a premeditated, active strategic path?

Yusen Dai: I've heard this was an actively chosen strategy. But I've also heard that Hunyuan, Tencent's own model, is aggressively hiring to expand its team. Looking at China's internet history, major platforms rarely rely entirely on third parties for critical infrastructure rather than building it themselves. So on one hand, I think Tencent's current decision is quite remarkable — perhaps this will open a new era. In the US, there are many such examples: Netflix has long used Amazon Web Services (AWS), even though Amazon has Prime Video, a direct competitor. Netflix still considers AWS the best choice commercially and technically. But in China, historically, if there was Alipay, you needed WeChat Pay — everyone wanted their own thing. Still, I think choosing DeepSeek was certainly a neutral decision, since the DeepSeek team has no intention of building a super app or going direct-to-consumer.

Q: So I think Pony Ma understands that Wenfeng Liang isn't particularly interested in building a high-DAU product.

Yusen Dai: Yes, so I think their goals are at least clearly aligned for now, and there's a foundation for cooperation. But whether Tencent will never want its own large model — that's hard to say. Technology changes too fast. Just as people said Microsoft was entirely dependent on OpenAI, Microsoft later seemed to plan training its own models and even invested in Anthropic. So these situations can shift. But the core question is who can consistently stay at the frontier. Over the past two-plus years, we've seen many who claimed to be building foundation models and challenging intelligence gradually fall behind. That makes sense — this requires talent, capital, and tremendous innovation.

Q: So when you said just now that among large companies, only ByteDance qualifies to do this, and among startups, currently only Moonshot AI?

Yusen Dai: If we're talking about pre-VC startups — the so-called "AI Six Little Tigers" — looking at the current landscape, only Kimi has the talent, team, capital, and users to sustain this capability. Even in a recent OpenAI paper, they simultaneously referenced the research of both R1 and K1.5. And around noon today, while you were on your way here, Kimi released Moonlight, its latest open-source small model. I think being able to continuously contribute to the technical community sets a fairly high bar for a team's capabilities and direction.

Q: Speaking of OpenAI's paper referencing both K1.5 and R1 — these two achievements were actually released on the same day. Right after they came out, I went to chat with people in the technical community. At the time, their feedback was that the recognition gap between K1.5 and R1 wasn't that large, but the ultimate impact difference turned out to be enormous. How do you view this?

Yusen Dai: I think open source is the key differentiator. DeepSeek-R1's work is indeed significant, and once open-sourced, everyone could use it — especially in the West, where it caused major repercussions.

In recent years, people in Silicon Valley had been questioning whether spending so much on pre-training was worth it. At least among secondary market investors, concerns gradually mounted about whether too much money was being spent. Then suddenly someone claimed you could train an o1-level model for $5 million — of course this was a misreading; the paper clearly stated this was only the cost of the final training run. But some wanted to make big headlines, sparking widespread concern in the US that caused NVIDIA's stock to plunge 16% on January 27. Once it became global news, its impact certainly couldn't be matched by Kimi simply publishing a paper or a technical innovation.

A classmate quite close to DeepSeek told me they felt OpenAI or Anthropic in the US could train a model like V3 for even less than $5 million, given their larger clusters and greater training experience. But many less-informed people saw this narrative and started comparing $5 million against others' $1 billion in funding. Now people are gradually realizing you can't compare them this way. Look, NVIDIA's stock has nearly recovered, right?

On training costs, industry insiders didn't actually find $5 million that shocking. People were probably more focused on innovations like MLA that reduced inference costs at the time. Moreover, model intelligence improvements and the decline in model training and inference costs — this has been continuously happening. After GPT-4's API launch, costs have already dropped over 90%, and they'll certainly drop another 90%-plus this year. This is inevitable. Chips will get more powerful, and people will find more optimization methods to reduce costs. So I think the primary focus now is still whether intelligence can improve. As long as intelligence improves, costs will rapidly decline — perhaps to one-twentieth or one-fortieth of current levels annually. So there's not particular concern about cost reduction. At least in the US, people believe this curve will materialize.

Q: So cost reduction is actually on a trajectory. Later, Anthropic founder Dario wrote a very long article. His earlier analysis was quite thorough, saying cost reduction follows the industry's larger curve.

Yusen Dai: Right, including on intelligence improvement. Of course, the latter part of his article got somewhat agitated, but I think his earlier technical analysis was quite correct. According to his argument, their alignment work on the Sonnet model took considerable time because they emphasize safety and such. Indeed, Sonnet isn't even considered a reasoning model, so they're still quite impressive.

I've heard they're about to release Claude 4. I think this is partly, perhaps, a benefit brought by DeepSeek — like the catfish effect. With a formidable, open-source competitor entering the model space, everyone has to accelerate. This is genuinely a good thing. And looking back, DeepSeek had another advantage: it was a completely fresh application, launching as the combination of R1 and search from a blank slate — that's a major characteristic.

There's another characteristic I only realized later. When people train reasoning models, they all benchmark against math and coding capabilities. Looking at papers from DeepSeek, OpenAI, or Kimi, they all use AIME, MATH, Code Bench for comparison. But after DeepSeek emerged, what stood out was its writing quality. I heard it was the first to specifically do alignment work on writing style, even recruiting people from Peking University's Chinese department for annotation.

When we first saw its responses, our initial reaction was that it seemed somewhat fanciful, constantly veering into quantum mechanics. For OpenAI, Kimi, and Doubao, they'd actually been trying to avoid this, because during model training, everyone fears the model talking nonsense.

But I think DeepSeek may have intentionally aligned its text output on one hand, and on the other hand, since it had positioned itself as a research lab, it didn't fine-tune much for so-called neutrality or truthfulness. So when it launched, people just started using it, and this characteristic unexpectedly became a good thing.

We noticed many people shared it because they found its answers and reasoning process particularly creative. I don't know if this was a happy accident, but in fact it also led to stronger virality.

Q: Have you talked to people in the circle about this? Do they think this was a happy accident? Was the writing ability intentionally trained?

Yusen Dai: I've heard some people say they may have indeed strengthened the model's writing ability, but others think this might be a result of insufficient alignment. So I think both scenarios are possible — I don't really have a definitive answer.

But purely from the results, this was a very important reason for its breakout. Because not that many people actually use it for math problems; most people use it for things like fortune-telling, and then suddenly discover that its results seem quite convincing. And things like MBTI tests — people don't think this is what a cutting-edge AGI model should be doing.

Q: Another point of curiosity about DeepSeek is how it actually makes money. We just talked about how Tencent, Baidu, and many other companies of various sizes have integrated DeepSeek. But my understanding is, it doesn't directly make money from these integrations, right?

Yusen Dai: If you're just using its model, it's already open source. The only way DeepSeek can directly make money right now is selling API access, and I've heard its API is profitable. Because they've done a lot of infrastructure innovation on the inference side, compared to other companies, their cost to serve their own model is lower.

Now many people want to use its API, but the problem it faces is insufficient compute. Because it still needs to do model training, and recently it seems to have closed API top-up entry — meaning don't give me money, I can't serve you. This is one manifestation of the business model. Many people ask if they can pay for a stable version, something like a GPT Plus subscription. So I've always felt that in the early stages of a technology revolution, don't rush to judge the business model by mature-phase standards. First create value for users and customers through technology, then extract a portion of that value as revenue. I think this will happen sooner or later, it just requires some patience.

Q: Did you have a clear understanding of this in 2024? Or was it only after the impact or inspiration from DeepSeek that you had a clearer, more firm idea?

Yusen Dai: I think this is also a continuous learning process. When our post-80s generation entered the industry, mobile internet was gradually rising, or the internet was already entering its second half. In the earliest days, like the 1990s, I was also an internet user then, but at that time I never thought about business models at all. I think we should often learn from history, and think about why many early internet companies were built on strong technology of that era.

Actually, looking back at the first problem Google encountered — it used PageRank, this new technology, to build a search engine with 10x better experience. Users loved it and spread it organically. But at that time it actually didn't know how to make money, because Google's search engine initially had no ads, the interface was very clean. After launching in 1998, a 2002 New York Times article said "The hardest thing to search on Google is its own business model," criticizing it for lacking a business model. But later, as we all know, it gradually found AdWords and AdSense in 2002, and after its 2004 IPO became the best "money-printing machine" today. This is a great example — if you asked Google what its business model was at the start, it didn't know either. But first it had a technology breakthrough, built a great product through technology, and then monetized the product's value.

Q: Does every technology breakthrough have such a natural progression? Or are we subject to survivorship bias, only seeing those technology breakthroughs that later achieved massive commercial success?

Yusen Dai: Of course, not all technology breakthroughs can make money. But I think it depends on which development cycle the technology breakthrough is in. I still hold that view — right now is a time when the slope of technological change is very steep. At such a time, if you force existing technology to monetize, it's like making a gifted high school student go earn money — he might only be able to do manual labor like moving bricks, and won't earn much. But if you nurture him more, wait until he becomes a PhD, then he can earn big money. So I think if technology development has already reached a plateau, like mobile internet, where technology five years ago isn't much different from now — that's when business models can bloom in a hundred flowers.

Let me give another example. Not just Google — when Facebook first appeared, it proposed a very cutting-edge product that triggered "viral" spread. But at that time nobody knew how Facebook would make money either. It tried banner ads, tried local ads, later did in-game ads, but none of these made much money. It wasn't until 2012, when it changed news feed from chronological sorting (like WeChat's sorting method) to recommendation-based sorting, forming the so-called information feed recommendation model. Only by achieving recommended feed sorting could ads be inserted. So it launched news feed ads in 2012, and went public the same year. Of course now information feed ads are also a super "money-printing machine," ByteDance's core business model. But Facebook launched in 2005, the feed launched in 2007, recommended feed launched in 2012, and it found its real business model at the same time — this took 6 to 8 years. At that time Facebook was always a company users loved but with an unclear business model, so great companies often go through such stages.

Q: Do you think ByteDance will go open source?

Yusen Dai: First, is open source something everyone must do? First, you need to be in a leading position for open source to have value. If you open source something mediocre just for the sake of open source, that's meaningless. Second, I think a slightly weaker form of open source is free. Free plus leading — I think that's already very powerful.

Is it necessary to go open source? I think this time DeepSeek had a "sweet benefit" — after going open source, it attracted high attention from the West. After making big news in the US, people back in China felt it was even more impressive, that it made Americans "break down." Of course, open source also has forms like partnering with WeChat, but I'm thinking this isn't just about open source — the company needs to persist in doing this. For example, suppose Doubao went open source now, would WeChat integrate it? I estimate not. So this isn't a simple question of open source or not. Suppose Doubao was as good as DeepSeek and then went open source, I estimate WeChat still wouldn't integrate it, and Alibaba's Qwen probably wouldn't either. This isn't saying their capabilities aren't good, but rather from Alibaba and ByteDance's standpoint, this is how it is. So I think what makes Wenfeng Liang and them impressive isn't just open source, but that they persist in open source, and their market positioning is one that doesn't make people feel threatened.

Q: Right, persist in open source, and maintain neutrality, without accepting particularly large investments from any major company.

There's also been a recent change — OpenAI is also considering open source. Sam Altman posted a tweet giving everyone two options: one is open sourcing o3 mini, and another is open sourcing a phone-size model, a model suitable for mobile devices. Which one would you be more excited to see them open source?

Yusen Dai: Of course, I think either would be great, but I'm definitely more interested in o3 mini. Because I think the use cases for mobile-side models may not be that large right now — what people need more is breakthroughs at the intelligence frontier. o3 mini is a very powerful model; after longer reasoning time, that is, in the current o3 mini pro, o3 mini high modes in GPT, its performance is very good. If a model of this level could be open sourced, and everyone could know how it was made and its characteristics, I think that would be tremendously valuable. And I've heard this model isn't very large in scale, so this might have a lot of reference value for everyone in model training and application.

Q: Have you heard how large it is?

Yusen Dai: From relatively reliable sources, I've heard its activated scale each time is 3.7B, which honestly shocked me a bit — it does seem a bit too small. But this size means they can indeed take a very large o3 (o3 should be quite large), turn it into a very small o3 mini, and then let o3 mini think for more time to get good results — this is indeed very impressive work.

Q: They actually shared their own concerns about not open sourcing before. They believed open source would weaken competitive advantage, for example giving Google an opportunity.

Yusen Dai: So I think this is what's great about Wenfeng Liang — he really shared many technical secrets with everyone, letting everyone get better. But from a pure commercial company perspective, there are indeed many concerns. After all, beyond the issues just mentioned, OpenAI also worries that powerful AI could be exploited by bad actors, which may also be a very reasonable concern.

Q: What impact do you think DeepSeek will have on companies that were already in the open source ecosystem wanting to take the lead, like Meta and Alibaba, which have always been open source?

Yusen Dai: I think it will definitely be an inspiration. Everyone discovered a more "intense" competitor has arrived. Originally the open source community, jokingly speaking, was a bit like "cyber Buddha," somewhat charitable in feeling. Whether Alibaba or Meta, they were big companies contributing compute for everyone to use, driving the whole industry's development. But now comes DeepSeek, which progresses faster and is more open — this is definitely both pressure and inspiration for everyone. But indeed, I think DeepSeek's neutrality is a rather unique advantage. Tencent can use it, Qwen can use it — this isn't just a capability issue, it's about where his interests lie.

Q: Didn't Apple recently have communications or cooperation discussions with DeepSeek too, though in the end it chose Alibaba.

Yusen Dai: Apple talked with many companies, and also talked with Moonshot AI. I think from Apple's perspective, choosing Alibaba is very understandable. It definitely needs to choose a partner with stable service, that can handle large-scale user volume well, whether in infrastructure, service, or technical experience — excellent in all these aspects.

Q: Actually Alibaba has been quite open in this round.

Yusen Dai: Qwen is quite compatible with Llama, and its product models are good, updates are very frequent, so many developers are using Qwen. Honestly, when using DeepSeek's R1, because there are many "hallucinations," if you use it to build applications, it may not necessarily be the best choice.

Q: Before DeepSeek went viral with the general public, I felt Qwen and DeepSeek had roughly comparable influence in overseas tech circles, since both were part of open-source model families.

Yusen Dai: Exactly. Looking back, no matter how good Kimi's benchmarks were, if it wasn't open to others—no open-source access, no overseas application services—there simply wasn't any awareness of it abroad.

Q: How did you discuss this internally before? Why didn't Kimi go open-source?

Yusen Dai: I think even now, open-sourcing isn't something that has to be done. As I mentioned earlier, it's just one option a company might choose under certain circumstances. For instance, you might consider it when there's no competitive pressure to keep things proprietary, no fundraising pressure—and what we're seeing now is an ex-post outcome. It took open-sourcing plus some fortuitous moments to get here. So I don't see it as a mandatory path. Of course, those who choose to open-source are impressive and deserve respect. But for a commercial company, the core is whether you can create user value and ultimately convert that into commercial value. So open-sourcing isn't the only way forward—it's just a very interesting, innovative path.

Q: But today, none of the companies exploring AGI would take user value as their core focus.

Yusen Dai: Many are still primarily driven by technical value. I'm just saying that during a period of technical growth, only when technical value improves can user value follow. So pushing the frontier of technology is absolutely critical. After large models emerged, there was probably a wave of so-called pragmatic investors or entrepreneurs who thought about making money with existing technology. But I think Kimi definitely belongs to the other category—it aims to advance the technological frontier. This circles back to what we said at the beginning: creating that astonishing, almost magical product experience, and ultimately capturing commercial value.

Actually, when Kimi took off in 2023, a big reason was that it was the first product to combine chat, search, and long-context capabilities. At the time, ChatGPT couldn't search, and its handling of long text, multiple texts, and multiple files wasn't great. So in its first two years, Kimi broke through precisely on the strength of its long-context technical vision and its fusion of search with chat—delivering a differentiated user experience.

Q: Was choosing long-context a non-consensus bet back then? Was it a difficult decision?

Yusen Dai: Long-context was definitely one option among technical choices, but whether to prioritize it above everything else—that wasn't a consensus. There was an anecdote floating around, I don't know if it was true, that after Kimi blew up, Baidu asked why they hadn't done long-context themselves, and supposedly it was because it hadn't been ranked among their first-tier priorities. There were simply many other higher-priority things to do at the time. Lots of people were working on CharacterAI-style emotional alignment, for instance. But Kimi firmly chose long-context and pushed it to the extreme. Because long-context unlocks two critical scenarios: processing multiple files, and search—like reading 100 web pages and summarizing them. Neither of these works without long-context.

Especially at the time, Kimi had just been founded and hadn't raised that much money yet. It was a young team, a small team, resource-constrained, and had to stay focused on one thing and choose the right direction. Actually, many factors behind DeepSeek's current popularity would have applied to Kimi in 2023 too. When resources are limited, you need to achieve a breakthrough at one critical point and deliver that stunning experience to users—that's how you break out. So I've reflected on this and found quite a few parallels. I'm not trying to flatter myself here; I genuinely think there are similarities.

Q: Does long-context still help with what Kimi is doing now?

Yusen Dai: For instance, when doing retrieval, Kimi actually performs better on truthfulness and accuracy. Of course, average users probably don't compare that closely. Honestly, many DeepSeek users right now don't even notice the "hallucinations," but you might get burned when you use it to write a report. I ran into this just yesterday—someone shared an article in a group chat, and everyone in there is pretty sophisticated, but I took one look and thought, "This reeks of DeepSeek."

Q: What direct impact do you think DeepSeek's breakout will have on the "Six Little Tigers"—the model startups that have been constantly compared to it?

Yusen Dai: Honestly, I think it's definitely had a clearing effect. Even before R1 broke out, several of the "Six Little Tigers" had already stopped training their own foundational models and weren't aiming for SOTA anymore. I think R1's emergence has made everyone realize: if you don't have a shot at SOTA, you're genuinely better off going vertical or building applications.

Q: Why did they give up?

Yusen Dai: Funding reasons, team reasons, positioning reasons. As Kimi's angel investors, speaking practically—looking at the K1.5 model's performance and what they have coming next—they'll show further improvements in the MATH and coding dimensions we mentioned earlier. From an academic contribution standpoint, K1.5's technical contributions like long-to-short in reasoning have been well-received. And today's Moonlight release, plus MoBA from a few days ago, also demonstrates that the Kimi team can sustain technical exchange and output with the broader community.

At the same time, Kimi's user base has reached the tens of millions in DAU and is still growing. Honestly, quite a few people, after trying both DeepSeek and Kimi, still prefer Kimi in many scenarios. For example, Kimi has fewer hallucinations and performs better in work contexts. In certain multimodal reasoning scenarios, like taking a photo to solve a problem, DeepSeek hasn't done that yet. So this may sound self-serving, but I genuinely believe that from the perspectives of team, funding, technical capability, and user product, Kimi is the only one among the "Six Little Tigers" with the sustained ability to compete for SOTA models. Of course, this path is extremely difficult—it requires money, people, all kinds of things—but I think it's at least worth trying.

Q: Will Kimi become more focused going forward? Will it cut certain things?

Yusen Dai: They've already cut quite a lot—for instance, their overseas business. Right now they're just going to keep pushing for SOTA.

Q: Have they officially stopped doing video generation?

Yusen Dai: At least from what I can see, knowing what not to do is very important.

Q: Most of the "Six Little Tigers" had already given up before DeepSeek emerged. Was this within your expectations?

Yusen Dai: Actually, by mid-2024 we already sensed this would be the outcome. By then it was already obvious with several of them—whether in terms of will or resources—that continuing would be very difficult. I think one nice thing about Kimi is that its team has been very stable. This has to do with how the team is structured—the co-founders have long histories of working together. If you look around, personnel changes at the various model companies have been quite significant. Entrepreneurship is really like walking a balance beam—your companions keep dropping away as you go. Often, just staying at the table is already quite an achievement.

Q: So far we've mainly discussed DeepSeek's impact on model companies, including the big players—both open-source and closed-source—and some startups. Next we can talk about companies in other parts of the ecosystem. For instance, in the more open-source trend that DeepSeek has brought, what kind of impact will there be? One category of company that comes to mind is AI cloud platforms. According to DeepSeek's announcements, during the upcoming Open Source Week they're planning to open-source some inference optimization techniques at the infrastructure layer—what impact might this have on companies like SiliconFlow and Infinigence AI in terms of their startup trajectory?

Yusen Dai: We're angel investors in Infinigence AI, and their business volume has been growing very rapidly—they've been getting a lot of demand. Especially various state-owned assets and government entities are desperately trying to deploy DeepSeek, so demand in that area has exploded.

They've done a lot of innovation, including doing inference on Huawei's chips, which is very hot right now—lots of people want to use it. I think the "heat" around open-source models has genuinely created great opportunities for AI Infra companies. What models were these companies supposed to serve before? If it was all closed-source, proprietary models like Doubao or Kimi, they couldn't really play much of a role, because ByteDance would handle its own serving. But in the long run, it still depends on whether they can continue serving customers well. After all, public cloud companies like Tencent Cloud, Alibaba Cloud, Volcano Engine—they have abundant capital, better Infra capabilities, resources, and customer service. So for customers, they're not doing charity work; whoever serves them well at a good price, they'll choose. So for startups, there are still plenty of challenges.

And with DeepSeek open-sourcing these "black magic" techniques, it means they actually have advantages on the serving side too—their costs for equivalent service may be lower than others'. In the short term, the surge in compute demand caught everyone off guard, so they couldn't handle it themselves and had to let others absorb it—that's perfectly normal. But if things stabilize, whether these startups still have advantages against the big public clouds and DeepSeek's first-party service remains to be seen. Overall, though, it's definitely created a lot of opportunities.

Q: Actually AI cloud platforms are sandwiched between clouds and models, right? They could get squeezed from both sides, but they might also gain opportunities from ecosystem shifts.

Yusen Dai: Right—if open-sourcing increases the choices at this middle layer, with different frameworks and different models to choose from, then this middle layer keeps getting better. But if it ends up converging like operating systems to just a few choices like iOS or Android, then probably the system providers end up providing it themselves.

Q: What impact do you think it will have on the broad mass of application-only companies?

Yusen Dai: I think it's definitely positive. It's another better, open-source, fine-tunable model available to use. In this process, if you want to do something office-related on the model's main highway, that's still quite difficult. But if what you're doing enriches the model ecosystem, that's different. I've always used this analogy: in the early days of a technology revolution, it's like the BlackBerry era. Because BlackBerry's technical capabilities were limited, the product-market fits you could achieve were very few. The BlackBerry era was basically email and messaging. Even if Yiming Zhang went back to that era wanting to build TikTok, he couldn't have done it—BlackBerry didn't have the conditions. But why did mobile internet flourish later? First because you had the iPhone, which was powerful enough to unlock many new scenarios. It had a good camera, good screen, good network, good chip—so it could unlock short video, mobile e-commerce, social networking, these scenarios.

After the iPhone came Android, which made the market more open. More phone makers like Xiaomi, OPPO, and vivo joined in, further spreading smartphone adoption. Sonnet, 4o, and o1 are a bit like the iPhone moment—closed-source technological advances that let many people build applications on top of them. DeepSeek might be the Android moment: it went from closed-source to open-source, and it's strong enough to give people more choices for building applications. So on one hand, technological progress brings better product experiences, leading to "killer apps"; on the other hand, it makes ecosystems more prosperous. Before, you could only do a limited number of things. It took the iPhone and Android to make something like TikTok possible.

Q: I'd also like to talk about the impact of o1 and R1 on infrastructure compute demand, which everyone is paying attention to. DeepSeek R1 was extremely hot for a while, and that connects to what we mentioned earlier about NVIDIA's stock price dropping. There's a view that because its training costs are low, it will reduce demand for compute. I also saw you posted some thoughts on this on Moments, and many people have different opinions.

Yusen Dai: I think compute demand has different structures. Originally there was just training and inference. During the 2023–2024 arms race phase, people summed it up simply as "scale is all you need"—the idea that if you just bought enough GPUs, you'd get better results. Of course, pre-training hadn't hit a wall yet, or people hadn't realized it had, so that kind of thinking held up.

But now we've found that massive short-term investment in pre-training has limited marginal returns. Grok 3 was trained on 200,000 GPUs—there was progress, but marginal returns are diminishing. So it's not that "scale is all you need" was wrong; it's just that the marginal returns of the miracles it produces are decreasing. But I think what will happen is that because model capabilities have reached the critical point for building Agent products, and keep breaking through, once Agent product forms can actually land, the tokens they use and the inference compute will increase dramatically. If you're just doing chatbots—chatting with ChatGPT, Moonshot AI, or Doubao—there's not that much to talk about, and it doesn't cost many tokens. When it can help you do more complex things, requiring more tools and reasoning, inference compute demand might not increase 10x but 100x or 1000x. This couldn't happen before because the technology hadn't reached that level. But now I think technology has hit this inflection point, and inference demand could rise substantially.

Q: Will this hundred-fold or thousand-fold growth in inference compute demand happen in 2025?

Yusen Dai: First, from the perspective of technology history, whether this happens in 2025, 2026, or 2027 really doesn't matter at all. Like autonomous driving—what matters is that it eventually happens; the exact year is less important.

But I think Agent products, at least from what I can sense, are about to break out of the tech bubble. Deep Research, for example, definitely requires many more tokens. That's why Sam Altman said GPT Pro, even at $200 per month, was still losing money—inference demand had increased so much. But I think there are two things here: first, the proportion spent on pre-training versus post-training inference will shift; second, this will indeed have structural implications for NVIDIA. As of February 2025, NVIDIA is still the highest-performing and most efficient choice for both inference and training. But we've also seen that after R1 blew up, domestic chips started optimizing specifically for R1, and this kind of targeted optimization actually works better.

Q: They're already using Ascend.

Yusen Dai: They've seen Ascend's 910B.

Q: And even with NVIDIA products, you can use FP4 inference optimization.

Yusen Dai: Right. I think there's always been this dynamic: when technology hasn't converged, GPUs have strong generality. Or rather, why does NVIDIA exist? Originally everything was CPUs, the most general-purpose. Then people wanted to play games, which had very specific demands, so GPUs were made specifically to accelerate games—and later, AI. Currently, GPUs remain the most general-purpose choice for AI training and inference. But if you're only serving one specific model, there are two approaches. One is like Ascend, where you can do specialized optimization; the other is doing something like Eclipse, what Broadcom and Marvell are doing.

Q: Or like Google with TPU, optimizing for their own needs.

Yusen Dai: That's also a form of specialization. Once architectures stabilize, the chip industry can usually achieve higher efficiency through specialization. So this involves whether the architecture will actually solidify, which I think is a point of intense debate. Right now, if the O1 and O-series path can go very far, then ASICs might gradually work. But from another angle, suppose next year or the year after, the foundational architecture changes—Transformers stop working and get replaced by something else—then building ASICs would have been a waste, and you'd have to fall back on GPUs. So there's a lot of uncertainty. But NVIDIA does have one problem: its market share is so high now that it's hard to go up.

Q: Right, it seems like it's already peaked.

Yusen Dai: Exactly. Its market share is over 90%, so there's room to go down. This possibility of decline worries a lot of people. On one hand, expectations for future compute demand are relatively high; on the other hand, expectations for NVIDIA's market position and the gross margins that come with it are also high. Once the market structure has problems, its gross margins could be affected—that's what people are worried about. But if you ask what everyone is actually doing now, including what DeepSeek most wants to get, it's definitely still NVIDIA products. Buy as much as you can, find every way to buy them.

Q: Actually, the most stable play in this wave is still Broadcom.

Yusen Dai: Broadcom or Marvell—both have performed quite well. But for ASICs: first, they basically won't be usable until 2027; second, there are situations where price changes could make the ASIC path not work. And actually producing ASICs and putting them into use involves many issues with production capacity, yield rates, and efficiency. It's not something you can just design and make, so there's a lot of uncertainty.

Of course, NVIDIA has also run into some problems, like liquid cooling issues and overall yield rate issues. Anyway, I think the landing of Agent products is definitely positive for compute overall—everyone's heard of Jevons paradox by now. But whether NVIDIA's market structure will change, I'd just say new possibilities have emerged. So for stock traders, the first reaction to DeepSeek might have been to sell first on the news, then add back when it seemed less problematic.

Q: We've talked a lot about future outlook—some things may happen this year, some may take much longer. To summarize, what do you think we'll most likely see in 2025?

Yusen Dai: I think we'll see more "Lee Sedol moments," where AI surpasses 99% of humans on certain tasks. This is already happening gradually. For example, coding—AI's coding ability should already be stronger than 99% of humans.

Q: Do you mean surpassing 99% of programmers, or 99% of humans?

Yusen Dai: I'm talking about humans right now, but I think surpassing 99% of programmers may happen soon too. In competition-level programming like Codeforces, AI already surpasses 99% of programmers. But competition programming and daily programming output are different—daily programming requires more context, reading various codebases. But I think cases where AI defeats humans, or even elite humans, in capability will become more and more common. We'll see more astonishing news along these lines. Also, I think more Agent products will emerge in more convenient, practical forms and become phenomenon-level products. Maybe not hundreds of millions of users, but I think they'll break further out of the tech bubble, reaching the kind of breakout that Cursor has achieved.

Q: What's Cursor's DAU now?

Yusen Dai: I'm not sure about DAU, but its annual recurring revenue (ARR) is around $100 million. DAU is hard to measure, so don't use DAU to evaluate AI products. What users are willing to pay for the value a product provides—that's probably what matters. I think model development speed will accelerate, and open-source sharing and experience sharing will increase, which is quite interesting. Actually, in China, we're now getting the feeling that the US had when ChatGPT first blew up, because now local governments everywhere are starting to use DeepSeek, and everyone is integrating DeepSeek. I think this is very important for raising AI awareness. People will realize: so AI is this powerful. Before, Moonshot AI, Doubao, and other models combined probably only had a few tens of millions in DAU, maybe under 200 million MAU. I think that only got about one-tenth of people using relatively advanced AI models. But if we can get tens of percent of people to try advanced models and feel AI's power, then whether from the perspective of entrepreneurs, users, new products, or resource and capital investment, I think the entire industry will experience a Cambrian explosion of ecosystem prosperity.

Q: It's 2025 now, and you've said that in some fields this year we may see "Lee Sedol moments" where AI surpasses 99% of humans, or even elite humans. I feel like the DeepSeek moment has made the whole industry's development speed faster. So what do you think happens if we achieve AGI faster, or unlock "Lee Sedol moments" in more fields? I'm having trouble imagining what changes—what will people do, how will social structures change.

Yusen Dai: I think we're living through a very interesting period in human history. Exponential growth is actually the normal state of world development, because every year we build on the previous year. But witnessing and personally experiencing exponential growth is quite rare.

Q: What exponential growth are you referring to? Economic output, or something else?

Yusen Dai: GDP growing 2% or 3% every year—that's exponential growth. But generally, you need a lifetime to feel that kind of exponential growth. This year versus next year might not seem very different. But with AI, specifically—from o1 to o1 Pro to Deep Research—I've distinctly felt its exponential growth within just a few months. That experience is quite special. And I think this will dramatically change how we form expectations about the future.

So a lot of people are asking now: what is AGI, and what happens once we achieve it? Personally, I do think AGI will have major impacts on productivity, society, even politics and culture. But as for exactly what those impacts will be when it arrives—I think we need to brace ourselves. Issues like safety, and how to handle social welfare when new technology emerges, I think people will only really take these seriously once they're actually happening.

Q: And who controls that capability really does shape the global order.

Yusen Dai: That's why the accelerationists argue that AI is going to advance regardless, bad actors will use it for harm, so the good guys should develop it faster.

Q: Like financial fraud, or those Deepfake AI porn cases that happened in South Korea.

Yusen Dai: So you need even more powerful tools to detect Deepfakes, because humans won't have the energy to verify everything anymore. I think this will definitely bring enormous impact. Actually, I've been thinking—on one hand, a lot of people will probably lose their jobs, I think that's fairly likely. Right now everyone's definition of AGI is basically how many people's jobs it can replace. If the point of AGI is replacing human labor, then achieving AGI is essentially equivalent to mass unemployment, isn't it? Of course, that's from a societal perspective. Some say material abundance will follow, everyone gets universal basic income, but I don't know how that actually plays out. I think there will definitely be massive shocks.

But from another angle, what we humans consider "reality" is going to change dramatically—whether it's video generation, image generation, or content generation. I was born in 1986. Back then, all the information a person could access was authority-certified: books, newspapers, nothing else could get published and distributed. The internet's huge significance was letting ordinary people's writing be seen. Now AI has become: whatever you want, it generates for you. I've noticed that even I, a lot of the time, have lost my judgment—I can't distinguish real from fake anymore. So in this kind of environment, how to adapt further and build your own cognitive framework, I think that's a crucial question.

Q: There's an internet meme that's getting more profound by the day—it basically means "you can't Photoshop video, so it must be real."

Yusen Dai: Right, now even video can be generated. I think this will profoundly affect how we socialize and how we perceive the world. I've noticed a pattern in technological development: the first wave is always the most capable people creating the most powerful technology. The second wave is using that powerful technology to build the strongest tools for the most capable people. Take computers—they were originally created for nuclear explosion calculations or codebreaking. These super tools designed for "superhumans" gradually become democratized, reaching ordinary people, then miniaturized for homes, then mobile, everywhere.

We're still in the phase where the most capable people are building super tools for elites. But I think this will ultimately benefit the masses. When we invested in Huiwen Wang's Lightyears Away, our slogan was "accelerate AGI to benefit humanity"—I think universal benefit is the inevitable end result. Though there will definitely be that William Gibson phase of "the future is already here, it's just not evenly distributed"—and right now it's certainly not evenly distributed. So whether it's DeepSeek going open source, or high-user-base products like Moonshot AI and Doubao, they're all accelerating a more even distribution of the future, and that carries significant meaning. I think new technology ultimately needs to benefit the masses, all of humanity, to have real value—not just be held by a few wealthy individuals or a handful of companies. That's the outcome I hope to see overall.

Q: I'm curious—what are you personally doing to prepare for AGI that may arrive sooner than expected?

Yusen Dai: Working out. I think in the investment world, identifying excellent entrepreneurial teams is actually crucial. With more technological innovation, entrepreneurs become even more important. Of course, Wenfeng Liang was an entrepreneur to begin with—it's just that he's so exceptional he could make money trading stocks and doing quantitative finance on his own. There are many potential Wenfeng Liangs out there who might lack startup capital. So I think VC is important at this moment, especially early-stage investment. Because theoretically, early-stage investors take the biggest risks. If everything were already certain, they wouldn't need us. But I think we're back in a period full of uncertainty—not everyone can show up with 10 billion in self-funded capital like Wenfeng Liang.

Q: What do you think the next generation should learn? How should they be educated? I think a lot of people are wrestling with this.

Yusen Dai: I think the most important thing is the ability to ask questions. For instance, I often encounter this situation: facing a highly capable Deep Research, what should I ask it? How should I direct it? As the head of an AI company, every day I need to think about what to have people work on, what this year's direction is, what this month's priorities are. This process actually requires enormous thought, because things don't push themselves forward—we need to actively set the direction. But our current education system mostly teaches students "how to do things," equipping them with skills.

Yet many of these skills can now be handled by AI, or accomplished by directing AI. So in this situation, what should we ourselves do? This becomes a vitally important question. Second, much of our current work is essentially "stitching"—copying, piecing together, and organizing various bits of information into a final report. But AI already does this better than humans. So we need to consider whether our output adds unique value to humanity or to the overall knowledge system.

Like our conversation right now, there may be some "stitching" involved, but at least it can produce some unique data. Can our work create unique information that doesn't exist in AI training data? Or are we merely repeating what AI already knows? This has major implications for the nature of both education and work.

Q: That Elon Musk quote I find particularly interesting, essentially: "I want to die on Mars, just not on impact."

Yusen Dai: Right, the key is don't crash and die.

Q: I have a somewhat personal question. You're doing investments now, about to study HI (Human Intelligence), and also researching the secondary market. Facing so many domains, how do you maintain such efficient learning speed?

Yusen Dai: Not particularly fast, or I'd have gone much deeper on DeepSeek by now (laughs). Actually, the day after V3's release—December 27th—I organized a discussion at home, inviting a dozen or so friends, including people from ByteDance and various AI research institutions, to discuss the latest developments in AI. It happened to coincide with DeepSeek V3's release, which was incredibly exciting. So our learning ability is still decent.

For example, the day after MLA was released, I thought it was extremely impressive. I was in the US at the time, discussing the technology with friends. I think interest is crucial—you only learn effectively when you're genuinely interested. I'm also somewhat "meddlesome" by nature. The day ChatGPT launched, I stayed up using it until 4am, feeling this technology was completely different. This habit probably traces back to getting online in 1998, and first using Google in 1999. Search engines were terrible back then, barely returning anything useful, while Google's results were completely different—it was a massive shock.

There have been many similar experiences. I started using Xiaonei the day after it launched, and later developed a deep impression of how internet entrepreneurship evolved. After ChatGPT's release, I experienced it immediately and organized a research group right away. Same with Devin—I thought it had potential, so I quickly organized discussions.

Looking back, the first wave of internet entrepreneurs were typically the earliest people online. Mobile internet pioneers were usually the first to buy iPhones. Even the earliest Tesla investors who made money did so because they bought Teslas first. So being willing to spend a little money, or even no money, to experience the future remains incredibly important. Devin's subscription is $500 a month—seems expensive at first glance, but for investors in my circle, that's maybe the price of a bottle of Moutai, and that cost lets us glimpse future trends ahead of time.

Q: Indeed.

Yusen Dai: So the most important thing is hands-on practice, actively reading papers, paying attention to what top researchers are doing. OpenAI, DeepSeek, and other frontier institutions—their high-quality information is mostly free and publicly available, worth learning from. Early last year, many in the secondary market believed AI demand would hit bottlenecks, that the industry might decline in 2025. But from my observations inside the industry, that was completely not the case. AI training is still accelerating, the arms race trend is clear, companies are purchasing compute at massive scale. I started investing in ASICs in the second half of last year. The logic then was that while ASICs might be important in the future, near-term realization was still uncertain. Similar stories are common in the industry—early on AMD was seen as challenging NVIDIA, now ASICs are seen as potentially threatening NVIDIA.

Q: ASICs have actually challenged NVIDIA several times already. Some companies from the 5G era were representatives of ASIC companies.

Yusen Dai: Right, every time it seems like there'll be an impact, but the effect ultimately proves limited. Though the secondary market tends to "hype first, ask questions later"—whether things actually materialize isn't necessarily important. One interesting thing about the secondary market is that it can serve as a tool to validate your understanding. For example, I knew DeepSeek was very strong quite early, but they didn't need external investment. In this kind of situation, the secondary market provides an opportunity to "place a bet"—like how training models need reward signals, market feedback can validate whether your thinking is correct. So I think the real value of the secondary market isn't making money, but providing a mechanism to continuously test and refine your understanding.

Q: How are you using AI tools for investment decisions now?

Yusen Dai: Deep Research gave me a very concrete case. Recently Trump was announcing new tariff policies every Friday, and I was researching US Treasury bond trading trends. I asked Deep Research: "When Trump announced tariffs in 2018, how did US long-term Treasury yields react?"

I had two hypotheses: one, that tariffs would drive up inflation, raising long-term inflation expectations and pushing Treasury yields higher; two, that market risk aversion would intensify, with investors selling stocks and buying Treasuries, causing yields to fall. Deep Research analyzed this in five minutes, pointing out that 2018 historical data showed Treasury yields dropped every time Trump announced tariffs, with markets favoring safe-haven moves. This analysis helped me make the decision to buy US Treasuries, which proved correct.

Q: That's indeed a good example of AI empowering decision-making.

Yusen Dai: Right, I ask it a question and get an answer in five minutes. If it were my assistant, or some friends with extensive secondary market experience, they might not get back to me until the next day with something like "it'll go up." In financial markets, rapid response really matters.

Q: You mentioned learning just now — it seems you have a strong interest in AI agents?

Yusen Dai: Yes, I love reading, which is why I talk about agents so much. I genuinely feel they've transformed my life. Sometimes when I'm reading, I'll come across an interesting idea and want to dig deeper. But if I looked up the sources myself, it could take a lot of time and disrupt my reading flow.

Here's an example. Reid Hoffman's new book, Super Agency, mentions the history of GPS development in the US. Initially, the US worried that GPS being too precise would compromise national security, so they deliberately introduced a 100-fold error margin, limiting it to very coarse applications. Later, they realized this was actually restricting GPS's commercial value, so the Clinton administration removed the restriction, fully opening up GPS and enabling applications like Meituan delivery and DiDi ride-hailing.

This example made me think about AI development: should we restrict it on national security grounds, or choose openness, mutual benefit, and ecosystem-building? So I had Deep Research investigate the background of the 2018 GPS openness policy and do a comparative analysis with current LLM policy. If I had looked this up myself, an hour might have passed. Instead, I just had Deep Research work on it while I kept reading, then reviewed the summary when it was done.

What I ultimately found was that the key to GPS openness was the US developing technology to directionally jam GPS signals — allowing localized shutdowns during wartime while keeping it open for civilian use. This answered a crucial question: how did the US government address national security concerns while opening up GPS? Researching this myself would have taken ages, but Deep Research handled it for me. That's why I'm willing to pay for it — from a time-value perspective, it's absolutely worth it.

Q: $200 per use — you think it's completely worth it?

Yusen Dai: Of course it's worth it. $200 for a research session, averaging $2 per use. The value is incredible.

Q: Do you have any other book recommendations?

Yusen Dai: I highly recommend A Brief History of Intelligence. The author is a tech entrepreneur who traces the origin of life on Earth all the way to GPT-4, summarizing five key breakthroughs in the evolution of intelligence and analyzing the driving forces behind each and their impacts. This was one of my top picks for 2024.

I also recommended it to OpenAI researchers, who found it illuminating after reading it. The book not only helps us understand the evolution of intelligence, but also makes us realize — we may be standing on the eve of the sixth great explosion, or perhaps we've already entered it.

Q: Any other recommendations?

Yusen Dai: There's a more specialized book called The First Eye. It covers the history of the Cambrian explosion. Life had existed on Earth for two billion years, but remained slug-like soft-bodied creatures. Then, over a few million years during the Cambrian period, life suddenly evolved multiple phyla in a massive burst of biodiversity.

Why did this happen? There are many theories — changes in atmospheric composition, seawater chemistry, and so on. But this book proposes the "light-switch hypothesis": certain organisms happened to evolve photosensitive cells, allowing them to perceive light and gain a survival advantage. As photosensitive cells proliferated, true eyes eventually evolved. And when the first eye appeared, the competitive landscape of the entire biosphere transformed dramatically — predators became more formidable, while prey evolved protective mechanisms like shells or greater mobility.

This theory reminds me of where AI development stands today. The release of DeepSeek and other advances makes me feel AI is in a similar "Cambrian explosion" phase. When competition intensifies, everyone must advance rapidly or be eliminated. It's like the Red Queen hypothesis from Alice in Wonderland: "It takes all the running you can do, to keep in the same place."

This competition drives technological progress, accelerating AI development faster and faster. But from an evolutionary perspective, this is both survival competition and the inevitable result of intelligence development.

Q: You mentioned the evolution of intelligence just now — is language one of its components?

Yusen Dai: Yes, language actually emerged relatively late in the evolution of intelligence. It's a highly condensed form of information expression. Current AI is primarily trained on language models because language itself contains extremely high information density.

But this raises a question: If AI becomes truly intelligent, might it reinvent a language of its own, no longer constrained to human natural language? In one of Liu Cixin's science fiction novels, alien civilizations might view human communication through language as profoundly inefficient.

So for now, AI relies mainly on language models, but it may transcend language in the future. AI's thinking speed far exceeds human capability; if forced to continue using human language, it may be limited by the mode of expression. Reviewing the history of intelligence evolution helps us understand AI's possible future directions.

Q: You mentioned reinforcement learning — what role does it play in the evolution of intelligence?

Yusen Dai: The book also explores the origins of reinforcement learning, analyzing it through extensive evolutionary biology case studies. I think this research offers significant inspiration for the AI field.

Q: Thank you so much, Yusen, for joining us. Today we started with the two key milestones of o1 and R1, discussing their impact on the AI landscape and the transformation that followed. In 2025, we may see more PMF breakthroughs for AI agents and more "Lee Sedol moments" to come.

Yusen Dai: Thanks for having me. I too look forward to AI's development in 2025. We're still on day one of the great AI intelligence revolution — there will certainly be more surprises ahead!

Recommended Reading