ByteDance Is Dancing, Alibaba Is Restless, the AI Six Little Dragons Are Shaken — A Year-in-Review of the Foundation Model Wars with *LatePost*

The "Crossing" year-end review series has finally arrived at the most closely watched topic in AI: large language models.

Just as we were planning how to summarize the development of LLM companies in 2024, LatePost published an article titled "China's LLM Survival War: Giants Circle, Startups Struggle", which chronicled how the six companies dubbed the "AI Six Little Dragons" went from being celebrated and closely watched at the start of 2024 to facing a dramatically shifted landscape — ByteDance made aggressive moves into large models, Alibaba actively deployed capital and already possessed its own Qwen LLM, and the six dragons each pursued different strategies amid personnel changes. The competitive dynamic became one of leading AI startups represented by the six dragons versus tech giants.

The article attracted widespread attention, going viral on WeChat Moments and Jike the day it was published. By the time we prepared to release this podcast episode, the original article had accumulated nearly 70,000 views. In our view, it offered detailed annotation and analysis for understanding Chinese LLM enterprises in 2024. So this week, we've invited the article's editor, Manqi Cheng, deputy editor-in-chief at LatePost[1], to use the article as a through-line for discussing LatePost's observations while writing it, plus additional "front-row impressions" of LLM companies.

Additionally, just as we were preparing this episode, Alibaba's open-source model Qwen released a new reasoning model. Both Qwen and DeepSeek, another Chinese company's LLM, were evaluated by foreign media as rivals to o1. We've also supplemented the episode with some information about Qwen and DeepSeek.

This episode is also a crossover between "Crossing" and "LateTalk[2]," with the content published on "LateTalk" as well.

Listen on WeChat:

Listen on Xiaoyuzhou:

China's LLM Survival War: Startups Struggling, or Opportunities Coexisting?

🚥 Koji

Before we begin, let me ask Manqi to introduce the main content of the article for those who haven't read it yet.

👩🏻 Manqi Cheng

Hello to all Crossing listeners. Since I've mostly played the question-asking role on podcasts before, I may share more thoughts today. If anything is inaccurate, please feel free to add corrections in the comments so we can form a more complete discussion.

Koji's observation is quite interesting, because ByteDance's English name literally contains "Dance," so "the elephant is dancing" — this metaphor comes from an investor's summary in the article, and I'm grateful for that observation. The article's theme is exactly what the headline says: startups struggling, giants circling.

After reflecting on this article, I think one core change is:

In this short span from last year to this year, the advantages that startups once prided themselves on — "organizational agility," "stronger technical teams" — seem to have become less pronounced under ByteDance's forceful catch-up.

ByteDance's rise has been extremely rapid, and the root of this phenomenon lies in the speed and direction of technological evolution.

Current market views on LLM entrepreneurship are sharply divided. Some remain bullish, while others have shifted from previous optimism to pessimism. The root of this divergence also lies in judgments about the pace of technological evolution. This is an important framework we can expand on in subsequent discussion.

If technology evolves faster, forward-looking technical judgment becomes particularly important, and this may be precisely where startups have an advantage. Because certain technical insights often reside with a small number of people, some startups may see technological shifts earlier due to the presence of a key individual. Even if Chinese companies aren't global leaders, being able to follow faster is itself an advantage.

The Information had a report yesterday about ByteDance's LLMs, mentioning that they encountered some difficulties in internal model training and even sent a team to Singapore to investigate why progress on math and reasoning capabilities wasn't going smoothly. By contrast, smaller teams like Moonshot AI and DeepSeek seem to have made faster progress on reasoning improvements. After OpenAI released the o1 direction in September, within just two months, Moonshot AI released the math model k0-math, and DeepSeek released r1-lite, both emphasizing stronger capabilities in math and reasoning.

On the same day we published the "LLM Survival War" article, we also published an exclusive: a researcher from Huawei's Noah's Ark Lab had joined Moonshot AI, specifically to lead exploration in the o1 direction. Although ByteDance also tried to recruit this person, he ultimately chose the startup. Additionally, we learned earlier that at an internal meeting early this year, Zhilin Yang mentioned that one of Moonshot AI's technical goals for the year was to achieve multi-stage reasoning. While I'm not certain whether his "multi-stage reasoning" at the time referred to what is now the o1 direction, the literal meaning seems somewhat similar. Because o1, through chain-of-thought approaches, can think step-by-step more like humans, potentially producing more complete or deeper results. Whether they saw this direction earlier, I'm not sure.

🚥 Koji

Just now Manqi mentioned that if technology is still evolving very rapidly, startups may have an advantage. Because relying on a few genius-type individuals, it's quite possible to make rapid breakthroughs and pull ahead.

After OpenAI released o1, I noticed that a domestic team called Monica[3] released their own technology Steiner-preview[4]. They have a young prodigy on their team, Peak Ji — if you know about him, you'd know he created the Mammoth Browser while still in high school, which became quite influential globally. This Steiner-preview is also Peak Ji's work, and you can now use this reasoning model inside Monica; it's a capability they developed internally.

Seeing such breakthroughs, I was genuinely surprised. If breakthroughs like this keep happening, small teams and small companies do have a chance to seize such opportunities.

👩🏻 Manqi Cheng

You're right, Peak Ji is indeed a prodigy — he made Mammoth Browser at 17 and started his entrepreneurial journey very early.

Returning to the earlier topic, if technology evolves faster, I think two important effects emerge:

First, startups may have advantages in technical judgment. Second, the integrated "model + application" model becomes more valuable.

This is a direction that many LLM companies have been emphasizing from last year to this year. Because model capabilities evolve faster, the response in product experience also becomes faster. If you control both links simultaneously, you can theoretically coordinate and optimize the system more quickly.

But if development moves in the other direction — that is, model capability growth slows — three situations may arise:

1. Experience differences brought by the model itself alone become less pronounced. As users, we can already feel this — the experience differences between Doubao, Kimi, and Tongyi are not that significant. In this case, product capability becomes more important. This is precisely where big companies have an advantage, because they've already accumulated product development infrastructure. More importantly, they have traffic and distribution channels. Our article also mentioned that ByteDance's efficiency in paid user acquisition may be higher than most startups', and Douyin has not accepted AI product advertising from outside ByteDance since April this year.
2. Those LLM companies with the highest valuations and most funding may face competition from "lightweight" companies. These companies might focus solely on applications, calling the best model APIs globally, or simply fine-tuning open-source models. These companies compete at the product level, and their valuations and baggage are much smaller.
3. If the pace of evolution slows, it's clearly unfavorable for fundraising. Investors' willingness to fund highly valued LLM companies depends largely on betting that they possess the most core technical teams and best talent, expecting them to gain leadership in the technology race. If this expectation comes into question, fundraising naturally runs into problems.

🚥 Koji

Actually, speaking of model-application integration, there's a miraculous company in China called DeepSeek[5]. They not only don't make applications, they've even publicly stated they aren't considering commercialization at this stage, aren't thinking about business models.

Because they believe that right now, all human resources, energy, material resources, and capital related to large models should be devoted entirely to technology exploration. They believe any factor that distracts from technology exploration — whether making applications, exploring business models, or considering early profitability — will dilute and scatter focus.

I think they're truly a magical force completely different from everyone else.

👩🏻 Manqi Cheng

Speaking of that, I recall that when we chatted with Zhilin Yang in June this year, he expressed similar ideas. He said the greatest energy should be placed on technology exploration and intelligence improvement, to pursue the limits of intelligence.

But on the other hand, DeepSeek is quite special — you could say they're also very strong in resources. Because they are their own financial backer — the High-Flyer Quant behind them itself does quantitative trading, with presumably quite substantial returns.

🚥 Ronghui

Speaking of large models, besides the companies called the "Six Little Dragons" or "Six Little Tigers," there's also Tongyi and DeepSeek. Tongyi is Alibaba's product, while DeepSeek is truly a miraculous company. I read some reports about DeepSeek and found that their founder said some quite prescient things back in 2023.

There's a claim (though they haven't directly responded to or confirmed it) that in 2023, there were no more than five domestic companies with over 10,000 GPUs. Besides several leading big tech companies, there was one quantitative fund called High-Flyer. This company possessed 10,000 NVIDIA A100 chips, and its founder was Liang Wenfeng.

He told media at the time that by next year at the latest, both big tech companies and startups would launch their own large language models. He also explicitly stated: no vertical models, no applications, just focus on research and exploration. Recently they've gained excellent international recognition after being covered by TechCrunch.

Open Source vs. Closed Source: Another Battleground for Large Model Development

👩🏻 Manqi Cheng

Speaking of which, there was an interesting piece of feedback after the article was published. A user on GeekPark with the ID "Lingchen 64" reposted it with this observation: When people internationally discuss Chinese large language models, Tongyi's open-source series and DeepSeek get mentioned most frequently. Domestically, the most attention goes to Kimi and Doubao.

This topic actually touches on an important factor that's currently putting pressure on startups: the pressure that open source exerts on closed source.

The leading large model companies that have raised substantial funding are mostly telling a story of "I have a better closed-source model." But with Meta and Alibaba — two major Chinese and American companies actively pushing open source — along with distinctive companies like DeepSeek contributing many excellent open-source models, many companies can now build directly on top of open-source foundations. This may create competition for large model startups and put pressure on them.

Returning to the core question: the speed of technological evolution. This is perhaps the hardest question for the market to answer. First, it exceeds most people's understanding — certainly mine. Even among practitioners, many top AI developers hold differing views on this, giving off a "watching gods battle" kind of feeling. I think this is precisely the core reason for the significant market divergence right now. The phenomena we describe in our article, including potential future changes, all stem from this.

🚥 Koji

When you were working on this article, what was your starting point? Did you see some specific signal at the time, or hear a particular story that made you want to produce this piece?

👩🏻 Manqi Cheng

Actually, the starting point was quite conventional. As business reporting goes, the main threads typically fall into a few categories: one is coverage of individual companies or figures, another is industry phenomenon pieces.

When we first planned this article before the October holiday, the main direction was to take stock of how China's leading large model startups had changed over the past two years, from last year to this year. At that point, we didn't have a particularly clear focus on their competition with big tech companies. It was only through conversations with more people, including reviewing extensive interview notes from last year to this year, that we gradually realized this might be a through-line that could connect many events. And this angle is relatively accessible for readers — fundamentally, it's a competition story. As you mentioned with specific cases, Doubao has indeed been making quite a splash lately.

🚥 Ronghui

Since publishing the article, have you received any feedback? Any comments from people, perhaps corrections to certain viewpoints in the piece, or did people generally agree with what the article presented?

👩🏻 Manqi Cheng

I'd like to ask first — what was your own feedback? How did you feel after reading it?

🚥 Koji

I reposted it on Jike immediately. Because this article distilled ideas that people had been harboring, even some consensus views, and expressed them in a very clear and forceful way. I think it crystallized consensus and formed a new narrative framework.

👩🏻 Manqi Cheng

Your repost even referenced The Fair's slogan, right? I remember it had something about "the direction of the tide" (laughs).

🚥 Koji

Yes, "the direction of the tide" (laughs).

👩🏻 Manqi Cheng

Right, we're more in the business of documenting — it's hard to say we can change the direction of the tide. I think describing these phenomena makes for an interesting retrospective.

There are a few points worth noting: First, as you mentioned, some content did resonate with people — there's shared sentiment; another is the China-versus-international attention difference I mentioned earlier, which is quite interesting. Our article didn't really go deep into the open-source versus closed-source question, which is also a topic very much worth exploring. Third, as you noted, some readers did feel certain content in the article wasn't entirely accurate and needed correction.

For example, regarding the QuestMobile 30-day retention data we used — we later pinned a clarification in the article comments. The main质疑 about these numbers was that they seemed too high: Doubao's retention in September was 34%, while Moonshot AI's Kimi and MiniMax's Hailuo AI were around 28%. For chat companionship or virtual social products, Maoxiang was in the 40s, while MiniMax's STARFIELD reached over 60%.

The reason for this discrepancy is that product operators typically focus on new user 30-day retention. QuestMobile's methodology calculates the proportion of users still active on day 30 relative to total users, computing this daily and averaging the results.

This data includes all users, not just new user retention, so it's generally higher than new user retention rates. Therefore, this dataset is better suited for looking at relative relationships between products, while the specific figures are indeed higher than the new user 30-day retention that investors or product managers in the market typically focus on.

Another interesting piece of feedback concerned the article's lead, which mentioned that "a big tech executive said the annual minimum spend for building large models is $2 to 3 billion." Shortly after publication, an investor told me that actual spending isn't that high, and provided his own calculation method.

🚥 Koji

Right, regardless of the specific numbers, the minimum spend for large models is indeed very high. This trend is also evident in overseas markets. Beyond the five companies — OpenAI, xAI, Meta, Google, and Anthropic — other companies are struggling to keep pace with the development rhythm. Even Cohere and Mistral seem somewhat out of their depth.

The second point mentioned earlier was the China-international attention gap. On one hand, overseas attention is on Qwen and DeepSeek — these past couple days, there have even been major Twitter accounts saying: "Look, the chip blockade isn't working, China is incredibly strong in AI." On the other hand, the Financial Times published an article in mid-November, but from a very different angle. They said the chip blockade indeed hasn't worked, and the reason is that Chinese companies, from Alibaba to ByteDance to Meituan, have all set up offices in the Bay Area and are heavily recruiting top talent. This shows that in the face of technological waves, many elements remain fluid.

Returning to this article — the opening is very engaging, describing last year's scene of investors scrambling to invest in large model startups. By this year, various rumors suggest some are already getting anxious to sell their existing shares.

Looking at the "Six Little Dragons" situation from early this year — Manqi, do you remember any particularly striking events or stories from that period?

👩🏻 Manqi Cheng

Right, "early this year" here should refer to last year. Large model entrepreneurship was indeed very hot then. Speaking of stories from that period — I know Zhilin Yang was relatively difficult to meet at first; investors found it hard to get meetings with him. Though this probably wasn't directly related to how hot the company was, but more due to his introverted personality.

According to his colleagues, Zhilin Yang once proposed:

Could we not meet with investors? I'll write a document, and they can just read that.

🚥 Koji

I think their recent press conference was brilliant. You could call it a clever PR presentation — Zhilin Yang spent over 40 minutes on stage solving math problems by himself, showcasing his image as a scientific researcher in full view of all the media, and反而 received very positive reviews and goodwill.

By the way, Manqi, weren't you there in person? Did you have that impression?

👩🏻 Manqi Cheng

I was there too. First, let me correct something: he didn't spend 40 minutes doing math problems, it wasn't that long.

Zhilin Yang spent 40 minutes total explaining k0-math, and the portion where he demonstrated math problems during the demo — my sense was about ten minutes or so. You can see some interesting shifts in the headlines from media coverage afterward: from "responding to everything" to "not responding to everything," to "unable to respond to everything."

Actually, there was a general feeling that he didn't address some market concerns beyond technology. This may have been related to the arbitration controversy reports that came out on Monday of the same week as the press conference.

🚥 Koji

Going back to early last year, what other stories left a deep impression on you?

👩🏻 Manqi Cheng

I'm particularly struck by stories from this recent period of hesitation, probably because it's more recent. Funding news has clearly slowed — even Alibaba, which was most generous in the first half of the year, started showing some hesitation by year-end.

Didn't Xiaohongshu also have rumors that an investment in a certain large model company fell through?

🚥 Koji

We saw on Xiaohongshu that it seemed to be StepFun.

👩🏻 Manqi Cheng

Right, that investment hasn't happened yet. Funding news is indeed not as fast or explosive as it was early in the year.

🚥 Koji

There's a story widely circulating online recently: Jack Ma reportedly asked at a meeting, "Why do we have to invest in every large model startup?" Everyone present looked at each other, no one answered. No one directly told Jack Ma, "It's so we can sell Alibaba Cloud." Jack Ma pressed further: "Is it because you don't understand?"

This story raises a question: From when did large model companies, especially this batch of startups, begin to seem less successful than imagined?

👩🏻 Manqi Cheng

Actually, it's not that development has been unsuccessful — every new technology goes through a process from frenzy to rationality after emerging.

Initially everyone is excited, but after the hype dies down, things naturally return to calm — this is normal. So rather than saying they're unsuccessful, it's more that they appear to be developing slowly compared to their own expectations and the performance of big companies.

From last year to this year, there's been an interesting twist:

ByteDance's large models and products, which were being collectively mocked last year, have performed quite impressively this year.

🚥 Koji

This reminds me of that episode between Crossing and ZhenFund partner Yusen Dai. The podcast title was "Large models are still elementary school students, please don't rush to put them to work and make money."

It's only been two years since ChatGPT 3.0's release, yet people seem to expect it to have already spawned the next Meta or Google. Looking back at these tech giants' development: Both Meta and Google went through six to ten years of growth before becoming remarkable enterprises in terms of revenue scale and profit levels.

The reason people think large model development is "unsuccessful" is mainly due to excessively urgent expectations. From another perspective, this isn't unsuccessful development — it just hasn't met people's inflated expectations. And whether those expectations themselves are warranted is debatable.

👩🏻 Manqi Cheng

It wouldn't be normal to be blown away every single day, right? People can't be shocked every day.

Taking Stock of the Six Little Dragons: Each Company's Focus and Progress

🚥 Koji

Manqi, when writing this article, you presumably did a comprehensive review of the current "Six Little Dragons" — six typical large model startups. Could you share what you've observed regarding each company's current development focus and progress?

👩🏻 Manqi Cheng

On Moonshot AI, I think their direction is pretty clear right now — they're focused on making Kimi the product work, and they're consolidating the brand around Kimi too.

These days when Moonshot AI refers to itself externally, it often uses "Kimi" rather than "Moonshot AI." Here's a fun fact that a lot of people probably know: Zhilin Yang's own English name is Kimi. So now the founder, the company brand, and the product are all called Kimi. Because of the name collision, to distinguish whether someone's talking about the founder or the product, they call Zhilin Yang "KK" internally, and say "Kimi" when they mean the product. Their biggest strategy, I think, is to nail Kimi and the productivity use case at this stage.

On MiniMax — they actually had products out before ChatGPT. Their first hit, Glow, launched in the second half of 2022, and they've tried quite a few directions. Right now their consumer-facing focus is definitely on Talkie, the international version of STARFIELD. At the same time, they've done a lot with their open platform, the API-to-B part, which should be a pretty important revenue stream for them.

01.AI has had a lot of adjustments this year too, and I think things are clearer now. One direction is consumer products for international markets. I saw someone on Xiaohongshu complaining that 01.AI doesn't have any products — that's actually a bit unfair to them, they just don't like talking about their product names. Since they're going international now, and they're a fairly large company, there's definitely some pressure there, but they actually have quite a few products. They recently launched two AI search products, which may be where their consumer focus is for the near term. They also recently spun off a gaming company that will be led by one of their co-founders, Ma Jie.

We reported exclusively before that 01.AI is doing B2B too, adding some new features on top of their existing API. APIs are pretty much the standard approach for most large model companies doing B2B. They recently held a B2B launch event in the China market, offering more complete solutions for specific scenarios. Though they don't necessarily do everything themselves — they partner with ISVs (independent software vendors) to keep things lighter and more standardized, mainly targeting marketing, livestreaming, and similar scenarios.

🚥 Koji

Right, I saw him putting heavy emphasis on virtual human livestreaming. In a recent episode of our "AI in China" podcast series, we found a founder who achieved PMF in the B2B space — which is pretty rare. That's Shijie Li of Jisi Technology, who mainly does virtual human livestreaming. That's definitely a direction where the use case and demand are pretty clear.

I just have some slight doubts — the total market size is probably decent for a startup going from zero to one, but for 01.AI, I'm not sure it can carry such heavy responsibility and Kai-Fu Lee's ambitions. I had some questions when I saw that.

👩🏻 Manqi Cheng

This actually brings us back to what we discussed at the beginning. If you're doing detailed products for specific scenarios, there are plenty of companies that don't make models at all — they don't have that many resources, and they don't face that many expectations, so they can operate very lightly and nimbly. These large model companies end up competing with those companies in specific scenarios. That's roughly the situation with 01.AI's B2B business.

On Zhipu AI, B2B and government have always been their focus. From what we understand, Zhipu AI's contract revenue this year is about the same as last year's. On one hand that might seem like no growth, but on the other hand, last year when they were doing customized model development, the per-project pricing was very high. Last year they had several major financial industry clients in China placing orders of 10 million RMB, while this year the same type of project might have dropped sharply to 1 million or even less. From that perspective, Zhipu AI may have expanded to more quality clients and revenue streams — I'm not entirely clear on the specifics.

🚥 Koji

I feel like Zhipu AI is putting enormous emphasis on talking about their AutoGLM right now. After the first release, they definitely got strong market feedback. Then just this week, they held another launch event, introduced a more formal version of GLM, and produced a very polished video. I think this is a major direction for them now.

👩🏻 Manqi Cheng

This is the one that can work with phones, right?

🚥 Koji

Right — the phone that appeared in the ad at their recent launch event was HONOR's Magic 7 series.

👩🏻 Manqi Cheng

I think they probably will have a partnership there — the market can look forward to that.

👩🏻 Manqi Cheng

There's also StepFun — relatively less information about them. Though they have made some consumer applications, which you can see from public information, like "Yuewen" and "Maopaoya" and such.

🚥 Koji

Speaking of StepFun, they've actually had some viral consumer hits recently. One is "Lyric Remix Machine," and another is "Soul Extractor" that they just released a couple days ago. What's interesting is that at the time, they emphasized that these products just used StepFun's API, without highlighting that these products were actually made internally — a pretty subtle strategy.

It's probably because they want to position themselves as a B2B company long-term, not a consumer company. So even when they make viral hits, they want to manage expectations — don't think StepFun is pivoting to consumer now — and they don't even let the growth effects of these viral products accrue to themselves.

👩🏻 Manqi Cheng

StepFun does have some very impressive talent. Xiangyu Zhang is there now — previously with Jian Sun, Shaoqing Ren, and Kaiming He. The ResNet they developed was probably the most influential achievement globally in China's computer vision field. Yibo Zhu is also at StepFun — he returned from Meta in early 2023.

Then there's Baichuan in the six dragons. The changes there are visible to everyone. Last year, Chuan Wang probably wanted to build a consumer super-product, but now their direction is clearer — focused on the healthcare scenario. Objectively this does avoid some competition with the giants, not fighting on the most contested battlefield. But healthcare itself has pretty high barriers and complex commercial relationships.

We're particularly focused on Moonshot AI and MiniMax — I think the comparison between them is quite interesting. As I mentioned, Kimi has a relatively clear through-line. When Zhilin Yang chats with us, or communicates externally, his logic is very consistent. He says they're going after the productivity scenario and pursuing maximum intelligence improvement. He believes productivity scenarios have a stronger relationship with IQ improvement, while entertainment and social products don't demand as much intelligence.

Productivity is indeed considered a high-value scenario. ChatGPT has somewhat validated this judgment. By end of August this year, OpenAI released official data showing ChatGPT reached 200 million weekly active users. At the beginning of this year, its WAU was around 100 million, so it doubled in under a year. In a sense you could say it's already the major product we've seen in the AI space.

🚥 Koji

Pretty incredible. And that doesn't even count all the wrapper apps — that's just their official client.

👩🏻 Manqi Cheng

Right, just their own numbers. I think this is a very valuable direction, and Zhilin Yang's ambitions are substantial. I looked back at our June interview — when he was talking about productivity products, he said: "It will eventually truly become a general entry-level product." This was in explaining why they're only doing Kimi — you can see he envisions finding an entry-level product for the future.

Of course, there's an obvious problem with this strategy: you're thinking this, but so is everyone else — isn't ByteDance thinking the exact same thing?

🚥 Koji

That's why Doubao is going all out.

👩🏻 Manqi Cheng

Right, because Flow is ByteDance's AI division and it has many products. But Doubao is clearly where they're putting the most resources — not just in the advertising spend that everyone can see, but also in organizational and personnel arrangements. Allen Zhu, previously founder of Musical.ly, is now head of all of Flow, and the product he directly oversees is Doubao. Doubao's web lead, app lead, and others all report to Allen Zhu. From investment to organizational structure, ByteDance clearly values Doubao very highly.

So the direction Kimi chose puts it in direct competition with ByteDance, or with the major companies' must-win territory. MiniMax's product strategy, I think, is more flexible — they'll try many directions.

Whether Glow's initial viral success was accidental or a deliberate strategy, looking at the results now, the directions they're pursuing aren't direct competition with the major companies. In products like STARFIELD, though ByteDance also made Catbox, Catbox's performance is currently not as good as STARFIELD's.

We have some specific data on this in our article. The product feedback is quite interesting — earlier this year I even made a short video about this. I think one reason Kimi became so popular is that its target users happen to be exactly the people most focused on AI — investors, analysts, journalists, these knowledge workers. When these people are both AI-focused and find the product genuinely good, it creates that wow factor. And these people are also the most vocal on social media — on Xiaohongshu, Jike, or Moments.

I experienced something quite amusing myself. During Qingming Festival this year I went hiking — organized by a VC friend, with many people who didn't know each other. We started introducing ourselves, and someone mentioned they'd been using Kimi on the way over. The self-introduction session turned into a Kimi usage sharing session, with everyone sharing how they use Kimi and how great they find it.

In March, April, and May, this organic word-of-mouth was genuinely very strong. Though partly because Doubao may not have been as strong then. STARFIELD's users are particularly interesting — I downloaded STARFIELD myself but couldn't really get into it, because I'm not the target audience. I think most investors probably aren't either. But if you look at these users' social media shares, you find they're actually deeply engrossed in this product.

There used to be lots of "raising digital kids" videos for Glow on Bilibili. You might click in and not even understand what the video is about, with comments full of what seems like insider slang. I've also seen people on Xiaohongshu sharing their emotional experiences raising kids on STARFIELD. These user groups have formed many small-circle discourses that are sometimes hard to follow. Those posts might have 8,000 likes and over a thousand comments — you can feel these users' emotional investment and stickiness are very strong.

This shows in the Day 30 retention data too. Though Catbox's retention isn't as good as STARFIELD's, it still has over 40% monthly retention — higher than Doubao. STARFIELD's own retention reaches over 60%.

🚥 Koji

So looking at it, these two companies have taken very different paths. Since you've both done fairly in-depth interviews with the two founders, do you think this is directly related to their personal growth or their core beliefs?

Manqi Cheng

First, I should say that after interviewing them, you probably have a more complete picture of who they are. I can only share my impressions. I think MiniMax founder Junjie Yan genuinely holds a certain conviction, which became the company's early internal slogan — "Intelligence with Everyone." He believes that if you're building an AI product, it should be accessible to the vast majority of people, so the barrier to entry needs to be very low.

In our interview, he said that people who type every day are essentially "everyone here in this room" — meaning knowledge workers or office professionals. If you want the product to reach more users, you need to add voice capabilities, or visual and other multimodal features. That's why MiniMax started working on these models from the beginning. And among domestic companies, their voice model is actually quite good — Koji, you might have more to say about that. I'm not sure if you've tried their products.

🚥 Koji

I think their video model has more buzz. Recently, various entrepreneurs and friends who've tested the major models, including my own experience, all agree that MiniMax's video model performance is genuinely outstanding. But there's a notable point here — MiniMax's video model team reportedly has only about a dozen people.

By the time this podcast episode is released, Tencent's Hunyuan video foundation model will have officially launched. I was also amazed during the beta testing — very high quality. ByteDance also has two teams working on video foundation models, fairly low-key without much PR, but people who've used them say they're good. So with giants like ByteDance, Tencent, and potentially Alibaba joining the fray, startups are indeed in for a rough time. The key is how to quickly open up a bigger market while still leading.

Manqi Cheng

What I just described was Junjie Yan's thinking. Actually, as this thinking develops further, it will also face competitive pressure — it's a matter of timing windows.

Zhilin Yang's thinking, at least what he expressed to us, is that he wants to build a product that significantly elevates intelligence, in a scenario that promotes skill improvement, while also having users. I think this resembles something like "climbing the technology peak while laying out commercialization along the way." Though as market conditions change, they may both adjust.

About these two people, I have a somewhat interesting observation, though it may not mean much. Some investors believe Zhilin Yang is quite decisive about direction — once he's set on the general direction, he'll push forward firmly. His WeChat signature is "A million different people from day to next", which in Chinese would roughly be "I change unpredictably from morning to night." This is a rock lyric, because he really likes rock music. When we asked him about this WeChat signature, he explained that "the long-term goal doesn't change, but I need to iterate rapidly."

MiniMax's product strategy appears more flexible, more adaptive. Meanwhile, Junjie Yan's WeChat signature is "To find more stable formulas." I find this forms a subtle contrast, though again, this may not mean much.

🚥 Koji

Pretty fun. I think recently we've also been doing some product experiments ourselves. One of them is having users upload their own Moments posts and various screenshots from their Moments feed, from which an incredible amount can be interpreted.

Manqi Cheng

Speaking of which, that suddenly reminds me — sometimes when you join a new social platform, you need to come up with a personal bio, and it would be nice if AI could help generate one.

🚥 Koji

We've recently discovered that even if a user only uploads one screen's worth of Moments screenshots, AI can fairly accurately infer many of this person's personality traits, even determining whether they have secure attachment or anxious attachment.

We're trying to develop a product that uses AI to help users better understand themselves, thereby helping with dating and finding suitable partners. After we built this demo, everyone who tried it was amazed. We'll officially release it in some time — welcome everyone to come try it out.

Manqi Cheng

But I can already imagine that after this product launches, people probably won't use it to understand themselves — they'll use it to understand other people. Like checking what's up with their crush by screenshotting their Moments and analyzing it; or figuring out the personality of someone they want to pursue; or even seeing what their boss's personality is like.

🚥 Koji

That's fine too, actually. It's essentially helping everyone improve their ability to understand others, to see others. If everyone's ability in this area improves, the world actually becomes more emotionally valuable.

🚥 Koji

Here we'd like to preview a very meaningful project that The Fair and LatePost will soon be releasing. We've united 20 entrepreneurs and cultural figures with considerable influence in the industry, preparing to jointly release a manifesto about AI, exploring how AI will help humanity and our society. This is a somewhat grand proposition, though it makes us seem a bit presumptuous. We've laid out ten key points that AI should achieve, and will release this manifesto together. We'll also introduce it in detail on the podcast when the time comes.

Manqi Cheng

A preview is good — once you preview it, it's set in stone, absolutely can't be delayed (laughs).

🚥 Koji

Alright, let's come back to our topic.

🚥 Ronghui

Recently we've noticed a phenomenon: many of the new batch of entrepreneurs announcing funding are people who left from those model companies founded last year. What do you think this phenomenon indicates?

Manqi Cheng

Right, we've written about this in some exclusive reports. For example, at Moonshot AI, some product leads left to start their own companies. They had previously been responsible for two overseas products: one called Ohai, and one called Noisee.

This is a good time to add: earlier we mentioned that Kimi only focuses on the Kimi product, but actually in this process, they've also tried some other product directions. But from what we understand, they didn't invest particularly heavily in these products, and by now these products may no longer be in active development or receiving continued investment.

Earlier still, they also had a product lead named Wang Guan who left to start a company, including Leon Ming, who we previously reported on as the Noisee lead. Later I found out he's very young, born in 1998. And his project reached a $50 million valuation in its first round — very sought-after by the market.

So I think this at least shows that right now, for AI applications, there are still quite a few people in the market willing to support them, and willing to support them at fairly high valuations. This also circles back to what we discussed earlier: if people believe that model evolution itself isn't happening so fast, they'll pay more attention to whether applications have opportunities, and will use lighter-weight approaches to build applications.

🚥 Koji

I think the things this batch of entrepreneurs from large model companies are doing are all quite interesting. For example, Wang Guan's one2x is doing AI video-related experiments, Dapeng remains firmly committed to productivity tools, while Leon Ming is doing AI coding, with a valuation reaching $50 million.

Manqi Cheng

The day Zhilin Yang released the k0-math math model, in the subsequent group interview, a reporter asked him about talent attrition. His response was quite interesting — he used the classic Zhihu answer format: "First ask if it is, then ask if there is." His meaning was that the company didn't have talent attrition, but rather because of business adjustments and no longer doing overseas products, they needed to focus their energy, so some people left. This was a relatively euphemistic way of saying "we don't really need these people right now." Overall, it's that some large model companies are adjusting their business, leading to personnel changes.

These people leaving to start companies, whether by their own choice or due to company adjustments, are all very well-received by the market. You can see this from their funding situations and investor attention. For example, Leon Ming coming out to do AI programming, first-round valuation reached $50 million. If you compare this to the hard tech entrepreneurship wave in 2015–2016, this valuation is quite high.

Wang Guan also raised funding not long after leaving. These founders are all relatively young. I think this phenomenon itself is also connected to the core hypothesis we discussed at the beginning — the question of how fast technology is developing. Some investors may have judged that the explosive growth of models has reached a certain stage and relatively slowed, and now perhaps there's more opportunity on the application side.

🚥 Koji

Let's go back to a number mentioned in the article: $2 to $3 billion. This was mentioned as the annual burn threshold for large model companies right now.

Earlier Manqi also said that after this article came out, you received some differing commentary on this number. Could you elaborate a bit here?

Manqi Cheng

Regarding estimates of investment scale, big tech executives believe $2 to 3 billion is needed. But one investor's calculation is that these startups need at most $200 to 300 million per year. His calculation basis is: a single model is currently in the tens of millions of dollars range, renting 10,000 H-cards for a year is only tens of millions of dollars. And training a model doesn't take a full year — typically a few months is enough. By this calculation, it's indeed far from $2 billion.

Of course this calculation has some premises: first, it refers to companies that continue doing model training, otherwise training costs don't need to be counted at all; additionally, regarding lower training costs, Kai-Fu Lee also mentioned that when they do model training, they achieve results comparable to xAI at 1% to 2% of the cost. But what may be ignored here is a cost: when there are no pioneers clearing the path, the cost of forward-looking investment and exploration can be very large. It's like running a marathon — when someone's ahead scouting the route, those behind do save some costs.

We can reference The Information's reporting on OpenAI's spending. In 2022, before it became hugely popular, OpenAI spent $540 million that year. A June report this year showed that OpenAI may spend $7 billion this entire year, of which R&D personnel costs alone are $1.5 billion. Of course, AI R&D personnel in the United States may be particularly expensive — that article's headline was "Why OpenAI could lose $5B this year."

This brings us back to the dilemma facing Chinese large model companies.

Under China's market capacity, it's very difficult to raise enough funding to pursue technological limits without worries — impossible to invest as much money as OpenAI. So you must have some commercialization and growth results. But simultaneously doing well at technology, product growth, and commercialization is itself a very difficult thing.

🚥 Koji

So this brings us back to what we discussed at the beginning — looking at Qwen, looking at ByteDance's models, and looking at DeepSeek. These players that step outside the Six Little Dragons, outside the startup framework — whether giants or those themselves holding major resources and substantial wealth — what kind of different forward-looking results they will produce representing China.

Manqi Cheng

I think at least ByteDance and DeepSeek clearly have this ambition at the model layer.

🚥 Ronghui

As Manqi just mentioned, some challenges are inherent to the models themselves — for instance, they're resource-intensive and costly. Many things in China may be difficult to benchmark against what's happening in Silicon Valley.

Beyond these factors, what organizational and human-resource challenges do startups face — and here we're specifically referring to these model companies?

👩🏻 Manqi Cheng

We have a lot of fragmented information about these organizations and people, but it didn't feature heavily in the article. From the perspective of early-stage employees, it's perfectly normal for a founder's priorities to shift rapidly — growth today, revenue tomorrow, this initiative yesterday, that pivot the day after.

These companies are juggling multiple considerations in their current situation, not just one. First, high valuations are a bottleneck — they themselves make subsequent fundraising more difficult. Second, the direction and pace of model capability evolution remain unclear, and product experience and growth narratives aren't compelling enough. Meanwhile, the tech giants are going all-out on product. On this point, I specifically used Doubao and Kimi today to look up Toutiao's historical numbers, since that data is hard to search for.

Toutiao launched in August 2012. Comparing its trajectory to what we've seen from large model products over the past year: by end of February 2013, seven months after launch, its DAU had reached 2 million; by end of 2013, DAU hit 10 million. Using this as a benchmark against Kimi among current startups, Toutiao's growth as a startup product was indeed much faster back then.

And we have to consider that traffic acquisition costs in 2013 weren't as high as they are now — they're significantly more expensive today. It was still the early mobile internet era. Baidu probably didn't fully grasp recommendation algorithms and may not have been closely watching this startup early on. In retrospect, Toutiao operated in a much more permissive growth environment.

As an aside, Doubao performed better in helping me look up Toutiao's historical DAU data. Both Doubao and Kimi gave the 10 million year-end 2013 figure, but when I pressed further on data sources, Doubao's explanation was more user-friendly. It directly cited which articles contained this information within its generated response. On mobile, you could see the sources without clicking external links — one step saved. Kimi didn't cite sources directly in its generated text, and later it said the year-end 2013 DAU was 1 million. When I asked whether it was 10 million or 1 million, it got somewhat confused.

So these companies currently face challenges requiring them to adjust and balance many factors, and adjusting one factor may impact others. For example, earlier this year there was considerable discussion within large model startups: reports indicated that at least two of the "Six Little Dragons" had halted pre-training.

In retrospect, was stopping pre-training necessarily wrong? Perhaps it was a rational choice. Startups shouldn't fear being proven wrong — among all the consequences this decision triggered, losing face is the least important.

Entrepreneurship inevitably involves being proven wrong, but these companies face practical questions: First, can their valuation hold? Because initially, investors were likely drawn to core technical advantages. Without pre-training, can that valuation be sustained? On the other hand, if they can build genuinely valuable products without pre-training, sustaining a multi-billion-dollar valuation is also fine. But this raises another problem: these teams were originally organized around large model technology as their core — whether this team can actually build hit products is something everyone remains skeptical about. So I think their current situation is genuinely difficult, with various problems all tangled together.

ByteDance: How Does an Elephant Learn to Dance in AI?

🚥 Koji

Hearing all this genuinely makes me worry for the "Six Little Dragons." Still, we want to send our blessings and encouragement — we hope to see all of them thrive. Because their success is tied to the health of China's entire venture capital ecosystem. Every role in this ecosystem gains more opportunities and stronger energy when they succeed, enabling everyone to do better in both their careers and daily lives.

Having discussed startups' challenges, let's turn to ByteDance. A major source of the pressure we've been discussing comes from a company like ByteDance — resource-rich, hardworking, and full of people who know how to "dance." We've seen some observers note that 2023 was the year of runway that ByteDance gave AI startups.

Manqi, based on your reporting and understanding, what was ByteDance doing in 2023? Why do people say it gave everyone a one-year window?

👩🏻 Manqi Cheng

Our article mentions that before ChatGPT, ByteDance's leadership was likely more focused on another AI evolution direction: AI for Science. This direction is certainly valuable — this year's Nobel Prize in Chemistry went to AlphaFold, which represents significant progress in one direction of AI for Science.

Here's an interesting tangent: quite a few major figures, both in China and abroad, have become particularly interested in biotech later in their careers — specifically the intersection of AI and biotech.

From publicly available information, ByteDance's accumulated capabilities in this area lagged behind other major domestic companies. In 2023 we published an article called "Big Tech's Big Models," surveying Chinese tech giants' moves and investments in large models at the time. The article included a timeline marking when Chinese tech giants began working on large models. Baidu was clearly the earliest, releasing a model in 2019 — this became the predecessor to Wenxin. It likely used Google's BERT architecture rather than GPT architecture at the time, but both count as large models.

Moving forward to 2021, Huawei released Pangu, Alibaba released two models called M6 and M6-plus — both later merged into the Tongyi series. SenseTime released its SenseNova large model. Even Tencent, which many perceived as lagging, had released a model before ChatGPT in 2022. They launched Hunyuan in April, and around October the year before last, WeChat itself released a large model called WeLM.

👩🏻 Manqi Cheng

Several points are worth noting about these public moves. One is that Huawei's 2021 Pangu large model was developed in collaboration with Zhilin Yang, who was then still at Recurrent AI, and people from BAAI. This helps explain why capital markets were initially so bullish on Zhilin Yang — his prior experience was directly relevant, and his technical background was genuinely strong.

Another example is SenseTime — MiniMax's Junjie Yan previously worked there. So I think ByteDance's relatively slower response in 2023 makes sense, as their prior technical accumulation in this area likely lagged other companies, and catching up takes time. From what we understand, before ChatGPT, ByteDance's AI Lab had an NLP (natural language processing) group of over 100 people, but only about 10 were working on large language model directions. So I get the sense they were racing to catch up on the technology while also adjusting their thinking.

🚥 Ronghui

Seeing this headline and article, I felt quite moved. It reminded me of a book — Who Says Elephants Can't Dance? — about IBM's transformation. There's a similar plot point: the person who made IBM "dance" was Lou Gerstner, an outsider who withstood enormous pressure to drive reform inside IBM.

When I was a reporter, I once interviewed an R&D leader at IBM China. He mentioned two important principles at IBM. If I recall correctly, Huawei later learned this methodology from IBM as well. One was always putting customers first as the top priority — something Huawei has consistently emphasized. The other was their heavy emphasis on their own R&D centers.

That leader told me something: it's precisely this emphasis on R&D and customers that keeps you continuously aware of what's happening in the market. Beyond organizational agility, this is what matters most.

So seeing this perspective in your writing, I want to ask: based on your understanding and reporting, beyond resource advantages, what has ByteDance done correctly to reach its current position?

👩🏻 Manqi Cheng

This timeline picks up from the previous question. Starting in second half of 2023, they began making changes, with results only gradually becoming visible this year. I think one key factor is the No. 1 person's emphasis on this initiative. This perspective comes from an investor who coined the "China's elephants can dance" metaphor. In our conversation, one insight was that a major factor is founder Yiming Zhang, who starting around late last year became more personally engaged and hands-on. A fairly direct change is that he's been broadly connecting with top researchers and personally recruiting teams.

That investor offered an interesting perspective: a large company's actions at a given stage are often closely tied to the founder's life stage — assuming the first-generation founder is still present and influential. Yiming Zhang is genuinely young — born in 1983. Looking at internet and mobile internet founders' ages: Tencent's Pony Ma was born in 1971, ten years older than Zhang; going further back, Jack Ma is from 1964, Robin Li from 1968. Mark Zuckerberg is even younger — born in 1984, a year younger than Zhang; Sam Altman is from 1985.

Interestingly, only two founders among the "Six Little Dragons" are younger than Zhang: Junjie Yan (born 1989) and Zhilin Yang (born 1993). Of course, age and life stage don't determine everything — this is just one angle for observation.

Yiming Zhang genuinely maintains strong learning and exploration potential. In the current market atmosphere, when he began taking this especially seriously, what specifically did he do? As I mentioned, he invested heavily in recruiting. Zhang has previously expressed that the two most important things in entrepreneurship are: first, accurate judgment and understanding of the situation, and second, finding the right people and team. Most problems can be solved through these two things. He cited this in his conversations with Zhilin Yang.

When we interviewed Junjie Yan in March this year, we found his assessment of the situation was quite similar. He believed that for himself, the most important thing was technical resources, because his product capabilities were already ready, and he had unlimited product resources to experiment with. His most important task was elevating technical capabilities. Later, we could see that in recruiting and R&D, he was indeed pushing in this direction.

After June 2023, ByteDance clarified its strategy: it would build its own large models and products in-house, and stop investing in Chinese large model companies. Before June 2023, they had considered investing in two companies: StepFun and MiniMax. From what I understand, these investments had progressed to fairly advanced stages, but ultimately neither went through. This decision also solidified a clearer internal consensus: focus on doing this well ourselves.

The third important development was the establishment of the Flow department and team, and the broader group did indeed provide this department with substantial support and resources. With this department in place, their approach looked quite similar to what Junjie Yan described: improving technical capabilities on one hand, while developing multiple products on the other. We can see they launched Doubao, Coze, Star Canvas, BagelBell, and Doubao Aixue (which was brought over from ByteDance's education business). Including BagelBell — previously a product under Douyin's Fanqie Novels — everything has now been consolidated here. Of course, besides Flow, other ByteDance teams are also working on AI, which is common at large companies.

Our article also mentions later that large companies might give startups an opening, because if a project goes for a period without clear feedback, a lot of tension can emerge inside a large organization. Things like turf battles between teams, or shifts in strategic priorities.

Speaking of which, something interesting happened recently: ByteDance is suing an intern for 8 million RMB in damages. When the news first broke, I don't know if you saw it, but the initial rumor was that the intern had deliberately sabotaged things because he was dissatisfied with his project not getting enough GPU support. I can't confirm whether this version is entirely true — I can only say this was how the story initially circulated. This person wasn't on Zhu Wenjia's large model R&D team, but rather on ByteDance's commercial technology team, under Liu Xiaobing's organization. The reality is that many departments at ByteDance are experimenting with AI.

🚥 Ronghui

Reading this report, ByteDance seems incredibly powerful — almost scarily so. Does it have any weaknesses?

👩🏻 Manqi Cheng

ByteDance also faces an external variable that we didn't particularly expand on in our article: it's already a massive company, and TikTok overseas — particularly in the United States — is facing some pressure. So pressure from abroad could affect the intensity and manner of their AI investments.

Although ByteDance now has many products overseas, according to their own statements, these international products don't actually use their own models behind the scenes — they use models developed by other companies like OpenAI.

Alibaba Cloud's AI Ambitions: Proxy Wars and Cloud Service Competition

🚥 Koji

While we're talking about ByteDance, there's another very noteworthy major company: Alibaba. They've been very aggressive not just in investing, but their own Qwen large model is also quite impressive.

Manqi, could you share some observations about Alibaba? Particularly any stories and findings from your interviews that left an impression?

👩🏻 Manqi Cheng

Starting from last September, Alibaba Cloud could be described as completely transformed. In the year prior, Alibaba had gone through its "1+6+N" organizational restructuring, and the cloud business had been in constant adjustment. But since Eddie Wu became CEO of the cloud business last September, he established a new direction: "AI-driven, public cloud first." This strategic direction very clearly summarizes Alibaba Cloud's development priorities. Last October we interviewed Zhou Jingren, CTO of Alibaba Cloud — I can put this article in the show notes, it was published on "LatePost." He very clearly articulated how Alibaba Cloud views this AI opportunity.

Alibaba Cloud's core goal is to become the computing infrastructure provider for the new intelligent era. First is the compute layer (IaaS), primarily large-scale GPU clusters. Above that is the model services layer (MaaS, Model-as-a-Service), including model building tools, fine-tuning tools, and so on. They're also building community ecosystems, like the "ModelScope" platform, where users can download various models.

Then there's the models themselves — the Tongyi open-source series. This part has received high marks for industry reputation and attention. According to their disclosed data, by this September, the Tongyi open-source series had been downloaded over 40 million times, spawning more than 50,000 derivative models. This figure ranks second globally only to Meta's LLaMA series, showing considerable international influence. I think Alibaba Cloud wants to build out this ecosystem well.

🚥 Ronghui

It has gotten quite a bit of attention — search on X and you'll see lots of commentary, much of it from influential industry KOLs.

👩🏻 Manqi Cheng

Then earlier today, Koji summarized this as a proxy war between ByteDance and Alibaba, right? Saying Alibaba invested in these companies. We didn't actually use this term in our own article. I saw it in a summary Lan Xi posted on Jike — I thought it was quite interesting.

If we're saying current competition is a proxy war, then I think the future market may indeed have one interesting angle to watch: whether there will be some competition between Volcano Engine and Alibaba Cloud.

I saw Zhu Yahui also mentioned this in his WeChat Moments — our article quoted him there too. His post was quite long, and we may have only quoted his assessment of ByteDance's earlier performance. Later on he actually wrote that at the beginning of the year, he had spoken with Alibaba Cloud executives, and he felt certain that ByteDance would go all-in on large model APIs, and would find ways to make certain vendors have to use their APIs — this is the key to Volcano Engine's rise. He said that Alibaba executive didn't quite believe him, but he sees this as the direction ByteDance will develop toward.

Regarding Alibaba Cloud and Volcano Engine, I've actually been in touch with some Alibaba friends from last year to this year. My sense is that Alibaba Cloud remains quite confident. Because in early 2023, Volcano Engine's compute momentum was quite hot for a while. I don't know if you still remember — they even held a press conference saying how many of China's large model companies were using their compute, and what percentage of models were being trained on their platform.

By this year, their momentum seems to have weakened somewhat. Alibaba people's explanation is: why was Volcano Engine particularly hot in early 2023? Because at that time, ByteDance did indeed have relatively large GPU reserves. And back then, everyone mainly needed to buy training compute, and needed it urgently — so their volume spiked quickly. But as this market develops further, the real bulk of compute demand actually lies in inference. Inference places much higher demands on the number of data centers, the breadth of your regional distribution, stability, and elastic computing capabilities.

Alibaba Cloud people believe that once you enter this kind of formal competition, Volcano Engine — let alone comparing with Alibaba Cloud — may struggle to compete even with Tencent and Huawei. They felt they were the best at the time, that Alibaba Cloud had accumulated many years of experience in infrastructure stability and engineering details. And after the o1 direction emerged, people summarized this shift as: from Pre-train to Post-train, you're looking at Inference Scaling Laws — scaling during the inference phase, where compute demands during inference also rise. I think how the relationship between Volcano Engine and Alibaba Cloud evolves going forward will be something to watch.

Other Tech Giants' AI Strategies: Tencent and Baidu's Approaches

🚥 Koji

Actually, Zhu Yahui also said in his WeChat Moments that he believes ByteDance will definitely follow the path of Google and Amazon. On one hand, they'll capture vendors' advertising spending; on the other, they'll eat into their cloud service spending. So ByteDance Cloud's market share will gradually equal ByteDance's advertising market share.

Besides ByteDance and Alibaba, do you feel anyone else is still at the table?

👩🏻 Manqi Cheng

I think Tencent should count. People may think Tencent wants to win by moving second — it dares to let others go first, and it can afford to. Whenever it chooses to get in the game, it has an entry ticket. Because it truly holds super apps in its hands, with enormous traffic — WeChat's traffic is essentially free.

I find that when observing whether a product or company has become infrastructure, one interesting data point is seeing what method people use to register for new services.

For example, if I'm overseas registering for a new product, I'll definitely use my Google account — everything can be registered and logged in with Google. Domestically, including for new products like Moonshot AI, I basically use WeChat one-tap login. It's already like water and air — infrastructure for the internet world.

🚥 Ronghui

Your internet passport.

👩🏻 Manqi Cheng

Right — your internet passport. So I think Tencent can basically join the game whenever it wants. I even already thought of a headline we could use when we write about this later, at least as a subhead: "ByteDance catches the cicada, Tencent waits behind."

🚥 Koji

So Tencent is the mantis, ByteDance is the elephant. Meanwhile, what have you observed Baidu doing?

👩🏻 Manqi Cheng

What have you observed Baidu doing?

🚥 Ronghui

This year's Baidu World Conference theme was "Applications Are Here." In an interview with Jazzyear, Robin Li said that compared to super apps, what's more important is continuously empowering millions of super-useful applications.

👩🏻 Manqi Cheng

I've had noticeably less information about Baidu's specific progress on large models this year compared to last. In our article, we wrote about some of Baidu's situation — namely that their organization is currently quite complex. For example, at Baidu, the team training foundation models, the team developing To C products using the technology, and the team providing To B solutions using the technology are in three different business groups, not three departments. This makes cross-departmental collaboration and communication quite cumbersome.

So one reason people previously rumored that Baidu had paused pre-training was that when the training team wanted to call on more GPUs, they might encounter difficulty accessing resources. Because the same compute, if deployed on To B business, could generate revenue. If reinvested to train new models, since performance improvements from training models themselves may not be that obvious now, the returns become uncertain.

I find Baidu a genuinely fascinating company. On one hand, it's genuinely technology-driven, with some genuinely forward-looking technical vision. Based on publicly available information, Baidu was the earliest among Chinese companies to work on large language models — it released one back in 2019. And it once had many exceptionally talented people: Andrew Ng was at Baidu, for instance. Among xAI's original dozen members, there's someone named Greg Yang who previously interned at Baidu. And Dario from Anthropic also interned at Baidu for a period. Baidu was once a stronghold for AI talent, which may be connected to how early it invested yet how difficult it found to stick with one direction. On the other hand, I think what may have drawn more attention to Baidu in AI this year is its progress in Robotaxi.

In June this year, the Luobo Kuaipao story really broke into the mainstream — it even made social news, with everyone paying close attention. I specifically went to check the Wuhan citizen message board, which was quite interesting. For ordinary people, whether complaining or expressing anticipation — saying things like "hurry up and open a Luobo Kuaipao stop near my home" — not a single person used the word "Robotaxi." That term probably only appears in media and investment circles.

What's rather interesting is that at the end of 2023, many messages on the board were asking: "The plan said you'd open a stop near my home, why hasn't it happened yet?" — all asking when stops would open. In short, earlier messages were all anticipating "come open a stop near me," while later messages were complaints — about how it was blocking traffic, or monopolizing gas stations so residents couldn't refuel. I find this phenomenon quite intriguing.

🚥 Koji

My hometown is Yongchuan District in Chongqing. Yongchuan was one of the earliest places where Baidu deployed Luobo Kuaipao back in 2021, second only to Beijing, Shanghai, and Guangzhou as a pilot area. So my fellow townspeople experienced Luobo Kuaipao quite early. My sense is that word-of-mouth was genuinely pretty good before that social news story blew up in June this year — many people found it convenient and were calling for "can you hurry up and open one near my home too."

👩🏻 Manqi Cheng

Right, I think this is something where after persisting for so many years, they finally seem to be seeing some change.

Traveling Light: New Opportunities for AI Application Entrepreneurs

🚥 Koji

After discussing the major tech companies, let's talk about some recent observations. Last Sunday, Crossing hosted an event in Beijing, and Manqi came too — we were quite surprised. After the event notice went out, 776 people registered within 48 hours. Ultimately, since the venue was at the Stanford Center on Peking University's campus, which could only accommodate around 200 people, that's all we could host.

Many friends were quite disappointed they couldn't make it, but on-site I genuinely felt tremendous energy. Everyone was incredibly enthusiastic — it felt full of hope. From what I observed there: first, most people were building applications; second, at least over a third, possibly close to half, were building applications for global markets.

👩🏻 Manqi Cheng

My feeling from attending that event was: we should interview more AI application entrepreneurs, people specifically focused on AI applications. That's a direction I think we should pursue more going forward. Because previously we probably talked more with large model companies.

Their situation is genuinely quite different. For example, many people I met on-site weren't even at the full-time entrepreneurship stage yet. Some were still searching for entrepreneurial direction; some might still be at large companies but working on side projects. So this is completely different from the typical VC-backed "go big or go home" entrepreneurial logic where "if you're not going big, you're dead." I think many of them aren't funded at all. I chatted with some people online — they're using their own money, or their product launched and quickly had cash flow, had revenue, operating that way. So I think they're in a "traveling light" state.

🚥 Koji

This is a very significant trend this year. First, with AI's arrival, super individuals or very small teams can now build what previously required $1 million in funding to accomplish — the startup capital needed has decreased. Second, with AI, people can start making money from day one, because paying for AI products in overseas markets is completely taken for granted. Third, with AI, Chinese entrepreneurs can more easily access global markets.

In the past there might still have been language barriers, even some fear around them. But today language simply isn't an issue. And there are many precedents — like Monica's Red Xiao, who says his English, though he passed CET-4, is actually quite poor, yet that didn't stop him from building an excellent global business. I think these factors combined mean more people today choose not to raise funding, instead bootstrapping their way through application development.

🚥 Ronghui

Reading this reporting, one feels Chinese entrepreneurs truly have it hard — blockades ahead, fierce competition behind, and everyone is working so hard. We've received this kind of education since childhood, and in this competitive environment, I think everyone, whether in large models or applications, is essentially exploring different directions toward the same goal. Everyone hopes this direction can develop together.

👩🏻 Manqi Cheng

I think it's hard to say who's better or worse — the logic is just different. For example, these smaller teams might indeed encounter a problem: they build a product whose demand might just be a one-time wave. Like it's particularly hot for a month or two, then nobody uses it anymore. It's not a sustained thing, not a real business — very possibly that kind of situation.

🚥 Ronghui

And although AI applications are booming, they also face a problem. Everyone was expecting 2024 to be the year AI applications exploded, yet by year-end there were many voices saying it feels like there wasn't really an explosion.

👩🏻 Manqi Cheng

Koji, do you think anything exploded? For example, does AI coding count as a small explosion? Does AI search count?

🚥 Koji

I think people's expectations for this were simply too high — as if a second ByteDance would appear tomorrow. With that kind of expectation, disappointment is inevitable. But being on the front lines ourselves, whether doing this podcast or working on our own projects, friends around us are all doing similar things. What I feel is an enormous number of new things springing up like bamboo shoots after rain, very hopeful, with so many new things constantly emerging — still overwhelming, really.

For example, I've been quite struck recently by Recraft. After using its interaction, I thought "holy shit, this is incredible" — Photoshop suddenly feels like something from the last era. If my kid were starting to use a computer for image editing, from day one I would definitely — not me teaching them, probably them being influenced by education or their friends — start with an interaction like Recraft. So natural, fitting human habits for operating images more easily and effortlessly.

👩🏻 Manqi Cheng

Can you talk about this? Is it for desktop, or mobile?

🚥 Koji

It's still desktop-based. First, it's an infinite canvas where you can generate images from text, or do localized adjustments to images. These functions were already possible in earlier WebUI or ComfyUI days, but those really intimidated ordinary people — the learning curve was the highest. Open WebUI and see 100 English parameters, anyone would be baffled. But Recraft not only has strong model capabilities, it's also made the interaction excellent. I've observed that netizens on Xiaohongshu pick it up quite easily too, and the final works are excellent in both quality and creativity.

Then there's Cursor — from day one of using it, I felt electric. I remember that day we had a Crossing Fellowship weekly Sunday online meeting. At that meeting there was a Gen-Z guy named Fanhan, one of our Fellows. He said that after using Cursor, he believed the learning barrier for programming going forward would no longer be programming itself, but overcoming one's own fear. I was genuinely moved hearing this. While that meeting was still ongoing, I downloaded Cursor and started using it — wrote my first demo in ten minutes: I input some random natural language, and it translated it into a string of Emoji expressing that natural language. I think these are all enormous changes.

We previously had a piece at Crossing called "The AI Applications Making Serious Money Overseas, Quietly", where we listed a dozen or twenty, all made by Chinese people. And after publishing that piece, friends kept telling us: "You missed this one, missed that one" — the misses kept growing. Many are low-key, they don't need PR and have no reason to do PR, because they're not fundraising, not targeting Chinese developers or users. So they just need very few people, don't even need to do PR for recruiting — teams of just over a dozen people. Looking at the application space, like how we held one event with 776 registrations, and over seventy people wanted to get on stage for open mic. I feel the ecosystem here is quite thriving — many people see opportunity, and some have already gotten positive feedback.

👩🏻 Manqi Cheng

This is quite interesting. I think this indeed will be fascinating to watch in the market going forward — besides the large companies and the most visible top-tier companies doing both models and applications, how applications emerge will be quite interesting.

🚥 Ronghui

What disrupted Yahoo wasn't its peers, it was Google. What later challenged Google wasn't search engines, it was Meta. The competitors who pose the greatest threat to you are never visible today — that's what's most terrifying.

👩🏻 Manqi Cheng

We've talked quite a bit today about AI software — whether large models or the applications we just mentioned, it's all software. Actually, the combination of software and hardware now, including attempts at new AI hardware entrepreneurship, is also quite active.

Entrepreneurs' Ordeal and Anticipation: Waiting Is Also Part of War

🚥 Koji

Manqi, after finishing this piece, what's your overall feeling about these large model startups?

Do you feel like everyone's working hard to find a way out, or is there an overall sense of struggling against the tide, or are internal emotions quite pessimistic, or the opposite of what we just described?

👩🏻 Manqi Cheng

I think they're definitely working hard to find a way out. At least at the founder level, founders don't have time to be pessimistic.

I think the ideal or typical founder profile is someone who's very alert, very sensitive to danger. But that danger drives more action, reaction, planning. It's probably more excitement than fear.

People who keep choosing to start companies, and choose the kind of direction where "I want to do something big" — they inherently have high risk appetite and achievement motivation. We can observe what kind of exits people are looking for. Beyond technology and product, some companies definitely want to go public earlier. On the business side, including financing channels, can there be some changes? I think everyone will have a lot of thinking and directions to try. So this story is definitely not over yet, which is of course good news for observers like us. We can continue to see what happens in the future, and I'm also looking forward to seeing some different plot twists emerge.

🚥 Koji

Just like that LatePost article quoted a line from Xing Wang: "Most people think war is made of fighting, but actually war is made of waiting and enduring." Although Xing Wang doesn't use Fanfou anymore, the various things he once posted there still pop up from time to time today.

👩🏻 Manqi Cheng

And looking at the backend data, this was the most highlighted sentence, highlighted about seventy-something times. He very incisively and concisely summarized this phenomenon.

🚥 Koji

Alright, let's wrap up here for today. Thanks, Manqi. As this large model adventure continues moving forward, I believe there will still be plenty of waiting and enduring for everyone. In this process, although it sounds cliché, I still hope everyone can enjoy the process, because the process itself is the reward.

👩🏻 Manqi Cheng

Thanks to Crossing for the invitation, thanks for the cross-show opportunity. Bye everyone.

---

Subscribe to the "Crossing" Podcast

🚦 We follow the new industry changes and entrepreneurial opportunities brought by the new wave of AI technology. "Crossing" is Steve Jobs' metaphor for Apple — standing at the intersection of technology and liberal arts, where great products are often born. AI is bringing change to all industries. We seek out, interview, and bring together "active actors" of the AI era, and together with them, explore and embrace new changes, new possibilities.

👦🏻 Host Koji: Co-founder of The Fair and Tangdao. I believe technology, especially AI, will fundamentally transform society and empower humanity in the future. Welcome to chat with me, bounce ideas around, and connect on the next possibility. Koji's Jike, Koji's website

👧🏻 Host Ronghui: Works at a tech VC, former Silicon Valley correspondent for CBNweekly. Ronghui's Jike

Join the "Crossing" Membership Group

☀️ First-hand AI news and insights

👫🏻 We encourage everyone to date/make friends/find future kindred spirits

🦀 Add our assistant on WeChat to join: Rwkfbcianvd, or scan the QR code below

References [1] Manqi Cheng: https://web.okjike.com/u/362209fd-6e58-4204-8b97-30a009b07b9c

[2] LateTalk: https://www.xiaoyuzhoufm.com/podcast/61933ace1b4320461e91fd55

[3] Monica: monica.im/

[4] Steiner-preview: https://monica.im/help/Changelog/new_in_version_7.2.0

[5] DeepSeek: https://www.deepseek.com/

[6] Koji's Jike: https://okjk.co/0JSUes

[7] Koji's website: https://koji.super.site/

[8] Ronghui's Jike: https://okjk.co/0cbnYV