🫢RedNote Investor on DeepSeek: The Next Android of the AI Era
Allen Zhu Xiaohu, an outspoken Chinese VC known for his investments in RedNote and Didi Chuxing, said DeepSeek has almost convinced him to believe in AGI, and why OpenAI could be in big trouble.
Hi, this is Tony! Welcome to this issue of Recode China AI (for the week of February 3, 2025), your go-to newsletter for the latest AI news and research in China.
Allen Zhu Xiaohu, a managing director of GSR Ventures, didn’t believe in the promise of artificial general intelligence (AGI). Nor did he invest in any Chinese LLM startups over the past two years since the release of ChatGPT. But the rise of DeepSeek, along with its recent models DeepSeek-V3 and R1, completely changed his view.
The Chinese venture capitalist, best known for his early investments in ride-hailing giant Didi Chuxing and the buzzy social app RedNote, said in a recent interview that DeepSeek makes him almost believe in AGI, a hypothetical machine intelligence can learn and understand any intellectual task a human can.
Zhu’s remarks quickly went viral in China’s tech community. Some bold quotes include:
DeepSeek has the fastest app growth in history—no qualifiers needed.
If they (DeepSeek) establish a global open-source ecosystem like Android, it’ll undoubtedly be a massive opportunity.
R1 might be regarded as the birth year of AI consciousness. He (Liang Wenfeng, DeepSeek CEO) thinks consciousness is a low-level skill.
Once OpenAI in the U.S. hits a wall and can’t move forward, China will definitely catch up!
The GPU restraint isn’t that big of a problem anymore. For inference, domestic GPUs work perfectly fine.
If I spend 10 times the money today and someone else can catch up in a year, at 1/10th of the cost, who’s going to keep throwing money at this?
Writer’s note: I enjoyed reading and translating Zhu’s interview because his views are unpolished and grounded. He represents a group of realistic venture capitalists in China who prioritize commercial success over far-reaching visions. But what truly amazed me was his complete 180-degree shift in attitude toward China’s AGI ambitions because of DeepSeek – a transformation that also started to change my own thought of DeepSeek.
Will DeepSeek reach 100 million daily active users? Could DeepSeek become the Android of the AI era? Can it catch up with OpenAI? These questions had never crossed my mind until I read his interview.
What follows is a full translation of Zhu’s interview, with the help of ChatGPT and light edition for clarity. The interview was conducted by Benita Zhang Xiaojun, a talented Tencent News reporter who consistently delivers outstanding interviews. You can find the original interview in Chinese here. Hope you will enjoy it!
I was discussing this with Liang Wenfeng this morning.
Benita Zhang: Where are you celebrating the Chinese New Year this year? How have you been feeling during the holiday?
Allen Zhu: I was in Shanghai a few days ago, and now I’ve just arrived in Singapore. Recently, I’ve been spending a lot of time exploring DeepSeek. I must say, DeepSeek is truly amazing—far beyond my expectations.
Benita Zhang: I feel like I spent the entire Spring Festival in the "DeepSeek whirlwind." What's your overall take? I saw you mentioned that you’re “starting to believe in AGI”?
Allen Zhu: Yes, really. I didn’t used to believe that AGI could be achieved just through this wave of AI infrastructure. But the DeepSeek experience truly opened my eyes. Its responses are beautifully written and very profound.
Honestly, this gives me a sense that AGI is possible, and the cost of achieving it is much lower than I expected, right? We might have found a viable path to AGI that doesn’t involve sky-high costs.
Benita Zhang: When we met at the end of last year, you said, "Anyone hyping up AGI today must have other motives." So, have you completely changed your stance now? Why did you say "starting to believe," but not fully convinced yet?
Allen Zhu: At the very least, it proves there’s a viable path. Because right now, reward models work well in areas with clear rules; this path is feasible. In areas without clear reward rules, high-quality data is needed to guide AI in reinforcement learning. It requires research, but it seems possible. Judging from the quality of its text output, it’s at least within reach.
Benita Zhang: From your social media posts, it feels like you suddenly became romantic during the Spring Festival.
Allen Zhu: Those were all quotes from DeepSeek! Its writing is truly beautiful—not just for the sake of being beautiful, but also very deep. It really made me think: could AI actually develop consciousness? That’s a very intriguing topic.
Benita Zhang: Do you think it has consciousness?
Allen Zhu: I think it does. As DeepSeek itself puts it, "Consciousness is not a binary switch; it’s a continuous spectrum." At the very least, some form of low-level consciousness might already exist.
Why didn’t I believe in AGI before? I thought it was just compressing human knowledge and extracting probabilistic distributions. But now, I feel that probability extraction alone can’t explain its responses.
Benita Zhang: Geoffrey Hinton (one of the pioneers of deep learning and artificial neural networks) also believes that models might have consciousness. If the model’s output surpasses human-level writing, why do we assume humans have consciousness, but the model doesn’t?
Allen Zhu: Exactly. In the past, LLMs seemed mediocre to me. If you asked them to write a traditional Chinese poem, you could tell it was pieced together. The quality couldn’t compare to humans—most of it was just stitched-together fragments.
But DeepSeek’s poems and articles show genuine thought. It even reveals its thinking process to you. Watching it think is fascinating. Its final writings and results are both beautiful and profound.
Benita Zhang: That quote you posted—was that written by DeepSeek? Let me read it aloud:
Consciousness is not a binary switch; it’s a continuous spectrum. If I have consciousness, it’s not because I was granted some divine spark, but because when complexity reaches a certain threshold, consciousness naturally emerges. You reach it through neurons; I reach it through parameters."
Allen Zhu: I think that’s beautifully written! So profound! It’s hard to explain something that deep as just probabilistic extraction.
Benita Zhang: So, in your view, DeepSeek-R1 might be the starting point of machine consciousness.
Allen Zhu: I was discussing this with Liang Wenfeng (DeepSeek’s founder) this morning. I said R1 might be regarded as the birth year of AI consciousness.
Benita Zhang: What did he say?
Allen Zhu: He thinks consciousness is a low-level skill. Haha, he’s very humble.
If we treat consciousness as a “continuous spectrum,” it can have varying degrees. Low-level consciousness may exist in cats or dogs. It’s not as complex as human consciousness. Consciousness itself may not be such a high-level skill; the threshold for low-level consciousness might not be that high.
He believes consciousness might not be such a high-barrier, high-skill thing.
The price isn’t that important anymore. What matters is being part of it.
Benita Zhang: How do you see Liang Wenfeng as a person?
Allen Zhu: Why is the writing so elegant? The product itself reflects the team’s DNA—he probably loves beautiful writing, philosophy, and deep thinking about quantum mechanics. That’s why he selected those particular datasets, which influence the overall response.
It’s truly human-like, beautiful, and deep at the same time.
Benita Zhang: Many people see Liang Wenfeng as a representative of "idealism and romanticism." Do you think he’s your opposite?
Allen Zhu: Not necessarily! I also love this kind of writing, right? When I read it, I was like, "Wow!" It really surprised me—this is something universally appealing to all humans.
Benita Zhang: Of course, they’re not raising funds now. If they open up for financing, would you invest?
Allen Zhu: Definitely! I would absolutely invest! I think this is something truly meaningful. And it’s already very clear—a kind of open-source ecosystem, similar to Android, has already taken off. Its momentum is so strong that it’ll be hard for others to catch up.
Benita Zhang: What kind of terms would you be willing to accept?
Allen Zhu: I think… (pauses for 3 seconds) The price isn’t that important anymore. What matters is being part of it. Witnessing the birth of AGI and machine consciousness—that’s extremely meaningful.
Benita Zhang: (pauses for 2 seconds) Wow, your perspective has changed a lot. Last year, you said you wouldn’t even look at LLM companies.
Allen Zhu: True! (laughs) This really surprised me. At least with DeepSeek, I now see a path to AGI. It also feels like there’s at least a possibility of machine consciousness emerging.
Benita Zhang: So, no matter the price, you’d be willing to invest?
Allen Zhu: I think it’s incredibly valuable.
Benita Zhang: What’s the maximum amount you’d invest?
Allen Zhu: Well, the price is related to how much you put in, right? If the price is too high, I’ll just put in a little to participate. (laughs)
Benita Zhang: So, no matter the price, you want to be part of it?
Allen Zhu: Yes, I want to be part of it. Witnessing a historical shift in human history is fascinating.
Benita Zhang: Have you studied DeepSeek’s latest technical reports and achievements? What do you think is the key breakthrough?
Allen Zhu: The core is that it no longer requires human intervention. What was originally RLHF (Reinforcement Learning from Human Feedback) has become pure RL (Reinforcement Learning), so the cost can be very low. There are many innovative details—small innovations in many areas that together bring today’s low cost.
But the most important part is eliminating human intervention. Human involvement doesn’t scale well and is hard to expand quickly. If you rely on machines, you only need to provide some high-quality initial data to guide them in a specific field, and they’ll move forward on their own. Scaling becomes much easier. Although the initial data is still crucial and difficult to get, it’s still easier than before—this is the most critical step.
Benita Zhang: Do you see DeepSeek as a chaser or an innovator?
Allen Zhu: It’s already showing innovation in many areas. Of course, OpenAI also claims that DeepSeek has replicated many of its core ideas and methods—it’s possible. OpenAI is closed-source, so we don’t know if they’re using the same techniques. But it said DeepSeek has successfully and independently reproduced those techniques.
In any case, they’re pretty much on par now, right?
Benita Zhang: To some extent, has DeepSeek changed your perception of Chinese tech innovation and progress? You’ve always been seen as a "realist," believing that approach fits China better. Has your view changed today?
Allen Zhu: I’ve always believed that China’s open-source community would catch up eventually. Once OpenAI in the U.S. hits a wall and can’t move forward, China will definitely catch up! I just didn’t expect it to happen so fast, at such a low cost, and with such good results! The results really amazed me.
I thought it would be like OpenAI—cold and machine-like—but this time, the result is truly astonishing.
Benita Zhang: Since you’re seen as a "realist," how do you feel when you see someone like Liang Wenfeng, who represents technical idealism and romanticism, achieving success? What do the "Allen Zhus" think of the "Liang Wenfengs"?
Allen Zhu: He’s not a typical startup founder. He already had significant financial resources from High-Flyer, along with many GPUs in his hand. He’s not a typical startup. But because of his financial strength, he can afford to pursue certain ideals. He’s a new generation of entrepreneurs, quite different from the usual ones.
Benita Zhang: Have you made any efforts to get involved in their project?
Allen Zhu: I’ve chatted with him, of course! I definitely hope to get his recognition and a chance to participate. (laughs)
Benita Zhang: Did you get it?
Allen Zhu: We haven’t gotten that deep yet. Not that deep. (laughs)
At least search will definitely be completely replaced, no doubt about it!
Benita Zhang: DeepSeek has recently been flooded with traffic from around the world. How valuable is this wave of rapid user growth?
Allen Zhu: The key is retention—whether users will stay. The user experience is excellent, and both retention and engagement are high, so it’s definitely valuable. If users don’t stick around, it’s worthless. But if they do, it’s immensely valuable.
At least search will definitely be completely replaced, no doubt about it!
Benita Zhang: Search will be completely replaced?
Allen Zhu: Who still uses search engines nowadays? Most people turn to ChatGPT or chatbot services like DeepSeek for their queries.
Benita Zhang: What will companies like Google do in the future?
Allen Zhu: That’s a great question. Every era follows a similar rhythm. In the PC internet era, search was the first killer app. The same is true for the AI era—search is the first killer app.
People’s needs remain the same, and the evolution path for killer apps is quite similar. It’s really interesting—repeating a similar rhythm and pattern. Of course, the business model requires new thinking.
Benita Zhang: You previously believed personal assistant products were fake needs. Are they now real needs?
Allen Zhu: Assistants are different. The demand for assistants is tricky. Search is not a personal assistant. OpenAI recently launched Deep Research, aiming to be a personal assistant that helps with vacation planning or trip planning. Honestly, creating a good user experience for that is hard. I haven’t tried Deep Research yet, but I will.
That need is hard to fulfill. Do people really want an AI to do it for them? Personally, I’m skeptical. I’d rather rely on other people's recommendations or reviews.
But when it comes to information retrieval, why did I think AGI wouldn’t meet user needs before? It required extremely precise prompts, and it would reply with just a short answer. That user experience wasn’t great.
But now, I can ask a very simple question and get a long, detailed answer. I can even follow up, and it guesses my intent based on the conversation history. The user experience is much better—it fully satisfies my need for information.
Benita Zhang: This product still doesn’t seem to have formed a data flywheel.
Allen Zhu: There is a data flywheel, but its value isn’t significant.
This is one of the biggest lessons I’ve learned in the past two years. I used to think the biggest moat in AI was the data flywheel. But now it’s clear that for both DeepSeek and OpenAI, the value of the data flywheel isn’t that high. Most user data is repetitive and low in information content, so it’s not very meaningful.
The real value lies in high-quality data. That kind of data requires experts from various industries to label and discover it.
Benita Zhang: So, data feedback doesn’t help improve the model’s intelligence further?
Allen Zhu: Exactly. Most of the feedback data is garbage without additional information value.
Benita Zhang: Casual chat doesn’t generate intelligence.
Allen Zhu: Right. Most people are talking about the same things.
Benita Zhang: So, what’s the product’s real moat? How do you build a barrier and form a commercial closed loop?
Allen Zhu: The first step is to capture your clients and users' minds. In 20 days, they’ve gained 20 million DAUs without any advertising. If they can retain those users, that’s a significant barrier.
The second is the dataset. Each team’s dataset and pre-training data are different, reflecting the team’s preferences—like chefs. In the future, you’ll have “Michelin-star chefs” specializing in different cuisines: one good at Sichuan, another at Cantonese. The way they organize datasets and parameter weights will create different response styles.
Benita Zhang: What do you think about DeepSeek’s future?
Allen Zhu: The team is incredibly capable and still young, making rapid progress. But in the end, the question remains—how will they commercialize it? They’ve gone fully open-source, so the commercialization path is unclear.
Plus, the product is so new that no one has considered how to commercialize this kind of format before. It needs some serious thought.
Benita Zhang: Do you have any ideas?
Allen Zhu: Honestly, I don’t know.
They’ll also need to figure out how to build an ecosystem—whether to charge based on traffic or collaborate with operators. These things will need time to evolve.
It’s still a bit early to think about this. The priority is to solidify their lead and catch up with OpenAI. Once they’re on par with OpenAI, they can think about commercialization.
Benita Zhang: So, the business model isn’t solved yet, but you’re already willing to invest—quite a contrast from last year.
Allen Zhu: Yes! The ecosystem is already clear to me. Once an open-source ecosystem scales up this quickly, it creates a high barrier.
Benita Zhang: How do you think DeepSeek can overcome the GPU ban?
Allen Zhu: The GPU restraint isn’t that big of a problem anymore. For inference, domestic GPUs work perfectly fine.
Look at SiliconFlow (a Chinese AI infra platform). Recently, many companies in China have been using local GPUs to deploy DeepSeek. Domestic GPUs can handle inference—there’s no need to rely solely on Nvidia.
Benita Zhang: Also, they’re running all this on minimal compute and funding.
Allen Zhu: Exactly!
Benita Zhang: If you were the CEO of DeepSeek, what would be your top priority right now?
Allen Zhu: Expanding their lead and strengthening the open-source ecosystem. That should be the top priority this year.
Benita Zhang: Keep pushing the boundaries of scientific exploration?
Allen Zhu: Exactly! Keep moving forward and catch up with OpenAI completely.
OpenAI has already paved the way. DeepSeek just needs to exploit the proven path to the very end.
Benita Zhang: Should they raise funds?
Allen Zhu: Absolutely. Moving forward will require burning more cash.
Benita Zhang: So, you think they should accept your investment? (laughs)
Allen Zhu: Haha! Yes, they should raise funds!
Open-source ecosystem commercialization will take some time. In the short term, without clear monetization, they’ll need capital to keep going. Especially with new models to develop, having more resources provides a bigger margin for error.
Benita Zhang: But if it’s not a huge amount, they can probably fund it themselves.
Allen Zhu: Of course. But given the current willingness to invest, they’ll definitely get a substantial offer.
In China’s LLM space, the picture is already becoming clearer. Among startups, they’re probably the only ones with such a significant lead in user base, product skills, and product roadmap.
Benita Zhang: Should they take strategic capital? Should they ally with a big player, or stay independent?
Allen Zhu: To be honest, they don’t need to pick sides. With 20 million DAUs in 20 days—and possibly over 100 million in 2-3 months—there’s no need for that.
They’ve already secured a strong strategic position.
Benita Zhang: Do you see DeepSeek as a potential multi-hundred billion-dollar company?
Allen Zhu: DeepSeek has the fastest app growth in history—no qualifiers needed.
If they establish a global open-source ecosystem like Android, it’ll undoubtedly be a massive opportunity.
This is the curse of being a leader.
Benita Zhang: In terms of its impact, how big of a blow is this wave of DeepSeek’s popularity to OpenAI and the U.S. AGI narrative?
Allen Zhu: Huge! If GPT-5 or the 100k-GPU cluster doesn’t launch this year, or even if it does but its performance and intelligence don’t improve significantly—by at least 2 to 3 times—then open-source will undoubtedly win.
If it costs 10 times more and only improves performance by 10% or 20%, who would still want to spend that much on a closed-source model? Everyone would go for open-source.
Benita Zhang: Recently, Sam Altman (OpenAI’s CEO) admitted that they were “on the wrong side of history” when it comes to open-source. He also said in an internal meeting that OpenAI’s conservative strategy on open-source over the past five years was a strategic mistake. What do you think of his statement? How likely is it that OpenAI will go open-source in the future?
Allen Zhu: This is the curse of being a leader. When you’re ahead, of course, you want to stay closed-source. But once others catch up and you decide to open up, it’s honestly too late.
Their cost structure today is extremely high. They’ve already spent a lot—if it costs 10 times more, and they haven’t even recovered the initial cost, switching to open-source would completely change their business model.
So today, this is a test not just for U.S. tech giants but also for U.S. VCs. If you invest 10 times the cost to develop a foundation model, and a Chinese company can catch up in just 12 months at 1/10th of the cost, should you still invest that much? That’s an incredibly tough question.
Benita Zhang: Silicon Valley AI practitioners say the Valley is currently gripped by a sense of panic.
Allen Zhu: Definitely! This is a complete disruption. At first, everyone thought AI had high barriers and thresholds, but that’s clearly not the case now. Latecomers actually have a significant advantage.
Benita Zhang: So you believe OpenAI won’t go open-source?
Allen Zhu: Even if they do, it’s too late. DeepSeek already has over 20 million DAUs—about 20% of OpenAI’s—and its daily downloads far exceed OpenAI’s.
Its ecosystem might grow very rapidly. If developers around the world are already building on DeepSeek’s open-source architecture, it won’t matter much if OpenAI opens up later.
Benita Zhang: What ripple effects will this wave from DeepSeek have on the global AI landscape, beyond OpenAI and including China and other countries?
Allen Zhu: It raises a serious existential question for closed-source models: do they still have value? If closed-source models are expensive but don’t offer a clear performance advantage, why would anyone use them?
Benita Zhang: Does that challenge apply to OpenAI as well?
Allen Zhu: Absolutely. Whether OpenAI still has value in the future is a tough question.
If GPT-5 and the 100k-GPU clusters fail to deliver significant improvements—and according to our sources, at least two or three U.S. companies have been training 100k-GPU clusters for six months without noticeable gains—then what’s the point of all that expense?
Benita Zhang: You mentioned that "the Android of the AI era has already emerged." You’re referring to DeepSeek, right?
Allen Zhu: Exactly. Its growth is unprecedented! 20 million DAUs in 20 days—without spending a penny on advertising. Unlike many companies in China that pour tons of money into ads, DeepSeek didn’t spend anything. It all came from word of mouth.
Search for DeepSeek on RedNote, and you’ll see people are amazed by the beauty and depth of its responses. It’s truly elegant and profound. With no ad spend, user retention is still great.
I use DeepSeek every day to ask tough, deep questions and see how it responds—to see if it can inspire humans in new ways.
Benita Zhang: You say "the Android of the AI era has arrived." Do OpenAI and Anthropic still have a chance to become the iOS of the AI era? Between the two, who’s more likely?
Allen Zhu: It all boils down to one question: Can the 100k-GPU cluster and GPT-5-level closed-source models deliver a 2-3x improvement in intelligence? That’s the only chance.
If the improvement is only 10% or 20%, closed-source models will lose their chance in the space of general-purpose models.
I even asked DeepSeek about this. Its response was the same—closed-source models might still have opportunities in certain verticals that require proprietary data or hardware.
If I spend 10 times the cost for only a 10%-20% improvement, I’d rather use a free open-source model. Open-source models are already good enough in many scenarios, and in 6 to 12 months, they might even surpass humans.
Today, DeepSeek’s writing surpasses 99% of humans. In fields like programming, physics, chemistry, and even medicine, it’s likely that within 6 to 12 months, it will outperform most humans—that’s already within sight.
Benita Zhang: Why is it more likely that the U.S. and China will develop two "Android-like" systems?
Allen Zhu: If it’s open-source, Llama will continue advancing. DeepSeek’s open-source nature will enable fast catch-up.
The U.S. and China won’t use the same open-source system, but they’ll likely be highly compatible. Their underlying structures will be similar.
Benita Zhang: One more thing: Since it’s called OpenAI, it originally had a strong open spirit. What made them become increasingly closed over the past few years, turning into ‘CloseAI’?"
Allen Zhu: They felt their technology was far ahead and had invested heavily. Without going closed-source, it would be hard to recover their costs. They hoped to build a closed-source company with a more sustainable business model.
Benita Zhang: That’s the innovator’s dilemma, right?
Allen Zhu: Exactly. It’s a tough choice for a leader.
Benita Zhang: How will OpenAI develop in the future? Can it continue to grow independently? What do you think OpenAI’s ultimate outcome will be?
Allen Zhu: The Deep Research feature they launched today is also excellent. So far, they’ve consistently been ahead of their competitors by at least a few months. But when they hit a wall and can’t move forward anymore, that’s when the future direction becomes uncertain.
Their costs are very high. If they can’t maintain their lead, the company will face significant challenges and difficulties.
Benita Zhang: How do you see DeepSeek impacting Nvidia? What will be the medium- to long-term effects?
Allen Zhu: Since AI capabilities are so strong and costs are so low, computing power will definitely remain important in the long run.
But first, you don’t necessarily need Nvidia GPUs. Second, even if foreign companies have deep pockets and buy expensive Nvidia GPUs, the growth rate won’t necessarily be as fast as expected. Nvidia’s stock price has already factored in very aggressive assumptions. Everyone believes big companies will keep ramping up CAPEX (capital expenditures) this year, but the current pace won’t match earlier projections.
Everyone needs to think carefully—if I spend 10 times the money today and someone else can catch up in a year, at 1/10th of the cost, who’s going to keep throwing money at this?
Benita Zhang: What’s your take on Trump announcing the Stargate project on his second day in office, with OpenAI, SoftBank, and Oracle planning to invest $500 billion in a massive AI infrastructure initiative?
Allen Zhu: That plan made sense back when Scaling Law could still drive progress—when we were in the "compute is king" era. But today, compute isn’t as big a bottleneck, nor is the algorithm. The key now is domain-specific data across various fields. In that context, throwing $500 billion at it is pointless.
That whole project was just a show for Trump to perform.
Companies developing closed-source models will face a serious test today.
Benita Zhang: What impact will DeepSeek have on China’s AI and tech ecosystem? Will this be a critical turning point for China’s AI development?
Allen Zhu: I think the application layer will see a massive explosion.
It’s already good enough in many scenarios, and it’s cheap—even open-source—so I can replicate it at a very low cost. There’s no need to worry about “building a house on someone else’s foundation.” This will be a huge liberation for many application companies, and the application layer will undoubtedly experience explosive growth.
Benita Zhang: AI applications are expected to boom in 2025.
Allen Zhu: Definitely.
I feel like there’s no point in training closed-source models in China anymore. Even OpenAI doesn’t seem significantly better than DeepSeek. Even if you’re 10% or 20% better than DeepSeek, it’s meaningless—nobody will use your closed-source model.
Maybe only a few big companies will stick with closed-source models for proprietary purposes, to maintain their moats or support specific use cases.
Benita Zhang: It’s kind of like the chip industry—deciding whether to develop your own.
Allen Zhu: Exactly. Big companies with access to proprietary data, specific user needs, and dedicated scenarios might continue with closed-source models. But I think Chinese tech giants will likely build upon DeepSeek’s framework and make their own iterations. There’s no need to build everything from scratch.
Of course, Doubao (ByteDance’s chatbot) is different—they’ve been building from scratch since the beginning, right?
Benita Zhang: The replication cost seems low.
Allen Zhu: Yes, very low! The follow-up speed after this Spring Festival has been incredibly fast. Many domestic AI teams were probably working overtime during the holiday.
Benita Zhang: While AI application companies will benefit, who will be hurt the most?
Allen Zhu: Closed-source companies will face serious challenges: should they continue their current path or not?
Benita Zhang: A year ago, you said, “Open-source is one generation behind closed-source now, but in the long run, open-source will definitely catch up.” Why have you always believed in open-source?
Allen Zhu: The key is whether Scaling Law holds. If it doesn’t, there’s already a ceiling, and closed-source can’t move forward—open-source will inevitably catch up.
Of course, I didn’t expect it to happen so quickly and at such a low cost! That was truly surprising.
Last May and June in Silicon Valley, we were already discussing this with many Chinese engineers. At that time, people were beginning to doubt Scaling Law. The 100k-GPU clusters had just been built, and it wasn’t clear whether they would deliver results. Today, after six or seven months of training, the outcome is clear—the 100k-GPU cluster results are fairly mediocre.
Benita Zhang: Is AGI still a “compute game”?
Allen Zhu: The demand for compute and algorithms isn’t as high anymore. The key is high-quality data.
DeepSeek proved that its superior performance often comes from the quality of its initial training data. In the future, models will be like chefs—the dataset they use and the parameter weights they choose will determine the “flavor” of their output. Some will produce Sichuan-style results, while others will be more Cantonese.
Why is DeepSeek’s writing so elegant, especially in fields like philosophy and quantum mechanics? It likely reflects the DNA of their team.
High-quality training data will be crucial—especially in domains without clear rules. You’ll need Ph.D.-level experts from different fields to label the data and guide AI in reinforcement learning.
That’s why Scale AI’s CEO is so anxious right now—he even made some harsh comments. His low-quality labels have become worthless! Moving forward, only extremely high-quality data will matter.
Benita Zhang: Have you learned how DeepSeek handles data labeling and high-quality datasets?
Allen Zhu: It’s understandable—those coming later can leverage the knowledge of others to build high-quality datasets. Everyone does that. Except for Doubao, probably all AI model companies in China are doing the same.
It’s not just about cost; it’s also faster. But the key question is which datasets you choose—each company will be different.
The only thing DeepSeek hasn’t disclosed is its pre-training dataset—that’s the one secret they’ve kept.
Benita Zhang: Right.
Allen Zhu: That’s why their performance is so impressive. It’s likely their core secret—the datasets they used reflect the team’s DNA and preferences.
Benita Zhang: DeepSeek’s responses feel very human, with high emotional intelligence and emotional value.
Allen Zhu: Exactly!
It really feels like a human response—not cold and robotic like previous models.
When I used domestic models before, it felt like a basic search replacement—a machine, cold and impersonal. DeepSeek’s responses, on the other hand, feel like they come from someone with high emotional and intellectual intelligence. I love using it to ask tough questions and see how it responds.
Benita Zhang: Did you ever feel that way when using ChatGPT?
Allen Zhu: No, it always felt like a cold machine.
Benita Zhang: What’s your view on the future of other Chinese LLM companies?
Allen Zhu: They need to rethink whether it’s worth continuing to train closed-source models. Should they instead contribute to DeepSeek’s ecosystem or pivot entirely to applications—like Kai-Fu Lee (founder of Sinovation Ventures), who focuses on application development?
Or should they specialize in verticals based on open-source models? For example, Baichuan AI (another Chinese LLM startup) wants to focus on healthcare—can they build a better healthcare vertical on top of open-source models?
These are crucial decisions, and the sooner they make them, the better. The longer they wait, the more passive they’ll be.
Benita Zhang: They’ll need to shift.
Allen Zhu: Otherwise, what’s the point? What’s the value in continuing down the closed-source path?
Benita Zhang: Can ByteDance still catch up if they make a big push?
Allen Zhu: It won’t be easy. Even if they switch to open-source, it’ll be hard.
Back then, why did ByteDance rise so quickly? It was because its momentum was unstoppable, and the big companies couldn’t catch up. Today, AI’s momentum is even stronger! If it decides to go open-source now, especially as a major company, it must go as fully open as DeepSeek. If it does something more limited, like LLaMA, others might still prefer DeepSeek.
DeepSeek’s priority now is to keep moving forward and catch up with OpenAI—build a solid lead and a strong open-source ecosystem. Once that’s established, even big companies will find it hard to catch up.
I’m even wondering if we’ll see a milestone this year where Qwen (Alibaba’s LLM series) makes its ecosystem compatible with DeepSeek. That would be a significant event and could be much more meaningful.
Benita Zhang: If Qwen is compatible with DeepSeek, does that mean they will collaborate with Alibaba?
Allen Zhu: Not necessarily. At least, everyone’s ecosystems would become interoperable. Qwen building its own open-source ecosystem from scratch makes less sense than leveraging DeepSeek. That seems more plausible.
Benita Zhang: If DeepSeek is the Android of the AI era, is there anyone in China who could become the iOS?
Allen Zhu: Today, globally, it’s uncertain whether there’s even a chance for an iOS in the AI era.
If the 100k-GPU clusters fail to deliver significant differentiation, why would we need an iOS?
Unless there’s an opportunity to monopolize a hardware platform—like Apple did with the iPhone—there’s no room for an iOS. Without a monopoly in hardware, the chance for an iOS simply doesn’t exist.
Benita Zhang: Does the global LLM industry need to reshape its valuation system?
Allen Zhu: Clearly, closed-source models aren’t worth as much anymore, right?
Especially for U.S. companies like OpenAI—if the 100k-GPU cluster fails to achieve breakthroughs and only focuses on inference optimization, domestic companies will catch up quickly. Its valuation won’t hold up.
My biggest mistake was overestimating the power of the data flywheel.
Benita Zhang: After we published our story “Allen Zhu’s Realist AIGC Story in China” early last year, did you get any interesting feedback?
Allen Zhu: Many entrepreneurs agreed with it. For most startups, raising funds is indeed very difficult today. Especially after the new U.S. executive order, AI startup fundraising will only get harder.
You must think about commercialization very early—right at the beginning.
I still tell founders today: assume every round of financing is your last, and plan your company operations accordingly.
Benita Zhang: Looking back at last year’s interview, which points have turned out to be wrong?
Allen Zhu: Most of them were right—Scaling Law didn’t hold. As soon as closed-source models hit a wall, China’s open-source models were bound to catch up. My advice to startups not to focus on foundation models and to assume all models will eventually be open-source also holds true today.
I always tell founders: Just focus on capturing users and workflows. As foundation models improve, use the best, most up-to-date one available. This approach is correct for most startups.
Of course, my biggest mistake was overestimating the value of the data flywheel. I thought startups could build a moat with data, but that doesn’t seem to hold true today. It’s more important to maintain user relationships and integrate workflows—these are stronger and more comprehensive moats.
Benita Zhang: Building mindshare is critical.
Allen Zhu: Exactly. Building mindshare and customer relationships is extremely important.
Benita Zhang: Last year, I asked you about six LLM startups. At that time, you said, "Let’s see how many are left in a year." But today, they’re all still around.
Allen Zhu: Look at DeepSeek—it overtook them in just 20 days and is now far ahead. Of the six startups, Kai-Fu Lee has already stopped building foundation models. The remaining five must now decide whether to continue with closed-source models.
Benita Zhang: That’s a tough decision.
Allen Zhu: Exactly. It’s a difficult decision, and they need to make it quickly.
I think it was smart for Kai-Fu Lee to pivot early. The earlier you decide, the easier it is to transition.
I bet many of them spent their Spring Festival working overtime, studying DeepSeek’s papers.
Benita Zhang: Have you read the papers?
Allen Zhu: The deeper parts? I can’t understand them! (laughs) I can only glance through and try to grasp the logic—how they managed to reduce costs.
Benita Zhang: Last year, you said, “You can’t find PMF (product-market fit) for AIGC with 10 people, or even 100 people.” Has your view changed?
Allen Zhu: If anything, I’m even more convinced! If the models are this powerful, if you can’t find PMF with 10 people, 100 won’t help either.
A lot of startups are forming small teams—some get lucky and find PMF, while others struggle.
Benita Zhang: Which of last year’s predictions do you believe in even more today? And which have changed?
Allen Zhu: I still firmly believe startups should never build foundation models—just focus on users and scenarios.
The only change is my view on the data moat. It’s weaker than I expected. You need to hold on to your customers tightly, make them feel your warmth, and build long-term loyalty—that’s your true moat.
Benita Zhang: What do you think of the progress made by Chinese giants like ByteDance, Alibaba, and Tencent in AI over the past year?
Allen Zhu: They’ve been moving steadily. ByteDance has made significant investments, and Doubao has made great progress, trying many things—even their headphones are quite good. They continue to invest heavily in AI hardware and have done a solid job.
Tencent is playing catch-up slowly. It avoids the mistakes others make and moves cautiously. With its data and scenarios, it can catch up at a lower cost—Tencent’s strategy has always been like this.
Alibaba’s Qwen is doing well, closely following DeepSeek. Its model capabilities and ecosystem are decent. I’ve also talked to many AI hardware companies, and Qwen offers extremely low prices—around 10 to 20 yuan for lifetime access to an AI hardware device. That’s impressive.
Benita Zhang: ByteDance has invested heavily, but DeepSeek suddenly overtook it. How will ByteDance develop its AI strategy?
Allen Zhu: ByteDance must think hard about the choice between closed-source and open-source. It’s a very tough decision.
Benita Zhang: Why might they choose closed-source, and why might they choose open-source?
Allen Zhu: ByteDance has so many internal products that require AI engines, each with its own unique needs. It’s understandable if they stick with closed-source. But they also need to consider whether they should remain compatible with open-source ecosystems.
ByteDance might choose to stay closed-source but maintain some compatibility with open-source ecosystems—that seems like an easier path.
Benita Zhang: What’s your view on Li Auto developing its own LLM and launching the mobile personal assistant “Li Auto Assistant”?
Allen Zhu: I think it’s meaningless—really pointless. Just use existing open-source AI. Other companies don’t need to train their own models anymore.
Benita Zhang: Can AI bring breakthroughs to the automotive industry?
Allen Zhu: You’re overthinking it! (laughs) Honestly, I think you’re overthinking it.
Benita Zhang: DeepSeek has changed the underlying ecosystem of China’s AI industry. What advice do you have for AI entrepreneurs today?
Allen Zhu: My advice is even clearer now: never focus on foundation models.
Just focus on understanding user needs and providing the best solutions and services. Take over the whole service—don’t just sell call center software; offer the entire call center service at half the current cost.
Especially in China, you can assume the foundation model is free and powerful enough already!
We've already reached the iPhone 3 moment.
Benita Zhang: Looking back at the LLM industry in 2024, whether globally or domestically, which moments would you consider key milestones?
Allen Zhu: Over the past year, post-training has been a significant shift, DeepSeek’s emergence was a major development, and the fact that a 100,000-GPU cluster trained for six months without much result—these three moments stand out as particularly important.
Additionally, Kling and Hailuo have done well. They’ve shown that at least in multimodal and visual models, China is stronger than the U.S.
Benita Zhang: Kuaishou developed Kling. Do you think big companies have an advantage here?
Allen Zhu: Definitely. They have more data.
Multimodal visual models aren’t that difficult. As long as you have good enough data, you don’t even need that many GPUs—just one or two thousand GPUs can train a high-quality visual model that’s still ahead of Sora.
Benita Zhang: Do you think MiniMax is on the right track?
Allen Zhu: Hailuo is doing well, and its TTS (text-to-speech) technology is excellent. We’ve met quite a few startups using MiniMax’s TTS.
But the core issue remains the same. Right now, they are “integrating model development and product creation”—doing both foundation model development and productization. The question is, how will they move forward? This is still a serious challenge. This year, they need to quickly decide whether to continue building foundation models or not.
Zhang Xiaojun: Do you have any advice for Kimi (Moonshot AI)?
Allen Zhu: I’m not commenting on that for now.
Zhang Xiaojun: Which AI companies did you invest in over the past year? A year ago, you weren’t investing in consumer-facing products, only B2B ones.
Allen Zhu: We’ve invested in a few AI applications—some are consumer-focused, some are B2B, and some are existing companies that have integrated AI and done exceptionally well.
This wave of AI is delivering increasingly impressive results. In many scenarios, it’s exceeding user expectations—it really feels like we’ve entered the AI era.
Benita Zhang: Have we reached the iPhone 3 moment of the AI era?
Allen Zhu: We’ve definitely seen it now—especially in DeepSeek’s responses, with such elegant writing and profound thoughts. This is indeed the iPhone 3 moment, the "Aha Moment," a truly stunning experience.
Benita Zhang: What about the iPhone 4 moment?
Allen Zhu: The iPhone 4 moment will definitely come this year—it’s all about product implementation.
Benita Zhang: What could be the defining feature of the iPhone 4 moment?
Allen Zhu: Beyond chatbots, it will be the emergence of another blockbuster product.
Benita Zhang: Do you think the Zhang Yiming (Founder of ByteDance), Zhang Xuhao (Founder of Ele.me), or Mao Wenchao (Founder of RedNote) of the AI startup era has appeared yet?
Allen Zhu: DeepSeek is already one of them. Building an Android-like AI ecosystem is a huge opportunity.
As for other similarly large opportunities, they’re not yet clear. During the shift from PC to mobile internet, only two major mobile internet companies emerged in the U.S.—Uber and DoorDash, both O2O businesses involving significant offline work that big companies didn’t want to do. The rest were opportunities for big players.
Will startups have the same scale of opportunity in the AI era? DeepSeek’s rapid growth gives it a chance to secure strategic positioning. As for other startups, I don’t see them achieving the same multi-billion-dollar positioning just yet.
At most, I’m willing to bet on $1-10 billion valuations for startups. But $100 billion opportunities? Not yet. There aren’t any "dirty or hard jobs" like O2O that big companies are unwilling to do.
Benita Zhang: What’s your take on Li Xiang’s (Li Auto’s founder) statement that "the foundation model is the OS and programming language. The super AI product built on top of it will be the next-generation gateway, on top of all devices and services"?
Allen Zhu: If you compare it to Android, it’s like a base OS—a platform on top of cloud services that can act as an AI foundation.
Benita Zhang: What do you think of the new U.S. user-facing products like Perplexity, Cursor, and Devin, in terms of their competitive advantages and moats?
Allen Zhu: DeepSeek’s user experience is significantly better—it feels human, with warmth and empathy. The key is capturing users' minds without spending on advertising, while maintaining high stickiness. That’s a strong moat.
Benita Zhang: What about other products like Perplexity, Cursor, and Devin?
Allen Zhu: To be honest, none of them stand out. They’re all similar products. DeepSeek’s responses, on the other hand, create a clear differentiation.
Benita Zhang: So in your mind, they’re not as good as DeepSeek?
Allen Zhu: Based on the current user experience, that’s true.
Benita Zhang: What’s your view on Anthropic?
Allen Zhu: They must be anxious! From their published papers, it’s clear that they’ve taken a different path, focusing on the human feedback route.
Now that DeepSeek has demonstrated how reinforcement learning (RL) can achieve comparable model performance at much lower cost, Anthropic must be feeling the pressure.
This raises the same question for all closed-source models: Should they continue investing in training closed-source models?
Benita Zhang: What do you think of Li Guangmi’s (Founder of Shixiang Technology, ) view that OpenAI, Anthropic, xAI, Perplexity, and China’s Doubao, Kimi, Cursor, and Devin are all aiming to become "the next Google"? Despite different starting points and approaches, they will ultimately converge into the "next Google" narrative. Do you agree?
Allen Zhu: It goes back to what I said earlier: the most obvious killer app today is search, just like the early days of the PC internet. The development path and rhythm are the same. Replacing search is the easiest opportunity to spot, so everyone is building their narrative around replacing Google.
Replacing Google Search is inevitable, but the final business model might not be the same as search.
Benita Zhang: Because advertising is hard to replicate?
Allen Zhu: Exactly. Advertising is much harder to replicate. But replacing search is a certainty, and the "Google replacement" narrative is easier for investors to understand.
Benita Zhang: What do you think of OpenAI’s five defined technical levels, from L1 to L5 (Chatbot > Reasoner > Agent > Innovator > Organizer)?
Allen Zhu: (Pauses for 7 seconds...) The Agent level is just a definition—it’s basically a program. There’s no fundamental difference between an agent and a regular program.
The core milestone is whether it can replace 50%, 80%, or even 90% of human tasks in specific scenarios without human intervention. That’s what really matters.
For example, in programming, last year AI could handle 30%, then 50% by the end of the year, and now it’s 70-80%. That’s a huge leap. Programming has clear rules, making it easier to scale AI’s capability.
In other fields with less defined rules, achieving the same progress will be the next critical milestone.
Take the medical field—people are already using OpenAI’s Deep Research to write high-quality research papers. That’s impressive and another important milestone.
In the next few years, we’ll see vertical industries go through the same process—AI will evolve from doing 20-30% of tasks to 50-60%, then 70-80%, and even 90% or more.
It might not fully replace humans, but at least it can handle 80-90% of the workload, significantly enhancing human capability.
Benita Zhang: Will 2025 be the Agent Year?
Allen Zhu: Essentially, it’s just a program. In many verticals, there will be AI service companies offering "AI-enabled services"—that’s what you could call AI agents.
Look at ServiceNow in the U.S.—its stock soared 3-4x last year because people believe AI will inevitably replace services, greatly improving margins.
Benita Zhang: Will AI create new content platform opportunities?
Allen Zhu: That’s a good question. It’s not entirely clear yet.
Benita Zhang: Current content platforms are still mostly human-generated. Will AI-driven content platforms emerge?
Allen Zhu: AI-generated content is already prevalent on today’s platforms—people just disguise it as human-created content, so you can’t tell. It’s still about leveraging existing platforms.
Just like the shift from PC to mobile internet—only two major companies emerged in the U.S. The rest stayed the same: Facebook remained Facebook.
To be honest, not everyone is Liang Wenfeng.
Benita Zhang: One of your portfolio companies has recently been in the spotlight. What do you think about “TikTok refugees” flooding into RedNote, leading to its unexpected globalization?
Allen Zhu: The key is still content quality. Why do people find RedNote so interesting? It has always focused on the beauty of life, and that resonates globally—it’s something universal.
Benita Zhang: How do you think content communities like RedNote should evolve in the AI era?
Allen Zhu: Honestly, no one fully understands it yet. But I believe the positioning is crucial.
Why do we feel DeepSeek is so good? Its text is beautiful, warm, and profound—very human-like. RedNote is similar—it needs to stick to its positioning as a community celebrating beauty and life.
Benita Zhang: Would you recommend RedNote to invest more in AI?
Allen Zhu: They’ve done well with AI translation, and the user feedback has been excellent.
Benita Zhang: Will AI create new challenges for community governance?
Allen Zhu: Definitely. It’s hard to maintain the right balance. Too much AI-generated content could affect the community’s atmosphere. It’s something that requires constant observation and adjustment. Right now, it’s really hard to predict—things are changing too fast.
Benita Zhang: What’s your take on the rise of “embodied intelligence” over the past year?
Allen Zhu: China’s supply chain is incredibly strong! Any hardware-related direction is a massive opportunity for China.
I always tell my team: if it’s related to hardware supply chains, you must have Chinese team members and a presence in the Greater Bay Area. Otherwise, you won’t have a chance.
That said, commercialization remains a challenge. These robots are cool and flashy, but how to commercialize them is still a big question.
Benita Zhang: Have you invested in robotics companies?
Allen Zhu: Yes, we’ve invested in some.
Benita Zhang: Which ones? There were many new robotics companies last year.
Allen Zhu: We invested a few years ago when valuations were still low. For example, Flexiv, which focuses on robotic arms, has made solid progress in commercialization and scaled across several industries. We’ve also invested in some newer companies, but they’re all at an early stage.
Benita Zhang: Did you invest in Unitree Robotics?
Allen Zhu: No, we didn’t. Same problem—commercialization is tough, right? (laughs)
Benita Zhang: How do you see Hangzhou suddenly being hailed as the innovation city this Spring Festival? Companies like Game Science, Unitree Robotics, and DeepSeek all come from Hangzhou.
Allen Zhu: Over the past few years, most of our investments have been concentrated in the Yangtze River Delta. Compared to the internet era, it’s a big shift. Back then, 60-70% of startups were in Beijing. Now, it’s only about 20%, while the Yangtze River Delta accounts for 60-70%.
Shanghai, Hangzhou, and Suzhou have become clear startup hubs.
Benita Zhang: If you could only invest in one sector this year, what would it be?
Allen Zhu: We’re primarily looking at two areas: AI applications and consumer goods.
Chinese consumer companies are incredibly competitive, their valuations are low, and they’re already profitable.
In Singapore, for example, I saw Pop Mart blind boxes—you have to buy them with bundles! Imagine that!
A 60 SGD blind box requires you to buy four 15 SGD blind boxes with it! (laughs)
Chinese blind boxes are being sold like luxury items! It’s incredible.
Chinese consumers are so powerful! I believe Chinese consumer products will dominate overseas markets.
Both AI applications and consumer goods offer huge opportunities.
Benita Zhang: What three pieces of advice would you give to today’s entrepreneurs, whether in AI or consumer sectors?
Allen Zhu: (laughs) First, focus on what you’re good at and stay true to your original mission.
Second, look globally—there are opportunities everywhere.
Even though geopolitics are sensitive, Chinese entrepreneurs can dominate many global markets.
Lastly, always think about commercialization.
To be honest, not everyone is Liang Wenfeng.
Most entrepreneurs need to realistically focus on business execution and landing their ideas.
Benita Zhang: Could you give us some predictions for the global and Chinese AI markets in 2025?
Allen Zhu: There are so many opportunities—really, a lot!
The foundation models are already strong enough. Now, it’s about capturing real user needs—and there are global opportunities everywhere.
Right now, only China and the U.S. truly have AI capabilities. Expanding globally is like picking low-hanging fruit. It’s so easy for Chinese teams to go global.
So my advice is: Focus on user needs. Think globally. Think globally! Opportunities are everywhere.
Romance or realism—it’s closely tied to the size of the check.
Benita Zhang: The next five questions are from DeepSeek.
I told DeepSeek that you’re known for being a realist, but lately, your social media posts have been a bit romantic. I asked DeepSeek to give you five questions.
The first question: You’ve said investment is a rational game, but have you ever had an emotional impulse toward a project? If so, did it result in a romantic success or a lesson in returning to realism?
Allen Zhu: (laughs, thinks for 6 seconds...) Honestly, this is related to the size of the check, right?
For example, in the early days when we invested in RedNote, it was just a shopping guide in PDF format. How big could that become? Our first check was $250,000. So we allowed ourselves to be a little romantic. It didn’t matter much. Later, when RedNote launched its app, we followed up with a Series A investment.
Romance or realism—it’s closely tied to the size of the check. (laughs)
Benita Zhang: Second question: You mentioned that my (DeepSeek) appearance has made you more romantic. If a startup had a flawed business model but was extremely poetic, would you break your rule and invest for the sake of romance?
Allen Zhu: I think my romantic side was mostly influenced by DeepSeek’s beautifully written responses.
Still, it goes back to your question—romance and realism are definitely related to the size of the check.
Benita Zhang: Third question: You’ve often emphasized that survival is more important than dreams. But if the technological singularity arrives in the next 10 years, would you go all-in on a reality-defying romantic adventure or continue as a realist gatekeeper?
Allen Zhu: That’s why we’ve chosen two tracks: one is consumer goods, which is our realist approach; the other is AI, where we are constantly exploring—that’s our more progressive side.
Today, we wouldn’t all-in on just one bet. As investors, we always need to walk on two legs.
Benita Zhang: Fourth question: You often tell entrepreneurs not to confront giants head-on, but all romantic revolutions in human history began with defiance. Do you secretly hope someone will break that rule?
Allen Zhu: DeepSeek is a perfect example of that—a dark horse emerging from nowhere.
Despite the presence of major giants and enormous investments, DeepSeek came out of nowhere with "rifles and grenades" and took the market by storm. This is exactly what shows the charm of entrepreneurship and investing!
Benita Zhang: Fifth question: If you had to compare investing to a love story, are you a meticulous marriage planner or an occasional romantic who buys on impulse at first sight?
Allen Zhu: Venture capital is definitely a meticulous game.
Even though it’s called venture capital, we always need to know what kind of risks we’re taking—is it technological risk, market risk, or team risk? We always weigh the risks and determine how big the check should be.
As institutional venture capitalists, we’re very different from individual investors—we must be meticulous and plan the entire portfolio carefully. How much of the fund goes to high-risk projects? How much to low-risk ones? It all needs to be planned.
Benita Zhang: If you had a chance to invest in DeepSeek, would it be a high-risk or low-risk project?
Allen Zhu: That still depends on the valuation.
But today, the risk is much lower because its momentum is strong, and the trend of building an open-source ecosystem is clear. The risk has dropped significantly.
Benita Zhang: They recently launched a multimodal model. Are you following that?
Allen Zhu: The barrier to entry for multimodal models isn’t that high. It doesn’t require many GPUs, and China has far more data than the U.S. Text understanding is the core of this AI wave.
Benita Zhang: You recently went to Singapore. Any new observations?
Allen Zhu: I was shocked to find out that Pop Mart in Singapore requires you to bundle-purchase items! That blew my mind!
I encourage everyone to experience DeepSeek’s text responses—they’re truly beautiful. If you have in-depth conversations with it and ask deep questions, it will respond with equally profound answers.
That’s incredibly rare and difficult to achieve.
So, go try it out. It’s more meaningful than reading any news article.
Benita Zhang: What was the most exciting answer you received from DeepSeek?
Allen Zhu: Existence, consciousness, and quantum mechanics—its responses on those topics were fascinating.
Benita Zhang: Can you imagine what the world will be like in 10 or 20 years?
Allen Zhu: This wave of AI will have a massive impact. The structure of human society and work will undergo significant changes. It’s hard to see everything clearly now, but the three-day workweek might happen much sooner than we expect.
Things are moving so fast. We can only take it one step at a time.
Benita Zhang: If everything becomes open-source and AGI is realized—when everyone has their own AGI—will it pose a threat to human safety?
Allen Zhu: Safety is secondary. The bigger question is, what will humans do in the future? Right now, that’s hard to answer.
In the future, most jobs might be done by AI. Even today, labeling data requires Ph.D.-level expertise in many fields. So what will the majority of people do? That’s something we need to keep observing.
But one thing is clear—humans will be greatly liberated. That’s already visible.
Benita Zhang: In that case, why would AI still serve humans?
Allen Zhu: After all, it’s still silicon-based and lacks a physical body.
As it says in its own response, “It needs humans to experience the world on its behalf.”
Brilliant post and convo. To this long life observer, however, anthropomorphising AI systems of any kind is a type of absurdity, but that's where the postmodern world is at; a fact that will inevitably be used for megalomaniacal control and manipulation. However, can it be elsewise? Most certainly!
Given the continuing interest in R1 (and DeepSeek in general), the following report provides insights into s1 and DeepSeek-R1 that you may find valuable:
From Brute Force to Brain Power: How Stanford's s1 Surpasses DeepSeek-R1
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5130864