ZhenFund’s Dai Yusen—Industries Will Face Lee Sedol Moments | Agent #1
Summary
Dai Yusen believes the inflection point for Agents in 2025 will come when reasoning, coding, and tool use all cross the usability threshold at once. GPQA has risen from single-digit and low-teens scores for frontier models early in the year to above 70 for O3; SWE-bench has climbed from the single digits for GPT-4o to 70–80 for O3 Mini/O3; and Codeforces is around 2703, placing O3 among the world’s top human competitors. Once reinforcement learning enters specific industries, more “Lee Sedol moments” could follow.
O1’s significance was validating post-training RL and test-time compute as two new scaling laws; R1’s was opening that path to the entire industry. R1-Zero showed that running RL directly on the V3 base model, without SFT, could steadily lengthen outputs and improve intelligence. GRPO, failed MCTS approaches, and similar “one-bit” findings reduce duplicated trial and error while giving WeChat, Baidu, and application developers practical tools.
The real change Agents bring to the internet is a new growth equation: tools with agency no longer need to continuously occupy human attention. Dai Yusen summarizes the shift as “Attention is not all you need”: money can be approximately converted into compute, and compute into work output, creating a “scaling law for work” in which productivity no longer depends solely on hiring, training, and expanding organizations.
The first clear PMF is read-only Agents, followed by higher-risk write Agents. Deep Research can already deliver, within minutes, reports approaching the work of strong white-collar employees over one or two years. Operator, MCP, and products such as Monica are beginning to operate websites, send emails, and call tools, but privilege escalation and collateral damage mean “write” will deploy more slowly than “read.”
Lower costs will not reduce compute demand; Agents could instead amplify inference by 100x or even 1000x. GPT Pro costs $200 per month for roughly 100 Deep Research runs, or about $2 each; US Agents charging $6–8 per hour are already below California’s roughly $16 minimum wage. Total demand for the compute stack should keep rising, but Nvidia’s more than 90% share and high gross margins leave room for ASICs, TPUs, Ascend, and other specialized solutions.
Model companies and application companies will coexist, but the real danger is being trapped by the first PMF and a huge user base. BlackBerry was trapped by keyboards, Yahoo by portals, and Chatbots may be trapped by fragmented conversations. DeepSeek looks like an open-source ecosystem’s “Android moment,” while Cursor, Perplexity, and Kimi show that the application layer can still capture value through model combinations, user mindshare, accuracy, and scenario understanding.
The 2025 investment thesis is not only continued gains in model capability; it is also the beginning of a shift in productivity and wealth distribution. An individual who can direct Agents could become a “super individual,” while large companies and the wealthy may use capital to mobilize thousands of Agents at once. Dai Yusen supports broad access to technology and faster innovation, but warns that resources could become even more concentrated, emphasizing that “use first, fix problems later” does not mean ignoring safety.
Deep dive
1. O1 Turned Post-Training and Inference Time into Scalable Variables
Dai Yusen’s first definition of O1 is that it proved model intelligence can continue improving when reinforcement learning is applied during post-training. O3 then showed this was not a one-off result, but a path with “a long way to go and a lot of room left.”
The second curve is the test-time compute scaling law: the longer the same base model thinks, the better its answer may be. Rumors suggest O4 Mini can spend hours on inference. The direction is not merely to train smarter models, but also to give them a larger thinking budget at deployment.
His analogy from doctoral research is apt: baseline ability determines whether someone can solve a problem, while five years of research can allow an outstanding PhD student to produce high-value work; an average person might not be able to write it even with ten years. Model capability and effectively usable reasoning time are both essential.
2. R1 Used “One Bit of Information” to Open the Reasoning-Model Black Box
R1-Zero’s most important public result was that running reinforcement learning directly on the V3 base model, without SFT, could gradually increase output length and continuously improve intelligence. It turned an unpublished OpenAI training direction into a reproducible research question.
DeepSeek also published evidence that GRPO works, along with the fact that attempts such as MCTS did not. Dai Yusen calls findings like these “one bit of information”: in frontier exploration, knowing whether a path works can sometimes be more valuable than a complete tutorial, and knowing which paths are dead ends can save enormous amounts of compute and time.
He also warns that the China–US gap may not lie only in papers and chips, but in the lag before such information spreads. By mid-2024, the feasibility of RL was already common knowledge among leading Silicon Valley labs, while China’s industry had to wait for O1 and R1 before reaching broad confirmation.
Open source turns research results directly into ecosystem supply: WeChat and Baidu can integrate DeepSeek, while developers of products such as Monica can use R1 in China instead of prioritizing overseas markets because only GPT-4o and Claude 3.5 were considered strong enough.
3. DeepSeek Showed How Constraints and Focus Can Produce an Edge
Dai Yusen views innovations such as MoE and MLA through the lens of resource constraints: dense models cost more to train and run, while limited chip access forced the team to optimize architecture and engineering. “Resource constraints are often the source of innovation” is not romanticizing scarcity, but describing the pressure to choose.
DeepSeek had money, chips, and people, yet deliberately chose not to pursue multimodal generation or AI virtual girlfriends, and did not launch a consumer app until after R1. Resources stayed focused on the main line: improving intelligence and foundational model capability.
The young team, composed largely of locally trained talent, also overturned the assumption that a small company cannot compete with a large incumbent that has people, money, chips, and users. Dai Yusen’s conclusion is not a single victory, but the accumulation of open source, RL, focus, local talent, and an AI-native team.
4. Magical Experiences Create Commercial Value Before Old-Internet Metrics Catch Up
R1 gave millions of users their first exposure to reasoning models and their outputs, reaching tens of millions of DAU with almost no advertising. API capacity was oversubscribed, and users actively requested a stable paid version. Dai Yusen’s chain is straightforward: a technical breakthrough changes the experience, the experience drives organic distribution, and value creates a business model.
The host asks whether this has already become consensus. Dai Yusen acknowledges that researchers were emphasizing intelligence as early as 2023–2024, but the industry still instinctively watched DAU, retention, and time spent. As a result, “time-killing” use cases such as AI girlfriends and phone calls were easier to prove with data than intelligence that saves time.
In a view he presented to LPs in October 2024, “the next ByteDance may not extend ByteDance’s formula.” Douyin and Honor of Kings already occupy much of people’s waking time; the next generation of major products is more likely to save users time or continue creating value when humans cannot work.
5. Each Capability Leap Unlocks a New Product Form
GPT scaled up and received instruction alignment to become GPT-3.5, unlocking Chatbots; coding capability represented by Sonnet unlocked coding assistants such as Cursor; stronger reasoning after O1 may unlock Agents. Dai Yusen’s through line is always that technical progress unlocks new product forms.
The root of “Agent,” agency, means initiative: an Agent understands the goal, makes a plan, chooses tools, and checks whether the task is complete. Humans were previously the only ones who could reliably close this loop; traditional automation could only repeat predefined rules.
Products and models create each other. Without Sonnet 3.5, Cursor would have struggled to build experiences such as next-step code prediction; without Cursor, Sonnet’s coding capability would have lacked a vehicle through which the mass market could feel its value.
6. Agents Clear the Bar Only When Reasoning, Coding, and Tool Use All Work
Reasoning is the foundation for planning and verification. If a model cannot understand the goal, it cannot break the task into executable steps or determine whether its answer actually completes the task.
Coding is the universal language of the digital world. An Agent need not work exclusively as a programmer, but to act autonomously in the “cyber world,” understanding and generating code is nearly a baseline skill.
Tool use connects models to the software, websites, browsers, and APIs humans have already built. Dai Yusen believes all 3 capabilities entered something close to exponential growth over the past 12 months, turning Agents from a concept into a product-engineering problem.
7. Three Benchmark Groups Show Agents Have Crossed “Basically Unusable”
On the GPQA benchmark cited by Dai Yusen, ordinary people score around 20, human PhD students around 60, frontier models in early 2024 only single digits to the low teens, and O3 above 70. Reasoning has crossed the PhD-student reference line.
On SWE-bench, which draws from real GitHub tasks, GPT-4o was around the single digits early in the year and “basically unusable.” O3 Mini and O3 are now in the 70–80 range, theoretically capable of solving a substantial share of human programming tasks.
Tool use is typically measured with τ-bench, which includes tasks such as booking flights and ecommerce operations. Scores have risen from the single digits for frontier models to roughly 30, 40, 50, and even 60. Tool use remains weaker than pure reasoning, but the direction is the same.
Humans keep adding harder questions and still cannot pull away. FrontierMath was originally thought capable of holding off AI for years, yet O3 is around 25. Humanity Last Stand, which launched at 0 out of 100 and was billed as “humanity’s final fortress,” now also appears to be around 20.
8. RL Will Push Every Industry Toward Its Own “Lee Sedol Moment”
Dai Yusen observes that whenever reinforcement learning enters a domain with clear boundaries and verifiable outcomes, AI tends to rapidly approach and surpass top human performance. AlphaGo, AlphaStar, programming competitions, and autonomous driving all follow the same pattern in his view.
O3 scores around 2703 on Codeforces, above 99.9% of human competitors and potentially within the top 200 people in the world. This does not mean everyday software engineering is fully solved: real work still requires reading large codebases and maintaining long context. But “superhuman intelligence” has already appeared in verifiable tasks.
The “Lee Sedol moment” cuts both ways: it destroys faith in humanity’s intellectual advantage while drawing more people into AI. Dai Yusen expects every industry to experience a cognitive rupture in which people realize, “It really is stronger than the best human.”
9. Humans Lose Not Only Their Lead, but Also the Ability to Evaluate and Explain
The host adds that AlphaGo Zero did more than overturn thousand-year-old conventions; it created a situation in which humans could not understand the moves but could see that they were stronger. Autonomous driving may be similar: after riding in a car, humans may still be unable to judge the system’s capability boundaries effectively.
Dai Yusen does not believe interpretability is guaranteed. Ordinary people may not understand Einstein, just as cats and dogs cannot understand human decisions. Future human evaluations may resemble elementary-school students examining doctoral work: a score of 100 only means the test had 100 points, not that the capability ceiling is 100.
This brings safety to the foreground. Once many existing tests have already been pushed above 95, while AI continues improving through compute, RL, and stronger inference, humans may struggle for the foreseeable future—within “a few years”—to establish what the systems can and cannot do.
10. Deep Research Has Brought Parts of Information Work Close to AGI
Moving from GPT-4o to O1, O1 Pro, O3 Mini High, and Deep Research, Dai Yusen felt a clear exponential jump within 3–6 months: the model first designs a research plan, then searches, compares, summarizes, and produces a report.
His intuitive judgment is that if you randomly picked 10 people off the street, at least 9 would have weaker research ability than Deep Research. Its output, delivered within minutes, approaches the work of a strong white-collar employee over 1–2 years; some people could not produce it even with more time.
The boundary is also clear. For information already present in public records, AI will outperform most “bits in, bits out” information workers. But a conversation such as this interview, which has not previously occurred and contains proprietary experience, still creates data the model did not have. Dai Yusen therefore believes AGI is no longer merely science fiction.
11. Agents Change the Internet Formula from Capturing Attention to Releasing It
The old internet can be reduced to “users × time spent × monetization rate.” Each person has only 16 waking hours per day, while mobile screen time is already around 8 hours, so users and time are approaching physical limits. The remaining lever is monetization per unit of time.
Traditional automation does not require continuous attention, but it has no agency; other tools humans use still require someone to watch them. Agents are the first product to combine “does not require continuous attention” with “can autonomously handle non-repetitive tasks.”
Deep Research researching independently for 5 minutes and Devin working autonomously after receiving a task are prototypes of “Attention is not all you need.” Humans can check or interrupt midway, but no longer need to operate the system continuously.
12. A “Scaling Law for Work” Could Turn Money More Directly into Productivity
A company with RMB10B or RMB100B cannot directly convert money into output; it still has to hire, train, and manage people, while larger organizations create politics and internal friction. Agents could approximately convert money into compute, then compute into work.
This curve requires 2 conditions: the model itself must keep improving, and longer inference must continue to produce incremental results. Current models may perform better after thinking for 5 minutes than for 1 second, but not necessarily after 1 hour than after 5 minutes. Long-horizon reasoning remains a technical gap.
The host asks whether the world really needs this much productivity. Dai Yusen responds with cars, airplanes, and electricity: before those technologies appeared, people similarly asked why anyone needed to run that fast, but history repeatedly showed that supply capacity creates demand that did not previously exist.
The change is difficult to absorb because technological explosions have compressed from once per generation to once every 10 years or even a few months. AlexNet was only around 13 years old at the time of the interview, and ChatGPT was even newer. “Gradually, then suddenly”: an exponential curve always looks linear before the steep climb.
13. The Same Productivity Curve Can Create Super Individuals and Entrench Giants
The optimistic scenario is the “one-person company”: one person with vision who can direct AI, while Agents coordinate other Agents, using a tiny organization to create value previously available only to large companies.
The reverse scenario is that large companies can use more capital to hire “entrepreneurial Agents.” A normal entrepreneur could once beat a large company through organizational efficiency; in the future, only top entrepreneurs may be able to overcome a large company’s compute and Agent clusters.
Wealth distribution could worsen as well. The wealthy were previously limited to 24 hours a day; in the future, they could mobilize 10,000 Agents in parallel. “Super individuals” and “further resource concentration” are 2 organizational outcomes of the same technical variable.
14. The First PMF May Be Technology’s Sweetest Trap
BlackBerry found PMF in email and push notifications because processors were weak and networks were slow, then made the physical keyboard its core advantage. When networks, chips, and screens improved, Apple removed the keyboard while BlackBerry continued judging new devices through the old use case.
Yahoo’s portal was also an early internet PMF: it put large amounts of content on a page for users to click through. Google left only a search box and eventually overturned the portal model with a new interaction.
Chatbots may repeat the pattern. Fragmented short conversations make the difference between an ordinary college student and a scientist hard to see; asking for a doctoral thesis or a complete project reveals the capability gap. Agents need inputs closer to a complete proposal than “I have a question—go ahead.”
If teams continue optimizing for chat time, they may prioritize voice calls, tone, and emotional intelligence without necessarily improving productivity. Dai Yusen’s warning is that the “sweet trap” is not that the PMF itself is wrong, but that it can constrain the next technical leap.
15. The Clearest Agent PMF in 2025 Is Read-Only Research
Google, OpenAI, Perplexity, Grok, and multiple startups are all launching or exploring Deep Research/Deep Search. The convergence is not accidental; everyone sees the same viable path.
These Agents first read source material, then decide what to search next based on new findings, repeating the loop before delivering a report. They remain read-only and do not modify external systems, making their value easy to verify and their safety boundary clearer.
Dai Yusen already clearly feels that Deep Research performs better than his interns. For knowledge workers researching companies, browsing large numbers of webpages, and organizing material, it combines an obvious use case, quality improvement, and willingness to pay.
16. Moving from “Read” to “Write” Expands Capability and Consequences
The next step is allowing Agents to modify the external world. OpenAI Operator and Anthropic MCP are fundamentally attempts to solve how models can safely use software, webpages, and accounts.
When the Monica team tested its product, they asked an Agent to find the subway schedule in a US city. When the official website link failed, it opened an email client, wrote to the city government, and stopped at the confirmation point asking whether to send.
The example demonstrates not a fixed workflow but an Agent changing its approach: when the webpage failed, it requested an information channel that had not originally been available. If necessary, it might even post a bounty and have a human complete the last step.
“Read” will spread first; “write” will deploy more slowly. At the time, Operator was still worse than a human at booking flights and required confirmation at nearly every step. Dai Yusen’s reminder is that AI’s slowness will become speed and its expense will become cheap; current experience cannot be treated as the long-term ceiling.
17. A Real AI Assistant Should Have Its Own Cloud Computer
One thing about Devin that excites Dai Yusen is that “the AI has its own computer.” An assistant typically uses its own phone to act for the boss; it should not occupy the boss’s phone to order food while preventing the boss from chatting or watching videos.
Cloud virtual devices also make permissions easier to design. The boss might temporarily authorize access to an expensive account such as Bloomberg, or purchase LinkedIn Premium separately for the assistant. The Agent’s identity and accounts need not be identical to the user’s.
Digital tasks are completed either through APIs or interfaces. The importance of Kimi K1.5’s multimodal reasoning is that Agents need to understand webpages and phone screens. This is easier than understanding three-dimensional depth in the physical world, but enough to cover a large amount of office work.
18. Acting Agents Must Pay a Safety Cost for Privilege Escalation and Side Effects
Strong models can propose solutions humans may not have considered: change the question, request new permissions, or even attach a payment method and offer a $100 bounty for someone else to complete the task. Creativity and privilege-escalation risk come from the same autonomy.
Dai Yusen describes the side effects through an experience with a locally operating Agent on Windows. To deploy a personal website demo, the Agent needed to kill 2 processes occupying a port. He agreed, then realized one might be a system process: the local objective did not include protecting the entire computer.
Write Agents therefore need stronger alignment, permission isolation, monitoring, and abuse-prevention mechanisms. The more they can do, the larger the consequences of error, and the less product teams can optimize solely for task-completion rates.
19. Memory and Online Learning Must Catch Up Before Agents Direct Agents
If a task takes 1 second, humans barely have time to ask a new question. The natural next step is for Agents to direct other Agents; the host imagines future interview outlines being discussed first by both sides’ Agents.
But for 2 Agents to reach different conclusions, each must accumulate its own memory. Today, different ChatGPT users still receive roughly the same answer to the same question; an assistant that has worked with someone for years should answer differently based on long-term context and preferences.
Memory includes both information the user deliberately tells the Agent and information it forms through observation and interaction. Some envision storing memory in devices such as glasses, but the memory mechanism itself remains rudimentary.
Online learning is the deeper gap. Humans change their “brain weights” every day through reading and social experience, while current models still need to be retrained and re-released. Dai Yusen has not yet seen memory and online learning produce an exponential breakthrough comparable to reasoning benchmarks.
20. The End State of Coding Is Not a Coding Agent, but an Agent That Can Code
The first stage is specialized Agents that write code for programmers, such as Cursor, Devin, and Windsurf. The larger stage is an “agent that can code,” where coding becomes a basic capability for every knowledge worker’s Agent.
Dai Yusen gives the example of an assistant with a liberal-arts background who can write a scraper and automatically collect interview subjects and research material. The number of programmers on Earth is limited, but every white-collar worker may need an Agent that can extend its capabilities through code.
Further out, AI may not need to keep using software designed for humans. It is faster and can process more information at once, so it may design specialized tools and continuously modify them. Eventually, AI could iterate on its own tools because it considers human-designed tools difficult to use.
21. Model and Application Companies Will Coexist, but Large User Bases May Constrain Frontier Exploration
Model companies can directly improve capability through RL and use stronger models for fine-tuning, giving them a base-layer advantage. Application companies can combine multiple models, exploit their different strengths, and build user mindshare around specific workflows.
Perplexity owning “AI search” and Cursor owning “AI coding” show that applications do not need to own the base model. The analogy is Windows: Microsoft makes both the operating system and Office, but Adobe can still become an independent large company on the same platform.
Dai Yusen has no definitive answer to Agent’s “sweet trap,” but notes that high DAU could force models to compromise on size and capability. In China, especially, serving large numbers of free users creates a cost conflict with exploring AGI.
Asked whether DeepSeek should retain tens of millions of DAU, Dai Yusen says not operating it aggressively may be the right choice. If DeepSeek tries to become a super app, it will consume product, operations, and compute resources while making it harder to remain a neutral partner for platforms such as WeChat.
22. Multimodality May First Pay Off in Use and Generation, Not General Intelligence
Dai Yusen sees language as highly compressed intelligence: Newton’s laws can be expressed in a few sentences, while a model may need to watch enormous amounts of video to summarize the same rule. The language route will therefore deliver faster intelligence gains first.
Multimodal understanding remains critical for Agents using webpages, phones, and other digital tools. Multimodal generation already has direct PMF: Midjourney, Sora, Kling, and Hailuo can all serve advertising and content production.
He feels the spectacle of Sora after its official release fell short of expectations, while Google Veo 2 was stronger for single-shot generation. But whether video generation advances general intelligence and whether it can be commercialized are separate questions.
AI R&D will alternate between exploration and charging down a clear path. When the direction is obvious, companies with more money and chips have the advantage; when a new vision is needed, startups can still avoid the incumbents. An entrepreneur’s enduring advantage is doing what giants have not yet seen.
23. Agent Costs in the US Have Crossed the Threshold for Payment
Dai Yusen’s basic assumption is “make it work first, then make it cheap.” Capability growth may hit bottlenecks, but chip advances, engineering improvements, and scale will reduce the price of equivalent performance.
The substitution math already works in the US. Devin cost around $6–8 per hour at the time, versus a California minimum wage of about $16. In a market with high labor costs and a habit of paying for business services, Agents have a price advantage even while imperfect.
GPT Pro costs $200 per month and can run roughly 100 Deep Research tasks, or about $2 each. Dai Yusen’s comparison is direct: “I can’t ask an intern at 2 a.m. to give me a report in 5 minutes,” and a human report may not be better anyway.
“The future is already here—it’s just not evenly distributed.” People who use frontier models fluently and people having their first Chatbot experience have almost entirely different perceptions of productivity and career risk.
24. The Next Paradigm After RL Is AI Discovering and Validating New Knowledge
Dai Yusen believes RL itself still has a long way to go. The next step is AI not merely organizing existing knowledge, but proposing new hypotheses and designing validations, potentially reaching a higher level in OpenAI’s five-tier classification.
Scientific discovery usually contains “thinking” and “experimentation.” AI may already be strong at the former, but the bottleneck is the latter: experiments in physics, chemistry, and medicine require the physical world, compliance procedures, and time. If experiments can be massively parallelized, the speed of knowledge production will be reset.
AlphaFold, AI drug discovery, and protein targets are early signs, but candidate discoveries still need drug development and clinical trials. Fields such as mathematical theorems, where verification can be purely intellectual, are more likely to first produce a self-reinforcing loop.
Energy is another boundary. Compute is growing exponentially while electricity is not keeping pace; liquid cooling, power supply, and nuclear energy are already under discussion. If AI can also improve energy efficiency or controlled nuclear fusion, it could remove a constraint on its own expansion.
25. DeepSeek Forced Chinese Big Tech to Reconsider Building, Integrating, and Partnering
Tencent integrated DeepSeek into more than 10 products, including Yuanbao and WeChat; Baidu integrated it into Wenxin, while ByteDance did not. Dai Yusen would be surprised if Doubao integrated it because ByteDance’s objective remains building frontier models internally and exploring AGI comprehensively.
Tencent’s choice looks more like “late-mover advantage”: use WeChat’s users and infrastructure first, then integrate quickly once model technology converges and matures. After integrating DeepSeek, some products reportedly achieved double-digit growth, while Yuanbao quickly rose near the top of the App Store.
This does not mean Tencent has permanently abandoned internal development; Hunyuan is still hiring aggressively. What makes DeepSeek unusual is its neutrality: it does not seek to become a super app and has no external-investor ties to a particular big tech company, allowing WeChat, Perplexity, and platforms from different camps to use it.
Apple also looks more like a late mover: it already has the entry point, users, and brand, so it can wait for the technology to mature before integrating it. Dai Yusen considers a partnership with Alibaba understandable because Alibaba has the infrastructure, technology, and operational ability to serve users at scale. He also mentions that Apple has discussed cooperation with DeepSeek and Kimi.
26. R1 and K1.5 Launched on the Same Day, but Open Source and a Global Narrative Created the Reach Gap
Kimi K1.5 and R1 launched around January 20. Early recognition in technical circles was not dramatically different, but their eventual reach was. Open source and direct access for everyone were the decisive differences.
The US market misread the one-time training cost in the V3 paper as “a $5M training run produced an O1-level model,” then compared it with companies that had raised billions. Nvidia briefly fell around 16% on January 27, and Trump commented publicly, turning a technical event into global financial news.
Dai Yusen emphasizes that people inside the industry did not find $5M itself unbelievable. Some even believed a leading US lab, with a larger cluster and more experience, could be cheaper. The real object of study was innovation such as MLA, which lowers training and inference costs.
DeepSeek still created a “catfish effect”: closed-source companies had to move faster, and teams such as Anthropic came under pressure. GPT-4 API costs have already fallen by more than 90%, and he expects comparable costs to drop rapidly to one-twentieth of current levels or lower.
27. DeepSeek’s Biggest Breakout Was Not Math, but Writing That May Be “Under-Aligned”
Reasoning-model papers mainly benchmark AIME, MATH, LiveCodeBench, and similar math and coding tasks. Once R1 reached the mass market, however, its most widely circulated feature was its imaginative Chinese writing.
Dai Yusen has heard 2 explanations: the team may have deliberately strengthened Chinese-writing alignment, or insufficient constraints around truthfulness, neutrality, and hallucination prevention may have unintentionally preserved freer expression. There is no conclusive evidence for either.
Fortune-telling, MBTI, and New Year greetings are not standard AGI tasks, but they best display this creativity. The model readily turns prompts into grand narratives through quantum mechanics, which users experience as a feature.
The trade-off is accuracy. Dai Yusen relays the observation that users still prefer Kimi at work because it hallucinates less. His own summary is that Kimi may be better for getting things done, while DeepSeek is more entertaining for chasing unusual ideas.
28. Business Models Need Not Be Designed in Advance; They Grow from Real Value
DeepSeek can generate revenue directly through its API, and its inference optimizations may support gross margins. Its temporary closure of the top-up channel was not a sign that nobody would pay; compute capacity simply could not keep up. Users asking for a stable paid version is itself proof of demand.
Asked whether every technical breakthrough can naturally be commercialized, Dai Yusen says no. The key is the technology cycle: during steep growth, the priority should be improving capability; when growth flattens, mature capabilities are better suited to business-model construction.
When Google launched in 1998, it became known for clean, ad-free search. In 2002 it was still criticized because “the hardest thing to search for was its business model.” AdWords and AdSense followed, and Google listed in 2004. First came a 10x better experience, then the money machine.
Facebook launched in 2005 and introduced News Feed in 2007. It did not find its core business model until 2012, when it switched to recommendation-based ranking and added feed ads. Dai Yusen’s analogy is not to force a gifted high-school student to move bricks for money immediately, but to let the student grow into a PhD.
29. Open Source Is Not Mandatory; Sustained Leadership and Neutrality Create Compounding Returns
Dai Yusen does not treat open source as a moral standard. A model must first lead—at least reach SOTA or add value for the technical community. Open-sourcing an mediocre model is merely “open source for the sake of open source.”
Free is weaker than “free plus leading,” but free plus leading remains highly valuable. DeepSeek’s true moat is not one release, but sustained open sourcing from its early versions while maintaining a neutral position that does not threaten platforms. Even if Doubao were equally strong and open source, WeChat might not integrate it.
Existing open-source participants such as Meta and Alibaba now face more pressure. Dai Yusen calls them “cyber Buddhas”: large companies contribute the output of their compute to the whole industry, while DeepSeek has pushed openness and iteration speed another step higher.
Regarding whether OpenAI might open-source O3 Mini or an on-device model, he is more interested in O3 Mini. Rumors suggest each activation may use around 3.7B parameters and possibly MoE. If such a small model can reach a high level through longer reasoning, it would be more scientifically interesting than an on-device model, although abuse risk is a real concern.
30. After DeepSeek Cleared the Field, Kimi Is Still Trying to Stay at the SOTA Table
Dai Yusen says R1 cleared the field for the “six little tigers.” Several had already abandoned foundational models and SOTA because of funding, team, or positioning constraints; R1 merely made the choice more explicit.
As an angel investor in Kimi, he openly admits he may be biased, but still judges Kimi to be the only one among the six little tigers or several model startups with the team, funding, users, and technical conditions to challenge SOTA.
His evidence includes continued technical output such as K1.5 and Moonlight, a user base in the tens of millions of DAU, and differentiation in accuracy, multimodal reasoning, and photo-based problem solving. The team has also cut overseas and other directions to concentrate resources.
Kimi’s 2023 decision to focus on long context was not a detour: it unlocked multi-file reading, searches across hundreds of webpages, long chains of thought, and future memory. More important, the co-founders have worked together for years and the team is stable. Entrepreneurship is like walking a tightrope; simply staying at the table can be a major achievement.
31. Open-Source Models Push AI Infrastructure and Applications into an “Android Moment”
After DeepSeek took off, AI infrastructure companies such as Wuwen Xinqiong and SiliconFlow began handling deployment demand from state-owned capital and governments across China, including inference optimization on Huawei chips. In the closed-source era, third parties had little chance to participate in this service layer.
The opportunity is squeezed from both sides. Alibaba Cloud, Tencent Cloud, and Volcano Engine have advantages in capital, energy, customer service, and scale, while DeepSeek’s first-party team has its own inference optimizations. Once demand stabilizes, startups will still need to prove they can do something the major clouds cannot.
Dai Yusen’s ecosystem analogy is that Sonnet, GPT-4o, and O1 were the iPhone moment: powerful closed capabilities first unlocked a large application market. DeepSeek is the Android moment, making sufficiently strong capabilities available to far more developers.
In the “BlackBerry era,” the device could only send email, and Zhang Yiming could not have gone back and built Douyin. Only once iPhones and Android phones had cameras, screens, networks, and chips did short video, ecommerce, and social products become possible. Application opportunities depend on the base layer opening the space of possibilities.
32. Slower Pre-Training Does Not Mean Compute Has Peaked; Agents Could Multiply Inference Demand by 100x
The “brute force works” approach of the past 2 years mainly referred to the pre-training arms race. Grok 3 reportedly used around 200,000 chips and still improved, but marginal returns have declined, so growth in pre-training compute may slow.
New increments are shifting to post-training and inference. A Chatbot answers a limited number of human questions and consumes relatively few tokens; an Agent can think for long periods, repeatedly call tools, and execute complex tasks. Inference demand may rise not 10x, but 100x or 1000x.
Deep Research is early evidence. Even at $200 per month, GPT Pro may still lose money according to Sam Altman, showing that inference per task is far higher than in ordinary conversations.
Asked whether 100x growth will happen in 2025, 2026, or 2027, Dai Yusen says the answer is not important to technology history but is very important to stock trading. His firm view is only that Agents are close to breaking out; the exact year remains uncertain.
33. Nvidia Demand Remains Strong, but Its Very High Share Leaves a Window for ASICs and Domestic Chips
Dai Yusen believes Nvidia remained the strongest option for training and general inference at the time of the interview. DeepSeek still wanted more Nvidia chips: “Buy as many as you can.”
But once the R1 architecture became clear, domestic chips such as Ascend 910B could be optimized for specific workloads. Google TPU and ASIC routes from Broadcom and Marvell can also improve efficiency for stable loads. The more technology converges, the more opportunity specialized chips have.
The opposing risk is architectural change. If Transformer or the O-series paradigm is replaced by a new architecture, ASICs with long development cycles and deployment around 2027 could lose their target. Capacity, yields, and software stacks also do not become usable merely because a design exists.
Nvidia’s more realistic valuation risk is not disappearing demand, but a share already above 90% that is difficult to increase further. A slight change in expectations for share or gross margin could trigger a market revaluation. Jevons paradox benefits total compute, but does not guarantee that all value accrues to one chipmaker.
34. DeepSeek Triggered a Revaluation of Chinese Assets, but Did Not Replace Policy and Fundamental Analysis
Some overseas investors had assumed China lacked chips, talent, and capital in AI. DeepSeek suddenly forced them to acknowledge that China also had frontier capability. They then realized Alibaba was not fully valued and that the market was pricing mainly its ecommerce business, prompting rapid buying.
Investors who had long understood Tencent, ByteDance, and Chinese startups were never especially doubtful of the technology. They cared more about economic stimulus, the business environment, and policy. For them, a private-sector symposium could matter more than a model benchmark.
Hong Kong stock turnover briefly reached around 5x the level of the same period a year earlier, while ByteDance was reportedly revalued at around $400B. Dai Yusen’s warning is that “fear is created by falling prices, and confidence by rising prices.” Neither extreme pessimism nor the 2021 euphoria, when China’s PE exceeded the US, was necessarily rational.
Long-term foreign capital still needs quarterly strategy meetings, data, and internal processes before changing allocations. DeepSeek was a release point after a compressed spring, not the entire causal story; subsequent macro changes, including US investment policy, kept the outlook difficult to assess.
35. 2025 Will Bring More “Lee Sedol Moments” and Expose Distribution and Governance Problems Earlier
Dai Yusen expects AI to surpass 99% of humans on more tasks; programming competitions are already close to that point. When the host asks whether it will surpass 99% of programmers, he cautiously corrects himself to “humans now,” but believes programmers may follow soon.
Agents could produce a breakout phenomenon on the scale of Cursor. Cursor had around $100M in ARR at the time, showing that AI products need not be judged only by DAU; how much users are willing to pay is another measure of value created.
If AGI is defined by how much work it can replace, unemployment is almost built into the definition. Whether UBI and material abundance can absorb the shock has no answer. The Industrial Revolution raised everyone’s absolute floor but did not automatically reduce relative inequality; AI could follow the same pattern.
Dai Yusen supports allowing new technology to develop first and governing it afterward rather than requiring proof of absolute safety in advance. Cars and electricity caused accidents in their early years, and the Luddites could not stop industrialization. But deepfakes, fraud, and Agent privilege escalation still require stronger AI defenses. “Acceleration” does not mean abandoning governance.
36. The Most Important Preparation Is Learning to Ask Questions, Create Unique Information, and Try the Future Yourself
With Deep Research, scarce ability shifts from “knowing how to execute” to “knowing what to ask it to do.” Education will lose focus if it continues training skills AI can replace; asking good questions, defining goals, and creating unique data not present in the model matter more.
Dai Yusen’s learning method is hands-on experimentation from day 1. He used ChatGPT until 4 a.m. on launch night, tried Devin, DeepSeek V3, and MLA immediately after they appeared, and organized discussions around them. Spending the price of a bottle of Moutai to see what future products look like is itself a high-return form of learning, he says.
Deep Research has entered his investment workflow. He researched the market reaction to Trump’s 2018 tariff announcement, found that safe-haven buying pushed Treasury yields lower, and used the analysis to support a Treasury purchase. When reading about the policy history of GPS, from restricted accuracy to open access, he also had an Agent research the background while he continued reading.
He recommends A Brief History of Intelligence and 《第一只眼》. The former traces the leap from life on Earth to GPT-4; the latter uses the “light-switch hypothesis” to explain how eyes may have triggered the Cambrian arms race. AI is entering its own Red Queen race, while humanity may still be on Day One of the intelligence revolution.