124. Yusen's Venture Capital Observations, Episode 1: 2026 Expectations, The Year of R, Pullbacks, and How We Bet

2025/12/13 · Source link · AI Summary archive

Summary

2026’s keyword is “Year of R.” Return, Research, and Remember. 戴雨森’s core logic: the progress of the new SOTA over the previous generation is slowing (think GPT-4 versus GPT-3, compared with GPT-5 versus GPT-4), while investment is growing exponentially—and roughly 50% of data-center investment is compute that will be obsolete within 4-6 years. “As everyone invests more, the focus on returns will only grow.”
The clear market call: US equities could see a meaningful pullback next year, potentially in the second half. The trigger would be a softening US labor market combined with AI returns falling short of expectations; “OpenAI’s own user and revenue growth will be a very important trigger for this broader concern.” He has essentially cleared out his secondary-market equity positions, but does not short: “Shorting is wanting other people to become miserable.” He only goes long, and can reduce exposure.
The bubble debate directly counters 朱啸虎’s claim that “there will be no bubble for three years; that’s pure nonsense.” “Every technological revolution in human history has brought a bubble, almost without exception.” AI is the most important revolution, so “it is perfectly natural that it could bring the biggest bubble in human history”—but we are not at the top yet. The peak comes when bad and fraudulent companies are also expensive; Nvidia’s short-term valuation is even low. The risk lies in a mismatch between the long and short ends: once long-term AGI expectations turn bearish, “even excellent short-term results may no longer be the main trading factor.” On the token thesis: usage rising 10x a year is inevitable, but “rising token volume does not necessarily guarantee the arrival of returns.”
Models have no secrets, and selling APIs is not a monopoly business. Silicon Valley has no non-competes and talent moves quickly, so “there are no secrets that can truly be kept for the long term.” Chinese open-source models have compressed the gap to 6-12 months at one-tenth to one-fifth the cost; “it is basically the age of Chinese models now.” First-party products are becoming increasingly important to model companies (Claude Code reached several hundred million dollars of ARR within months, and Dario was willing to compete directly with its largest customer, Cursor), while thin-shell applications where “the model gives you code and you give the user code” will face sustained pressure.
All four monetization paths are slower than consensus expects. Subscriptions are hard to reprice (Netflix has barely raised prices in 20 years; “the model you bought for $200 last year can now only be sold for $20”); advertising and e-commerce are existing pools of spending (Google took 4 years to find AdWords and Facebook 6 years to develop feed ads); replacing programmers is value deflation, not wage transfer (Jevons paradox); and enterprise adoption is slower than expected (Office Copilot remains below expectations). Sequoia’s David’s $200B question has inflated into a trillion-dollar question—the elephant in the room, with no good answer yet.
The China-US AI valuation gap is the biggest option in global allocation. Thinking Machines’ $50B valuation “is greater than the combined valuation of all Chinese AI startups”; Mistral at $14B versus Kimi below $4B; high-growth US applications receive 30-100x ARR, versus roughly 10x in China. The allocation is a barbell: OpenAI/Google/Anthropic at one end for stability, Chinese companies at the other for maximum optionality and potentially 100x returns. “I think the gap has already passed its widest point.”
For founders, the era of growth through negative-margin token resale is over. Silicon Valley now expects applications to have SaaS-like gross margins above 50% and retention; founders need a high-quality view of the SOTA 6-12 months out (杨植麟 on long context, Manus on agents), must target global markets from day one, and must “stay at the table and keep turning over cards”—Manus only worked on its third attempt. Memory, multimodal generation, and voice will be the key battlegrounds in 2026.

Deep dive

1. 2025 in Review: AI Met Expectations, but the Biggest Upside Surprise Was China’s Venture-Capital Revival

戴雨森’s year-end self-assessment: AI’s development “basically matched what I had expected”—the chain from technological progress to product deployment to monetization has been validated. ChatGPT revenue grew rapidly, Anthropic’s API revenue grew rapidly, Cursor and Claude Code approached $1B in coding ARR, and Manus and Genspark were moving toward more than $100M in ARR as general-purpose agents.
The real upside surprise was the revival of China’s venture-capital ecosystem. A sharp rebound in the secondary market created a positive exit effect; DeepSeek emerged at the start of the year; and robotics, optical modules, and other semiconductor-linked industries proved globally competitive. “The number, amount, valuation, and intensity of competition for investment projects all improved substantially.”
The US was equally exuberant, producing new types of companies such as Thinking Machines, which reached a $50B valuation “with no product, or basically no product,” along with a group of $10B-scale startups valued at roughly 50x ARR.

2. The “Smart Elementary School Student” Starts Making Money: The Agent Year Delivers

Last year’s analogy was that “AI was still learning rapidly. It was a very smart elementary school student, but an elementary school student might not yet be able to earn money by going to work.” In 2025, that student started handing in assignments.
The three key technical advances identified at the beginning of the year all arrived: improved reasoning from the O series through thinking-time scaling, better programming ability represented by Sonnet 3.5, and Computer Use. Together, they unlocked the Year of the Agent: L3 coding agents such as Claude Code and Codex that can program autonomously for more than an hour, alongside general-purpose agents such as Manus and Genspark.

3. Three Calls from Last Year: Time-Saving Beat Time-Killing, Humanoids Continued to Disappoint

Chinese internet companies made money over the past decade through time-killing applications, but “compared with applications that kill time, AI applications that save time are actually more important.” A year later, companion applications had not reached expectations, while the major products were largely productivity-oriented.
Caution on humanoid robots proved justified: “Changing the digital world has always been somewhat easier; we have consistently overestimated our ability to change the physical world.” Optimus and others fell short on manipulation, and embodied intelligence “will definitely be pushed further out,” although VLA and products such as Sunday offered some interesting early signals.
General-purpose hardware intended to replace the smartphone—Humane, the Friends pendant, Rabbit—remains difficult. The original call stands.

4. Agents Are Still Stuck in the Prosumer Market

Deploying agents inside large companies requires extensive data access, privacy controls, and permissions. “Agent penetration inside large companies is actually very low.” Current agent products serve several hundred million knowledge workers worldwide, professionals, and small and midsize teams. The episode repeatedly returns to the same point: rapid penetration of the early market does not mean the mainstream market has arrived.

5. The Model Layer: Thinking-Time Scaling and More “Lee Sedol Moments”

The technical pattern is bottom-layer breakthroughs followed by scaling, which lifts performance and expands use cases. O1, released in October 2024, unlocked the reasoning path: GPQA rose from the 20s to nearly 90, surpassing human PhDs; SWE-bench rose from below 5 for GPT-4 to nearly 80; and Humanity’s Last Exam moved from single digits to the 30s and 40s.
By year-end, OpenAI and DeepMind’s general models both achieved IMO-gold-medal-level results, while DeepSeek Math V2 delivered open-source IMO-gold-medal-level performance. “More Lee Sedol moments are arriving.” In programming, mathematics, and commonsense reasoning, AI is surpassing the strongest individual humans, which “will bring many changes to how we understand the world.”

6. Multimodality: From the Ghibli Moment to Nano Banana Pro

GPT-4o created the Ghibli moment at the start of the year: once image-generation models became better at following instructions, user scenarios expanded sharply. Nano Banana and Nano Banana Two Pro can turn long text into infographic posters—“even explain it to you like Doraemon.” VO3 and Sora 2 delivered synchronized audio and video, longer durations, and better physical consistency.
Sora 2’s significance is not retention—“it is not a place for consuming content”—but that it let many people see for the first time that “the boundary between the virtual and the real is dissolving.” It was “a moment that felt almost magical.”

7. In One Year, Open Source Became China’s Model Market

In 2024, people worried about China’s shortage of chips and the concentration of talent in the US. DeepSeek showed that China could reach near-world-SOTA performance at one-tenth or less of the cost. In the second half, Qwen, Kimi K2, GLM, MiniMax, and ByteDance all showed this was not unique to DeepSeek: within months of a US SOTA release, China could produce an open-source model with similar performance, one-fifth to one-tenth the cost, and genuine usability.
At the start of the year, the open-source ecosystem was still led by Llama. “Now it is basically the age of Chinese models.” Cursor’s Composer One is widely believed to have been trained on Chinese open-source models, while all-in investor Chamath said publicly that he had shifted substantial usage to Kimi K2 because it “performed very well and was very cheap.”

8. Models Have No Secrets: Benchmark Saturation and “High Scores, Low Ability”

Why does the ranking keep changing, with no one opening up a meaningful lead? “Information and talent move quickly in this industry, especially because Silicon Valley has no non-competes. There are no secrets that can truly be kept for the long term.” Everyone has roughly similar data and tens of thousands of GPUs.
Benchmarks are also saturating. The latest example may be Opus 4.5 (heard in the recording as “OPENS 4.5”), at around 80 on SWE-bench, versus roughly 78 for Gemini 3 Pro. “It looks like only a two-point difference,” but developers find the former much more usable. 戴雨森’s Tsinghua analogy is that some people are “high-scoring but low-ability.” Model benchmarks need to evolve; once they saturate, the real differences appear only with developers and users.

9. The “More Compute Solves Everything” Thesis Is Under Pressure, Alongside Ilya’s Alternation Theory

The crude 2023 interpretation of scaling laws was: “Put in enough GPUs, train a model, establish a moat, and capture all the money that follows.” Both parts are now under pressure. More compute has not produced a single clear winner; the leading group is still trading places. Nor has the capital intensity created a moat that later entrants cannot cross—Chinese companies are closing the gap rapidly.
Ilya’s framework became the episode’s theoretical anchor: “AI history has always alternated between scaling and research.” Improvements within the current paradigm have become marginal, making whoever finds the next scalable paradigm important again. Demis’s latest view is that one or two more research breakthroughs are needed to reach AGI.

10. Three Narratives Broken by Chinese Open Source: 100,000 GPUs, Talent, and Distillation

Narrative one: “You cannot build a good model without a 100,000-GPU cluster.” When catching up, many experiments do not need to be repeated; fewer resources force more disciplined optimization, which “is what produced efficiency breakthroughs such as MLA.”
Narrative two: “All the best talent is in Silicon Valley.” It became “US Chinese versus China Chinese.” “They were sitting at the same table studying together.” Top Silicon Valley talent has even begun returning to China—not often, but enough to be unimaginable two years ago.
Narrative three: “It is all distillation.” Distillation is certainly an advantage for catch-up players, but attributing everything to it obscures genuine technical innovation by Chinese companies: DeepSeek’s MLA and DSA sparse attention, and Kimi’s iterations on the Muon optimizer (heard as “Moon”). The calibrated conclusion: completely overtaking the frontier in the short term is difficult, but compressing the gap to 6-12 months “looks realistic for now.”

11. Selling APIs Is Not a Monopoly Business

The live example is Manus. At launch it used Claude 3.7’s API exclusively—“at that time, only Claude 3.7 could make it run.” It soon found that another model (heard as “Dreamline,” identity unclear) also performed well and was cheaper, so it shifted a large volume of calls quickly. Once Codex arrived, Anthropic was no longer uniquely dominant in coding. “For an application, switching between APIs is actually quite easy.”
The reversal benefits application companies, which can “select the best features from every provider” and use different models for different tasks. That is itself evidence that the moat around selling APIs is thin. “It is definitely not a monopoly business.”

12. Model Companies Must Invest in First-Party Products

The clear trend this year is that Claude, originally “an application it did not spend much effort on,” produced several hundred million dollars in ARR within months of Claude Code’s launch. Owning an application means more user data, more opportunities to build memory, and a stronger model.
达瑞 said in an interview that the company would build first-party products in strategically important areas, “even if that meant competing directly with its largest customer, Cursor.” Being both a customer relationship and a competitor is one of the year’s most interesting dynamics.

13. The Factory Framework: Cursor’s Dilemma

戴雨森’s central analogy is to treat an application as a factory that buys tokens from a model company and processes them into output. In coding applications, “the model company gives it code, and most of what it gives the user is also code. The application’s processing is limited, so the moat is relatively thin.” Applications that perform complex processing and deliver a complete result directly have stronger moats.
The battlefield is no longer Claude Code versus Cursor. It is Claude Code, Codex, Antigravity, and a series of first-party model products versus Cursor, with Lovable in some verticals. The disclaimer remains: “AI is extremely difficult to predict. For any specific prediction like this, you should be prepared to be proven wrong at any time—three years ago nobody expected ChatGPT either.”

14. Why Doesn’t ByteDance Open Source?

“I may not be the best person to answer this question,” but the view is that ByteDance has no financing or customer-acquisition pressure requiring it to build ecosystem influence through open source. “For ByteDance, what may matter more is whether products such as Doubao and Jimeng are products users enjoy using, so whether it open-sources its models may not be that important.” Open source is a way to treat others well and expand influence—a means, not a belief system.

15. Applications: From the “BlackBerry Moment” to Coding Crossing the Chasm

Last year, AI applications were still in the BlackBerry moment rather than the iPhone moment: the technology was not strong enough and the product form was uncertain. This year’s narrative delivered. After Sonnet 3.5, coding applications “crossed the chasm into the mainstream market.” It is “hard to imagine a programmer with even a little information work who does not use AI coding tools,” and the category moved from $100M in ARR toward $1B and beyond.
Better reasoning and fewer hallucinations also drove vertical adoption. In the US, that includes high-value sectors such as law, represented by Harvey, and labor-intensive white-collar work such as customer service (the recording names products resembling “Cara” and “Daihong”). In China, examples include 真格-backed Heihu, AI plus factory manufacturing for rapid response to small orders—“very suited to China’s conditions”—and AI plus education (heard as “Yuan Weiwu”).

16. The Application Moat: From Sashimi to a Manchu-Han Banquet

The first generation of AI applications in 2023 was merely a UI around a model—“like sashimi: fish with wasabi and soy sauce.” Advanced applications now add a context layer outside the model, including real-time context, proprietary industry data, user preferences, and memory, as well as an environment-and-tools layer that gives agents tools and a changeable environment. “Sashimi has become a Manchu-Han banquet.”
The evidence: Manus outperformed all advanced models on Scale AI’s Remote Labor Index. ChatGPT’s model may not be absolutely the best, but retention keeps improving; many analysts attribute this to memory. “The memory layer is actually at the product layer.”

17. Chinese Applications Go Global from Day One

Unlike Meituan, Pinduoduo, and ByteDance, which required intensive local operations when expanding overseas, LLM-based applications “are international from day one.” Languages are native capabilities, and “the needs of an American, a Brazilian, and a Japanese person making a PowerPoint may be almost identical.” More than a dozen of the Chinese companies on Andreessen’s AI Applications Top 50 list were built in China.
Chinese founders also operate with remarkable efficiency. Plaud’s global revenue had passed $100M before it ever raised financing. Shenzhen-based Polo, recently backed by 真格, makes online video generation and reached $20M in ARR without raising money, compared with US peer Higgsfield at roughly $100M in ARR. “Chinese founders are still extremely execution-oriented and very capable.”
This is not an argument against raising capital. “Financing is a time machine for a startup: can you turn something that originally took a year into 3 months through financing?” The first-principles definition of a startup is a company that grows very quickly.

18. Agents Are Not the Year of the Agent; They Are the Decade of the Agent

小珺’s instinct was: “I personally don’t feel there are that many agents.” 戴’s response was that a first year is still only the first year. Citing Karpathy’s blog and the history of autonomous driving: “Agents are not The Year of the Agent; this is the Decade of the Agent.” Karpathy drove an L4 vehicle for several miles more than a decade ago and thought practical deployment was near, but very few companies have delivered robotaxis. “It can quickly give you an interesting demo,” but autonomy is always a long process.
The optimistic quantitative view is that all major model companies are collecting agentic data. People use fewer than 100 common applications on computers and fewer than 20 on phones. “If the marker is AI being able to use the software humans use on computers and phones with considerable freedom, agent capability may improve substantially in a year.” Handling exceptions, understanding tasks, planning, and feedback still make the decade a more accurate frame.

19. Coding Agents Are an Open Secret; Startups Need Data Models Do Not Have

“Every model company is working desperately to improve coding ability. Coding is a language.” It will be difficult for a startup entering now. The opportunity lies in moving beyond the current paradigm: tasks humans have never written, or weak back-end systems. “Everyone doing web coding is decorating the outside of the house; is the plumbing inside the house good enough?” Overall, however, the opportunity still favors the large companies.
The vertical thesis is even harsher than 朱啸虎’s version. “Other people not having the data is obviously the right answer. The problem is identifying what data they truly do not have, and that is not easy.” Healthcare is a counterexample: most medical data is public, and clinical consultations largely reproduce public knowledge; current models are already strong. Education is the positive case: two students learning the same quadratic equation need different teaching methods, and that data is not inside the model. A vertical is defined by delivering value the model does not possess, not by packaging and reselling the model’s value.

20. The Contrarian Opportunity in General-Purpose Agents: OpenAI Built One, but Didn’t Do It Well

朱啸虎 says startups should only build vertical products and stay three blocks away from big tech in the AI era. 戴’s response is direct: “When everyone believes there is no opportunity in general-purpose products, there may be a contrarian opportunity.” When investing in Manus, many investors thought OpenAI would make it impossible. But OpenAI did build one, and after ChatGPT Agent launched, people found it was first, not very usable, and second, not especially strong.
There are two micro-level differences. A chatbot is not the best interface for an agent—Claude Code and Codex are standalone applications. Model companies can theoretically do everything, but the difference lies in how many people they devote to a specific direction and whether that is enough. Manus’s wide research can run 100 parallel agent tasks at once, “which no other agent application can do.” Generality needs infrastructure that gives an application an advantage the model company does not have—making one plus one greater than two.

21. The Netscape-Yahoo Argument: You Cannot Skip the Earlier Dumplings Because You Know the Ending

“If you were an angel investor in Cursor, would you feel very sad, or would you actually be happy because you could make a lot of money?” Netscape and Yahoo were the best-known companies of the early internet and were eventually displaced. “But can you say that investing in Yahoo or Netscape was meaningless? I don’t think so.”
The methodological conclusion is: “You cannot say this is what the end state will look like, and therefore refuse to explore the process from here to the end. I ate the last dumpling and got full, so should I stop eating the earlier dumplings?”

22. 真格’s Pace: The Best Companies Are Funded in Bearish Periods

真格 invested in roughly 20 companies in 2025 and kept a steady pace: “When the market is very bearish, we still believe excellent people are starting companies. When the market is very hot, we also believe there are not that many truly excellent people.” Most of the AI companies now performing well were funded in 2023 or earlier: Manus in 2022, HeyGen (heard as “Heijian”) in 2021, and Genspark and the AI education project in 2023; Kimi was also funded in 2023.
His self-assessment includes some regret: the fund could have taken larger stakes in leading companies. Manus is doing well, but 真格 is not the largest shareholder in the other companies. The fund remained cautious on robotics; now that the sector is hot, some companies are preparing to list. “Whether the capital market is hot is not something we can really judge.”

23. The China-US Valuation Ledger

In venture rounds, Thinking Machines raised at a $10B valuation in its first round and $50B in its second, with essentially no product. “The $50B valuation should be greater than the combined valuation of all Chinese AI startups”—model companies together are worth just over $10B and application companies just over $10B.
At the model layer, Mistral’s latest round valued it at $14B even though it has essentially stopped doing pretraining; its benchmark peers are Kimi, Qwen, and DeepSeek, while Kimi’s last round valued it at only slightly above $3B. At the application layer, a Manus-level company—$100M in ARR within months, gross margins in the tens of percentage points, and 20% monthly growth—would receive a $3B-$5B valuation in the US. Polo reached $20M in ARR without financing and was valued at $80M in its first Chinese round. The US pays 30-100x ARR; China pays 10x.

24. The Barbell: Maximum Stability Plus Maximum Optionality

For a US investor able to bet across the full spectrum, “a typical strategy is a barbell.” One end holds the safest names—OpenAI, Google, and Anthropic. The other holds the optionality of Chinese companies: “Valuations are very low. If China produces the ByteDance of the AI era, there could be opportunities for more than 100x returns. Is another 100x return possible from OpenAI now? It may be difficult.”
The conclusion is convergence. The China-US valuation gap should narrow from both sides: two years ago the trade may have been one to 100; now perhaps five to 100; ultimately, 30 to 100 or 50 to 100 would both be reasonable outcomes to expect.

25. AI Venture Capital Resembles a Semiconductor Revolution More Than the Mobile Internet

Mobile internet was essentially the internet becoming mobile: distribution changed, but the underlying technology—TCP/IP—did not. “An app from 10 years ago can still be used today.” AI still extends mobile internet at the distribution layer, but its technology resembles the integrated-circuit revolution: heavy capex, intensive research, and clear generational cycles.
The structural difference is that internet platforms have network effects and therefore a first-mover advantage: invest heavily in the leader first. AI is a research contest. “You explored ahead of me, but I can use my capital and data advantages to surpass you later”—a second-mover advantage. Traffic is also a mature market: everyone is trying to take time from Douyin and Honor of Kings. The era when Instagram was acquired for $1B with 13 employees is over.

26. The Next Zhang Yiming May Already Be a Name You Know—But First He Has to Compete with Zhang Yiming

“If a new generation of epoch-defining founders like Zhang Yiming, Wang Xing, and Huang Zheng emerges 10 years from now, we may already have heard his name.” This is one reason everyone is suffering from FOMO and investing broadly. But 戴 punctures the assumption: “There was no Zhang Yiming competing with him then. Everyone was trying to become Zhang Yiming. Now everyone first has to compete with Zhang Yiming.”
A footnote to the idea that outcomes are never obvious in advance: ByteDance understood information distribution and recommendation, but executed through massive trial and error. Many people once looked down on Xiaohongshu and thought it could not become large. Building an app is easier now—web coding is enough—but giant-company founders are “young and formidable”; Sergey Brin personally directed code red. “AI has another interesting feature: when everyone believes AI will disrupt them and that it is a huge opportunity, the difficulty for startups only rises.”

27. Year of R, Part One—Return: The Investment Math Is Getting Harder

The bet over the past few years was that “investment in AI would continue to outperform expectations,” which attracted capital through the prospect of enormous returns. But AI investment is not like building highways or laying fiber. Roughly 50% of data-center investment is compute, and it will be obsolete in 4-6 years. “This investment needs to show realized returns over a shorter payback period.”
The pressure is coming from both ends. At the model layer, the improvement of the new SOTA over the previous generation is slowing, investment has grown dramatically, and nothing has stopped Chinese open-source companies from approaching rapidly at low cost. Marginal returns are falling. At the application layer, last year’s low expectations and high growth have turned into high expectations this year—Meta’s daily bidding for talent is evidence of how high future expectations are—while the AGI timeline is moving out.

28. The Subscription Ceiling: Netflix Has Barely Raised Prices in 20 Years

Subscriptions are the largest source of AI revenue. OpenAI’s ARR is approaching $18B, with nearly 600M DAU and 1B MAU. “As an application that helps people acquire knowledge, its penetration is not low anymore.” Penetration among knowledge workers willing to pay for AI is already relatively high.
The pricing paradox is that “the model you bought for $200 last year or the year before can now only be sold for $20.” Better intelligence and falling token prices undermine the story that smarter means more expensive. Netflix is the comparison: 20 years ago it cost around $17 per month; today the high tier is just over $20 and the low tier is a few dollars. The value delivered has increased enormously, but the price has barely moved. With Gemini, Meta, and xAI competing, large subscription price increases for ordinary users are difficult.

29. Advertising and E-Commerce Are Existing Pools—and Slower Than Expected

“Advertising plus e-commerce is mostly a redistribution of an existing pool.” Online advertising is proportional to GMV, and neither pool will suddenly double. “Dividing the existing pie does not create new productivity.” That is one reason Meta, Google, ByteDance, and Tencent are highly alert to large-DAU chatbots and are investing heavily.
The process will not be fast. Google took years to find AdWords and AdSense; Facebook took 6 years to invent feed ads. Advertising also conflicts inherently with subscriptions: if users pay, why should they still see ads? The recent report that OpenAI entered code red and would focus on model capability rather than advertising runs against the market’s expectation that ads will contribute meaningful revenue in 2026. “I think the pace will fall far short of today’s optimistic expectations.”

30. The Programmer-Wage Fallacy: Replacement Does Not Mean Capturing Their Wages

The popular narrative is that more than 100M programmers worldwide earn $100,000 a year, creating a $5T pool for AI to capture. 戴’s rebuttal is the episode’s sharpest point: “If you replace many programmers, it does not mean you can capture those programmers’ wages. It means the tasks those programmers used to perform become less valuable.” Work worth $100,000 can be done by a $200 AI. “It is a process resembling deflation in value,” like telephone operators whose work and wages disappeared together.
Jevons paradox cuts both ways. Lower token costs will inevitably increase usage, but “10x more usage versus a 10x fall in price” is not automatically positive. Most non-SOTA tasks will be priced at a flat rate; only frontier tasks can command high marginal usage fees. “In the short term, the outlook is not actually that optimistic.”

31. Enterprise Services Are Slow, and the Trillion-Dollar Question Has No Answer

Office Copilot is the most direct use case, but “penetration and revenue have consistently been below expectations.” Fortune 500 companies adopt new technology more slowly than people think; data, privacy, and permissions are all bottlenecks. Once the early market reaches a certain level of penetration next year, AI enterprise-service growth may also slow.
The ledger has grown from Sequoia’s David Cahn’s $200B question to a $600B question. “Now it is not just a $600B question; it is a question of more than $1T.” The industry would need to generate hundreds of billions of dollars in additional annual returns. “This is an elephant in the room, an increasingly serious problem, and nobody has a particularly good answer.”

32. The Signal for Founders: Negative-Margin Growth Is Over

The early-year Cursor narrative was that growth mattered more than margin, even if growth came at a negative margin—buying $1 of tokens and selling them for $0.50. Silicon Valley’s focus has shifted to “good margins, good retention, and quality growth.” Model companies can be understood to burn cash, but application companies process tokens and sell added value. “People will still expect applications to have SaaS-like gross margins above 50%.”
The outlook is dialectically optimistic: “When everyone needs to look at returns, companies that truly create value will stand out more.” When returns did not matter, abundant capital created indiscriminate competition and forced good companies to compete on the same terms.

33. Year of R, Part Two—Research: New Labs Are Silicon Valley’s New Bet

Where will returns come from? “Over the past two weeks, people have broadly pointed to another R: research.” That follows Ilya’s alternation theory. Silicon Valley is producing a new species of startup: research-led organizations centered on researchers and differentiated from the leading model companies. Thinking Machines, Ilya’s SSI, Reflection, and more recent names such as Humans&, Periodic, and Isara (heard phonetically) have launched with billion-dollar valuations and hundreds of millions in funding.
Research requires a permissive environment, freedom to explore, and no KPIs or deadlines. The essay Staney (likely) wrote at OpenAI, “Greatness Cannot Be Planned,” reflected the research-lab model. “But OpenAI is no longer a research lab. It is primarily a product-driven startup driven by users, revenue, and matrixes.” New labs are fundamentally trying to become the next OpenAI and occupy the next paradigm shift.

34. The Second Half of Benchmarking: The People Who Set AI’s Problems Matter Most

Benchmarks are optimization targets for training and move in tandem with capability improvements. SWE-bench is the coding example. “The people who set problems for AI are very capable.” But existing benchmarks have largely been exhausted, and “it is getting harder for humans to benchmark an entity smarter than themselves.” 戴 strongly agrees with 姚顺雨’s idea of a second half: new benchmarks are needed to measure and guide training.
This can become a moat. “A benchmark is a bit like the national college-entrance-exam committee. Once everyone has exhausted the existing questions, deciding what questions to ask next becomes an important internal moat for each company.”

35. Year of R, Part Three—Remember: Real Memory Does Not Need a Notebook

Memory is the key to differentiation in AI applications. “If you have no memory and no personalization, the answers you get from asking AI and the answers I get from asking AI are the same.” Even at its current primitive stage, ChatGPT’s answer to him is much more personalized than Gemini’s because it has 3 years of chat history.
But today’s memory is based on retrieval. “Imagine that your friend’s understanding of you comes entirely from carrying a thick notebook recording every sentence you have ever said. Real memory surely means he does not need to carry that notebook.” It should operate at the weight level, through online learning, and genuinely understand you. “Ultimately, the AI each of us uses may be a model specifically tuned to us.” This is a core battleground for research.
Memory is also the prerequisite for proactive agents. “A good secretary should be able to read the boss’s mood and solve problems proactively before the boss says anything.” The key battleground for AI applications in 2026 is acquiring as much user context as possible and turning it into good memory.

36. The Fourth R That Does Not Quite Fit: Multimodality

“I spent a long time thinking about whether I could express it with an R, but I could not find one—it would be forced.” Multimodality has two branches: reasoning and generation. On the generation side, the analogy is that GPT-3.5 unlocked ChatGPT and Sonnet 3.7 unlocked Cursor. “What large application opportunities could advances such as Nano Banana Pro and VO3 unlock?” Generating images inside a chatbot is not a natural experience. In verticals, could it disrupt Canva? Creator and advertising-material generation represent $10B- and $100B-scale opportunities.
On the reasoning side, a breakthrough may be close. On Zero Bench, a visual-reasoning benchmark involving tasks 100 humans can do but AI basically cannot, the best current model scores below 10. “But I hear that frontier labs may have models capable of scoring 60 or 70.” A breakthrough in visual reasoning would substantially expand interaction and what AI can do.

37. Responding to Sequoia’s “Year of Delay”: Agree on Delay, Be Careful with Extrapolation

On David’s two-part forecast—data centers delayed and AGI pushed out, but applications continuing to grow rapidly—the physical-world delay is inevitable. Building data centers and power in the US is difficult, and upstream companies such as TSMC are unwilling to increase capacity tenfold for downstream enthusiasm.
The AGI delay is consistent with 戴’s view that a research-paradigm breakthrough is needed. But “continued adoption growth” must be disaggregated. Only two products have truly crossed the chasm: chatbots and coding; AI coding penetration among programmers worldwide is probably above 20%. AI already has distribution channels, unlike Pinduoduo, which had to wait for smartphones to reach lower-tier markets, so the early market spread extremely fast. “That rapid early-market expansion can easily make everyone overly optimistic about mainstream-market progress.” When early users are exhausted and mainstream users lack motivation, growth can fall off a cliff. That is the chasm.

38. The Bubble View: The Biggest Revolution Deserves the Biggest Bubble, but We Are Not There Yet

“Every technological revolution in human history has brought a bubble, almost without exception—canals, railroads, highways, the internet. AI may be the most important technological revolution in human history, so it is perfectly natural that it could bring the biggest bubble in human history.”
The peak is marked not merely by good companies becoming expensive, but by bad and fraudulent companies becoming expensive too. The internet bubble is remembered for companies whose business models did not work but whose valuations were enormous. We are not there yet; Nvidia’s short-term valuation is even low because growth is so fast. “I do not think it is objective to say there is no bubble.”
The trading mechanism is a mismatch between long and short horizons. In early 2023, many sophisticated secondary-market investors shorted Nvidia because of its inventory, but were blown up when the long-term picture changed. Today the reverse is possible: the short term is highly optimistic, with strong Nvidia expectations for next year and the year after. If the AI payback period lengthens and AGI slows, “even excellent short-term results may no longer be the main trading factor” once long-term expectations turn bearish.

39. The Specific Pullback Path: Second Half, Labor Market, OpenAI as Trigger

“My own feeling is that next year, if we use US equities as the representative market, there could be a meaningful pullback, with the most significant timing potentially in the second half.” The conditions are a softening labor market combined with AI returns falling short. “OpenAI’s own user and revenue growth will be a very important trigger for this broader concern.”
Bearish sentiment has not yet fully transmitted. “If it had, the stock market would not look like this.” The current market is trading concern about OpenAI: after Oracle and others announced large OpenAI deals, people asked how it could finance $1.4T of commitments with a valuation of only $500B and perhaps $100B raised. After Gemini 3 launched, OpenAI shadow stocks such as SoftBank and Oracle fell sharply. Oracle fell back to where it traded before the huge IPO-order announcement—“the market believes it has erased all of that.”
The stance is strong opinion, weakly held: “If a major new monetization scenario appears next year, or model capability improves substantially, I will be very happy to change my view.”

40. Good Bubbles and Bad Bubbles: The Taxonomy of Boom

A good book read this year, Boom, divides bubbles into two types. Bubbles based on the belief that the future will be much better than the present—canals, the internet, AR/VR, and AI—are good bubbles. “That is what we do every day as VCs: invest in the future.” Bubbles based on the belief that the future will be exactly like the present—subprime loans that could supposedly always be repaid, Chinese property that would always rise, and Moutai—are bad bubbles.
The difference lies in externalities and participants. Good-bubble investment becomes fiber and startup capital, and the participants are mainly professional players such as VCs and founders. Bad bubbles involve ordinary people and do not encourage changing the future. “Bubbles also reflect people’s expectations and ambitions for the future. Good bubbles are sometimes necessary.” The task is to avoid obvious bubble losses in specific investments and startups.

41. Direct Response to 朱啸虎: 10x Token Growth Does Not Justify Valuations

小珺 relayed 朱啸虎’s view: “At least for 3 years, I do not see a bubble. When everyone is talking about a bubble, the bubble is definitely not here; these arguments are pure nonsense. Token consumption exploded more than 10x this year and will explode more than 10x again next year.”
戴 does not evade the response: “Token consumption will definitely continue growing very rapidly. A 10x increase every year is inevitable. But token prices are also falling.” Rising token volume cannot justify saying current valuations are appropriate. Usage does not necessarily guarantee returns—on the internet, there have been many cases where bandwidth usage rose but profits did not. As for the claim that “if everyone says there is a bubble, there is no bubble,” ask investors and you will find that the question remains highly contested.

42. Flat Positions, No Shorts: Values and Leverage

On his personal exposure: “I am basically flat right now.” But he is explicit about not shorting: “Shorting is wanting other people to become miserable. I do not think that is a good thing from a values perspective.”
Technically, “shorting is essentially leverage. You can only make 100%, but you can lose infinitely, and you will be nervous.” His approach is therefore to go long only, while reducing position size when necessary.

43. How to Invest Through a Pullback: Do Not Time the Market; Invest in People as Margin of Safety

A prolonged secondary-market pullback would transmit to private markets. US primary-market financing is already optimistic enough to value an angel round at $10B; upside is limited, but downside remains substantial. 真格’s response is unchanged: “We cannot determine market prices. Our principle is always to invest in the best founders, and to invest in the first round. That is our margin of safety.” It echoes Buffett’s idea of buying good assets at a reasonable price.
History supports the approach. Amara’s Law—heard as “Mara Law”—works every time: overestimated in the short term and underestimated in the long term. “The pullback will come, and the long-term development will also come.” Great companies are often founded early: Amazon and Google were founded before the internet bubble burst, but Amazon’s stock fell 95% and nearly went bankrupt. Survival itself becomes an opportunity when macro conditions clear the field. The mindset: “When everyone is extremely optimistic, prepare for pessimism; when everyone is extremely pessimistic, prepare for optimism.”

44. The Extreme Case: If the Paradigm Does Not Break, What Can Existing Technology Support?

小珺’s stress test was what happens if technology stops improving and there is no new paradigm. 戴’s answer is not that improvement stops, but that it shifts from exponential to incremental linear gains. That would be enough for L2: a white-collar worker sitting at a computer and following a process should be replaceable, because most office work solves problems similar to ones already solved.
L3—AI making autonomous decisions and bearing the consequences—requires much longer. “L3 and L4 autonomous driving are actually relatively simple tasks in a two-dimensional world, and that took more than a decade.” So the expectation that AI will make people unemployed or eliminate the need for work does require further improvement in model capability.

45. Four Rules for Founders: Foresight, Global, Stay at the Table, Vertical

First, make high-quality judgments about technology. “Your application is being built for the SOTA 6-12 months from now.” 杨植麟 anticipated long context in 2023 and agentic systems in 2025. Before building an agent, the Manus team researched AI-controlled browsers; Pig’s arrival brought a deeper understanding of the model frontier. An application founder does not need a technical background, but must have foresight—or hire someone who does.
Second, build for the global market. 肖弘 had never been to the US before Manus launched. The key is not living abroad but understanding global product design, aesthetics, and operations.
Third, stay at the table and keep turning over cards. “New cards are dealt every year.” Manus was the third attempt: Benchmark, then Monica, then Manus. Typeless moved from a Shopify plugin to Max AI to voice input. The company heard as “Heijian” moved from e-commerce model face-swapping at Surreal/诗云科技 to virtual humans. “Do not become fixated on one thing. Keep iterating. If you stay at the table, many new opportunities will appear.”
Fourth, a vertical moat can include distribution. Abridge does hospital transcription in the US, and major investor Epic supplies software to most US hospitals. “A hospital cannot simply install a chatbot and start using it.” A unique distribution channel is an important source of vertical value.

46. Reading People: Four Quadrants, Age Is Irrelevant, Invest Only in Leaders

真格’s internal framework has four quadrants: prodigy, veteran driver, operator, and scientist. “杨植麟 was a prodigy then and is a veteran driver now.” In response to the rumor that it invests only in founders born after 2000, it did have nearly half of its founders in that cohort this year—they are more AI-native and need smaller teams—but it also invested in 景鲲 and 张怀林, both born in the 1970s. Founders of companies nearing listing are also in their 70s; 肖弘 and 杨植麟 were born in 1992 and 1993. “Age really is not a distinguishing factor.”
The harder filter is leadership. “In every technological revolution, the people who truly make big money are generally the first wave, or even the first founder to propose the new form.” In the US, acquisition exits mean companies outside the top two can still make money. “In China, if you are not in the top 3, or even number one, perhaps nobody has much chance of making serious money.” The internal keywords are dark horses, leaders, and key makers. There are already 150 humanoid-robot companies; “there are always many companies following the crowd.”

47. Teams: Assembled Coalitions Are Cracks; Roommates Are a Vote of Trust

As the pace of change accelerates, team composition matters more. Be wary of teams assembled opportunistically—people who are unprepared and pushed into a startup by the wave. The counterexamples are consistent: Meituan’s core team included Wang Xing’s classmates and roommates; ByteDance’s founding team included Zhang Yiming’s roommates; miHoYo was built by classmates and roommates; Kimi’s 植麟 and Tim were Tsinghua classmates who even played in a band together.
After watching the film F1, the observation was that “starting a company is like Formula 1 racing, an extreme test of a team. Any crack in the team, lack of understanding, or insufficient trust can easily cause the whole thing to break.” After coming out of top schools and big tech, the question is how many excellent people around you are willing to work with you while everything remains uncertain. That is fundamentally a test of leadership and persuasion.

48. Manus’s Personnel Story: Every VC Saw Monica, but Nobody Invested

Pig built Mammoth Browser in high school, received investment from 真格, and later sold it to 4Paradigm. He joined 真格 as an EIR to help review companies. “He is more like a chief scientist—extremely passionate about technology—but he really does not want to manage people.” Zhang Tao was introduced through the blogging community in college and worked with 老王, Wang Huiwen, at Light Year Beyond. After the acquisition, he waited for “the truly big opportunity that belonged to him” and joined the Monica team as product lead in 2024.
“To be honest, basically every VC in China knew Monica at the time. Many had seen it, but nobody invested. Everyone thought it was just a wrapper.” 真格’s logic was that the team would eventually build something. 戴 offered what he considered an important product suggestion: Manus originally had a Devin-style three-column layout. “I felt ordinary users would find three columns too complicated, so I suggested something more like a chatbot.” The area where AI worked would be collapsed by default and expanded on click.

49. Reflecting on Missing 王星星: Investing in People Can Become a Case of Moving the Boat to Find the Sword

A colleague had met 王星星 in the past but failed to advance the relationship. “This was clearly a major flaw in our philosophy of reading people.” The revised filter is obsession. “王星星 was obsessed with building robots. His robot videos on Bilibili were already popular. He must have stood out very early.”
But honesty matters: “Often we know where Unitree and 王星星 are today and then reason backward to what they were like then. Investing in people is often moving the boat to find the sword—you can only infer early traits from founders who are successful today.” Another red line is character, as is half-hearted entrepreneurship. The key question is intrinsic motivation. For him, “a product that makes ordinary people powerful is great,” and not wanting to be a pawn in a large company made entrepreneurship the natural answer. 肖弘 similarly sees entrepreneurship as “a predetermined way of life.”

50. Extroverts, Introverts, and White, Black, and Gray Horses

小珺 used 朱啸虎’s claim that introverts are not suitable for investment as a prompt. 戴’s observation: “Many great founders are introverts, but they learn the skills of extroverts.” Deep thinking is needed for non-consensus judgments; negotiation and recruiting require extroversion. “Wang Xing, Zhang Yiming, and Huang Zheng were probably all introverts.” Among $100B companies, 马老师, Jack Ma, is an extrovert.
The Club Deal risk framework: white horses are consensus good companies—expensive but with high win rates—and can make money. Black horses attract no investors and are cheap, but can also make money. “The one we must always watch out for is the gray horse: a white-horse price with a black-horse win rate.” 杨植麟’s company had the lowest valuation among large-model companies, so it could be viewed as a black horse. Red was also a black horse; some even called him a grassroots founder. “A Huazhong University of Science and Technology graduate in his 20s who built and sold a SaaS company for several hundred million—how is that a grassroots founder?” The best outcome is a company that is actually a white horse bought at a black-horse price.

51. Four Bets for 2026: Token Usage, Multimodality, Memory, Voice

Four views shared at an AGM with LPs: when token prices fall 10x and usage rises 10x, what makes people use more tokens? Agents are the main driver—when Manus launched early in the year, it used “1,000 times more tokens than a chatbot.” The other directions are video and image generation and understanding, personalization through memory, and voice. “Voice is the most natural medium for communicating with people and computers.” Siri and Alexa were early versions of a killer application, but AI was too stupid at the time. This year, Plaud, Granola, Wispr Flow (heard as “Whisper Flow”), and Typeless reached a magic moment in voice interaction.
The Typeless experience is concrete: press one key, say “I have 3 points,” and it structures them as 1, 2, and 3. Select text and say translate, and it translates. “The larger vision is to do what Siri never did: put only one key between you and a very powerful AI.”
Why have time-killing applications not arrived? To take time from Douyin or Honor of Kings, you face extremely strong competitors. Productivity applications compete against a lack of intelligence. Emotional projection requires changing human nature, which “may not happen in our generation, but in the next.”

52. Surviving an Age of Abundant Intelligence: Agency, Taste, and OOD

“We need to learn how to live in a world where intelligence is abundant and plentiful.” Throughout human history, intelligence was a luxury; only top companies and kings could deploy large numbers of smart people. Now $20 can buy substantial intelligence. As execution becomes cheap, agency and taste become more important: what do you want to do, and which of 3 AI options do you choose?
The mindset must shift from linear to parallel: assign different tasks to different agents and mainly direct the AI laborers—“be a good AI boss.” Move from “I am only missing a programmer” to trying more things. Use disposable code for disposable problems. Always use the most advanced tools to experience the future. He used ChatGPT until 4 a.m. on launch night; “the earliest ChatGPT users, earliest iPhone buyers, and earliest Tesla buyers all experienced the future ahead of time.”
The deepest lesson comes from AI training itself: OOD, or out of distribution. “If what you do and what you think are basically well in the distribution, AI can replicate you very well. Picasso and Mozart produced highly different data, which is why they became masters. How much of our contribution as a collective of humans lies outside the existing distribution?”

53. Bonus: The Super-Entrance and Satya’s AGI Standard

A pre-recording conversation moved into sharper territory. ChatGPT wanted to tell the story of the super-entrance? “The super-entrance is a concept Chinese investors and founders know well. The US does not really pursue large and comprehensive products.” The best monetization for a large-DAU product is advertising, but “when everyone spends huge amounts training their own models, the 80% of the pie you are dividing is still the same pie.”
Citing Satya: what is AGI? “AGI should make global GDP grow at 10%—you have to create a new pie, not simply divide the existing pie held by the giants.” Advertising and e-commerce resemble dividing an existing pie. Why are people willing to invest $100B? The AGI assumption is extremely powerful productivity gains—new drugs, new science—plus investments that create barriers others cannot enter. “But now we see, first, the superintelligence has not arrived; second, you cannot stop a Chinese company from quickly building an open-source model with 80-90% of your capability at 10% of the cost. Should you keep investing at the same pace? I think next year is a critical year.”

54. Bonus: The Container Thesis—Chatbots Are AI 1.0; Do Not Define Yourself by the Previous Era’s Form

ChatGPT “looks quite like an AI 1.0 product”: one question and one answer, general but not proactive, not sufficiently multimodal, and awkward for complex tasks. The historical pattern is that technological evolution requires new products and interactions. PageRank came with a minimalist search box; recommendation engines came with Douyin’s swipe-up and swipe-down interface. “Only a new container can give users a completely different experience.” A chatbot was the best container for language models at the end of 2022. “Will it still be the best container 5 years from now? Not necessarily.” The champion of version 1.0 may not win version 2.0; Yahoo, the strongest product of Internet 1.0, was displaced by search.
OpenAI’s own five-step ladder is Chatbot, Reasoner, Agent, Organization, Innovator. The best form for an agent may not be a chatbot. Manus offered a preview of the agent era: a visible sandbox, a canvas, and an environment the agent controls itself.
On AI browsers—an area Manus explored and abandoned: “When you call it a browser, you must meet people’s existing expectations of what a browser is.” Arc shows how difficult that is. In the mobile era, attempts to build “mobile versions of something”—a browser or search engine—did not succeed; Douyin, a new species, did. Do not define yourself by the product form of the previous era. You can meet the same need under a different name.

55. Bonus: Is OpenAI a Bubble? Seven Parts Beer, Three Parts Foam

There are two types of bubbles: something with no value being inflated, as in Web3 or the metaverse, and something valuable but overvalued. “OpenAI is clearly the AI application with the most users and revenue today. There is substantial real value underneath it, even if the valuation may have considerable room for adjustment. This is at least not a bubble with no beer underneath.” What is the ratio of beer to foam? Sometimes 70% beer and 30% foam; beer without foam is not enjoyable. Some sectors may be pure foam with no beer, which is dangerous.
The spending paradox is that if everyone sells ads, “OpenAI selling ads is clearly uneconomical. Douyin has a very low marginal cost, while you have inference costs, expensive people, and expensive training. Why would advertisers sell there?” Inference costs will fall as models reach the right level—translation does not require an increasingly powerful model—but fixed training costs will not fall much. “This may remain a higher-marginal-cost business than the internet for a long time.” The user’s $20 may not change, but “you certainly cannot say that $20 3 years from now will buy the same intelligence as today’s $20.”

56. Bonus: ChatGPT versus Google—Brand and Memory Against Resources and Distribution

“I think Google could very possibly collapse.” But the other side is immediate: Sergey Brin entered founder mode, merged DeepMind and Google Brain, and committed fully to Gemini. “It showed the very strong competitiveness of an incumbent. Imagine if it had reacted one step slower—it might now be very difficult to recover.” Not every incumbent can do this.
The opposite conclusion is not settled. “I do not think Gemini has already proven that it can stop ChatGPT’s growth.” ChatGPT retains users much better than Gemini, due first to memory and second to brand. For ordinary users, ChatGPT may simply be synonymous with AI, just as Yahoo was synonymous with the internet in its early years. The tornado theory from Crossing the Chasm says the number-one product entering the mainstream market receives enormous organic traffic, making the number two difficult to challenge. He already uses ChatGPT instead of a search engine, turning to Google only when he needs a specific link.
The threat to Meta is less immediate. “We still cannot see ChatGPT affecting people’s social relationships, which are the foundation of Facebook and WhatsApp.” Over the long term, however, an application with 1B DAU and high time spent must put pressure on the time pool supporting every large-DAU advertising business.

57. Bonus: Tools Have No Network Effects and Converge on the Best One

Why do AI products generally lack network effects? “Tool products generally lack network effects.” Excel and Word are exceptions because file formats create network effects. Search engines rely on scale effects plus brand: more queries improve long-tail optimization, while switching costs are nearly zero. “There is no reason to use the second-best search engine,” so the leader captures 90%. Chatbots may develop a similar structure unless memory creates a flywheel: Gemini may eventually have the stronger model, but without your history, ChatGPT may still offer the better experience.
This is where the assumption from 3 years ago collapsed. People thought investing $10B or $100B to obtain a model others lacked would create permanent leadership. Now a Chinese company cannot be stopped from producing a comparable model. Models are becoming commodities, while OpenAI never enjoyed Google’s earlier blank competitive field; from day one it faced the Mag Seven, ByteDance, and other giants. That is why Chinese model companies are heavily undervalued: only a handful of companies worldwide can produce models scoring in the 80s or 90s; in the US they are valued at $100B, while in China they may be worth only several billion.

58. Bonus: Anthropic, Cursor, and the Pioneer Becoming a Casualty

On Anthropic: “When it has no product of its own, it faces substantial risk.” API switching is fast; building a good car is not enough if the engine can be replaced without much impact. Claude Code is critical. Recent developer feedback suggests Opus 4.5 (heard as “OPENS 4.5”) may be far better in actual use despite only a modest SWE-bench improvement. Gemini 3 scores highly but may perform worse than Anthropic’s models in practice—benchmark high scores, low ability, again.
On Cursor: “It still feels difficult.” Codex, Antigravity (heard as “Windows Server Team” moving to Google), and Claude Code mean all 3 major foundation-model companies now have coding agents of their own, distributed through the cloud. Perplexity is the warning: it pioneered AI search, but once ChatGPT and Gemini incorporated search, “its ceiling became obvious and it could gradually be eroded by model-native products.” The shared disease is that the user’s output is not much different from what the model itself provides. Applications need to build out the incremental value of the context layer and the tools-and-environment layer; otherwise the wrapper remains thin.

59. Bonus: The Path Forward for China’s Large-Model Companies: All Must Solve R

On Kimi, DeepSeek, MiniMax, and Zhipu: “This question would offend too many people, haha.” But the direction is clear: pure model companies will remain difficult. The model must ultimately become a product, whether through overseas expansion or products with Chinese characteristics. “I may not have made money yet, but I can see how money can be made.” China also has indirect monetization models in which “the sheep pays for the pig.”
The table metaphor closes the discussion. “There will be many new AI games every year. If you miss this one, wait for the next. 肖弘’s first card was Monica; his second was Manus.” In the late mobile-internet era, there might be only one card a year, or none, which is when founders become truly stuck. Research-oriented new labs—Mira, Ilya, Fei-Fei—are different: “more money is not always better, and many people even question whether research labs are suitable for investment.” China has few such opportunities.

60. Quickfire: Beef Noodles, Sinkholes, Noses, and a Reading List

Changde beef noodles use round noodles, unlike Changsha’s flat noodles. His favorite place is the cenote caves of Cancun, Mexico, where a beam of light falls from above. “Plato had the metaphor of the cave. If you could cut an opening in the cave and let a beam of light in, that would be an enlightenment for humanity.”
A piece of trivia: you can actually see your own nose, but the brain continually erases it. After wearing inversion goggles for several days, the brain automatically “turns the image right-side up.” “The world we see is not the real world; it is the world after our brains process it.”
He reads roughly 100 books a year. Annual recommendations include A Brief History of Intelligence, from marine archaea to the birth of ChatGPT-4, which he read in January and recommended to 姚顺雨; and The First Eye, whose light-switch theory describes trilobites evolving eyes in a world where every creature was blind—“suddenly one organism evolved eyes,” pushing the entire ecosystem forward, much like technological progress drives applications. Crossing the Chasm is also essential. He only dares list familiar papers—AlexNet, Transformer, and Bitter Lesson (heard as “Beat Lesson”). “We do not pretend to understand the technology itself. We mostly talk with different outstanding people and try to identify the common patterns.” Reading the R1 paper (heard as “R Y”), he found the result-based reward “theoretically elegant,” but admitted, “I honestly do not have the ability to judge it.”

Closing Takeaway

The most important bet from today’s perspective is that the China-US AI gap—technologically, in applications, and commercially—will narrow. The gap has already passed its widest point.

2026 AI New-Year Conversation: the year of R | A Conversation with ZhenFund's 戴雨森 Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer