Summary

广密’s verdict on the “AI bubble”: it is not a bubble but an AI war. “This is a war or arms race that no tech giant can afford to lose”; AI should be treated “like a new form of national defense or nuclear weapons,” so “many people may spend their last dollar rather than leave the table.” The commercial logic: models white-label cloud providers, as HP and Dell were back in the day, while the agent entry point short-circuits super apps—“in the Bay Area, Uber often takes 50% of the fare,” and once information is symmetric, intermediaries like this “shouldn’t be making that much money.”
The landscape will be defined by two alliances—Nvidia GPUs versus Google TPUs—with GPT, Claude, and Gemini alternating at the top. That should be the norm over the next year: under the current pre-training-plus-RL paradigm, no one can pull decisively away because the know-how is too evenly distributed. The strategy is “whichever side falls, that is when you add.” He currently sees Nvidia and OpenAI as undervalued—Rubin can leapfrog TPU v8 by a generation, and Nvidia has locked in TSMC’s 1.6nm capacity—while Google is “not cheap” after its PE expanded from 18-19x to 28-29x.
OpenAI’s $1.4T commitment does not work on near-term math. It implies more than $200B of annual cash burn on average versus a 2030 revenue forecast also in the low-$200Bs, but its cash flow should be fine for the next 2-3 years. The visible businesses—subscriptions, advertising and commerce, and API—could reach $200-300B in annual revenue, roughly the scale of Meta or ByteDance; the less visible option is agents consuming the wage budget of 1B white-collar workers, a $1T-$10T revenue opportunity. The core conclusion: Google and OpenAI together could exceed $10T in market capitalization.
The third paradigm, online learning, is already taking shape and would be a nuclear-level breakthrough. Pre-training data is oil, 70-80% consumed; RL expert data is renewable energy; online learning is nuclear fusion. “If it breaks through, it will be unbeatable, and humanity will enter the silicon-based age.” Signals could emerge in 2026, perhaps by summer; OpenAI is furthest ahead, followed by SSI and Thinking Machines, while Google is not pursuing it top-down yet.
The sharpest take of the episode: robotics, world models, and even multimodality may be fake problems. Online learning may be “the only real problem that matters”—without generalization, AI will repeat the path of autonomous driving: “as much human labor as you put in, as much intelligence as you get out.” The breakthrough in robotic world models may not come from today’s leading researchers; today’s group may end up like the previous generation of NLP researchers.
The AGI discussion returns to reality: “the model is the product, and the data is the model.” Today’s model is still “a giant compressor”; a rumor has Sam saying internally, “forget AGI for now.” Partial L3/L4 automation for knowledge workers is “extremely certain,” but reaching R4 across the whole workforce will be difficult. The ARR meta-conclusion: the more top-tier the company, the cheaper it is and the less bubble-like it is—OpenAI at $20-21B, Anthropic at $9-10B, and Cloud Code, Cursor, Search AI, and Mokr each above $1B, with some names unclear in the audio.
The investment strategy is to bet on the steepest part of the technology growth curve: the leading models, their compute infrastructure, and the spillover benefits of their technology. The AGI index shifts from 40% OpenAI, 40% ByteDance, 10% Google, and 10% Anthropic to 25% OpenAI, 25% ByteDance, 10% Google, and 10% Anthropic, plus 10% each in Nvidia and TSMC. The biggest 2026 expectations are an online-learning breakthrough and “agents taking the web—that is the real web3.”
China: The host says Doubao has reportedly surpassed 100M DAU; 广密 sees its performance as better than expected and says “today’s model does not represent Doubao’s true level.” 2026 is a critical window for consumer chatbots. The probability that a Chinese team will build the world’s leading AI company within 3-5 years rises from 0-5% to 20%; the bottleneck is capital, not talent—“there is too little market-driven private capital.”

Deep dive

1. Not an AI Bubble but an AI War: New National Defense and Nuclear Weapons

广密 opens with a reset: there was reason to worry for a while, but the current conclusion is not to worry too much—“you can’t call it an AI bubble; it looks more like an AI war,” even “a war or arms race embedded in the national strategies of both China and the US that neither can afford to lose.” AI “has to be viewed as a new form of national defense or nuclear weapons,” so the giants “will spend their last dollar rather than leave the table.”
Why can’t anyone afford to lose? On the B2B side, developers used to choose AWS or Azure first; now they choose among GPT, Claude, and Gemini. Cloud providers are being white-labeled, “just like HP and Dell back then—they have been pushed into the background.”
The consumer-side inference from the Doubao phone: AI agents become the traffic entry point and short-circuit today’s super apps. Platforms monetize information asymmetry—“in the Bay Area, Uber often takes 50% of the fare, leaving the driver with less than half”—but once agents make distribution fully symmetric, intermediaries such as Uber and Ctrip “shouldn’t be making that much money.” The same applies to traditional search. The giants therefore need to defend and attack at the same time.

2. The Fuse: Sam’s $1.4T and “I’ll Find You a Buyer”

The bubble thesis began in October, when Sam proposed $1.4T and 30GW of compute buildout. Amortizing GPUs over 6 years implies OpenAI burns more than $200B a year on average, while its public 2030 revenue forecast is also only in the low-$200Bs, with cumulative revenue of just $500-600B over the next 5 years. “From a business and cash-flow perspective, the math clearly does not work in the near term.” “To be honest, nobody knows today why Sam wants to spend so much.”
But one point is clear: OpenAI should have no cash-flow problem over the next 2-3 years. It is raising another $100B at a valuation above $800B, and most of the $1.4T investment comes in 2028 and later.
Another source of confidence volatility: when the host repeatedly pressed Sam on how he would pay for $1.4T, Sam replied, “If you want to sell your shares, I can help you find a buyer. Enough.” 广密’s verdict: “That was a very bad answer.”

3. OpenAI’s Revenue Stack: From $80B in Subscriptions to a $10T Wage Budget

The 3 visible revenue pools are subscriptions, advertising and commerce, and API. Under an optimistic case of 4B MAU, 2.5B WAU, and 1.5B DAU before 2030, a 10% paid-conversion rate yields 400M paying users at a $200 annual fee, or $80B—comparable to Office’s 300-400M paying users. “My intuition is that ChatGPT should be bigger than Office and Netflix.” Advertising and commerce could generate $50-100B from 2B+ WAU at $20-50 ARPU; API could add another $50-100B. “The model is the new cloud,” analogous to Amazon building its own cloud after launching e-commerce. Total annual revenue could reach $200-300B, “roughly the size of today’s Meta or ByteDance.”
The real imagination lies in what is not visible. “Go back to the first few years of AWS or Douyin; nobody would have thought either could reach $100B in revenue.” Ten months ago, nobody believed AI coding could grow 10x in a year and “somehow produce $10B in revenue.”
There are 2 possible paths. First, agents replace labor end to end: 1B white-collar workers, each represented by an agent worth $1,000-$10,000, means “$1T-$10T in revenue.” Software development is the fastest case: a roughly $150B global market, with AI coding already above $10B this year and penetration below 10%, potentially reaching 30-40% over the next few years. Second is the more expansive vision of producing high-value incremental content at scale that humans cannot create—AI scientists, room-temperature superconductors, “curing every disease humanity has”—which requires more reliable models and the fundamental breakthrough of online learning.

4. The Giants’ Game: Anthropic’s Value as a Chess Piece and the Two Alliances

Because the stakes are so high, “nobody will let Google become dominant on its own, and nobody will let OpenAI become dominant on its own.” Nvidia, Microsoft, and Amazon will continue supporting OpenAI and Anthropic. Anthropic has “the value of being a chess piece for a giant.” The logic extends further: “The stronger Google gets, the more an anti-Google alliance will form; the stronger OpenAI gets, the more an anti-OpenAI alliance will form.” After all, wasn’t OpenAI founded to challenge Google?
Hardware breaks into 2 camps: Google TPU versus Nvidia GPU. Google integrates its models, chips, and cloud end to end, “more like Apple in the AI era.” OpenAI and Anthropic remain the top students in Nvidia’s ecosystem. Google still has the stronger talent density; Nvidia “looks more like Android in the AI era.”
As long as demand is far above supply, both deserve a place in the portfolio. “Alternating leadership—whichever side falls, that is when you add.” The current view is that the Nvidia line, including OpenAI, is undervalued: GPUs remain better overall than TPUs, although they are more expensive and “TSMC only has so much capacity.” Rubin “can leapfrog TPU v8 by a generation,” and Nvidia has already reserved future TSMC 1.6nm capacity.
Asked whether Chinese companies can serve as strategic chess pieces, he says the major Chinese giants are all strong, but no independent model company has a clear lead. “ByteDance and Alibaba both look strong, and Tencent is catching up after hiring away the top talent.”

5. Model Market Structure: The Top 3 Take 80-90%; True SOTA Is Measured by Revenue, Not Rankings

The 3 leaders’ annualized revenue by year-end should be roughly $21B/$9-10B/$5B for GPT, Claude, and Gemini. Together they account for about $36B, or 80-90% of global AI-native revenue; all other foundation-model companies combined generate just over $40B. xAI’s Grok is “maybe somewhere between $300M and $1B.” A 3%, 5%, or 10% lead can translate into dramatically higher earnings—like elite athletes: “Cristiano Ronaldo and Messi are not dozens of times better than other players, but they earn dozens of times more.”
Alternating leadership is a consequence of the current paradigms. Under pre-training plus reinforcement learning, “nobody can throw their know-how far enough ahead; talent moves, information moves, and to pull away you need a new paradigm.” He does not believe Google has a decisive lead yet; xAI and Meta “have not truly taken a seat at the first-tier table.” The implication is that model companies do not have especially high moats, but they do have extremely high barriers to entry—capital and talent.
Meta illustrates what SOTA is worth. Launching another SOTA model “is nowhere near enough if all you are doing is handing in an assignment.” You need a breakout product at ChatGPT scale to support $120B of capex next year, against roughly $50B in cash and $60B in next year’s profit.
The harsher take: “Gemini 3 and GPT-5.2 both look highly suspected of score inflation—high scores, low capability.” True SOTA should be defined by revenue, customer count, and tokens generated, not by a benchmark leaderboard.

6. Divergence Is an Escalation: Strategic Bets and “Vertical E-Commerce or Xiaohongshu?”

Last episode’s “models are diverging” and this episode’s “alternating leadership” are not contradictory; they are sequential. General capabilities remain in a race, but each company has a strategic bet. Gemini 3 is in a different league on multimodality, with text apparently deprioritized. Opus 4.5 remains the best coding model and “is still undervalued”—it handles longer tasks, is more reliable, and uses fewer tokens. ChatGPT is in a different league on the consumer side, at roughly 485M DAU, 5-6x Gemini’s level, betting on 1B-2B DAU as a personal assistant and personal friend.
Anthropic’s strategic lesson for founders is to focus on B2B, abandon B2C, and focus on coding and agents—“otherwise Anthropic would have died early in such intense competition.” Everyone is optimizing from the end state backward around their chosen bet; fundamentally, they are optimizing data.
广密’s framework: foundation-model companies resemble general e-commerce platforms 10 years ago, where scaling data is most efficient. ChatGPT resembles early Amazon, which started with books to build logistics and warehousing before scaling horizontally into more SKUs; today it is making PowerPoint decks, Excel sheets, and investment research in-house. So are Perplexity and Cursor building vertical e-commerce businesses or a Xiaohongshu? Vertical e-commerce can make money but must remain flexible about exiting and historically has not become a major platform. Xiaohongshu built high-value consumer content assets on top of a general platform. “There is no answer yet; everyone is still exploring.”

7. Google + OpenAI = $10T: Who Is Overvalued and Who Is Undervalued?

The core view is that the 2 companies together can exceed $10T in market capitalization, and OpenAI is “massively undervalued”—“the only company in the past 20 years with a chance to challenge Google” (even Microsoft never truly did). With annualized revenue above $20B this year and $40-50B next year, “even if you say the multiple is high, it is not that expensive.” It also has the highest probability of producing the next paradigm shift. Its financial losses are not the main concern: “The wealth accumulated by capitalist countries over the past few decades is more than enough. What else can people invest in today besides AI? I expect all the money to be poured into AI.”
Google will not become dominant on its own either. This is not a zero-sum game; chatbots are creating new markets, much as Douyin competed with Youku by expanding the category. But Google is expensive in the near term: launching an SOTA model pushed its PE from 18-19x to 28-29x, and “Google is not cheap today.”
The data supports ChatGPT staying first over the long term: it could reach 1.4B MAU in 3 years, versus roughly 450M MAU for WeChat after 3 years. The final split—70/30, 60/40, or 50/50—is still unclear, but he leans toward ChatGPT remaining the long-term leader at 1B-2B DAU. It could build an entirely new search engine and ultimately absorb traditional search advertising; traditional search could also incorporate chatbots. Both paths require watching.
That does not mean Google is weak. It has Waymo, a stake in SpaceX, and a drug-discovery business (the audio says “Ansmorphic,” likely Isomorphic Labs). “Google, along with Nvidia, may still be among the strongest candidates to become humanity’s fastest route to a $10T company.” Moving from $3T-$5T to $10T is “far easier than taking a company from tens of billions to $100B.”

8. The Truth About Gemini 3 Traffic: A Monday-to-Friday Workhorse

The model really is SOTA for the first time in a meaningful way—“if it did not improve, that would be more worrying”—but Gemini 3’s impact on DAU growth is even weaker than Nano Banana’s was at the time.
Usage patterns differ. ChatGPT sits on the first screen of the phone as a life assistant and is used on weekends; Gemini is largely “a Monday-to-Friday workhorse,” a Web-based productivity tool for front-end development and PowerPoint creation. Gemini MAU is about 20-25% of ChatGPT’s, DAU/MAU is only 10% versus roughly 25% for ChatGPT, and monthly session volume differs by 3-4x. Geographically, ChatGPT holds North America and Europe, the “highest-value regions,” giving it a strong monetization base. Gemini is taking a “rural encircling the cities” route through Brazil, India, Vietnam, and other Android-heavy emerging markets, where MAU is around one-third of ChatGPT’s.
At the market level, ChatGPT’s share of total Google Search traffic has shifted from 5:95 at the start of the year to 15:85. “It is no longer marginal traffic; it has become a mainstream traffic platform.” Google’s long-term crisis is unresolved. People have simply stopped treating it as Nokia because the narrative has reversed: either ChatGPT truly takes search advertising share next year, or Google reinvents itself and “continues as the king of the AI era.”

9. The Third Paradigm: Oil, Renewables, and Nuclear Fusion

Responding to Ilya’s claim that scaling has stopped: “Pre-training scaling is indeed nearing the end, but it would be more accurate to say online learning is just beginning.” Pre-training is at 70-80%—there is not much more data to find, and 50-60T may be the ceiling for many model companies; active parameters are shrinking toward highly sparse, large-MOE architectures. RL is at 30-50%, and everyone is scaling data. Under these 2 paradigms, “the landscape will not change dramatically; the competition is now face-up.”
The internal metaphor has 3 stages. Pre-training data is oil: fossil fuel, abundant but finite, with 70-80% already used. Human-expert data distilled through RL is renewable energy: useful but limited in total volume, driving revenue booms for expert-outsourcing companies such as Search AI and Mokr, whose names are unclear in the audio. Online learning is nuclear fusion: “there has been no breakthrough yet, but if it breaks through, it will be unbeatable, and humanity will enter the silicon-based age.”
The reason it matters is fundamental. Today’s large models are “static, frozen products with no data flywheel”; they cannot learn from daily interactions. The mobile internet became 10x larger than the PC internet because interactions generated durable data. Online learning lets a model learn while reasoning and interacting: “AI will absorb human intelligence.” It still needs infrastructure—longer context, LoRA, and parallel sampling across multiple models. Conversations with leading researchers remain optimistic: signals of a breakthrough “should appear in 2026, perhaps in summer or around then.”
OpenAI is “still far and away the leader,” with the largest investment, followed by Ilya’s SSI and Mira’s Thinking Machines. Google “does not appear to be pursuing online learning top-down across the organization today.” The lineage is revealing: Anthropic was OpenAI’s earliest scaling team; SSI was its earliest pre-training team; Thinking Machines is the original ChatGPT and post-training team. “OpenAI remains very strong even after disintegrating 3 or 4 times.” SSI and Thinking Machines “should show something in 2026.”

10. The Paradox: Today’s Researchers May Become the Previous Generation of NLP

The episode’s sharpest judgment is that robotics, world models, and even multimodality may be “fake problems”; online learning may be the only real problem that matters. Only once autonomous learning breaks through are the other problems solved at a fundamental level. Without generalization, AI will repeat the autonomous-driving path: 10 years of incremental data collection, with “as much human labor as you put in, as much intelligence as you get out.”
His own deliberately provocative inference: “The breakthrough in future robotic world models may not come from the people sitting here today. The people sitting here today may end up like the previous generation of NLP.”
Is the third paradigm inevitable? His causal argument is that GPU shipments, the compute available to each researcher, and both the capabilities and headcount of researchers have exploded over the past 3 years. “It is like Hollywood: the entire industrialized creative system has formed, and producing new works is now a matter of time.” That is also the passion behind the podcast: encourage Chinese technology companies to buy more cards and give young researchers more opportunities to grind on compute. Ilya emerged in Silicon Valley because large companies made long-term investments without regard to short-term ROI. “China’s Ilya, the next generation of Ilya, will very likely emerge over the next 3-5 years. There are already many young researchers showing that potential, and they are very young.”

11. Back to AGI Reality: The Model Is the Product, the Data Is the Model

After seriously studying frontier-data-labeling companies over the past 6 months, the strongest takeaway is: “If this kind of data is not in the model’s data distribution, this kind of task simply does not work.” “Today’s model is still a giant compressor; data matters too much.” Models already know more than most people, but agents have never seen the environments where people actually work: how a print-shop clerk uses Photoshop, how a salesperson uses Salesforce, or how a bank teller or dermatologist works. Those workflows require distilling gold-standard expert data and applying RL for generalization. “Then you suddenly realize this looks a lot like autonomous driving,” with a very long tail. The summary is: “The model is the product, and the data is the model.”
The rumor, relayed nearly verbatim: “Sam has also been saying internally recently, forget AGI for now.” 广密 admits his estimate of the time required for AGI “has definitely lengthened”; agents need more time to land.
The practical endpoint is that partial L3/L4 automation for knowledge workers is “extremely certain”: information retrieval, coding agents, PowerPoint and Excel agents, and investment-research agents. “There are more gold mines like coding to be found around knowledge workers.” But reaching R4 across the workforce will be difficult, and whether agents can truly capture wage budgets “is hard to say.” If the cycle drags on, it becomes a cash-burning marathon: Google has major advantages; Meta’s investment is so large that “it is hard to say”; Anthropic has chess-piece value; OpenAI “can still raise enormous amounts of money because US capital is relatively abundant.”
Frontier labs will therefore diverge more sharply. “General-purpose chatbots and general-purpose agents may be a false proposition.” Companies need differentiation on top of general capability: ChatGPT as a friend, Anthropic as a co-worker, Gemini as a multimodal assistant. “Everyone has to optimize downward into their own data.”

12. The New-Lab Map of Silicon Valley: One Horizontal, One Vertical

The Bay Area trend can be summarized as “one horizontal, one vertical.” Horizontally, labs distill human expert knowledge and expand into more industries—Anthropic believes this path can reach AGI. Vertically, they pursue the next paradigm: memory, learning while reasoning, always-on proactive agents, and potentially new hardware. Mark Chen’s picture of the next product is a system with dramatically stronger memory that learns deeper questions from a user’s prompts, reflects and associates in advance, and “gets smarter the more you use it.”
The named labs include SSI, led by Ilya and focused on autonomous model learning, perhaps predicting a longer logical trajectory rather than the next token; Thinking Machines, which “has a big chess game underway”; Periodic Labs, led by Liam, OpenAI’s former post-training head who briefly took over after Mira and Barrett left, together with the head of DeepMind’s chemistry-model work, building foundation models for materials and room-temperature superconductors. Materials experiments have shorter feedback loops than drug discovery and shorter RL cycles, although “you still cannot see when the results will arrive.” Isara, founded by Eddie from OpenAI, whose surname is unclear in the audio, is working on multi-agent collective intelligence and training environments. Another xAI-founded company is building “empathetic models,” in a direction similar to early Inflection.
Two especially interesting cases: General Intuition, formerly a gaming highlight-reel recording platform, has accumulated 3.8B game clips. Unlike YouTube, it records only users’ peak moments during gameplay, naturally generating high-quality, user-labeled action-response data. Nvidia’s recent investment Reflection “looks like a US version of DeepSeek”: with Chinese open-source models filling US enterprises, some believe the US should not become overly dependent on them and should support a US open-source model to fill the gap.

13. Robotics: The GPT-0.5 Moment and the Data-versus-Hardware Fight

The field is still far from the GPT-4 moment. Today is “the GPT 0.5 to GPT 1 moment”; the next 2-3 years may reach GPT 2-3. Architecture, data, and methods have not converged. The key difference from language models is that language models unify first and diverge later, while robotics is fragmented from day one, with neither a common base nor common hardware. The main real-world applications today are “folding clothes and making coffee”; there is still no broad generalization. Some researchers even believe “hardware may account for 70-80% of the factors behind success,” which is why algorithm companies are starting to build hardware themselves.
Data is where the bets and differentiation lie. Google and Pai, pronounced “Pai” in the audio and likely Physical Intelligence, have teleoperated real-robot data. Sunday, pronounced “Sandy” in the audio and unclear by name, is Tony 子豪’s team; it uses gloves and crowdsourcing to collect low-cost household data and “wants to enter homes directly to do housework.” Generalist has collected what it calls the largest real-robot dataset so far, with 270,000 samples, and claims to have found a robotics scaling law: below 7B parameters, adding data has limited impact; above 10B, more data improves the model dramatically and, importantly, predictably. “That is a signal worth taking seriously.”
The preferred names are Google, with the brain, and Tesla, with the physical hardware, among the large companies; and Pi, Generalist, and Sunday among startups. Pi has a world-class team and strong research culture, and recently released the RECAP reinforcement-learning framework, adding a value function and credit assignment so each step receives a 0.7 or 0.8 score. “Like chess, it knows which steps help it succeed,” allowing learning from success and failure, although “the academic component is still stronger today.” Dyna, pronounced “Danna” in the audio and unclear by name, is more practical, focusing on folding clothes and tissues. It is “less cool on AI research than the others, but its operating capability is actually quite strong.”

14. The ARR Landscape: The More Top-Tier, the Cheaper

The meta-conclusion comes first: “The more top-tier the company, the cheaper it is; the more top-tier the company, the less bubble-like it is. The companies more likely to contain bubbles may actually be further down the list.”
Ranked by annualized revenue: OpenAI at roughly $20-21B; Anthropic at $9-10B, with Cloud Code alone above $1B, though the product name is unclear in the audio; Cursor above $1B; the data-labeling pair Search AI and Mokr both above $1B, with names unclear in the audio; Midjourney at $700-800M; Together AI at roughly $300M. Below that: Glean and Replit at $200-300M each, ElevenLabs near $300M, Lovable above $200M, OpenEvidence, the “ChatGPT for doctors,” at roughly $150M, Synthesia at $150-200M, Harvey at $150-200M, Fireworks AI at $100-200M, 黑镇 above $100M, with Abridge, Ciera, Minus, and Jane Spark also on the list; several names are unclear in the audio.

15. Investment Strategy: Bet on the Steepest Curve; Add Nvidia and TSMC to the Index

The one-line strategy: “AI investing means betting on the steepest part of the technology growth curve.” There are 3 lines: the 2-3 most advanced global model companies—like investing in e-commerce platforms, where there is no small-but-beautiful vertical platform; the compute and silicon infrastructure required by the leading models—“CATL and Nvidia are the same kind of company, and people still underestimate the size of this trend”; and the spillover benefits from leading-model technology. More companies like Proprietary Cursor and Moccour, whose name is unclear in the audio, could rapidly reach tens of billions through the gaps created by technological spillovers. At the extreme, there may be “opportunities on the scale of Pinduoduo and TikTok,” as AI infrastructure matures, just as Pinduoduo benefited from mature Chinese e-commerce infrastructure.
Platform research is mandatory: ChatGPT, Google Gemini, and ByteDance—“the probability that these large companies rise another 3-5x is higher, and these are the easy questions.” At the same time, look for gaps where the platforms have not executed well. Build a portfolio rather than make a single bet: “Betting on only one is also very risky; it is better to invest in several because leadership alternates.” Deep research can itself be the core strategy: many people have not actually done focused, top-down research in depth.
The AGI index is rebalanced from 40% OpenAI, 40% ByteDance, 10% Google, and 10% Anthropic to 25% OpenAI, 25% ByteDance, 10% Google, and 10% Anthropic, plus 10% each in Nvidia and TSMC, “with something in every name.” After the host pressed him, he emphasized that this was not investment advice. Nvidia is undervalued today “because Google has risen”; the 2 will alternate leadership.
The 2 biggest 2026 expectations are an online-learning “nuclear-level paradigm breakthrough,” bringing more proactive agents and changing product forms, and the Doubao-phone-inspired “agents take the web—I think that is the real web3.” Startups finally have a chance to flip the table; fundamentally, this is the endgame battle for control over traffic allocation. He also believes multimodality “could make this a big multimodal year.”

16. China’s Table: Doubao Beats Expectations; The Decisive Moves Are Elsewhere

小珺 says he has heard that Doubao has surpassed 100M DAU. 广密 replies that it is “still better than expected,” and that “today’s Doubao model does not represent its true level.” Talent density is high; a new frontier model could change everyone’s impression of Doubao. 2026 is a critical window for China’s consumer-chatbot competition. Tencent and Alibaba will invest aggressively, and “only these few giants will compete for the traffic entry point.” Independent model companies are copying Anthropic on coding, as Kimi is doing, or copying Minus, whose name is unclear in the audio, on agents.
The next decisive moves are 2-fold: first, an end-to-end product bet—“Anthropic’s bet on coding found it a strategic niche, while ChatGPT’s end-to-end bet is on being a friend for consumers”; second, the next paradigm—“whoever gets online learning working first. Once autonomous model learning works, the final problem is just adding compute; it can run very fast.”

17. The US-China Narrative and Chinese Founders: Self-Reliance Makes Everything Stronger

The industrial contrast is that American capitalism “only goes after the tip of the pyramid.” Buffett-style capitalists pursue returns; Cook-style managers moved manufacturing to China, hollowing out the industrial base. “US manufacturing is money printing—it printed countless dollars and Treasuries, and all of that is being invested in AI today.” It resembles the US-Soviet rivalry: “China is like the US back then, hiding behind the scenes,” industrializing Silicon Valley’s 0-to-1 innovations, from electric vehicles and smartphones to 3D printers and action cameras. “AI really will determine the future strategic landscape and competitiveness of China and the US.”
His 3 recommendations to Chinese founders are to commit to globalization, especially the US market, which can contribute 60% of revenue and 70-80% of profit; use China’s engineering-talent dividend while deemphasizing identity and geography, without awkwardly packaging the company as American or Singaporean—“building a good product is the only rule; self-reliance makes everything stronger”; and take whatever early capital is available, because trust chains with top Silicon Valley VCs are difficult and Chinese VCs are an important source of funding.
Why the obsession with “a Silicon Valley in China”? In 2011, as a college freshman, the shock of Singles’ Day packages piling up outside the school gate; later, the 0-to-1 innovation of O2O companies seen at Sequoia. “That kind of innovation atmosphere can now only be felt in Silicon Valley.” A major reason the gap has widened is “capital’s bias—the market-driven private-capital pool is too small.” His team counted more than 100 AI application unicorns globally over the past 3-4 years, but only 3-5 Chinese-founded teams, excluding foundation-model companies. “That is far fewer than in the internet era. It is a great shame.”
His long-term expectation is that the probability a Chinese-founded team builds the world’s leading AI company in 3-5 years has risen “from perhaps 0-5% to 20%.” The talent base is strong, but capital, compute, and an innovative environment are still required. If “leading” is defined by scale, revenue, and talent, it could be ByteDance. He hopes China can eventually produce 10 ByteDance-scale companies capable of matching or surpassing Meta and its US peers.

The Great Security Update: AI ∧ Formal Methods with Kathleen Fisher of RAND & Byron Cook of AWS All-In x Kill Tony: A Hilarious Holiday Special