Innovators
4 Curated Dialogues
广密
Key Views & Dialogues
136. Global Foundation Model Quarterly, Episode 9: A Conversation with 广密—Coding as AGI’s Second Act, the Truth About Silicon Valley’s Big Three, and Models Becoming the Next-Generation OS
广密 sees Opus 4.5 to 4.6 as a GPT-3-to-GPT-4-scale leap, while Coding becomes AGI’s second accelerator after GPUs and compresses development cycles from weeks to days. AI economics may shift from DAU to top-tier token usage, but the Big Three remain unsettled as compute constraints, Harness efficiency and white-collar disruption shape the next catalyst and risks.
View summary (7)
- 广密 sees the jump from Opus 4.5 to 4.6 as the quarter’s only generational leap worth highlighting: AI has moved from Chatbot’s first act into Agent’s second act, where it can complete high-value tasks. He sees the step as close to GPT-3 to GPT-4; model progress over the past quarter may even have exceeded all of 2025, while Miso and Spark would need to meet expectations to deliver a “true GPT-5 moment.” Another major leap could still arrive before June or July this year.
- Coding is no longer a vertical application but the second-largest accelerator of AGI after the GPU. At frontier labs, 70%-80% of a system’s code may still have been written by humans last year; this year, the share may already be below 1%. Claude and Codex are considered close to the level of a company CTO or Meta L8/L9 on many tasks, compressing the path from idea to working code from 2-3 weeks to 1-2 days or 2-3 days. “Language is the world, code is the solution” (语言即世界，代码即方案): if code can express most solutions in the digital world, much of the office white-collar workload could be automated.
- The core metric for AI businesses may shift from DAU to token usage among high-value users, with 1M-2M users at the top of the pyramid potentially worth more than 5M-6M subscribers. 广密 said Anthropic’s ARR has surpassed OpenAI’s and projected that the 2 companies could reach $80B-$100B by year-end and $150B-$200B or more next year; if that plays out, a teens-multiple PS valuation would not look unreasonable to him. The real constraint may not be demand but compute: “The biggest bottleneck to $100B of ARR this year may simply be compute.”
- Silicon Valley’s Big Three will continue to “each have their moment for 100 days,” with the 2026 hierarchy still far from settled. Anthropic’s strengths are top-down focus, data culture and a high-priced market at the top of the pyramid; OpenAI misread Coding because of ChatGPT’s overwhelming success but still has roughly a 50% chance of becoming the ultimate winner in 广密’s view; Google is 3-4 months late in the short term but has the strongest long-term foundation thanks to TPU, cash flow, Workspace and a systematic organization. “Today’s winning formula may be the poison of the next era.”
- The endgame for model companies is not a larger SaaS business but a next-generation operating system that supports an Agent ecosystem and automates global GDP. 广密 frames the roadmap as 3 acts—Chatbot, Coding Agent and Automated AI Researcher—and believes a company may announce AGI by year-end or early next year. His investment view is correspondingly extreme: if the 3-5 companies that keep delivering SOTA models ultimately automate 30%-50% of global GDP, each could reach a $10T valuation.
- White-collar deflation, an employment rupture and a painful window for America’s middle class may come before the technology boom. 广密 offered an aggressive but explicitly conditional estimate that “30% of jobs could disappear this year”; entry-level programmers, consultant demand, SaaS and IT outsourcing could all come under pressure, while even top AI researchers worry about losing their jobs in 1-2 years. Models make knowledge and intelligence cheap, while creativity, taste and IP may appreciate; his advice is blunt: “AI replaces people who do not embrace AI” (AI 取代的是不拥抱 AI 的人).
- Application opportunities remain, but the spillover window from stronger models will shorten; Harness and positive token ROI are the more durable frameworks. Products such as Cursor and Manus could be squeezed by upstream model companies building their own Agent products if they cannot access the strongest models; meanwhile, a good Harness can let ordinary or open-source models handle high-value tasks. For a One-person company, the hardest test is not a demo but the economic loop: “Spend $100 on tokens—can it make $110?” Outside models, 广密 is most bullish on robotics and AI for Science, which could see a step change over the next 6-18 months.
127. Cross-Year Frontier-Model Quarterly Conversation: 广密 on an AI War, Two Alliances, and the Third Paradigm—Online Learning
AI looks less like a bubble than an unwinnable arms race, with models white-labeling cloud providers and agents potentially compressing information-asymmetry profits in search and travel. GPT, Claude and Gemini may alternate at the top as Nvidia GPUs and Google TPUs form two alliances; commercialization and 2026 online learning remain pivotal catalysts.
View summary (8)
- 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.”
112. A Conversation with Guangmi on Large-Model Quarterly Reports: Divergence and Convergence, All-in-One Bundles and Vertical Integration, L4 Experiences and the Mining Window
Models may converge toward a Google Gemini–OpenAI duopoly, while Anthropic’s coding-and-agentic focus lifted ARR from under $100M to roughly $950M in 2024. Claude Code overtook Cursor within 3-4 months, highlighting the shift toward downstream application value. ChatGPT’s brand and retention form a nontechnical moat, while L4 subscriptions, platform integration, and startup token losses shape the next battleground.
View summary (7)
- Frontier models are diverging; there may soon be only 2 general-purpose players. Guangmi thinks the only companies still building truly general-purpose models may eventually be Google Gemini and OpenAI; Anthropic has nearly abandoned multimodality and to C, going all in on coding and agentic products—ARR rose from under $100M at the end of 2023 to roughly $950M in 2024, with some guessing annualized ARR could exceed $12B in 2025 and reach $30B-$40B in 2026. At a $170B valuation, “I don’t actually think it’s expensive, because there’s still more than 10x growth in a year.” Kimi and DeepSeek are moving toward the Anthropic playbook: “I think that’s very smart.”
- The leaders on the consumer side are converging, and ChatGPT’s moat is nontechnical. As the brand moves downmarket, retention “may be second only to WeChat.” Model-training spend converts into brand the way Netflix’s content spend does: “The money you pay for the model is also money you pay for the brand.” His $100 portfolio has changed from the prior quarter to $40 OpenAI, $40 ByteDance, $10 Anthropic, and $10 Google, because “the moat around coding and the moat around the model are nowhere near as high as ChatGPT’s moat.”
- The miners are moving downstream to harvest the value. Model companies spend tens of billions mining intelligence, then see “the front-end application Cursor making money hand over fist… I’m not stupid either.” Claude Code launched 8-9 months after Cursor and overtook it in 3-4 months; it could soon reach $1B ARR and command a standalone valuation of $20B-$30B. The mining window has compressed from nearly 2 years for Perplexity to 9 months for Cursor and 3-4 months for Manus, while startups facing token loss rates of 1:3 to 1:10 “will be dragged to death.”
- L4-level experience is the product North Star. The 2 best agents today are ChatGPT Deep Research and Claude Code, both delivering end-to-end aha moments and prompting users to upgrade from $20 subscriptions to $200. The next mine is Office/Workspace: PPT is exactly where coding and agentic capabilities matter most, expanding the user base from 30M coders to 500M-600M office workers.
- Google is back in favor. Product forms will ultimately converge into all-in-one bundles, and ChatGPT will inevitably run ads—its new commercialization CEO came from Instacart and previously built Meta’s ad system—while “Google is still the best advertising platform in the world.” Google can claw back ground through the scale effects of end-to-end vertical integration, from chips to the OS; the biggest uncertainty is whether it can change user habits.
- Founders may be desperate, but they still need to find “today’s Huang Zheng.” The giants “understand it, take it seriously, and can keep up,” so they are difficult to beat head-on. Yet in 2015 many said Taobao was unassailable, while Pinduoduo was collectively passed over in its A and B rounds. In venture investing, “diversification has failed; concentration is what works”—many Silicon Valley VCs missed OpenAI, Anthropic, and Mira’s $10B angel round because gross margins were not high enough.
- The bubble’s watchpoint is the grand narrative. “Just watch what the smartest cohort is using”—quant traders and top programmers overwhelmingly use Claude. The biggest fear is that roughly $300B of AI capex this year becomes “doll-within-a-doll revenue” with no value created at the front end. From the Bay Area, America’s 2 core advantages are financial seigniorage and Silicon Valley’s technology lead: “If China breaks through advanced lithography within 2 years, I don’t know whether global capital will flow back to China.” One veteran’s summary: “Jewish finance, Chinese AGI”（犹太人的金融，华人的AGI）.
97. 2025 Q1 LLM Earnings Review: A Conversation with 广密 on the Biggest Non-Consensus Call, AGI’s Main Line and Summit
广密’s biggest non-consensus call is that pre-training remains the only path to higher intrinsic capability, while synthetic CoT and fused training-inference systems could reopen scaling. He places AGI in late 2026 or early 2027 and sees Agent inference and AI for Science driving compute demand, though application moats depend on capturing research spillovers.
View summary (8)
- 广密’s biggest non-consensus call for 2025 is that pre-training is far from over—and is the only way to raise a model’s intrinsic ceiling and produce new capabilities. Six months ago, he put the odds of pre-training ending at 50%; he now stresses that post-training and RL can only elicit and reinforce existing capabilities. “The O series is basically test prep,” while “only pre-training can fundamentally turn an elementary-school student into a middle-school student.” High-value CoT synthetic data generated through RL, along with the fusion of sampling, inference and training on GPUs, could reopen the data and scaling opportunity.
- He has pulled forward AGI’s arrival to late 2026 or early 2027, and says the probability of achieving it within 2 years is, in his view, “100%.” The key is not waiting for an unknown paradigm, but treating coding as a dynamic, operable digital environment with real-time feedback, while code becomes the model’s “hand” for changing that environment. Agents may soon operate computers and phones better than 99% of people. His AGI definition remains “three 90%s”: reaching expert-level performance in 90% of industries, completing 90% of economically valuable work, with “computers and digital environments” added as the boundary.
- The fundamental split between OpenAI and Anthropic is traffic first or intelligence first, O-series first or base model first. OpenAI is betting on reasoning models and a billion-user ChatGPT; Anthropic is betting on pre-training, coding, agentic systems and the enterprise ecosystem. 广密 acknowledges that OpenAI currently leads in traffic and AI-native revenue, but favors Anthropic’s long-term value. The evidence includes Claude becoming Cursor’s default model, with Cursor paying Claude Sonnet more than $10M a month in token fees—roughly 4–5x what it pays OpenAI.
- Agents have not yet exploded at scale because model capability, inference supply and product infrastructure remain constrained—not because demand is lacking. Manus consumes roughly 700,000–800,000 tokens per task on average, versus thousands for a chatbot and tens of thousands for AI search. Complex tasks also require long-context reasoning, tool use, instruction following, and a closed loop of planning, coding and action. 广密 estimates that agentic inference could consume roughly 1,000x as much compute as chatbots, reinforcing his long-term bullishness on GPU demand alongside the return of pre-training.
- On the “AGI mountain,” ChatGPT is only base camp, Agents are the main course, and AI for Science is Everest. 广密 expects an AI for Science breakout in 2026–2027: virtual laboratories could first expand the sampling space before wet-lab experiments, while foundation models for proteins, RNA and genomes may develop superhuman capabilities. If AI can cure cancer and more diseases, health-related consumption could be 10–100x spending on smartphones. By contrast, he has lowered his near-term expectations for robotics: general-purpose datasets and generalizable architectures remain immature, and collecting 100M hours of useful data could cost hundreds of millions of dollars.
- The investability of application companies depends on whether they can become “Prometheus stealing fire for intelligence” and capture research spillovers, rather than building moats through old-style product growth. “A model company’s secret has no shelf life”: learning systems continuously internalize features, forcing wrapper scaffolding to reset repeatedly. Cursor, Perplexity and Manus will ultimately need post-training, RL and even end-to-end training if they want to scale. 张小珺 asked whether SaaS would still be necessary once Agents reached undergraduate capability with infinite memory, online learning and full software access. 广密’s answer: he genuinely does not know what SaaS companies would be for.
- China’s certainty of reproducing AGI depends on whether the current pre-training and RL path is sufficient, and whether it can be done without hundreds of thousands or even millions of GPUs. If pure replication requires only 30,000–50,000 or 50,000–100,000 GPUs, while exploration costs 5–10x more than replication, he believes China’s odds of achieving AGI are very high. The US–China gap is narrowing in chunks every 3–6 months. DeepSeek showed how a small team, strong infrastructure and willingness to make bold bets can rewrite the field. ByteDance is his favorite Chinese large company because of its capital, talent density and full-stack infrastructure; open source can also “break through geopolitical containment.”
- The most important investment assets are not static products, but compute, talent density and an organization’s willingness to decide what to give up. 广密’s hypothetical portfolio is Anthropic 25%, Baidu 25%, OpenAI 10%, Mira 10%, Ilya 5% and Manus 5%, with the remaining 15% undecided; if DeepSeek raised money, he would invest up to the fund’s 25% single-position cap. He later said he would invest the same amount in Manus and Cursor, although Cursor was absent from the earlier list. He believes the Agent bubble “has not even started.” The real black swans could be a radically more efficient architecture, multimodal ChatGPT or a new Agent form. Before 2030, several companies could exceed $10T in market cap, and “AGI bringing about a Renaissance in science” is his biggest bet for the next 3–5 years.