97. 2025 Q1 LLM Earnings Review: A Conversation with 广密 on the Biggest Non-Consensus Call, AGI’s Main Line and Summit
Summary
- 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.
Deep dive
1. Pre-Training Has Become the Biggest Non-Consensus Again
广密’s opening judgment was unambiguous: “The biggest consensus today is that pre-training is over. But the biggest non-consensus today is also that the pre-training opportunity remains enormous—perhaps it has only just begun.” The next SOTA base model could still beat GPT-4.5, Grok 3, Claude 3.7 Sonnet and Gemini by a wide margin.
He revised his view from 6 months earlier, when he believed there was a 50% chance pre-training was over and shifted his focus to the RL reasoning paradigm. Now that everyone is chasing the O series and reasoning models, he is instead calling for a renewed focus on pre-training: “People still need to take pre-training seriously. I think this is the most fundamental thing.”
The core distinction is between raising a capability ceiling and eliciting capabilities. Post-training and RL can strengthen and release capabilities that already exist, but cannot replace pre-training in creating new intrinsic capabilities. “The O series is basically test prep.” Reinforcement learning on top of a weak base model is like an elementary-school student drilling test questions—it saturates quickly. “Only pre-training can fundamentally turn an elementary-school student into a middle-school student.”
2. OpenAI’s Pre-Training Slowdown Reflects Both Strategic Payoffs and Organizational Attrition
Strategically, the O series may deliver more benchmark gains in 2 months than pre-training delivers in 1–2 years. At the same time, ChatGPT growth has been “terrifying—and accelerating.” These 2 fast-payback curves naturally absorb management attention, but that does not mean OpenAI has formally abandoned pre-training.
Organizationally, Dario took a group of people with him, Ilya left, and Mira took core post-training staff as well as members of the early core inference team. Existing pre-training personnel were repeatedly reassigned to post-training. 广密 therefore believes the visible decline in priority may have been driven more by organizational turmoil.
His external assessment is that Anthropic and xAI currently have stronger pre-training teams than OpenAI. OpenAI is not unaware of the issue; “the people who were good at pre-training have all left.” Once pre-training is downgraded, switching back from RL or the O series later will be much harder than catching up in reasoning.
3. Synthetic CoT and the Fusion of Training and Inference Could Reopen Scaling
The old argument was that internet data had been exhausted. But RL can now generate more high-value synthetic CoT data and feed it back into pre-training; RL could even be introduced during the pre-training stage. 广密 sees this as a response to the data bottleneck, not evidence that scaling has ended.
The real engineering challenge is integrating training with RL inference. GPUs must sample and train at the same time, requiring previously separate training and inference frameworks to work together. The integration is extremely difficult. Weak infrastructure can force researchers to spend enormous amounts of time debugging, directly slowing AGI progress.
广密 suggests plotting each path as a resource-input-to-intelligence-output curve: pre-training, RL and data each have different slopes and ceilings. The data curve may have particularly high returns, so taking pre-training seriously does not mean simply adding parameters; it also means rebuilding the data and training systems.
4. The Tool-Use Leap Will Create the Next ChatGPT Moment
Today, models operate computers at roughly “our parents’ level.” The next stage could look more like a proficient programmer. Office and Windows contain millions of functional points. Most people cannot write Excel formulas, but an Agent could theoretically call these functions fluently and handle most digital operations.
广密 calls Manus the “ChatGPT moment” for Claude’s tool-use capability. Anthropic may have trained specifically around hundreds of tool calls, but focused on the model and did not move first on the product. Manus, by understanding Claude’s capabilities and executing aggressively, gave the public its first magic moment with the technology.
张小珺 asked why the model companies themselves did not deliver this breakthrough. 广密’s explanation is that neither model companies nor cloud providers expected Agents to arrive so quickly; inference, environments and product containers were not ready. The Manus team built the “shell” first and waited for model capabilities to cross the threshold.
5. The AGI Timeline Has Been Pulled Forward to Late 2026 or Early 2027
广密 says he has “unprecedented confidence” in achieving AGI within 2 years, assigning it a subjective probability of 100% and placing the target between late 2026 and early 2027. The closer AGI gets, he believes, the less mysterious it becomes. Following the existing pre-training and RL path should be enough; no entirely new route is required.
The definition proposed in earlier episodes has not changed: reach expert-level performance in 90% of industries and complete 90% of economically valuable work. The new addition is an environmental boundary: the benchmark applies to computers, phones and broader digital computing environments, making the AGI path clearer.
This remains a conditional but forceful call. If the existing path cannot continue improving generalization, online learning, unified multimodality or an unknown architecture may become necessary breakthroughs. 广密 is highly confident about AGI, but much more cautious on ASI, true long-term memory and superhuman general intelligence.
6. Coding Is Both AGI’s Best Environment and the “Hand” with Which Models Change the World
A good environment must be dynamic, operable and capable of providing real-time feedback. Coding, gaming and science meet these conditions, with coding the most general cyber environment. AlphaGo had a game board, search companies had web indexes, and e-commerce had product-search environments. Code can express a much broader range of digital-economic activity.
“The model’s hand” is more than a metaphor. Humans use their hands to collect information, operate in and change the physical environment. Models generate and execute code to collect information, process it, take action and receive feedback. OSWorld, computer use and tool-use scaffolding are all strengthening that hand’s understanding and control of the environment.
广密 speculates that most future economic growth could come from the digital economy, describing its relative expansion versus the physical economy as potentially “thousands of times” larger. Agents could also dispatch people and physical entities in the real world. Within 2 years, they may operate computers and phones better than 99% of people and cover 99% of routine human digital behavior.
Coding may therefore function more like a search or recommendation engine than a programming interface. Consumer products will abstract away the code barrier, just as Douyin does not call itself a recommendation engine. The final form could be a new content feed or interaction system, but “the next Baidu or Douyin” has not yet been defined.
7. Anthropic’s Coding Lead Is Being Validated with Real Money by Developers
广密’s confidence in Anthropic comes from 3 consecutive bets: pre-training base models, coding and agentic systems. Claude 3.5 Sonnet unlocked Cursor, and Claude Sonnet later became its default coding model. He sees this as direct evidence that base-model capability can translate into a developer ecosystem.
The episode cites a direct figure: Cursor’s Agent already supports more than 150 commands and could reach 400–500 by year-end. It may pay Claude Sonnet more than $10M a month in token fees—roughly 4–5x what it pays OpenAI, the runner-up. “That’s developers voting.”
He distinguishes competition-style coding from real-world coding. The O series scores quickly on math and code benchmarks, but looks more like solving contest problems. Anthropic has invested more deeply in real software-engineering data. OpenAI may still “pull out a big move” in the future, but its failure to close the gap over roughly the past year suggests a strategic and organizational problem rather than a lack of technical understanding.
His view of the market is extremely aggressive: coding will certainly produce opportunities on the scale of Douyin, Pinduoduo and WeChat. If Anthropic maintains its lead, its valuation could reach $300B–$500B, while Cursor could reach $100B. A further breakthrough in AI for Science could create a path to $1T or even $10T.
8. OpenAI and Anthropic Have Become Two Different AGI Organizations
OpenAI has 2 core bets. The first is to achieve AGI through the O series or reasoning models. The second is to turn ChatGPT into a billion-user killer app and take on Google directly. It is more consumer-facing, product- and traffic-oriented, with a more bottom-up innovation culture.
Anthropic continues to bet on backbone pre-training, strong base models, coding and an agentic ecosystem. It is more enterprise-oriented and more like a top-down research organization. Dario’s pre-training background gives him an instinctive belief in scaling; the path dependence of different founding teams has hardened into corporate strategy.
The internal balance of power at OpenAI matters as well. The Frontier team that built the O series has risen, with Mark Chen becoming a key figure second only to Sam and Greg, and resources shifting accordingly. 广密 summarizes it this way: “Strategic choices are fundamentally expressions of different organizational capabilities”—and of the influence of specific people.
He has no definitive answer on the O series’ ceiling. Its benchmark growth is rapid, but the market for math is limited and coding is more competition-oriented. Whether it can independently expand the intelligence frontier still lacks real-world evidence. Combining strong base models with the O series may work, but that does not make the 2 interchangeable.
9. Silicon Valley’s Real Argument Is Whether Traffic or Intelligence Matters More
Veteran investors such as Roelof of Sequoia US and Vinod of Khosla Ventures favor classic consumer-internet logic: OpenAI should become a large platform because model training is too capital-inefficient. 广密 says bluntly, “I think they are wrong.” In his view, they are using the taste of old business models to understand a new research-driven paradigm.
He worries that OpenAI may put a billion users ahead of AGI first, research first and model-training first too early. A few thousand researchers can leverage the productivity of billions of people, cure cancer and reduce disease. Chasing platform profits first could mean missing “the Renaissance of science brought by AGI.”
Which path wins depends on whether base models can create a meaningful gap. If intelligence leads only marginally, brand and traffic still matter. If a new model is 10x better, users will migrate because “traffic has no loyalty.” Google repeatedly catches up with ChatGPT on individual metrics but has not taken its traffic because it has never created a sufficiently large model gap.
广密 also rejects framing the disagreement as a simple right-or-wrong question. OpenAI and Anthropic could both achieve AGI, like climbers reaching Everest by different routes. OpenAI is closed-source, but it continues to publish a technical path others can follow, and its contribution to the industry deserves respect. The question is whether each team dares to push its own bet to the limit.
10. AGI’s Main Line Is a Mountain Map; the Summit Is Not the Chatbot
广密 repeatedly returns to one overarching thesis: “The improvement of intelligence is the only main line. Intelligence itself is the biggest application.” ChatGPT reached GPT-3.5 and unlocked general conversation; Claude 3.5 Sonnet unlocked coding; the next stages are coding Agents, general Agents, AI for Science and robotics.
The route already includes milestones such as GPT-3.5, GPT-4, ChatGPT, Claude 3.5 Sonnet, the OpenAI O series, Deep Research, Cursor, Devin and Manus. Multimodality, online learning and multi-agent systems could also enter the main line. Every few hundred meters higher, users see scenery that did not exist before.
“ChatGPT is just the appetizer. The Agents that follow are the main course,” while AI for Science is Everest. If intelligence can discover drugs, cure cancer and treat more diseases, it will be not only a scientific breakthrough but also a new engine of economic and consumer growth.
He does not put every generative capability on the main line. Video generation and text-to-image have had magic moments, but may not advance intelligence in the orthodox sense; multimodal understanding is closer to the core path. GPT-4o’s text-to-image generation is an early step toward a unified all-modal architecture, but the research and compute costs remain high.
11. Every 10x Improvement Turns Old Moats Back into Beaches
ChatGPT, AlphaGo, Claude 3.5 Sonnet, Sora and Manus share one trait: they create magic moments that users can immediately feel. Once the capability gap is large enough, traffic migration and application explosions follow without first proving a stable business model.
广密 therefore revised his view of ChatGPT’s brand moat: “You can’t build a moat while you’re still on the beach.” Multimodal ChatGPT, an Agent version of ChatGPT or a broadly usable Manus could all break existing user mindshare through a 10x experience gap.
Applications provide the environments or containers that capture research spillover. DeepSeek saw terrifying traffic growth because it was the smartest model at the time, potentially reaching more than 100M DAU, but chose to keep building intelligence. Given its organizational culture and the economics of chatbot monetization, 广密 believes giving up the push for product scale “may have been the right decision.”
This does not mean stopping application development. It means prioritizing “research-driven applications.” Founders should build the container before the next model leap, while product growth should not crowd out model training. When the next 10x technology arrives, accumulated traffic may still fail to hold.
12. AI for Science Is Closer to the Next Summit Than Robotics
广密 acknowledges that his view of robotics has changed. Language models have Common Crawl, while robots rely on people operating dozens of devices to collect data at costs of tens to more than $100 per hour. If 100M hours of useful data were required, the bill could reach hundreds of millions of dollars; validating the scaling law itself would be extraordinarily expensive.
There is no consensus at the architecture layer either. The underlying algorithms have not converged, and general-purpose generalization has not appeared. His more natural route is to let digital-world AGI mature through 2D screens and operating systems, then extend multimodal capabilities into 3D and adapt AI to different hardware, rather than binding it prematurely to one robot body.
AI for Science already has a clearer container: Agents can build virtual laboratories and handle much of the repetitive work of a biochemistry PhD, increasing the batch size of the lab. Wet-lab experiments remain irreplaceable for now, but the sampling scale and exploration space before them can expand dramatically.
Foundation models for proteins, RNA, genomes and even virtual cells may be approaching their own “GPT-3 moment,” developing superhuman perception in fields where experts still rely on trial and error. 广密 therefore sees 2026–2027 as a possible breakout window for AI for Science.
13. The Phone Will Remain an Organ in the AGI Era; Glasses Are an Enhancement
The maturation cycle for robotics may look more like autonomous driving. It was heavily hyped in 2015–2016, yet 10 years later has still not reached mass deployment. Until a generalization-driven magic moment appears, the technology will continue through peaks and troughs. Even if virtual Agents can complete large amounts of work, whether physical robots are needed at comparable scale remains an open question.
On smart glasses, 广密’s view is that they “absolutely cannot replace the phone; at most, they can be an enhancement.” Today’s glasses are closer to a GoPro with a camera. If glasses, a watch, a ring and a phone all pushed the same spam notification, they would add friction rather than value.
Cloud intelligence will still require stable endpoints for context input and output, and the edge-cloud combination will not disappear. “The phone has already become a human organ,” so in the AGI era it will be enhanced, not easily displaced.
14. The Essence of Intelligence Is to Expand the Space for Survival, Exploration and Automation
广密’s macro framework has 3 words: survival, exploration and automation. Exploration creates experience, knowledge and organizations. Automation transfers and standardizes best practices while continuously raising productivity. Money functions like society’s reward model, incentivizing more effective exploration.
AI matters because it expands exploration capacity. In the past, he might spend a morning searching and reading hundreds of web pages to study a topic. Now ChatGPT and Deep Research let him pursue 3–5 questions in parallel. This is not merely a time saving; it multiplies the capacity of investment research.
He agrees with Dario’s view that the exploration of the next century could be compressed into 5–10 years. Countless Agents could simulate Earth, build virtual laboratories and express the results of automated exploration in code. But “the essence of intelligence” has no single answer; individual experience, taste and methods of distillation will still create differences.
15. Token Consumption and Superhuman Behavior Are Better Measures of Intelligence Than Old Benchmarks
One quantitative metric is the actual token consumption of each task. An ordinary chatbot conversation may use several thousand tokens, Perplexity-style AI search tens of thousands, while an average Manus task reaches 700,000–800,000. More tokens do not automatically mean more intelligence, but they show that the product is “squeezing intelligence dry” more thoroughly.
A Manus task often contains dozens, hundreds or even thousands of steps. It reads large numbers of web pages, repeatedly calls tools and feeds the result of each step back into context. Token use therefore does not rise once; it accumulates repeatedly over long chains, placing especially high demands on long context.
The more fundamental signal is that models are beginning to produce behavior beyond human cognition—for example, code that people cannot understand but that demonstrably works. 广密 expects more such results, while traditional human benchmarks may fail to see the “dark matter” behind model behavior.
16. Agents Are Still Supply-Constrained; the Breakout Will Resemble the Big Bang
Only a limited number of products currently approach the Agent category: Deep Research for text tasks, Devin and the future Cursor coding Agent for coding, and Manus, Computer Use and ChatGPT Operator for general-purpose work. Model companies and cloud providers have not prepared enough inference supply.
The 3 most important capabilities for Agent deployment are long-context reasoning, tool use and instruction following. The system also needs planning, coding and action. A model must know which steps are executable and which are not, while continuing to follow the original instruction through a very long task.
广密 describes Claude as having a “clear lead by a wide margin” in these agentic capabilities, closely tied to its coding advantage. The environment also matters: virtual machines, operating systems, browsers, tools, context and memory together form the world in which an Agent can act.
A friend who is also an “AGI fundamentalist” compared the inflection point to the Big Bang. Today, it is still possible to draw a single-point roadmap; once intelligence approaches AGI, applications will bloom and expand in every direction. 广密 accepts the description: slow early adoption does not mean growth will remain linear after the threshold is crossed.
17. Long Context Is Not a Longer Window; It Is Rebuilding Memory and Information Compression
A long-horizon task requiring 1,000 steps may depend on large amounts of output from every earlier step, while tool returns can rapidly push the context toward 1M tokens. The real challenge is enabling effective reasoning and self-reflection, not merely stuffing more text into the window.
Training data is the first missing piece. Historically, there has been almost no ready-made data involving simultaneous reasoning over 1M tokens. Costs also rise roughly quadratically with length. 广密 therefore believes Google’s One Million Token is closer to “finding a needle in the ocean” than a solution to true long context.
The long-term direction may be long-term memory replacing simple window expansion. Data in personal social, travel and e-commerce software is like money held in a bank, but there is still no “online banking” interface. If access could be standardized, the context layer could produce an “Alipay-scale opportunity.”
His infrastructure analogy is: intelligence is electricity, long context is a low-loss long-distance power line, MCP is the standard outlet, and each application is an appliance. Linear attention and other compression methods remain insufficiently dynamic. The ideal state may be writing information into model parameters and weights, but there is no standard answer yet.
18. Online Learning and Unified Multimodality Are the Most Important Candidate New Paradigms
The core of online learning is not the word “real-time,” but enabling a model to explore autonomously under conditions of survival and incentives, abstracting good workflows and automating them. It is closer to how humans learn and could allow Agents to accumulate experience rather than forget everything when each task ends.
The hard questions are the reward and update boundaries: when to update memory or weights, and how to define the objective. None has an answer. Chatbot interactions also provide sparse signals. Personalized LoRAs, memory Agents and multi-agent systems may offer partial implementations, but 广密 acknowledges they may not be enough.
Fusing pre-training scaling with RL scaling and applying RL during pre-training could count as “half a paradigm-level innovation.” A unified architecture with multimodal inputs and outputs offers even greater possibility. He views GPT-4o’s text-to-image generation as one step in that direction, but says it is “extremely compute-intensive.”
19. The Return of Pre-Training and Agentic Inference Will Drive GPU Demand Together
广密 sees 2 main compute-demand drivers. The first is the return of the pre-training narrative. The second is sustained inference and multi-agent execution for complex Agent tasks. The latter could consume roughly 1,000x as much compute as a chatbot, which is why he remains “very optimistic” about long-term GPU demand.
He believes NVIDIA’s biggest strategic move would ultimately be to become a cloud provider like AWS or Azure, while still preserving its existing cloud partnerships and therefore being unable to fully pursue that path. He considers CoreWeave’s GPU-cloud software capabilities clearly weaker than AWS and Azure’s. Having GPUs alone does not make a complete cloud.
贾扬清’s Lepton took a full-stack approach and correctly identified NVIDIA’s need for cloud software capabilities, making its acquisition by NVIDIA strategically logical. The value of the deal is not one tool; it is the software stack needed to move NVIDIA toward a complete cloud platform.
20. AI’s $20 Monthly Price Is Wildly Misaligned with the Value It Creates
广密 does not understand why products universally copy SaaS pricing at $20 per month. Traditional SaaS does not continuously consume large amounts of tokens, while AI subscriptions resemble “every household paying a $20 monthly electricity bill.” The more heavily users consume them, the worse the unit economics become.
If an investment researcher earns RMB1M a year and a mature Deep Research product can raise that person’s capacity to the equivalent of 3–5 people, paying RMB300,000 a year would be reasonable. More broadly, professional pricing of $1,000 or $2,000 per month may gain wider acceptance as the value is validated.
Model training is equivalent to investing upfront in the R&D cost of humanity’s next 10 years. Traditional software had to develop features one by one; now 70%–80% of the capability is already embedded in the model. OpenAI’s tens of billions of dollars of investment, leveraged by a few thousand people across hundreds of millions of users, is a “huge deflationary force” in an inflationary world.
广密 gives a rough picture of today’s highly imbalanced profit pool: NVIDIA captures more than 80% of profits, AWS captures 30%, Anthropic is loss-making and Cursor may even have negative gross margins. Over time, profits should migrate toward the model and application layers. If consumer pricing continues rising, OpenAI’s business model may not be inferior to Anthropic’s API model.
21. Learning Systems Will Eventually Swallow Feature Systems, but Wrappers Still Have a Window
“A model company’s secret has no shelf life,” because talent continues to circulate through the Bay Area. The era of publishing a bare model may be ending. Future moats are more likely to come from 2 paths: becoming the Cloud or becoming an OS with a developer ecosystem. Cloud GPUs are the hardware layer of that operating system.
For application investment, 广密 only wants to prioritize products driven by research. Founders should become “Prometheus stealing fire for intelligence,” translating spillover from model research into new experiences. If a company uses product growth to drive research backward, or treats user scale as its moat, he sees limited value.
The user-data flywheel is still weak. Chatbots mostly accumulate preference data, which can improve MAU but may not improve model capability. The truly scarce resource is capability data, and models are already smarter than most casual-chat users. “Casual conversation does not produce intelligence.”
A model is a continuously improving learning system; a wrapper is a feature system. More and more scaffolding will be internalized, forcing external systems to reset repeatedly. Wrappers can still make money for a period, but both execution and exit must be fast. Applications that want to scale over the long term will ultimately need the ability to tune models downward, conduct post-training and RL, and even train end to end.
22. Cursor, Perplexity and Manus Are All Racing Model Upgrades
Manus and Perplexity are both high-execution teams and both “people who opened Pandora’s box.” The difference is that Manus is like a token container: the stronger the model becomes, the more of its capabilities Manus can absorb. Perplexity gets less direct marginal experience improvement as models continue to strengthen.
Perplexity captured the window created by model companies’ failure to build a good search product for 2 years, allowing it to “grow quietly.” Manus faces more intense competition and does not have the same long runway. 广密 advises Manus to prioritize research quickly; it will ultimately need to enter post-training and RL.
Deep Research could take 2 paths: become a more advanced Deep Search product and continue replacing Google, or evolve into a professional Agent such as a co-scientist, co-analyst or co-worker. It could also integrate Operator, coding and more tools, eventually becoming the prototype of OpenAI’s general Agent.
Devin’s problem is that its task boundaries can easily be consumed by the model. When the model does not work, Devin does not work; once the model works, Devin’s incremental value falls. Cursor currently offers stronger interaction value, but Claude 3.7 Sonnet’s built-in scaffolding once conflicted with Cursor’s external scaffolding, forcing weeks of fixes. That experience also pushed Cursor to recruit design and training talent and experiment with its own end-to-end model.
23. OpenAI’s Lead Continues to Strengthen, but Its Biggest Risk Is Losing the Base-Model Lead
Despite criticizing its strategy, 广密 acknowledges that OpenAI’s practical lead over the past 2 years has strengthened. ChatGPT is approaching 700M MAU and has more than 400M weekly active users, continuing to increase its concentration in chatbots. By his estimate, OpenAI’s Sora captures roughly 80% of revenue in the AI-native market.
ChatGPT continues to add capabilities while Google Search has not improved by a comparable degree. Gemini often catches up with or even surpasses OpenAI on individual metrics 2–3 months after an OpenAI release, only to be “sniped” by Sam; its traffic remains small. OpenAI’s success has also killed off substantial startup space, while its ecosystem has not produced companies comparable to Cursor or Manus.
广密 believes GPT-4.5’s significance in pushing model size to a new level may exceed O3’s, because only by climbing higher can a model see new capability “scenery.” He guesses GPT-5 will be a hybrid model combining a GPT base model with the O series, potentially delivering a GPT-3.5-to-GPT-4-scale improvement in the summer. But the cost of applying RL directly to a large model remains a constraint.
OpenAI’s biggest risk is not near-term financing—it has just raised roughly $40B at a valuation of about $300B—but losing focus after growing to nearly 3,000 people. If a competitor first builds a materially stronger base model and developers then build a super app around it, ChatGPT traffic could migrate.
24. Microsoft, xAI, Mira and Ilya Are Betting on Different Time Horizons
OpenAI’s support for Anthropic’s MCP is viewed by 广密 as a sign that MCP could become “the TCP/IP protocol of AI Agents.” It also shows that the 2 companies, despite competing, have differentiated paths. OpenAI is more consumer-facing, Anthropic more enterprise- and ecosystem-oriented, and Sam continues to show respect for the old team.
If OpenAI had $500B in cash, it could become a new Azure-style cloud provider and enter fundamental conflict with Microsoft. 广密 believes a split would hurt Microsoft more: it “got up early,” only to let developer products be captured by Cursor and others. If foundation models are the next Windows, Microsoft may be losing its own operating system.
xAI’s strengths are writing, extraordinary execution, infrastructure and production capabilities, with Grok 3 in the first tier of base models. The merger with X primarily adds traffic and product distribution. X has more than 500M MAU, comparable with ChatGPT’s 600M–700M, but the old Twitter may not be Grok’s final product form.
Ilya’s ASI is closer to “creating an Einstein,” while AGI would simply let an average college graduate operate a computer. The former must generate new knowledge and light up the technology tree, which still requires a breakthrough. Mira’s team has assembled core OpenAI post-training and infrastructure talent and may first conduct post-training on DeepSeek V3, then move into mid-training, pre-training and products. It also represents a more open value system that puts AGI’s power in the hands of the community.
25. China’s Catch-Up Depends on Infrastructure, Open Source and Small Organizations Willing to Bet Big
广密 believes that if the US achieves AGI along the current path, China’s replication is nearly inevitable. The gap has narrowed rapidly over the past 2 years, with many capabilities catching up every 3–6 months. The US has greater concentration of top scientists, while China has hundreds or thousands of first-rate researchers supplied by its enormous education system.
If hundreds of thousands or millions of GPUs are not required, pure replication could take 30,000–50,000 or 50,000–100,000 GPUs. The enormous GPU fleets of US companies also support exploration and serving, while exploration may consume 5–10x more resources than replication. If this condition holds, China’s certainty of achieving AGI will be very high.
DeepSeek’s foundation is the solid infrastructure accumulated over time by High-Flyer Quant, while a paradigm-level scientific innovation from DeepSeek has yet to appear. Its impact on ByteDance was to pull the organization back toward the intelligence main line and reduce its excessive focus on traffic-driven growth for Doubao. 广密 most hopes ByteDance becomes China’s long-term AI lab in the mold of Google DeepMind.
Open source is “borderless” and may help China break through geopolitical containment. China has historically been good at scaling Silicon Valley’s 0-to-1 into 1-to-100, but DeepSeek showed that abundant capital, risk-taking, strong infrastructure and a small team can produce 0-to-1 innovation. If a good product develops a clear lead, founders do not need to be scared off by geopolitics first.
26. The Final Portfolio Is a Bet on Organizational Capability and a Renaissance in Science
Assuming every deal were investable, 广密 would allocate 25% each to Anthropic and Baidu, 10% each to OpenAI and Mira, 5% each to Ilya and Manus, with the remaining 15% undecided. He later said he would invest the same amount in Manus and Cursor, although Cursor was absent from the earlier portfolio list. If DeepSeek raised money, he would invest up to the fund’s 25% single-position cap.
At the time, the respective valuations were approximately $61.5B for Anthropic, $300B for OpenAI, $10B for Mira and $30B for Ilya. The episode later described ByteDance as also worth roughly $300B, but with strong revenue and profits that could fund AGI internally. Mira has a higher probability of success or acquisition by a major company, while Ilya offers greater upside from scientific breakthroughs and greater failure risk.
Organization and culture are “the core competitive advantage second only to compute.” A good AGI organization should be AGI first, research first and researcher first; it should have a true AI leader and reward its own bets. Youth, intelligence, imagination and execution matter more than seniority, while talent density matters far more than headcount.
The key question for judging an organization is “what is it willing to give up?” 广密 agrees that “we should no longer hold technology companies that are not all-in on AGI.” The Agent hype has only just begun. The future will include both large numbers of wrapper companies going public and Ilya-style new architectures, multimodal ChatGPT or entirely new Agent forms that destroy existing valuations.
His biggest bet for the next 3–5 years is “AGI bringing about a Renaissance in science.” Before 2030, several companies could exceed $10T in market cap. The next Google might be OpenAI, the next Windows might be Claude, the next Office might be coding or a general Agent, and AI for Science could produce the next Pfizer or Eli Lilly.