OpenAI’s 姚顺雨: A 3-Hour Interview on 6 Years of Agent Research
Summary
The AI Agent thesis is entering “the second half,” with competition shifting from training stronger models to defining more valuable tasks, environments, and rewards. 姚顺雨 believes language priors, reasoning, and reinforcement learning have finally formed a general recipe reusable across tasks: in the past, “you built different weapons for different monsters”; now “you have a machine gun,” and the question is where to aim it. As an investor, he would focus on verifiable tasks, real digital environments, and user context.
The core opportunity for startups is not to defend against model capability spillover, but to use it to create new interaction paradigms beyond ChatGPT. 姚顺雨’s view is that an old interface with a new model can be swallowed by the model company, while a new interface without a steadily improving model cannot stand on its own; both are necessary. Cursor shows that a “non-human” Copilot interaction can become an independent product. ChatGPT and other Super Apps will create path dependence, so betting on products with a different super-app form remains an important non-consensus position.
An Agent consuming 500–1,000x more tokens than a chatbot is not the primary problem; the real issue is whether the value created can cover the cost. “If something is very valuable, I spend $500 but it makes me $1,000, that’s not a problem.” Model prices will fall, and simple tasks can migrate to smaller, more specialized models. The real bottlenecks remain finding PMF and identifying tasks that can deliver enough economic value.
The fundamental advance of language Agents over AlphaGo-style systems is cross-environment generalization enabled by reasoning, not a higher score on any particular benchmark. 姚顺雨 uses DeepSeek as an example: RL on math and coding may also improve creative writing, whereas AlphaGo can only play Go. He still draws a boundary: generalization is a spectrum, logical tasks transfer more readily, and reasoning about human relationships may transfer along another path; this does not imply that every game or task will interoperate.
He sees long-term memory, intrinsic reward, and multi-agent systems as the 3 most important fundamental research lines for the next phase. Long-term memory lets a system accumulate knowledge like a scientist researching for 20 years; intrinsic reward lets it explore without external feedback; multi-agent systems expand a single “college-student intern” into an organization. The biggest commercial obstacle today is often not that O3 is insufficiently intelligent, but that an ordinary employee has 7 days or 3 months of accumulated tacit context that the model lacks.
A genuine data flywheel is far scarcer than the market narrative suggests, while memory may become the most direct commercial moat. A flywheel requires a company to train its own models and obtain clear rewards from interaction that align with product value. 姚顺雨 sees Midjourney as one of the few cases approaching this condition, while most applications still depend on foundation-model progress. ChatGPT’s memory and the personal and enterprise software connected through MCP could turn research advantages into user stickiness: “Essentially, environment is always the outermost part of the memory hierarchy.”
If allocating $50B, 姚顺雨 would consider both the model layer and companies that own user context, specialized environments, or distribution platforms. Model companies will probably continue to create enormous value, while the context accumulated by WeChat-like platforms resembles oil companies before the invention of the automobile. The future may also bring stronger centralized companies alongside more distributed Agent networks. His core worldview is not that centralization will replace diversity, but that “becoming more centralized” and “becoming more diverse” can happen simultaneously.
Deep dive
1. Non-consensus value judgments hidden inside the “model student” path
姚顺雨 describes his first 28 years as “extremely well-behaved”: he attended Tsinghua’s Yao Class from 2015 to 2019, pursued a PhD at Princeton from 2019 to 2024, and then naturally joined OpenAI, which he considered the best place to do research.
His rebellion first appeared in his taste in research topics. Yao Class favored theoretical computer science, but he felt that reducing the complexity of a graph algorithm from n^2.83 to n^2.82 had little real-world significance; determinism and community recognition do not mean a problem is worth solving.
He is not inclined to prove himself by skipping grades or being deliberately contrarian. He would rather bet on the upside: when “there happens to be something with an extremely high ceiling,” choosing incremental, low-risk work may mean missing the enormous space opened up by the era.
2. An embedding demo moved him from theory toward deep learning
A multimodal word2vec demo in 李建’s class in 2016 gave him his first shock: embedding operations linking king, man, queen, and woman could even extend across to image representations, showing him that learned representations could carry abstract relationships.
At the time, Tsinghua—and at least Yao Class—lacked deep-learning faculty and resources. In 2018, Yao Class required overseas research, so he went to MIT to work with 吴佳俊, where he began systematic exposure to deep learning, initially in computer vision.
His intuition for moving into language was straightforward: vision is difficult to use as a direct route to general AI, while “anything can be represented with language.” Language is not merely a vertical perception module; it is more like a central layer connecting knowledge, reasoning, learning, and action.
3. Language Agents revived AI’s original ambition
Looking back at AI in the 1960s, 姚顺雨 says its original goal was to solve vision and language, then combine the modules into an Agent smarter than a human. Because that goal was too difficult, research gradually fragmented into increasingly narrow vertical tasks such as translation, classification, and question answering.
Scaling laws and a series of breakthroughs after 2015 created an opportunity to move from vertical thinking back to general thinking. Language is special because “language is a tool humans invented to achieve generalization”: it may not be the most efficient medium for any individual task, but it can connect almost every task.
4. The first lesson in building Agents: “Use GPT, not BERT”
When 姚顺雨 entered the field, roughly “95% of people were working on BERT and 5% on GPT.” BERT was good at choosing among given options to classify sentiment or entailment, while an Agent needs to freely generate actions that did not previously exist.
His example is a text game: the system may need to “kill the monster with the sword,” “go to the third room,” or “use the golden key to open the door of the first room.” This is neither a finite action space such as up, down, left, and right nor a multiple-choice problem a classifier can enumerate.
The second lesson is that the environment determines the ceiling of capability. A Zork-style text game is harder than classification, but still resembles Go: RL can brute-force a single game to completion while transferring almost nothing to other tasks. “When you have a very bad task, you can never learn something very good.”
5. WebShop solved the environment first; ReAct connected reasoning and action
World of Bits in 2015 had already proposed treating computers and the internet as environments, but technology at the time prevented the idea from being fully realized. 姚顺雨 and his advisor returned to it in 2021; precisely because models were still immature, “not being mature” made it a good time to start.
In 2022, they built WebShop. Instead of continuing to optimize a closed text game, they tried to construct an Agent environment closer to the real internet, with practical value and evaluable outcomes.
GPT-3 and Chain of Thought then created a methodological opening, while ReAct organized reasoning and action in alternating steps. 姚顺雨 still considers it one of his favorite projects because the method was simple enough and general enough.
At the time, academia was reluctant to accept prompting as research: there was no mathematical formula, training run, or major engineering project, so it looked “too soft.” His counterargument was that if academia continued competing with model companies on training, it would likely just repeat work those companies had done several years earlier. Studying how to use models offered more unexplored space.
6. The definition of an Agent changes by context; language Agents differ through reasoning
In an NLP context, an Agent means moving from merely generating text or dialogue to using calculators, the internet, and other tools. In the older AI sense, any system that makes autonomous decisions, interacts with an environment, and optimizes reward can be called an Agent—including AlphaGo, Waymo, and robots.
The fundamental advantage of a language Agent is not its name but its reasoning. Faced with an unfamiliar game, a person may think, “The light is off, so there may be danger; the light is behind me, so I should move backward first,” connecting common sense, observation, and action through intermediate steps rather than guessing an action directly from complex text.
姚顺雨 believes reasoning is a key reason humans can do so many things: language models provide a sufficiently strong prior, the prior supports reasoning, and reasoning allows skills to generalize across environments.
7. Agents have passed through three waves: rules, deep RL, and language reasoning
The first symbolic wave tried to encode human reasoning as rules and produced expert systems. By the 1980s, people found that no matter how many rules they wrote, they could not cover the special cases of the real world; a system for diagnosing heart disease could not naturally transfer to lung disease.
The second wave of deep reinforcement learning used endlessly repeatable virtual environments and clear rewards to produce successful systems such as AlphaGo, Dota Agents, and robotic hands. But every environment required dedicated engineering: Go capability could not transfer to other games, much less to real-world applications.
The third wave is based on large language models. Reasoning lets Agents enter real digital environments such as coding, the internet, and computers. 姚顺雨 stresses that methods and task environments are 2 mutually reinforcing evolutionary lines; it is a mistake to focus on model architecture while ignoring upgrades to the environment.
OpenAI’s 5-level framework follows the same logic: language priors first support chatbots, then reasoners, and only later generalizable Agents. He is unsure whether innovators or organizations represent Level 4 and Level 5, believing the 2 are more likely orthogonal and to advance in parallel.
8. ReAct remains the dominant skeleton, but the next bottleneck may not be reasoning
姚顺雨 observes that most general-purpose Agents still use a ReAct-like structure: reason first, then generate an action. Specific tasks add workflows, but “the simplest thing may still work best.”
There is no consensus on the capability taxonomy. Agents can be classified by tools such as coding, web access, and computer use, or by capabilities such as multimodality, long context, and reasoning. If forced to choose, he currently cares most about context processing or memory.
Memory is not merely about extending the input window. It supports lifelong learning and online learning: after entering an environment, the system continues accumulating knowledge and updating itself instead of arriving at every task as if it were its first day on the job.
9. Code is the “hand” of a digital Agent
姚顺雨 calls coding one of AI’s most important affordances. Humans use hands to make hammers, pens, and chopsticks; digital Agents use code to operate machines. Web pages, novels, and GUIs were designed first for humans, while code is the interface naturally defined for machines.
This explains his early motivation for InterCode: if code execution returns results, why let a model generate a program only once instead of feeding execution feedback back into the loop to create a multi-turn Agent environment? This line later extended into SWE-bench and SWE-agent.
In 2022, he wondered, “It’s obvious that coding Agents are the most important—why is nobody doing it?” His resulting research judgment was that if you are convinced something matters but nobody is working on it, that is not necessarily a bad signal; it may be an opportunity before consensus has formed.
10. API and GUI do not have to be either-or; the real obstacle is complete context
On whether future AGI should be based on APIs and code or on human-facing GUIs, 姚顺雨 uses the analogy of cars and roads: should cars be modified to fit every road, or should every road be rebuilt to fit cars? The more likely answer is to “meet in the middle,” with both ends evolving together.
Many services lack APIs, but having an Agent use code, screenshots, and the frontend simultaneously may not be harder than adding APIs for the entire world. The final system will probably “do everything”: code will matter, but it will not eliminate GUI.
广密 asks whether a general Agent capable of operating almost every task on computers and phones will appear in 2027. 姚顺雨 does not directly endorse the timeline. Models’ coding, math, and logical abilities have in some sense surpassed those of most people, but “almost every task” is not well defined, and missing complete context is often more fatal than capability itself.
11. “The Second Half” declares that the bottleneck has moved from methods to problem definition
The phrase “The Second Half” originated with an invitation to speak at Stanford and became 姚顺雨’s summary after 6 months at OpenAI. People were accustomed to emphasizing training, algorithms, and models, but the central bottleneck had shifted toward defining good tasks and environments.
His core metaphor is that in the past, facing many monsters meant building many different weapons. Now there is finally a general method: “I have a machine gun, so the question I need to think about is where I should aim it.”
“The second half” does not claim that the methodological problems of audio, multimodality, and robotics have all been solved. It refers specifically to the main line from language priors to reasoning to Agents, where a generalizable unified recipe has emerged.
This also changes what is scarce in research. Once well-defined exams and games can be solved quickly by RL, the difficulty is no longer simply finding answers; it is discovering problems worth solving that can also create real-world value.
12. A good reward must evaluate outcomes and be white-box verifiable
Since WebShop, 姚顺雨 has preferred 2 principles: reward outcomes rather than process, and calculate rewards through clear rules rather than human preferences or another model’s black-box score.
The success of math and coding comes from this structure: “If the answer is 3, then it is 3.” A program either solves the problem or it does not. Rewarding the process can invite reward hacking; optimizing for human or model preferences can produce elegant but ineffective code.
The hardest part is not turning Amazon or Facebook into an environment, but satisfying difficulty, real value, clear rewards, and low noise at the same time. Engineering can be piled on; task and reward design determine what RL ultimately learns.
13. Agent tasks should be decomposed by reliability, creativity, time, and scope
One axis is reliability versus creativity. Customer service may require at least 99 error-free outcomes out of 100; proving the Riemann hypothesis, writing difficult code, or producing literature may allow 100 attempts as long as one is exceptional. The 2 require completely different optimization objectives.
The second axis is temporal depth. Cursor may modify a file in 3 seconds, while other work takes 30 minutes, 3 hours, or even 3 days. As duration expands, long-term memory becomes a core capability.
The third axis is task breadth. Fixing one bug differs from building a Windows-scale repo from scratch, corresponding respectively to an individual and a team or company; the latter naturally leads to multi-agent systems. Human career levels do not map directly onto Agent difficulty.
14. pass@k rewards inspiration; pass^k tests commercial reliability
Coding commonly uses pass@k: run the same task k times and measure the probability of succeeding at least once. pass@100 is well suited to asking whether 100 attempts can produce one good answer, but it obscures the requirement that every attempt succeed.
The τ-bench metric that 姚顺雨 helped develop uses the mirror metric pass^k: the probability that all k consecutive attempts succeed, or equivalently the risk of at least one failure. Customer service and similar businesses need to optimize this side, not merely success rate—or pass@1.
People overlook reliability because they are still “doing benchmarks” rather than taking responsibility for real applications. Once you adopt a business mindset, the metric the business needs is the metric you should optimize directly.
15. Model capability spillover is startup fuel, not a sin
Responding to concerns that foundation models will swallow Agent startups, 姚顺雨 says the opposite: “Startups should worry that models do not spill over capability.” Without continuously improving underlying capability, many new products cannot exist at all.
The optimal combination is a new interface plus continued model spillover. If a product still uses the ChatGPT-style interface, even a model update offers little reason not to be replaced by ChatGPT. If the interface innovates but the model stops improving, the new interaction cannot keep delivering value.
The boundary therefore may not be “model companies do general-purpose work while startups do vertical work,” but interaction design. Cursor’s value is not owning a chat box; it embeds AI in an IDE and creates a form of collaboration between humans and the digital world that did not previously exist.
16. Super App path dependence leaves a window for new interactions
张小珺 asks whether ChatGPT can also adopt startup interface innovations. 姚顺雨 acknowledges that it can, but says Super Apps are double-edged: once a successful form exists, research, organizations, and resources are reorganized around it.
His historical analogy is Google in 2020. It had the Transformer, top talent, and unlimited resources, yet the most natural choice was still to improve search first. Similarly, once a company owns a Super App, most resources will continue to revolve around its existing interaction rather than evenly exploring every possible interface.
A Her-style voice assistant is valuable and highly natural, but it is still essentially a human-like assistant—a position model companies already occupy. The non-obvious space is interaction that does not resemble a human and may initially feel unfamiliar. Google versus Yahoo’s “yellow pages” is an earlier example.
Canvas is a direction he considers worth exploring: generate the most suitable frontend online for a specific task, context, and personality. It is more flexible than a fixed chat box and more difficult to execute well.
17. Most applications lack a data flywheel, and owning data does not equal owning trainable assets
姚顺雨 believes most AI companies have not yet built a data flywheel; growth mainly comes from better foundation models. A genuine flywheel first requires the company to train models itself, then obtain data and rewards from interactions that clearly distinguish good from bad and align with product value.
Midjourney is the relatively successful example he cites: users choose the images they prefer, so the reward is directly aligned with product value. As preference modeling improves, both the model and the company improve, with no obvious misalignment between the training and commercial objectives.
An application must also be sufficiently outside the model company’s main line. Coding is a core capability every major company will strengthen. Whether Cursor should do pre-training depends on how much gap remains between closed and open models and how many resources it has to close that gap. He would train models, but would not unconditionally redo pre-training.
18. Memory, intrinsic reward, and multi-agent systems point toward innovators and organizers
姚顺雨 sees the Agent ecosystem as 2 trees: one is the evolution of fundamental research, the other the evolution of applications and interaction methods. Cursor mainly innovates on the latter, and the 2 lines require different people to explore them.
The 3 fundamental research priorities are memory, intrinsic reward, and multi-agent systems. Memory and intrinsic reward jointly support the innovator, while multi-agent systems support the organization.
He uses a long-term researcher as an example. Researching a theorem may take 20 years, requiring long-term memory. Before the final proof, there may be no award, salary, or immediate feedback; the researcher can continue only through an internal judgment of value.
Today’s Agents are more like “ordinary college-student interns” who can work on a computer. The boundaries human society truly admires lie at one end with Einstein-, Van Gogh-, and Beethoven-like creators, and at the other with Musk- and Jobs-like creators of organizations.
19. AI’s utility problem may fundamentally be a context problem
姚顺雨’s question is: if models are strong at exams, reasoning, and games, why have they not created economic value commensurate with their capabilities? His belief is that the root cause may not be insufficient intelligence, but missing complete context.
Companies write some knowledge into Google Docs, Notion, and text records, but a boss’s habits, team tacit coordination, and historical trade-offs remain distributed across human brains and cannot be handed over in a single briefing. This keeps people who can enter an environment and accumulate context difficult to replace.
An ordinary employee may be worse than O3 at math and logic, yet possess after 7 days on the job background that O3 lacks and therefore perform better. “You are not as smart as O3, but you have this context” is one of the most overlooked advantages in current human-model comparisons.
Long-term memory may ultimately be implemented through weight updates, language memory, embedding memory, or a new architecture. 姚顺雨 believes the problem “will inevitably be solved,” but remains uncertain about the specific mechanism and timeline.
20. Intrinsic reward must transfer an infant’s curiosity into the “text game”
An infant bites toys and repeatedly manipulates objects without promotion, bonuses, or external rewards. Curiosity, a sense of control, security, and novelty drive active exploration and learning.
The difficulty is that infants first learn a physical game composed of vision, touch, hearing, and skeletal control. Adults act in a “text game” composed of language, institutions, companies, and interpersonal relationships. The 2 forms of intrinsic motivation may not transfer directly.
Traditional RL can use a world model to define novelty in mazes or robot simulations. How a system should judge what is worth exploring and what constitutes real progress in a linguistic world remains an open question.
21. Learn from human capability gaps, not by copying the brain
姚顺雨 once worked in Josh Tenenbaum’s cognitive science lab. The early idea was to study why humans can few-shot generalize and then transfer the mechanism to machines. He later found that the scaling laws, RL, and learning methods that actually work may differ substantially from human learning.
A more robust approach is to first ask, “What can humans do that machines cannot?” and then solve that gap from first principles. Humans can accumulate context during a 3-month company internship—that is an objective gap—but a system does not necessarily need to copy the biological structure of the hippocampus or episodic memory.
Whether an Agent should become more human-like should also be judged by utility. Playing Go need not resemble human play, and driving may benefit from different rules. But when entering a company, understanding a boss, and completing ambiguous tasks, humans currently perform better; research should focus on bridging that gap.
Long-term memory does not automatically imply an “ID card.” If a lonely person needs a friend, a human-like Agent has value. If the system is experienced as a platform, recommendation page, or game, users may not perceive a “person” at all, making personification unnecessary.
22. AI may strengthen centralization, liquidity, and diversity at the same time
Discussing Agents and crypto, 姚顺雨 first says he does not know much about crypto, then abstracts the question to whether technology will become centralized or decentralized. The centralization case is that OpenAI and Anthropic could become $1T, $10T, or even $100T companies, concentrating compute, platforms, and resources.
The decentralization case is that if intelligence becomes as cheap as electricity, it will reduce information, cognitive, and capability gaps while empowering large numbers of individuals. The 2 forces do not necessarily have to produce a winner and a loser.
He sees society as a network. History has both increased the concentration of resources and widened inequality, while also increasing the speed and possibility with which ordinary people can move from the periphery toward the center. Ancient clans, aristocracies, and caste systems were more rigidly fixed.
The future may also bring distributed multi-agent networks. Each person’s Agent could hold local information and exchange information, transact, or collaborate with millions of others. Central nodes would have incentives to continue concentrating power, while peripheral participants would have incentives to protect their own information and rights.
23. OpenAI surpassed DeepMind through a different bet
姚顺雨 emphasizes that this is his own summary and speculation, not the company’s official narrative. OpenAI began as a bottom-up research lab, with RL naturally at the center because DeepMind was the most prominent company at the time and AlphaGo its signature project.
But “you can only surpass the previous hegemon if you have a different bet.” If OpenAI had done only RL, it might have surpassed DeepMind locally, but the public would still have associated the field first with DeepMind. GPT ultimately became that different bet.
The degree of non-consensus was greater than the retrospective narrative suggests. His advisor Karthik was the second author of GPT-1 and still doubted that the result was good enough or worth the cost. Even being inside OpenAI and directly involved in GPT did not mean one already believed scaling up was the optimal direction.
姚顺雨 sees one of Ilya’s key contributions as pushing the organization to go all in. GPT-3 also required someone willing to commit massive resources. An organization does not need universal consensus; it only needs enough people to form a conviction that allows a non-consensus experiment to happen.
24. GPT did not replace RL; it gave RL a generalizable prior
OpenAI did not abandon RL after choosing GPT; it merely shifted the allocation of resources. Later, ChatGPT still relied on RLHF and alignment. History was not one route being completely rejected, but several lines coexisting while more promising directions attracted larger bets.
Traditional RL theory says that with a good enough environment and reward and enough samples, no prior is needed. But in internet and language environments, brute-force learning might take “10^30 years,” which is equivalent to impossible on the scale of the universe’s age. Pre-training makes language RL practically feasible.
The new change is transfer. RL on math and coding in DeepSeek may also improve creative writing, while AlphaGo learning Go does not make it play chess. 姚顺雨 sees this as a fundamental difference, while stressing that generalization remains a spectrum and transfer strength depends on the nature of the task.
25. New scaling dimensions will come from memory and organization
姚顺雨 expects more moments like GPT-3, but not necessarily from making pre-training simply larger. Long-term memory will create new forms of test-time compute scaling, while multi-agent systems will add a scale dimension based on parallel collaboration.
As the number of skill dimensions increases, the question will become how to select and weight different skills for specific applications.
On “ultimate generalization,” he cites a phrase: “overfit the reality.” If a system already covers everything in reality, debating whether it is memorizing or generalizing becomes less meaningful in practical terms.
26. Agent cost should be measured against the value created
An entrepreneur raised the concern that an Agent could consume 500–1,000x more tokens per user than a chatbot, and that multiplying this by millions of users would make the compute burden unsustainable. 姚顺雨’s response is that cost itself is not the issue; the issue is when cost cannot justify performance or value.
“I spend $500, but it makes me $1,000” is commercially viable. Simple tasks can use smaller models to lower costs, while high-value tasks such as investment and deep research can use larger models. The first priority is finding an application worth paying for.
He also does not believe Agent founders necessarily need to be researchers. A research background may provide understanding, but it may also leave someone “holding a hammer and looking for a nail.” Product sense, identifying a real pain point, and finding PMF matter more than coming from the NLP community.
27. Manus shows that product innovation can also be learned in reverse by model companies
The host mentions Minus and Jazz Park. 姚顺雨 says he has tried Minus but not Jazz Park, then describes Manus as “pretty interesting,” praising its product sense and polish. The original text alternates between the names Minus and Manus, and the 2 should not be treated as identical.
The traditional narrative is that large companies innovate first and startups copy later. Manus suggests the reverse path is also possible: a small company creates a new Agent interaction or product form, and a model company later borrows it while adding its own underlying capabilities.
The relationship between model companies and application companies may therefore be “mutual copying,” not one-way absorption. An application company without underlying capabilities can be copied, but a model company can also borrow the interactions created by startups.
Asked whether Manus should immediately become vertical, his answer is not either-or. Keep a sufficiently general product shell with a high ceiling for imagination, while periodically finding killer apps such as PPT or deep research—just as iPhone, iPad, ChatGPT, and WeChat gained momentum through specific applications.
28. DeepSeek’s impact was both the productization of Chain of Thought and a revaluation of open source
姚顺雨 sees DeepSeek’s long Chain of Thought display as a breakthrough in product form. Technical accumulation is like a flood reaching a dam; it needs a moment that makes ordinary people feel something “magical.” ChatGPT, iPhone, and DeepSeek all served as release points.
The second shock was a renewed examination of open source. The US industry had assumed that open models would lag closed models significantly because model training is not like Linux, where thousands of people can contribute independently; it relies more on a small number of top researchers, huge funding, and highly concentrated organizations.
Doing open source well is not trivial. It requires resources, talent, organizational culture, and commercial justification. Relying entirely on philanthropists with hundreds of billions of dollars to fund it indefinitely is a low-probability scenario, but DeepSeek has prompted a reexamination of these default assumptions.
He also specifically praises DeepSeek’s organization, engineering, and infrastructure, rather than reducing its success to a single algorithm or a cost story.
29. The intelligence frontier will be defined by multiple systems, not one strongest model
姚顺雨 used to imagine AGI as “the smartest person in the world”: one model with all knowledge and capabilities. He now believes different interaction forms will create different kinds of strength, and the intelligence frontier may not be rankable by a single model.
A companion-like friend may not need to be strongest in math, physics, and memory. Being too rational and never making mistakes might make it feel fake. There is no single answer to whether it is “stronger” than ChatGPT because the value functions differ.
OpenAI may become the Google of the new world—an extremely important node without monopolizing every form. “Becoming more centralized” and “becoming more diverse” are not contradictory: the largest company can become more dominant while the number of industries, interfaces, and systems continues to grow.
姚顺雨’s core bet is “a product form with a different super app and different interaction methods.” If you do not believe this, the world is left with only a few players such as OpenAI and Anthropic. If you do, applications, platforms, and new model paths still have substantial room.
30. Memory can turn model leadership into user stickiness
Comparing memoryless chatbots is essentially a comparison of whose model is stronger, with the lead fluctuating whenever a new model is released. Once memory is added, context accumulated on a platform improves the experience and raises retention.
姚顺雨 therefore sees memory as a potential commercial moat for ChatGPT. The more a user interacts, the deeper the accumulated context and the more accurate the retrieval, the more likely research advantages are to become product stickiness rather than merely a one-round benchmark advantage.
He admits that he does not know the internal details behind ChatGPT’s “memory updated” mechanism. He only speculates that the system is better at extracting or retrieving relevant information from large volumes of dialogue, without presenting a personal inference as an official explanation.
31. MCP and external environments belong to the same memory hierarchy
张小珺 points out that personal and enterprise context is scattered across many software products and asks whether MCP is fundamentally also memory. 姚顺雨 says “in a sense, yes”: it lets an Agent enter external environments where context is stored.
He quotes an idea from von Neumann’s The Brain and the Computer, written shortly before his death: “Essentially, environment is always the outermost part of the memory hierarchy.” Beyond CPU cache, memory, and disk, USB drives, the internet, and the physical environment remain outer layers of memory.
The same applies to humans. Beyond working memory and long-term memory inside the brain, notebooks, Google Docs, and Notion are the longest-term and most external forms of memory. The long-term memory problem for Agents cannot be discussed only within model parameters or the context window.
Long context is one way to implement long-term memory. With 100M, 100B, or even unlimited context, a system could remember everything through a route different from humans. But 姚顺雨 does not believe this proves it is the only or most suitable solution.
32. “Needle in a haystack” is only a necessary condition for long-term memory
Current long-context evaluations often hide one fact inside extremely long text and ask the model to retrieve it. 姚顺雨 calls this necessary but not sufficient: if a model cannot do it, memory is out of the question; if it can, that is still far from proving it can operate, accumulate, and update over the long term.
Without better tasks and evaluations, it is difficult to meaningfully debate whether linear, sparse, or hybrid methods are superior. The problem is not only architecture; researchers have focused on an easily quantified necessary condition without defining a harder, more valuable sufficient condition.
He especially wants to overturn the assumption that averaging parallel samples is a valid evaluation. Averaging one person’s performance on the first day of employment across 100 parallel universes does not measure how that person grows after working for 1 day, 30 days, or 1 year.
The real world has no standard answers or fixed rewards, which is precisely why it is difficult. Once exams and games are well defined, existing general recipes can often solve them quickly. The North Star for Agents should return to real-world value, not leaderboards that are easier to optimize.
33. Two complementary types of Agent products will emerge over the next 12–24 months
姚顺雨 expects chatbots such as ChatGPT, Grok, and Anthropic Claude to evolve naturally into Agents. Chat will remain a subset, but the default interaction will increasingly favor systems capable of executing tasks.
The second opportunity is a Cursor-style Copilot. Rather than taking over everything from a remote virtual machine, it embeds an Agent inside existing software, computers, and industry settings. The first type creates a new environment; the second expands the user’s existing digital environment. The 2 are orthogonal and complementary.
Even if an application founder cannot do pre-training, they can control 2 key interfaces: how an Agent interacts with the digital world through MCP, APIs, and similar tools, and how humans interact with the Agent. Accumulating user intent, context, and an ecosystem also requires substantial product, infrastructure, and engineering innovation.
Asked whether a massive number of Agents 2 years from now will need dedicated virtual machines, browsers, identities, and economic systems, he expects the next 2 years to remain relatively centralized, with a few Super Apps building their own infrastructure. Beyond that, the uncertainty is too high: “No one can predict what will happen 2 years from now.”
34. Pre-training and RL will recombine according to application economics
姚顺雨 summarizes the second half with the phrase “the world’s affairs move from division to unity, then from unity to division.” A decade ago, translation, games, ticket booking, and mathematics belonged to separate communities; today, one method can handle them for the first time, much like Newtonian mechanics unified previously scattered phenomena.
The Agent is the goal; pre-training and RL are tools. Pre-training provides the prior needed to handle the language world, while RL improves behavior through environments and rewards. Debating which will replace the other mistakes technical means for the final product.
Closed, high-value environments with abundant feedback may rely primarily on RL. AlphaGo, Google Ads, and TikTok recommendations are close to this structure. Long-tail tasks require entering unfamiliar environments and adapting online, where the generalization value of pre-training is higher.
Whether to restart expensive pre-training ultimately depends on whether additional value can cover the cost. Only when genuinely different Super Apps require different capabilities and models, and the commercial value is large enough, will training from scratch become economically rational rather than a strategic posture.
35. $50B should buy intelligence, environment, and context at the same time
If he became Berkshire’s CEO and had to allocate $50B to AGI, 姚顺雨 would consider model-layer companies such as OpenAI and Alphabet, believing they would “probably” continue to create greater value rather than claiming a certain return.
Another category would be companies that accumulate user context or build specialized environments. If AGI is a system, it requires not only intelligence but also an environment and an understanding of users. Companies with large amounts of user data or context resemble oil companies before the automobile was invented: the oil already existed, but the major application had not yet emerged.
WeChat-like platforms are therefore “easy to defend and hard to attack,” with networks, scenarios, and large amounts of context. As intelligence becomes cheaper and more democratized, scarcity may shift toward platforms, environments, and user relationships.
His first recommendation to the CEOs of traditional companies is to learn AI personally. If the CEO does not understand it, the organization may be misled even after hiring experts. The next step is to create new value from existing scenarios and resources, not mechanically copy a ChatGPT unrelated to the company’s own strengths.
36. WeChat can avoid rushing, but must watch for discontinuous innovation
If he were responsible for WeChat, 姚顺雨 would not rush to Agentify every feature. He would first learn and observe new interactions. From a position that is easy to defend and hard to attack, launching incremental innovation a few months earlier or later may not change the landscape.
The real danger is a substitute with a completely different form, just as WeChat defeated QQ not by building a product more like QQ. Established giants should watch for the user network itself being rewritten by a new interaction, not merely for competitors adding similar features.
The host counters that if long-term memory and multi-agent systems mature while an Agent system is not built on WeChat, the old network could lose value. 姚顺雨 accepts this conditional argument: the outcome depends on how much of future interaction consists of human friends, Agent friends, and professional interactions.
37. Safety includes at least commercial reliability and broader risks
姚顺雨 believes “safety” often mixes different questions. The first layer is product safety: if an Agent deletes data or keeps making errors, the product has no commercial value. Even a company focused only on profit must solve this.
The second layer lies beyond product responsibility and concerns more ideological or human-scale safety. The industry has no clear consensus on whether additional responsibilities are needed or how to define the boundary. He admits he worries about it, but also finds it difficult to give a definitive judgment.
His current priority is controversial but clear: AGI has not yet been achieved or generated enough value. Before knowing how to make it useful, discussing whether it is “too powerful” or “too widespread” does not fully make sense. For applications with clear value, safety as a necessary condition will continue to receive investment.
38. The consciousness question temporarily loses to the definition question
姚顺雨 does not believe “consciousness” has an operational definition yet. If a system can process sufficiently complex context, has high autonomy and decision-making power, continuously processes information, generates ideas, and chooses consequential actions, it “might” objectively be called conscious.
This is not a claim that models must develop consciousness, but an attempt to push the debate back toward definition. Which observable capabilities constitute consciousness remains undefined; without that definition, a simple yes-or-no answer based on machine principles is premature.
39. Great methods often grow from good tasks; ReAct instead searched for its task
姚顺雨 summarizes his 2 research lines as simple, general methods and tasks that create value in the real digital world. Computers, coding, and the web sit between game-like exams and the physical world, forming a “virgin territory” that had not been sufficiently explored.
Most great methods are inspired by concrete tasks. Translation, for example, naturally requires nonlinear attention to different positions in a sequence and was well suited to giving rise to attention. 姚顺雨’s unusual experience was to first believe a method was sufficiently general, then search for a task that could demonstrate its initial signal.
The hardest part of ReAct was not writing down a “think–act” framework, but finding an environment that could show its potential when the available model was still something like early GPT-3. He calls this a “method-task fit” analogous to startup PMF.
Research therefore cannot remain trapped inside RL, NLP, or a single community. ReAct combined language tasks with game environments; by connecting the boundaries, it allowed a method to reveal its generality.
40. His radical task expanded from “building a scientist” to “building an organization”
At Princeton, the most natural ultimate task for 姚顺雨 was to create an Einstein-like scientist: an AI capable of discovering the next theory of relativity or a grand unified theory. That alone would mark AGI, perhaps even ASI.
After moving to Silicon Valley and entering companies, he began to see human organizations as an equally important research object. If AI could create a $1T company, that would be a different but equally difficult achievement from discovering relativity.
Organizational mechanisms resemble general methods. Equity structures, incentive design, and organizational architecture are not individual products, but they can continuously produce many different great products. Innovators and organizations therefore need not come sequentially; they are more likely 2 routes that can be explored simultaneously.
41. His fixation on generality runs through research, reading, and rap
姚顺雨 enjoys books, films, and different places, and once competed in informatics and won a national silver medal, but unlike some classmates he did not focus exclusively on grinding one skill to its limit. He gradually realized that even the smartest individual can master only a small fraction of human knowledge, and wanted to create “something more general and more cutting-edge than himself.”
Language may not be the best representation for an individual task such as Go, but it is a general representation formed to connect all cognitive tasks. AI may create a more efficient new language, but humans already possess a strong prior and want to understand, monitor, and control machines, so English or an English-based representation may remain dominant.
Rap taught him a similar lesson about distinctiveness. Everyone has a personal flow, personality, and way of thinking about life: “It may not be the best thing, but it is something different.” When GPT-3 first appeared, he tried generating lyrics with both rhyme and content and found that this remains difficult.
The difficulty is not merely next-token prediction. A good, distinctive flow lacks a clear reward, while repeating one style too often destroys its value. More fundamentally, great rap comes from life, and AI “does not have a life yet.”
42. “Let others do what they can do” defines his personal bets
The sentence that influenced 姚顺雨 most from a mentor was: “If someone else can do it, then it’s okay to let them do it.” This does not claim that any work is absolutely irreplaceable; it reminds researchers that if someone else is already doing the same thing, they can use their energy to explore a direction nobody has yet validated.
He does not rule out starting a company and acknowledges that many people at OpenAI are considering it, but wants to think clearly first about what is worth doing. Even if copying the assistant form might be easier to monetize, he still finds it more interesting to create something completely different, even if it ultimately fails.
Asked what bet he has recently made based on taste, he only says, “I’m doing something.” His public advice to Agent founders remains simple: technology is a tool; the most important thing is to understand the incremental value delivered to users, and to keep betting on different Super Apps and different interaction methods.