A Conversation with 杨植麟 One Year Later: K2 and Agentic LLMs

Summary

• K2’s core bet is token efficiency. 杨植麟 sees the data wall as “an objective fact” and high-quality data as “close to a constant,” so the breakthrough is not training faster but “getting more intelligence from the same amount of data.” The Muon optimizer delivers roughly a 2x gain under compute-optimal scaling—“if you have 30T high-quality tokens, it is equivalent to 60T”—and is “at least in publicly available information, the first non-Adam, matrix-orthogonalization-based optimizer used to train a model at this scale,” a non-consensus bet that has now been proven out.
• Agent generalization is the biggest unsolved problem today. It is also his most important current bet. Existing RL is “what you plant is what you harvest”: scoring 99 on one type of math competition may improve performance on other math problems by only five points; benchmarks are scarce and often break down, and “scores often do not reflect capability.” The answer points toward “using AI to train AI”—having models participate in their own development, “using L4 technology to solve L3 problems”—which means OpenAI’s L1-L5 ladder is not a linear staircase.
• The two forms of test-time scaling are diverging, but will ultimately converge. One is “brain in a vat” pure reasoning—the o1 route, turning pass@k into pass@1—and the other is the agent route, with repeated tool interaction. Cloud models show that reasoning need not be especially strong for agent performance to be strong: “behind them are two different scaling dimensions.” But to solve the most complex tasks, “in the end you need both; it is only a matter of time.”
• First-party products are replacing the scaffolding logic. Building products on someone else’s foundation model is essentially “reverse-engineering the model’s training process”—guessing Anthropic’s tools, system prompts, and context engineering—while model companies design the environment first and train inside it, so “the model will naturally perform better in your environment,” potentially raising the ceiling. “The model is the product” still holds in his view: “once the model is trained, your product is already done.”
• His open-source view has publicly reversed. Last year he said leaders would not open-source and only laggards would; this year, when the host confronted him with the old quote, his answer was direct: “because we have not achieved complete leadership yet.” The community can help serve a model, but “to make the model itself better, only the original developer can do it.” The opportunity is downstream: fine-tuning specialized agents on K2. His end-state view is that global model companies will consolidate from hundreds to “a few”—“that now looks highly probable.”
• Commercialization: invest first, and the certainty is actually higher. The clear business models are APIs and first-party products, but the top priority remains the model; the market is already validated and growing rapidly—“leading companies have ARR in the several-billion- to tens-of-billions-of-dollars range, and can double or triple every one or two quarters.” Why has no data flywheel emerged? FLOPs-based scaling is too powerful—RL is on-policy and uses negative gradients, making it more efficient than pretraining—and large models remain sensitive to noisy user feedback.
• Manage organizations with RL, but beware reward hacking. Research, model training, and team management share the same structure: use RL—set goals and give rewards—as the main mechanism, with limited SFT as a prior to prevent people from “flying too far,” but “too much SFT makes people lose their initiative.” The biggest risk in RL-based management is being hacked. His latest understanding of the CEO role is to “balance RL, SFT, and reward hacking.” AI’s meaning, as Kimi put it for him: “an amplifier of human civilization.”

Deep dive

1. Another Stretch Up the Snowy Mountain: More Clarity, and New Unknowns

• A year and a half after their last conversation, 杨植麟 opened with “one day in AI is one year in human society,” though the feeling of climbing a snowy mountain “has not changed much.” The progress is tangible: two years ago, models “could barely write an article clearly”; today they “can work continuously for several hours to complete a potentially complex coding task.” New scenarios have been unlocked, but higher up the mountain the view is still familiar: large numbers of unknown technical problems remain.
• What has become clearer is how different RL paradigms produce reasoning and agentic capabilities. Two years ago, “it was not that clear”; the main path was pretraining plus RLHF. Those questions now have answers, but new ones have emerged: RL ultimately depends on good evaluation and verification. Math problems and coding tasks with test cases work reasonably well, but “for more complex end-to-end tasks, it is not easy to find ways to evaluate or measure them.”

2. Two Sentences to Carve in Stone: Problems Are Inevitable, but Solvable

• He repeatedly reads The Beginning of Infinity. The 2 sentences he most wants carved in stone are: “problems are inevitable” and “problems are solvable.” Before the Enlightenment, society was static: thunder was blamed on the Thunder God, snow on the gods being in a bad mood—bad explanations, with only a tiny number of people creating knowledge. Society then became dynamic: every solved problem creates new ones “because the boundary of your knowledge is expanding.” AI R&D is exactly this process: solve many RL problems, and new problems around evaluation, measurement, and verification appear.
• The mountain may have no summit, and his attitude is: “I hope it never ends—that is what the beginning of infinity means. It may be an infinite mountain.” Nor does he necessarily have to climb it alone. K2 is already being used for model-training and data-processing work that “would previously have required people to write code manually; if someone could not code, they could not do it.” The model itself becomes an amplifier for the climb.

3. AGI: From Destination to Direction

• What are they pursuing? “The process of climbing, or being able to keep climbing higher.” AGI is not a particular step on the staircase: “it is not that you reach this step and suddenly achieve AGI overnight.” Translation is already better than 99% of humans, and at the current rate of improvement, many math competition and programming problems “will soon be fully solvable.” But it will be difficult to declare at a particular moment that AGI has arrived.
• The company is named after a moonshot, but he draws the distinction himself: “a moonshot ends the moment you stand on the moon.” AGI requires continuous progress. Technology’s impact on society will also unfold over a much longer horizon: as with the steam engine, society may take decades or centuries to absorb it. “Some jobs will no longer be necessary, but you will create new jobs … everyone will become a superhuman.”

4. One Paradigm: Strong Reasoning Models Turn pass@k into pass@1

• The first paradigm-level change of the past year was strong reasoning RL, represented by o1. He breaks “reflection” into 2 capabilities: generating new conjectures and implicit self-verification. Rather than explicitly training a verifier, the model generates a conjecture, verifies it, and generates another during reasoning. It “effectively tries several times … turning pass@k into pass@1,” much like the way people conduct research or solve problems.
• 小珺 described this as a free, nonlinear exploration process. 杨’s correction is worth recording: under serial sampling, “it is actually close to a more linear process,” because each new conjecture is based on conjectures already ruled out. Parallel sampling can test multiple paths simultaneously, “but some recent papers argue that serial reasoning has a higher ceiling. That may be related to our experimental findings.”

5. Brain in a Vat and the Agent Leaving the Fishbowl: Both Are Test-Time Scaling

• A reasoning model is a “brain in a vat”: “Imagine a fishbowl with a brain placed inside. It has no connection to the outside world and just thinks inside its own head; it can still solve a problem.” An agentic-RL-trained model thinks while acting—“calling search for a while, using a browser for a while, writing a few lines of code”—with its next action determined by state updates from environmental feedback.
• Both point to the same thing: test-time scaling. In the conversational era, a single response might contain a few hundred tokens. Now, whether you increase the number of turns or add more reasoning tokens within each turn, “both are ways of scaling up token usage so that you can complete more complex tasks.” For example: “clone a code repository, translate it into a new language, debug and test it, fix every bug, and get it running normally”—an end-to-end task lasting several hours with no human involvement.

6. The First-Party Product Logic: No More Reverse-Engineering Anthropic

• The third trend is model companies building first-party agent products. Earlier products that built scaffolding on foundation models, such as “Minus,” were essentially “reverse-engineering the model’s training process”: guessing which in-house tools Anthropic used, which system prompts worked, and which context-engineering methods worked. “It was an inverted process.”
• The first-party logic is entirely different: design the tools and context engineering first, then train the model in that environment. “Your model will naturally perform better in your environment.” The company can also train end to end and adjust the tool design in response. “The second approach may have a higher ceiling.” Clock Code and ChatGPT Agent are examples. Kimi has made relatively little first-party investment so far; “our main line is still the model.”

7. L1 to L5 Is Not Serial: Agents Can Move First

• 小珺 asked about the logic behind OpenAI’s L1-L5 classification. 杨’s breakdown: the ceiling of agents depends on very strong reasoning, “but it does not seem to be strictly necessary.” Reordering the development path—building agents first and narrow long-CoT reasoning later—“may actually work.” The cloud route may already be ahead on this point: “reasoning performance is not particularly high, but agent performance is high.” Behind this are 2 different test-time scaling dimensions: scaling the number of turns and scaling pure thinking.
• But the paths ultimately converge: “if you want it to solve the most complex tasks, your reasoning ability also has to be very strong … you still need both; it is only a matter of time.” He sees the levels as several important technical milestones, but not a completely serial sequence. All the capabilities must continue improving.

8. The Mark of Innovation: K2 Participating in K3’s Development

• Why does innovation come after agents? “The key point is when your model can actually participate in model development. We hope K2 can participate in developing K3”: generating new ideas, running experiments, analyzing results and drawing conclusions, iterating the next version, or optimizing infrastructure performance—all require strong agentic capabilities.
• Organization will be a multi-agent system, and an early version already exists: fork many different agents from one agent, with some writing tests, some writing documentation, and some designing the overall software architecture, then run them serially or in parallel and merge the results. The challenge in end-to-end training is avoiding overfitting to a few agent types. Innovation and organization “are not necessarily linear; the 2 may happen together.” Is organization the peak? “I do not think so. It may really have no summit.”

9. 2 Key Turns of the Year: K1.5 Validated RL, K2 Targets the Best Base Model

• Looking back at the key decisions from 2024 to 2025, he identified 2. The R&D paradigm shifted from focusing on pretraining and SFT to “focusing on pretraining and RL,” requiring new talent reserves and a different way of working. The second was the shift “from conversation to agents,” which “significantly changed how we actually work.”
• K1.5 was primarily a technical validation of RL. They entered the path early and found that they “did not need much process reward or a value function”—these could even have side effects in training. Using end-to-end reward directly “was enough to train it very well,” while the team accumulated RL infrastructure and algorithmic know-how.
• K2 has 2 goals: a very strong base model and strong agentic capabilities. The first requires answering where the pretraining bottleneck lies. “The growth of high-quality data is indeed very slow … high-quality data is close to a constant.” The core question therefore becomes token efficiency: “given the same amount of data, can you grow a bigger brain?”

10. Muon: Treating 30T Tokens as 60T

• He first draws a line between 2 things. Optimizing training-system performance “is highly valuable, but training faster by itself cannot raise the intelligence ceiling, because you still have the same number of tokens.” Unlike Adam, which has been used for a decade and treats each parameter independently, Muon—proposed by Keller and previously adapted by 月之暗面 in Moonlight for language models at a meaningful scale—accounts for dependencies among matrix parameters and learns more efficiently. “Under compute-optimal scaling, there is basically a 2x improvement. If you have 30T high-quality tokens, it is equivalent to 60T.”
• 小珺 pressed him with the original question: “But it is still the same amount of data.” 杨 replied: yes, “but after learning, its brain grows faster. You feed it the same amount of data, but it absorbs it better; the compression ratio rises faster and the loss falls faster.” Increasing sparsity while adding more parameters can also improve token efficiency. “Experiments can verify that it is genuinely better.”

11. Rewriting the Entropy Problem in Data

• The true essence of high-quality data may amount to only tens of billions to hundreds of billions of tokens. Training repeatedly on the same data does not necessarily generalize well and can create overfitting, so K2 uses existing models to rewrite existing corpora, “so they can be absorbed better by the model and generalize better.” He disclosed only that they had found one approach that worked well in experiments. “There is a lot of room in this space, with many research opportunities.”
• 小珺 relayed a pointed objection: rewriting and expansion are useless—if the model can write it, the knowledge was already inside it. 杨 acknowledged that “this is a very good question”: “theoretically, it depends on whether you have a new input of entropy.” That imposes requirements on the rewriting method, and “our rewriting method today may not be the best.”

12. Agent Generalization: The Biggest Limitation of RL Today

• The episode’s central technical judgment is: “for an agentic model, perhaps the biggest challenge today is actually generalization.” Existing RL often trains against single-point tasks and metrics: “you train it on 3 benchmarks, and it improves on 3 benchmarks.” But “improving the metric does not mean your model’s generalization has improved.” The problem is more severe in agent training than in conversational models.
• Evaluation itself is a bottleneck: there are not many useful agent benchmarks, and a score “often does not reflect the capability.” A potential answer is “a more AI-native way of training AI.” If an AI can conduct high-quality alignment research, “in theory it may generalize better.” This is “the next hundreds of steps up the mountain.”
• RL generalizes better than SFT—on-policy samples plus negative gradients, “at least based on the evidence”—but only to a point. A model may score 99 on one type of math competition while improving on other math problems by 5 points, but it is hard to get to 99 there too: “what you plant is what you harvest.” Will this ever be escaped? “Problems are inevitable, but problems can be solved … you will only get better generalization; it may not have an end.”

13. Jumping the Gun and the “Proven Non-Consensus”

• 小珺’s framing was that K1.5 followed OpenAI while K2 was trying to jump ahead. 杨 did not fully accept the framing, but offered concrete evidence: “we certainly borrowed from many technical directions, but we also wanted our own innovations—at least in publicly available information, the first to use a non-Adam optimizer based on matrix orthogonalization and train a model at this scale.” Their agent-data approach was also relatively early.
• Last year he described a “probabilistic non-consensus.” This year’s version is improving agent generalization and the optimizer itself: “before we built it, no one should have been working on this, because everyone used Adam. So it became a non-consensus that we actually made work.” That is what delivered a better base model and more room to extract the ceiling in post-training.

14. The Result Is Predictable Before You Press the Training Button

• Project selection is not a flash of inspiration: “we are always training the next-generation model. The decision is which of the many new tools in the toolbox to take out and use.” The technical buildup began a year earlier. Technologies such as Meme Clip require a long validation cycle, from early small experiments to trillion-parameter models. “It is not that I decide today to train and start from zero; that is not something you solve in 2 or 3 months.” Research and training are handled by the same team: if you encounter a problem during actual training and did not understand it beforehand, you have no way to solve it.
• The only surprise during training was Muon’s max-logit explosion at scale. “You cannot predict it from small-scale experiments; the explosion simply does not occur there.” After revisiting the issue, they fixed it with a new clipping method. Otherwise there was no aha moment: “you already know during the process how the model is training.” What matters more is predictive capability: “predict everything that can be predicted, and your scaling success rate will be relatively high.” Most of the know-how is in the paper. “We are very open and still want to share with the community.”

15. Agent = Multiple Turns + Tools; With Generalization, Vertical Agents May Not Be Needed

• His definition of an agent has converged on 2 features: multiple turns and tools. Multiple turns are a form of test-time scaling. Tools are “how you connect the brain to the outside world”: a search engine connects the model to the entire internet, while code gives it automation across the digital world, where “basically all automation can be described in code.”
• The key implication points directly at the vertical-agent market: “if a general-purpose agent can generalize to long-tail tools—your company’s internal databases, personal documents, custom APIs, canceling a ticket, placing an order—then the various vertical agents everyone has discussed may not be all that necessary.” You simply give it different tools each time.

16. The Goal Is Not a Human, but a Universal Constructor

• 小珺 said agents are essentially trained on the sequence of “using tools over multiple turns” that people go through every day, simulating human behavior. 杨 corrected her on the spot: “simulating human behavior this way may not be accurate. Its main purpose is generality. It is a universal constructor that can complete almost any task. Its similarity to human behavior is simply a fortuitous result, not the purpose for which the system was designed.”
• His analogy remains unchanged: “an airplane is designed to serve as a means of transportation; it is not designed to fly like a bird.”

17. Medium-Difficulty Tasks Are Implicit Curriculum Learning; Coding Is the Hand of the Digital World

• Why train K2 on medium-difficulty tasks instead of designing the hardest possible tasks? “You cannot immediately ask it to prove a mathematical problem no one has ever proved. Sample efficiency would be extremely low, it would learn almost nothing, and your compute would be wasted.” With a good sampling strategy, “RL is essentially an implicit curriculum-learning mechanism”: start at the right difficulty and gradually raise it. But he added a concern: today’s tasks mostly come from human data or human design. The AI-native component remains limited, bringing the discussion back to generalization.
• Coding AGI is a subset of general AGI—“possibly a very important subset.” It is easy to verify and therefore easy to learn, and it represents automation in the digital world: creating a new tool is essentially writing code, while context engineering is also code underneath. But it is not the endpoint. Many non-programmers—lawyers, product managers, designers—now use cloud code because models have some degree of generalization and are not merely code writers. “We still want to build general-purpose models.”

18. Long Context Cannot Come at the Expense of Intelligence

• “Many tasks cannot be solved by 128k or 256k of context at all; you need millions or more, but your brain still has to work extremely well.” The tension is that high compression requires a large model, while the model must also remain strong. “These 2 things naturally conflict.” Some architectures improve long-context performance while hurting short-context performance, requiring an architectural balance. “These problems will gradually be solved; there are solutions.”
• He singled out one path: “if you use pure linear attention, it affects intelligence because the architecture introduces certain biases.” Another opportunity lies in RL training itself. When training complex multi-agent systems, “end-to-end reward alone may not be enough. How intermediate rewards are generated, and whether we can escape human design, will be a very interesting challenge.”

19. The Open-Source Reversal: “Because We Have Not Achieved Complete Leadership”

• 小珺 brought up last year’s prediction: open source would lag closed source, leaders would not open-source, and “only laggards would do it.” In retrospect, he was wrong. 杨 answered plainly: “because we are not yet absolutely and completely ahead.” The original judgment still partly holds: the community can contribute on the inference side and “give you more people serving the model for free,” but “to make the model itself better, it seems only the original developer can do that.”
• The new opportunity is downstream: agentic post-training based on K2. “If you are a startup building a legal agent, you can absolutely train a specialized agent on top of K2 using your customized tool set.” But this empowers downstream applications and “is hard to turn into an improvement to the base model.” They will continue opening things over time, “but not necessarily everything in full.” Is open source faith or strategy? “Objectively, it is both.” Ultimately, the aim is to use it to help technology “reach a better level more safely and quickly.” The same applies to Chinese companies open-sourcing collectively: mutual borrowing can accelerate reaching SOTA.
• His end-state view is that the market will consolidate: “from hundreds to dozens to a few. A few may be the final stable number; that now looks highly probable.”

20. The Model Is Still the Product; General-Purpose Models Bring System Complexity

• “We used to like saying the model is the product. Nothing has changed.” An agent product must combine the model, tools, and context, but “when you train the model, you basically have to build the entire stack … once the model is trained, your product is already done.” The interface layer is embellishment. Asked whether the company needed to evolve from “China’s OpenAI” into “China’s Anthropic,” he rejected the frame: China and the US are different, and today the company needs to take a global view. “Being China’s anything is not necessarily a valid concept. We will just keep climbing.”
• The complexity of the new super-system—the analogue to Google’s search system in the PC internet era—comes from generality itself. On one hand, it becomes simpler: everything goes into one model, with no need to maintain a stack of routing policies. On the other, putting coding agents, search agents, and other tasks into one model creates conflicts: the tools and data patterns differ. Multimodality is a typical case. “You do not want 20 experts in an MoE dedicated to multimodality, because what you get is a stupid multimodal model. We want an intelligent multimodal model.” The goal is to share one brain and activate the intelligence learned from text in multimodal modes. “The later you get in post-training and the deeper you go into RL, the more serious this problem becomes.”
• The new super-node is the system running in the data center—what Jason often calls an AI factory—combined with endpoints and new forms of interaction. Interaction changes as models evolve: coding began with Copilot, then Cursor, then cloud code. “When you have a new generation of model, you discover that the interaction can change.” Today’s cloud code interaction is not the final form either.

21. The Data Wall Is Real, but Models Are Improving Faster

• “The data wall is definitely an objective fact.” The way through is token efficiency plus scaling more compute across a range of RL tasks. One of the episode’s most important observations is: “somehow, we observe today that the rate at which models improve is not slowing down. I think it may even be accelerating.”
• Why has no AI product created a data flywheel? “Because compute-based scaling is so powerful.” RL is on-policy training with negative gradients, and its scaling efficiency is higher than pretraining. “The improvement from directly scaling FLOPs is enormous, so the gains from other sources look very small.” The other issue is that large-model learning is sensitive to noise, unlike recommendation systems; user-feedback noise has not been solved. New interactions can reduce signal noise, but “your interaction cannot surpass the model’s capabilities. It should be designed within the model’s capability range.”
• Does that mean there is no point building consumer products if user data cannot be used directly? He disagrees. User feedback cannot be fed straight into training, but having a meaningful user base tells you the distribution of demand, which can then be abstracted into evaluations to optimize the model. “If no one uses it, you do not even know what direction to optimize.” Another inflection point has arrived: OpenAI’s consumer users generate a large share of its commercial value, and if many users are professionals using agents, they themselves represent significant productivity value.

22. Next-Token Prediction’s Flaw Must Be Activated by RL

• 小珺 relayed 张翔宇’s view: as models scale, conversation, knowledge, and emotional intelligence improve, but reasoning—especially math—rises, plateaus, and then declines with further scaling. Large models tend to “skip steps and be dishonest” in math, a fundamental flaw of next-token prediction. 杨 largely agreed but offered the remedy: “that is why you need RL scaling. Without RL, it is hard to call the model very intelligent. But a larger base model still has a higher RL ceiling, because it has more knowledge. The point is essentially to activate a reasoning paradigm and unlock that knowledge.”
• What about world models—building worlds versus building people? He tied the answer back to the main line: “if you have a good world model to simulate these things, that is a way of using AI to train AI, and generalization will be much better. It may be one possible path to better generalization.”
• He recalled that when an investor first met him, he talked not about GPT-4 or GPT-5 but about what GPT-10 might look like; he no longer remembers exactly what he said. Today, “we have actually achieved many of the things we imagined before. You really can complete tasks lasting several hours.” The next step is to build more evaluations and break through generalization. “That still seems relatively clear.”

23. Business Model: Invest First, and the Certainty Is Actually Higher

• The 2 clear business models are APIs and first-party products. They “will both be explored,” but “the top priority today is still to keep making the model better.” The market is growing rapidly: “leading companies have ARR in the several-billion- to tens-of-billions-of-dollars range, and are growing quickly, doubling or tripling every one or two quarters.” What if Kimi cannot make money? “You still invest first. Whether you can make money depends on how good your model ultimately is. Focus on making the technology better, and the certainty around everything else is actually higher.”
• The boundary between model companies and agent-product companies—are companies like “Midas” customers or competitors? “It is still very early. In the short term, the relationship will be more collaborative than competitive.” Take Claude and Cursor: Cursor must keep adjusting, either by making up for model capability—“my technology R&D curve is still very steep”—or by offering tools or environments others cannot provide, such as controlling offline service access so that its ordering or transaction tool cannot be replicated.
• The end-state mechanism is straightforward: “generality works because its overall commercialization efficiency is higher.” Even content platforms may ultimately achieve higher commercialization efficiency by connecting to general-purpose agents than they do today. “But it may take a long time. During that window, specialized agents will also have room.” Buying traffic is unlikely to win the war during a period of rapid technological change; it is only an auxiliary tactic. “The biggest variable today is still intelligence. Still climbing.”

24. Manage Organizations with RL; Do Not Let SFT Kill Initiative

• His biggest organizational insight of the past year is that research, model training, and team management share the same structure. Knowledge does not come directly from experience. “Humans existed on Earth for a very long time, but 3 or 4 hundred years ago no one said the Earth was a sphere.” Knowledge comes from conjecture plus verification. Neural-network training works the same way: observing hundreds of thousands of internal metrics does not directly give you knowledge; you propose a conjecture about why something is happening and design an experiment to test it. Management is the same. “You need to manage with RL, not SFT. The team tells me this every day.”
• He described the operating details precisely. SFT directly tells people what to do; RL gives them goals and rewards. The right approach is mostly RL, plus some SFT as a prior: “with PTX loss, you do not want the model to fly too far.” But “you have to restrain your own hand. Too much SFT makes people lose their initiative and become unable to innovate.”
• The risk has an equally clear counterpart: “the biggest problem with managing a team through RL is that you can easily be hacked. Everyone may look like they are getting great results, but in reality they have not achieved what you ultimately want.” The solution is to establish more observation metrics and avoid overfitting as much as possible. Defining rewards also requires understanding how the details work; otherwise, “reward hacking” will emerge. His latest understanding of the CEO role is how to “balance RL, SFT, and reward hacking.”

25. Amplifier, Fear, and “Standing at the Beginning of Infinity”

• Why solve AI at all? He had asked Kimi, and adopted its answer: “it is an amplifier of human civilization. I think that makes a lot of sense.” It takes 20 or 30 years to reach the frontier of any advanced field; AI may learn it overnight, after which people can pursue new breakthroughs. AI could become a meta-science. Could it destroy human civilization? “You cannot say that risk does not exist,” but alignment and social mechanisms are both areas where work can be done. He also quoted Kimi: “if you give up on this, you are giving up on the ceiling of human civilization.” He added that the meaning of life is creation, experience, and love. AI may do a large share of the creation, but the latter 2 will remain centered on people. Adjusting social mechanisms “will not be done in 1 or 2 years; it may take 10 or 20.”
• His emotional volatility remains low. 小珺 would not let him off the hook: “I keep asking you about all kinds of emotions, and you always say you are fine.” He admitted that “remaining unmoved by gain or loss is hard,” but the goal is to “avoid emotional decisions.” Fear exists, but he focuses more on what can be done at the current step. Has the probability of success increased or decreased from a year ago? “It has increased. Every time you climb higher, the probability of success rises, because some people will stop climbing.” His biggest growth over the past year was recognizing that problems are inevitable, will always exist, and must continuously be solved. “That may also be the most interesting part.”
• The rapid-fire takeaways: the must-read book is The Beginning of Infinity. The papers that shaped AI include backpropagation, Transformer, and GPT-3, plus ResNet as a building block—“it may be the foundation of generalization”—and Adam; now Muon may belong on the list as well. His most important current bet is “a generalizing agent—use innovation, use L4 to do L3.” His closing self-deprecation: “My brain is already mush; I have talked through a year’s worth of thoughts. The limitations of carbon-based life.”