Tsinghua's 刘子鸣 on the Neo-Lab Frenzy, AI for AI and Interpretability

Summary

• China’s neo-lab capital frenzy has arrived. In 2026, a wave of research-trained professors and students—first-time founders with no product—raised funding at valuations of hundreds of millions of dollars. The 2 hottest areas for capital are world models and Auto Research (AI for AI). 刘子鸣 returned to China in March and participated in a financing round in late May; his reaction was, “It was just crazy.” The private market was “extremely formal”: “Everyone was probing each other for information, and investors were moving very quickly.” The fundraising playbook is “a seed round, then an add-on, then another add-on—you can keep adding rounds indefinitely.”
• The core call: the next Transformer should not be discovered by human guesswork. 刘子鸣, first author of KAN, is taking an unconventional AFAI route by training a “meta-model”: feed it a model, dataset and optimizer, and have it output a training curve. “A world model predicts the next state, a language model predicts the next token, and we predict the next curve.” Screening 100 architectures becomes 100 forward passes instead of 100 experiments. The conviction came from a 60-day self-experiment: he predicted training curves before running experiments every day, then “suddenly grokked it” after 40-50 days—even achieving accurate cross-modal predictions—and came to believe that a meta-model exists.
• LLMs are the anti-bitter lesson, as 谢赛宁 puts it. Their success comes from language, a modality pre-compressed by millions of years of human evolution, not from the model itself. The Transformer “won the hardware lottery”; it is powerful not because it is called Transformer, but because it enables efficient information propagation across context. The investment implication is that LLM convergence is now clear, while architecture design matters most in modalities that have not been compressed well—vision and world models, precisely where the current paradigm cannot reach and new labs still have room to win.
• His approach is orthogonal to, not in conflict with, 田渊栋’s RSI. Recursive is pursuing “more diligent AFAI”—a coding agent generates 100 ideas and finds one that works—while he is pursuing “smarter AFAI,” generating 3 ideas and getting one to work. The 2 approaches should eventually converge somewhere. His sense of urgency is unusually direct: “If their route ends up looking similar to mine, then it will definitely be because 田渊栋 built it.”
• Data is the real bottleneck. Research lacks the structured, GitHub-scale corpus available to coding. “People treated exposing their chain of thought as a form of shame: ‘I dreamed it; I didn’t derive it.’” His answer is the OPHIS methodology plus a proprietary chain-of-thought collection system: 20 people logging 10 entries a day produces 10,000 entries in 2 months, “more or less enough to start,” allowing him to distill a “子鸣 model” that can give students feedback on his behalf.
• The endgame is a vibe-training autopilot. Users provide only the objective and budget—“I only have 100 yuan”—and the system designs, trains, deploys and delivers the model end to end. “The goal isn’t to build another OpenAI ourselves, but to incubate OpenAIs for every vertical across every industry.” The roadmap calls for 6-12 months of research, which cannot be skipped, followed by productization around the 1-year mark. The China-U.S. funding gap is stark: American investors gave his postdoc adviser 3 years without commercial pressure, while “domestic investors wish you could be profitable the same day.”
• Mechanistic interpretability is being revalued. Anthropic’s neuron-level approach contains “some wishful thinking”—change the random seed and the story no longer holds—but under an AFAI framework, a field once thought to require hundreds of years could plausibly reach major milestones within 3-5 years. That is also why 唐杰’s claim that 智谱 will commit trillion-scale resources no longer sounds absurd. The signature AFAI milestone is a knockout architecture in a single vertical: something like ResNet, which beat the runner-up by 10 points—“use your fists to knock the enemy out.”

Deep dive

1. China’s neo labs in 2026: the private market enters “formal” mode

• The phenomenon is straightforward: research-trained professors and students have become the new darlings of the private market. They are generally young, graduates of elite universities, first-time founders—and can raise funding at valuations of hundreds of millions of dollars without a product. The 2 hottest destinations for capital are world models and Auto Research, or AI for AI.
• 刘子鸣’s timeline captures the market’s temperature. He returned to China in March, joined Tsinghua in late April and participated in a financing round in late May. “My visceral reaction to the whole process was—it was just crazy.” The private market was “extremely formal”: “Everyone was probing each other for information, and investors were moving very quickly.”
• The financing structure has changed as well: “You go from a seed round to a seed round plus, then another plus round, then another—you can keep adding rounds indefinitely.” Investors who missed the previous round simply get another chance in the next one. Investor categories remain crude: AI for science, AI for AI and auto research “all get lumped together.” Founders and investors “aren’t even talking about the same thing”; the reason they use cliché terms is that “investors need something they can understand.”

2. From physics to AI: the feast is over

• His decision to leave physics began with the high-energy community’s verdict that “the feast is over”—there are fewer important problems left for young researchers. He says he was “beaten up by both theoretical and experimental physics”: he could not compete with his idol 姚舜宇, who had published 2 PRLs as an undergraduate, while his magnetohydrodynamics experiments left him “with my hands and face covered” and made the lab more chaotic whenever he entered it.
• He encountered GAN in 2017, during his sophomore year: “This thing hit exactly where my excitement was. The math wasn’t that hard, and what it produced was incredibly fancy.” Still unwilling to abandon his physics background, he spent 5 years working on AI for physics, specifically high-energy phenomenology.
• ChatGPT’s arrival at the end of 2022 reversed the roles: “There aren’t that many things in physics as sexy and worth doing as ChatGPT.” He moved from AI for physics to physics of AI—using the methodology of physics to understand and design AI.

3. Max Tegmark: bird people recruit bird people

• When he applied to MIT’s physics PhD program in 2020, his research proposal said, “I’m going to work on AI for Physics.” “Only Max saw my materials and said, ‘I want to recruit this person.’” Max Tegmark had been promoting AI for physics since 2012 and was among the earliest advocates of mechanistic interpretability at the end of 2022—both examples of being early to a field before it was obvious.
• He uses Freeman Dyson’s distinction between bird people and frog people to describe the mentor-student pairing. Bird people “know a little about every field and can see structural similarities between fields, then transfer a method from one field to another.” Max is an ENTP; “an hour of talking with him can cover 5-10 topics.” He is also one of the originators of the parallel-universe theory.

4. How KAN was born: answering a warning from 1989

• KAN began as a side project Max did not approve of. He sent 刘子鸣 Tommaso Poggio’s 1989 paper, which argued that the Kolmogorov-Arnold theorem could not produce a strictly smooth representation in the worst case and therefore could not become a real algorithm. 刘子鸣’s physics-style response was: “We don’t care about the worst case. We care about the typical case… We humans can live in this universe, so this universe is not the worst possible universe.” He added: “All models are wrong but some are useful.”
• They built a prototype in 2 weeks, but Max still was not convinced: “You’ve still built a black box.” 刘子鸣 spent another week hand-crafting visualizations—“Back then I was still someone who didn’t believe in coding agents; now it might take an hour”—and the structural diagrams changed Max’s mind: “This is awesome. We can actually see what this network is thinking inside its head.”
• The deeper starting point was an obsession dating to 2021: could he build something that was both a neural network and a symbolic formula, analogous to quantum mechanics’ wave-particle duality? Only days after writing out the equation did he realize that it was the equation Kolmogorov and Arnold had written in 1957. The universal approximation theorem became the theoretical backing, finally putting to rest “the obsession I had been carrying all along.”

5. After KAN: scale model design itself

• KAN was an example, not a product. “After seeing it succeed, I wasn’t thinking about how to apply it to different scenarios. I was thinking about how to reproduce the fact that I can propose a new architecture.” “Designing models should be a science. It shouldn’t work the way it does now: the Transformer was pulled out of thin air, GAN was pulled out of thin air, ResNet was pulled out of thin air, and KAN was pulled out of thin air.”
• That led to his central doctrine: “We shouldn’t scale compute or data; we should scale AI model design.” The objective is to distill the thought processes of himself and the world’s top AI researchers into models, “even producing things better than the very best human scientists.”

6. The Transformer won the hardware lottery; LLMs are the anti-bitter lesson

• Asked whether new architectures can adapt to current hardware, he turns the objection around: “That only reinforces the point that the Transformer isn’t that impressive algorithmically—it won the hardware lottery.” He agrees with 谢赛宁: “Large language models are the anti-bitter lesson. The reason they succeeded is language, not the model. Language is a gift of millions of years of human evolution. Whatever model you use, as long as it can absorb that gift, it will be a good model. In a parallel universe, it might have been another model.”
• The broader demystification is simple: “The Transformer is good not because it is called Transformer, but because it enables efficient information propagation along the context dimension. ResNet is good not because it is called ResNet, but because it enables efficient gradient propagation along the depth dimension.”
• Model design becomes more important after the LLM era. Vision and other modalities have not been compressed as effectively, creating demand for models that can learn abstract representations. Evolution made humans “particularly good at identifying the most important signal in a huge amount of information: there’s a piece of fruit over there, and if I pick it I won’t go hungry today; there’s a tiger over there, and it must not see me.” That is precisely the abstraction capability current AI still lacks.

7. A single through line: grafting neural and symbolic worlds

• His retrospective summary of his research is: “How do you graft these 2 worlds—neural and symbolic—together?” AI for science uses neural methods to supplement science, which is too symbolic; science of AI goes the other way: “I’m not satisfied with AI being a black box. I want to use symbolic things to describe, understand, improve and design my neural things.”
• Citing Elias Askever’s “Intelligence is compression,” he says symbols represent compression in its most extreme form. But his own view has evolved: “If you had asked me last year, I would have said the endpoint must be symbolic. The more I do, the more I realize that maybe all you need is a predictor. The ultimate purpose of physics is prediction, and that predictor could simply be a neural network—as long as it predicts reasonably well in certain situations.” Symbolism “is still, so far, a personal obsession.”
• That is why his AFAI differs from the mainstream: “Most people think about using AI directly to improve AI. I say stop—there’s another step in the middle. The full name of my AI for AI should be AI for physics of AI for AI.” Physics of AI is needed to design AI, but doing that research manually—“old-school research”—is too slow, so AI must automate it.

8. Astronomy’s 3 stages: AI may not have reached even the Tycho era

• His favorite analogy runs from Tycho Brahe’s star records to Kepler’s fitting of elliptical orbits and other empirical laws, then to Newton’s compression of those laws into the theory of gravity. AI is broadly at “stage 1.5”: there are many experiments but few empirical laws. Scaling laws are among the few, occupying a role similar to Kepler’s laws. “We are nowhere near AI’s Newton.”
• The more pessimistic version is stage 0.5. Tycho at least pointed his telescope at the sky; “we converged on the Transformer too early,” stared at one patch of sky and decided “this patch of sky is incredibly beautiful,” then lost the motivation to explore other architectures. Our data works for only a small number of models; we are nowhere near seeing the entire sky.

9. Kong Yiji’s gown: MIT conservatism and Stanford’s mind-opening

• MIT’s 2024 bubble was an “everyone thought the agents being built in the Bay Area were scams” environment. Even after receiving his Tsinghua offer in May 2025, his mindset was still: “I want to do more influential work in academia and not get mixed up with industry.” “Looking back, that was naive. I still hadn’t taken off Kong Yiji’s gown.”
• The MIT faith is worth preserving: spend 5 years writing down the physics of AI, then discover that the vast resources being deployed today are wasted because first principles can derive the optimal architecture. He now mocks his own optimism: “A few smart people could write down the physics of AI, then save the world—or kill the world, I don’t know.”
• Six months at Stanford “felt like my mind had been opened.” His friend 高文浩, now working on AI for chemistry, decided “子鸣, I’m going to start a company” after just 2 weeks in the Bay Area and began teaching him entrepreneurship instead of meeting weekly to discuss academia. The mechanism was disarmingly honest: “You especially believe in peers whose style resembles yours. When you see them do something you had considered dangerous or impossible, you start wondering: am I not aggressive enough? Am I too deep inside my comfort zone?”

10. The mentor’s transition: from pure scientist to entrepreneur

• He watched his postdoc adviser Andreas Tolias take a new lab from secret planning to public launch. The route is neuro for AI: use the representations and structures in monkey and mouse brains to constrain artificial neural networks. Tolias is “one of the world’s leading people in neuroscience foundation models.”
• The reason for the transition was “very simple.” Experiments involving monkeys require expensive equipment, while the work inside a company is not much different from work at a university. The investor promise was: “For 3 years, you don’t need to think about commercialization. Just focus on R&D.” The data requires people to raise monkeys, so the timeline cannot be compressed.
• 刘子鸣 declined the invitation to join: “It felt a little like living under someone else’s roof. The idea for the technology has to be native to me.” He recognizes this as “the academic’s need for control,” and allows that he may later realize “this is also a kind of gown.”

11. Why neo labs emerged: avoid being a foot soldier and target pain points the paradigm cannot reach

• Top talent has 3 basic options: stay in academia, join a Big Tech model shop and “be a foot soldier,” or start independently. The proliferation of new labs reflects a period in which the paradigm is gradually converging and talent increasingly does not want to serve as a foot soldier at a Big Tech company.
• The second condition is a clear pain point that the current paradigm cannot address, with early work requiring substantial research rather than an immediate contest over resources and engineering. Auto Research is one example: unlike coding, research has no high-quality structured dataset like GitHub. “We still talk about research intuition and research taste. Using those words assumes research hasn’t been structured. Once something can be structured, standardized and scaled, the mystique disappears.” World models are another: they require novel architecture design. The host summarizes the underlying opportunity as “training a brain for robots.” Those are currently the 2 hottest areas for capital.

12. OPHIS: invent a language for research first

• His 3-step framework is: structure research, collect structured data, then train a research language model. The methodology is OPHIS: Observation, or what phenomenon was observed; Problem, or what problem it suggests in the system; Hypothesis; Intervention, or the next action; and Speedup, or whether it actually improved training. “Every research paper should be structured this way.” Real research jumps back and forth between nodes, “but that’s fine. Everything should at least be tagged as O, P, H, I or S.”
• The vision is straightforward: “Once we have a language for research, we can train a research language model just as we train a large language model. Research can then be automated and scaled.” A blog post covering the methodology and initial results is due for release.

13. Why return to China: bring in the prodigies

• Why not start a company in the United States? “Auto Research needs high talent density. I worked with 20-30 undergraduates at MIT, and the average level was far below that of undergraduates at Tsinghua and Peking.” The project needs prodigies to join; “I need a lot of prodigies.”
• He therefore joined Tsinghua and can “also poach some students from Peking University,” allowing him to assemble not just the smartest students in China but the smartest students in the world. He currently works with more than 10 “extremely, extremely top-tier” students at the Institute for Advanced Study in Intelligence.

14. 田渊栋’s RSI: more diligent AFAI versus smarter AFAI

• His view of RSI, based on Recursive’s latest blog, is that the team is focused on improving coding agents: “more diligent AFAI—extremely strong execution, iterating nonstop.” His own goal is “smarter AFAI”: coding agents previously generated 100 ideas and got one to work; he wants to generate 3 ideas and get one to work. The approaches are orthogonal and “will eventually be integrated somewhere.”
• His sense of competition is unusually explicit: “田渊栋 himself is quite supportive of science of AI. If their route ends up looking similar to mine, then it will definitely be because 田渊栋 built it.” The 2 have previously discussed grokking, and their “taste is actually very similar.”
• His technical criticism of the RSI route centers on memory. He suspects the system emphasizes maintaining a useful record—if 5 learning-rate adjustments failed yesterday, remember not to try it again today. “But if you don’t understand why it failed, the previous experience may not apply. Maybe it failed yesterday for reason A, but today the reason is B, in which case adjusting the learning rate is actually correct.” The dividing line is experience-driven versus mechanism-driven research.

15. The 60-day self-experiment: train yourself like a model

• His conviction in a meta-model came from an intensive retreat early this year. Every day, he randomly selected a dataset and a model, then asked himself before running the experiment: what will this training curve look like? He did this for roughly 60 days, updating his blog frequently—“sometimes twice a day.” At first, he could not even predict the direction; after 40-50 days, “I suddenly grokked it, suddenly had the realization.” He could even predict the result of a vision experiment after doing a language experiment a few days earlier.
• How did it happen? His honest answer is worth keeping: “I don’t know. Maybe it was a form of visual reasoning; maybe there was reasoning in latent space. I can’t explain it.” The ascetic practice convinced him that a meta-model exists: “If you can foretell what will happen next, you must have captured some kind of world model”—a form of prophetic ability.

16. The meta-model: predict the next curve

• The definition fits in one line: “A world model predicts the next state, a language model predicts the next token, and we predict the next curve.” The input is a model, plus a dataset and optimizer; the output is a training curve. It is “a world model about models.”
• The motivation is self-externalization: “I can’t write down my know-how. That knowledge exists in my brain in the form of some kind of electrical waves. I want to copy myself.” A human can run only 1 or 2 experiments a day, while a model can see thousands of configurations in a batch. “This meta-model doesn’t need to sleep, eat or drink, or have emotions like I do.” He grokked the pattern after 40-50 trials; the meta-model may need 400-500, but spending somewhat more compute is not a problem.
• The use case goes directly to AFAI’s bottleneck. If there are 100 architectures and no way to know which to try, “running the meta-model forward 100 times to make 100 predictions is extremely cheap.” It can rank the most promising ideas before experiments begin and help AFAI’s iterative system iterate faster. The real experiments still have to be run.

17. The training recipe: the data is the model, and the goal is to escape local minima

• The recipe is unintuitive but central: “Our training data is the model.” Train a batch of models, collect model-and-training-curve pairs, and train the meta-model with supervised learning. It requires meaningful compute, but unlike a large model it does not need a massive cluster or heavy inter-GPU communication because it is running large numbers of small models.
• The small models are not scaled-down Transformers; diversity is the point. “The meta-model needs to have seen many architectures that look extremely strange. Maybe one of those architectures is what helps us escape the local minimum where the Transformer is stuck.” The ultimate goal is clear: “The next Transformer should not be discovered by a human pulling an idea out of thin air; the meta-model should predict it for us.”
• The intermediate plan follows Rich Sutton’s bitter-lesson logic: when data is scarce in the short term, inject structure; when data is abundant in the long term, structure becomes unnecessary. Research has little structured data, so the pragmatic near-term path is to structure it, collect it and train a research model. The long-term North Star is the end-to-end meta-model. He jokes that this runs counter to industry fashion: “It’s a bit like doing world models in the short term and VLA in the long term.” The route is “pushing forward while exploring”; it changes weekly, sometimes daily. “If this interview had happened last week, I wouldn’t have talked about this section.”

18. Chain of thought is the scarcest data—and showing it was once shameful

• “Maybe only 1% of AI papers are structured well.” One example is 朱泽园’s Physics of LMs, where the observation, decision and experimental reasoning chain are clear; 99% of papers are not. The cultural root is sharper: “For a long time, people treated exposing their chain of thought as a form of shame. You wanted to seem inscrutable, just write down the formula and say, ‘I dreamed it; I didn’t derive it.’ That made people think you were brilliant.”
• The answer is a self-built collection system spanning devices and software. The process is mandatory: before starting the next experiment, a researcher must write up what the previous experiment did and what conclusion it produced. Only then can they move on. The collector is also the environment—a “research version of Minecraft”—which can later support next-action prediction. When a student gets stuck, the “子鸣 model” can be called to provide feedback at his level without him being present.
• His estimate of the required data is explicitly a guess because collection has not begun: 20 student employees multiplied by 10 entries a day yields 200 entries daily; “two months gets you 10,000 entries,” “more or less enough to start training a model to distill ourselves.”

19. China’s AFAI map: drawing a line between the project and paper agents

• The landscape includes 孙天祥, an early entrant whose belief is that “language is basically the entire world.” His FARS system—pronounced like “FARS”—is a paper agent that automatically produces papers; he recently joined Baidu’s foundation-model team. 陈永超 of Tsinghua’s Institute for AI is also working on AFAI, though his current route looks more like a paper agent. RSI is as described above; “it’s still unclear what their science of AI actually looks like.”
• His distinction from paper agents comes with self-deprecation: “I was a paper machine during my PhD. I know that only a few points in those papers have real value.” “The first principle of research is expanding your own knowledge graph. A paper is simply the by-product dumped out when the knowledge graph is large enough that your brain can no longer hold it.” The objective is to accumulate an insight database for the physics of AI and use it to design better architectures systematically and scientifically.

20. A 3-part physics of AI and the phase diagram: no one says where their trick works

• He breaks physics of AI into 3 finer dimensions. Spatially, it means looking beyond end-to-end behavior and dissecting internal neurons and representations. Temporally, it means studying evolution rather than only the final state: “Nothing in biology makes sense except in light of evolution.” Across parallel universes, it means controlled experiments and phase-diagram sweeps, extending 朱泽园’s methodology while paying homage to Max’s parallel-universe theory—“for a bacterium in one petri dish, another petri dish is a parallel universe.”
• The phase diagram addresses the industry’s current disorder. The no-free-lunch theorem says no trick is better on every task. “The reason things are so confused now is that everyone says their trick works, but no one says in which phase their trick works.” Mapping the conditions under which each trick and architecture applies is the scientific approach, rather than building something and immediately selling it.

21. Anthropic’s mechanistic interpretability: demystify first, then revalue

• The criticism is informed by direct experience: his group did extensive work using the same approach at the end of 2022, and a student later joined Anthropic. The Claude research on how it does addition extended a paper from the group. “SAE and transcoder methods try to lift superposed representations into a high-dimensional sparse space and extract atomic concepts. I personally think there is some wishful thinking in that. Change the random seed and the whole story may be wrong.” Neuron-level explanations are highly specific, but not necessarily robust.
• The AFAI framework changes the valuation. Mechanistic interpretability is “the biology of AI.” Under the old manual approach it might take hundreds of years, “because we had no way to automate it”; once automated, progress could accelerate sharply. “Half a year ago, when I saw 唐杰 say that 智谱 would invest trillion-level resources in mechanistic interpretability, I would have wondered what they were doing. But under the AFAI framework, those North Stars that looked extremely remote could plausibly be reached within 3-5 years.”

22. The milestone: knock out the enemy, but AGI is still missing a piece

• The signature AFAI milestone is a new architecture in a single vertical—“it doesn’t need to be general”—that lifts performance “not by 1 percentage point, but by 10.” The benchmark is 何恺明: ResNet beat the runner-up on ImageNet by 10 percentage points. “You either use your fists to knock the enemy out, or, if you can’t knock them out, you have to persuade them with words.”
• The boundary is clear: AFAI is necessary but not sufficient. True AGI still lacks the ability to abstract data. “Humans can continually learn because humans can abstract. Current AI still can’t. If it appears to, it is mostly because the data itself has already been abstracted, as in natural language.” In response to the host’s challenge that AFAI amounts to “standing on your right foot with your left foot and flying into the sky,” he narrows the definition to proposing good architectures in verticals—a goal that does not require solving continual learning.

23. The product: vibe training and model parity

• The commercial answer is a training autopilot, with a copilot as the near-term version. The user states the objective: “I only have 100 yuan; within that budget, deliver the best model you can.” The system then designs, trains, deploys and delivers the model end to end. “We have vibe coding now; the future will be vibe training.”
• The ecosystem is explicitly a platform play: “We’re not building another OpenAI ourselves. We’re incubating OpenAIs for every vertical across every industry.” Just as LLMs unified NLP’s many tasks, the end-state system would design each vertical’s own “Transformer.” Asked what use there is in everyone training models, he gives an unusually unvarnished answer: “To be honest, I haven’t thought it through either. Once the technology becomes cheap enough, certain needs will emerge naturally.” He invokes the question of the household computer—“Why would a household want to keep a giant computer at home?”—and the example of training a companion robot dog.
• Democratization depends on cost. Once the physics of AI is understood, training costs should fall: “We don’t have the optimal recipe today. There may well be a great deal of room for optimization.”

24. The essence of a new lab: lab first, company after the grokking moment

• His cleanest definition of a new lab is that “it is neither a lab nor a company.” For 6-12 months it operates like a research institute; once it groks the problem and has the realization, it should operate like a company. This is the divide between the pre-OpenAI and post-OpenAI eras. The research-driven phase “cannot be skipped”: founders cannot rush into a direction that is not good enough just to meet a short-term deliverable.
• The China-U.S. investor temperature gap is stark. American investors gave Tolias 3 years without requiring commercialization; the host says domestic investors wish founders could be profitable the same day, and 刘子鸣 replies, “The same day is still impossible.” His roadmap emphasizes R&D and demos for the first 6 months, then “will definitely begin productization and commercialization” at the 1-year mark, funded by continued fundraising. The shift is notable: a year ago he thought 3-5 years of R&D were necessary; confirming that the meta-model is feasible has accelerated the timetable.
• After several months of starting a company, his condition is “basically working whenever I’m not sleeping.” He is back-solving the route from goals several years out, which feels “a little out of my distribution.” Can he be both professor and entrepreneur? “I want both. If I have to choose, we’ll see.”

25. Research taste as a fig leaf + rapid fire

• Asked how to train an epiphany into a model, he first dismantles the myth: “I also used to believe in research taste and research intuition. I’ve demystified them now. Research taste is actually a fig leaf: we’re smart enough to come up with an idea, but not smart enough to explain in words how we came up with it.” “A good warrior wins without a celebrated victory—the best people generate ideas as effortlessly as they breathe.” He has trained himself to keep asking how he generated a good idea and concludes: “If you interrogate yourself deeply enough, you can always articulate something.” The alternative is to measure brainwaves directly, but that requires advances in neuroscience and the resolution of ethical questions.
• The rapid-fire answers are revealing. The book that changed his life is Sapiens: “Our society is really driven by stories. Technology matters, but the narrative behind it matters enormously too.” The paper that influenced him most was GAN, which “opened the door to the AI world.” His key current qualification is: “After AI for coding, the next thing will be AI for research—but it has to be done through the physics-of-AI approach.”