178: Talking RSI with 田渊栋: How Will Model Self-Evolution Arrive?

Summary

• 田渊栋’s startup was triggered by a hands-on experiment last August: after co-writing a Grokking paper with GPT-5, he realized the model was already capable enough that he might not need to hire interns. Ideas, research directions, and theorem proofs could be worked out through conversation, with the coding handed to AI; the research validation loop went from days or weeks in the advisor-student model to minutes—“Rather than be replaced, I might as well start a company.” Recursive Super Intelligence emerged from 4 months of stealth in early June with a $650M Series A at a $4.65B valuation.
• The central debate is over Silicon Valley’s popular assumption that the first company to reach RSI will pull irreversibly ahead during a window of opportunity. 田渊栋 thinks the logic is “still broadly correct,” but the decisive variable is the shape of the curve: intelligence gains will follow an S-curve with plateaus, not a smooth progression. If scaling laws hold perfectly and the curve is smooth, startups have no chance; but scaling requires 10x compute and 10x data for linear gains, and “you can’t have all the electricity on Earth supplied to one person for self-evolution.” The real bottlenecks are already visible, leaving room for breakthrough-driven paths.
• RSI is “much bigger than a coding agent”: coding agents are merely executors, while open source is flattening that capability. He feels the gap between GLM-5.2 and Opus 4.6/4.7 is narrowing—“4.6, 4.7, and 4.8 aren’t that different”—while RSI’s real deficit is research taste: the abstraction and creativity to “see precise, detailed connections from a small number of examples.” Large labs may prioritize coding businesses as they push toward IPOs, creating an opening for new labs; “everyone I talked to has left.”
• 田渊栋’s read on Anthropic’s When AI Builds Itself is deliberately cooler: Claude’s 800-hour effort narrowed the weak-to-strong gap by 97%, but it is still productivity enhancement rather than recursive self-improvement. Humans completed 23% of the work in a week; the method remains conventional and score-driven, making it only the first rung of self-evolution. OpenAI’s timeline is an AI research intern by September this year and a genuinely automated AI researcher by March 2028.
• Recursive Super Intelligence’s first 3 results all reached SOTA using the same general-purpose system. NanoChat’s 5-minute training run beat the community on BPB, the speed run moved from the community’s 79-second SOTA after 2 years of work to 77 seconds, and operator optimization beat Double AI—a specialist GPU team led by former Mobile AI CEO Sasha—by 10% on Nvidia’s new SOL Exact Bench. The team had no GPU specialist; the starting point was “1 person plus AI reaching SOTA.”
• Beyond scaling, he is betting on interpretability and considers himself a minority voice. While “most people may already feel hopeless,” he believes there must be a good principle that makes models work; AI will eventually become a science, as alchemy became chemistry, through a Tycho-Kepler-Newton progression. The next Newton “could be AI, could be humans plus AI.” His organizational view is equally blunt: “this wave of AI is anti-big-tech,” teams start breaking down after 150 people, and Llama 4 is the example.

Deep dive

1. Rather Than Be Replaced, Start a Company: A Paper Co-Written with GPT-5 Changed the Calculation

• The turning point came last August, when 田渊栋 used GPT-5 to complete a Grokking paper. “After I finished it, I realized the model’s capabilities were basically strong enough that I felt I didn’t need to hire interns.” Ideas, research approaches, and theorem proofs could be developed through dialogue, while the AI handled the code.
• That led to the line that runs through the entire episode: “Maybe I’ll be replaced in 5 years, and that point may already be close… Rather than be replaced myself, I might as well start a company.”
• Recursive Super Intelligence spent 4 months in stealth before announcing a $650M Series A at a $4.65B valuation in early June. Its initials, RSI, match “recursive self-improvement,” and “in terms of meaning, they are very close.”

2. Why RSI Is Suddenly Hot: Stronger Coding Agents and Models Crossing a Threshold

• 田渊栋 attributes the shift to 2 factors. First, coding agents have become much stronger. Second, models have made breakthroughs in understanding AI models themselves, making it possible to use a model as a researcher—to identify weaknesses in a model and optimize it. “That loop will emerge”; recursion is the loop moving forward.
• The prerequisite was repeated several times: “The condition is that the model itself has to be strong enough.” When he started the company last year, “most people hadn’t even thought about this concept”; it only began entering the consensus in April and May this year.
• The efficiency gain is measured in orders of magnitude. Traditional research runs on an advisor setting the direction, a student executing, and feedback arriving over days; a paper can take weeks or months. Now a single validation loop can be compressed to minutes: “Have a cup of tea, finish the meeting, and the result is already there.”

3. The Position Has Not Changed: Frontier Models Still Cannot Replace Top Researchers, but Recursion Comes First

• Compared with his interview last May, he remains clear: “I still think large models can’t replace top researchers.” The change is that models “can genuinely be used” to do research work.
• He qualifies the OpenAI-style “automated AI researcher” narrative: full automation “may not happen very soon.” Humans will remain in the loop, but at a higher level. Low-level work will no longer require people, while “people may need to use more brainpower on higher-level work.”
• His conclusion remains cautious: “I think recursion will happen first. As for full automation, I don’t think we can see it yet.”

4. Who Sees the Signal First: The Balance May Shift Back Toward Experienced, Hands-On Researchers

• He sees a gap between senior figures who “don’t actually take a model and use it on a very concrete project” and recent graduates who treat models only as execution tools and “won’t tackle harder problems.” Without hands-on use, it is difficult to feel the direction of change.
• The past 3 years belonged to young people: “Don’t overthink it… skill up and just do it.” Under RSI, however, execution moves to AI while humans handle judgment, direction, and research taste, so “the balance may shift back toward more experienced researchers.” When someone who has been a manager or director for 10 years suddenly starts coding, “you feel like this guy’s youth has come back.” He acknowledges that this is just his view and remains debatable.

5. The Divide from AutoML 10 Years Ago: Who Defines the Solution Space

• 田渊栋 lived through the previous wave of architecture search and AI-guided optimization. Its central problem was that “we didn’t have a model capable of modeling high-level human knowledge.” Humans specified every detail of the search space; AI was relatively dumb and searched within the boundaries people provided.
• The qualitative shift this time is that large language models “understand AI itself, or the architecture itself.” They can partly replace the human task of defining the problem and “find the structure of the solution space themselves,” making the space both much larger and more intelligent.

6. Cooling Anthropic’s When AI Builds Itself: This Is Productivity Enhancement, Not Yet Self-Evolution

• Anthropic’s June paper highlighted a Claude agent that independently completed an open-ended AI safety project end to end, working for 800 cumulative hours and narrowing the weak-to-strong performance gap by 97%; humans completed 23% in a week. 田渊栋’s assessment was: “Probably not yet… It’s basically still productivity enhancement.” The algorithms remain “fairly conventional,” combining many standard ideas to lift the score.
• He does not dismiss scores as a metric: “Other than scores, there’s currently no other way to measure how capable a model is.” Break the task into subtasks and see whether AI improves on them; “this is the first layer of self-evolution.”

7. Code Can Self-Evolve Along a Path; AI Research Cannot—That Is the Real Divide from the 2 Major Labs

• At a high level, “they’re all more or less the same; I don’t think there’s a particularly large difference.” The difference lies in the difficulty. Anthropic demonstrated self-evolution on existing projects—writing a Linux system or a CI compiler from scratch. “These things all have a path… They’re just extremely tedious and exhausting.”
• His focus is on the harder class: “For many difficult AI research problems, the truth is that humans don’t know how to solve them.” Problems that require a breakthrough discovery remain very difficult for AI.

8. The Field Is Taking Shape: “Everyone I Talked to Has Left”

• New entrants are appearing in clusters. Mirandil (likely) announced a $200M seed round on June 25, with a team from DeepMind and Anthropic; Sakala AI (likely) established an RSI team in Tokyo in June. Mirandil’s founder contacted him during the preparation phase. People have also left Apple and Amazon; David, who led Amazon’s San Francisco Lab, has also departed. “I don’t know what they’re preparing to do now.”
• Why leave rather than build inside a major lab? OpenAI and Anthropic “are both trying to push toward an IPO.” Their immediate priority is to make money from their coding-agent businesses, while “RSI is, after all, a fairly distant direction.” When he spoke with large companies last autumn, RSI was still niche; they wanted him to “provide support” for an existing agenda.
• How serious are the 2 major labs? “Right now, everyone may still be talking about a concept.” As for how to turn it into reality or define a concrete metric, “not much is being said.”

9. RSI Is Much Bigger Than a Coding Agent—And Open Source Is Flattening Coding

• He rejects the inference that coding plus agentic capabilities are the big labs’ main line and RSI merely its natural extension: “I think RSI is much bigger than a coding agent.” Its problems, difficulty, and possible paths are much larger than those of today’s coding agents; this is not just about market size. “A coding agent is only an executor.”
• 曼祺 asks whether having coding as the initial foothold is insufficient to build RSI. “Yes.” The existence of open-source models means “this has become a capability everyone has.”
• His firsthand model read is worth recording. GLM-5.2 and Kimi 2.7 still “definitely have a gap” with closed-source models, but “should be getting close.” GLM-5.2 and Opus 4.6 are “about the same”; 4.6, 4.7, and 4.8 show “not much difference, only a version-number difference.” The only meaningful distinction with 4.8 may be a longer context window, so it no longer forgets after compacting midway. GLM-5.2 is strong at execution, “but sometimes it may get stuck in an infinite loop, or show all kinds of glitches.”

10. The Missing Capability: Taste for Extracting Precise Connections from Sparse Samples—Grokking as the Test

• Beyond coding, RSI most needs “the taste of a human researcher” and “the ability to abstract a problem”—something models generally lack today. It is larger than reasoning and closer to creativity: extracting concepts from a confusing surface, summarizing them, and applying them to a new problem.
• Models are powerful because their data is effectively endless. Scientists confronting frontier questions often have very few samples; this is where humans retain a significant advantage over large models.
• His concrete example is understanding emergence itself. Starting in 2013 and 2014, researchers tried the renormalization group, spin glass, neural tangent kernel, Bayesian methods, and Gaussian processes. “All these angles have been tried, but none has produced a particularly satisfying result.” Finding a new explanation is difficult to achieve by simply transferring existing logic onto AI.

11. The Advantage of RSI: It Is a Staircase, Not a Self-Driving Car’s Zero-or-One

• 曼祺 asks whether AI can help AI evolve faster without solving the deepest theoretical problems. 田渊栋 says yes, and sees this as RSI’s structural advantage: “It isn’t like a self-driving car, where it’s either 0 or 100. It has many steps.” Optimizing algorithms and improving performance at an early stage already creates practical value; a second step that uncovers something deeper creates more.
• He does not avoid the difficulty of the highest step: “We want AI to become a brain like Einstein or Newton, acquiring deep knowledge from very few samples.” Even if that goal is not reached, “the applications of the thing itself are numerous.”
• In addition to taste, RSI needs agentic behavior—using tools, retrieving knowledge, and completing tasks—and information-gathering ability. But “many open-source models or existing solutions are already there,” and the deep-research direction already has substantial work behind it.

12. Breaking Down the “Winner Gets Stronger” Assumption: The Decisive Variable Is the Curve Shape

• The Silicon Valley thesis is that whoever reaches RSI first will pull farther and farther ahead, creating a window of opportunity. He first concedes that “the logic is probably still broadly correct,” then identifies the weak point: progress may come through breakthroughs rather than gradual improvement. Starting from a given system, “there will be an S-curve: it rises quickly at first, then hits a plateau until the next breakthrough”—S-curve, plateau, then another S-curve.
• The implication goes straight to startup survival. If scaling laws are “100% correct,” there would be no plateaus, and “no startup could possibly catch OpenAI and Anthropic,” because the major labs would always be further ahead and moving faster. He is not convinced. Scaling laws “may be correct, but the x-axis may be exponential”: 10x compute and 10x data for linear growth, with that 10x resource requirement already hitting bottlenecks.
• The hardest practical constraint is blunt: “You can’t have all the electricity on Earth supplied to one person for self-evolution.” There will inevitably be competition and limits. He acknowledges that the plateau thesis “is probably not very mainstream yet.”

13. The Space of Intelligence Is Only Beginning to Dawn: Cultivation Societies and Housing-Price Systems

• The second variable is the size of the intelligence space. “It’s very large, and we’re just beginning… My book is called In the Dawn, and we’ve only just seen a little starlight; the sun hasn’t even risen.” His humility is self-deprecating: humans are “pitiful monkeys evolved from creatures on Earth,” and what we have always seen has been relatively shallow.
• His end-state vision is a private, novel-like one: entering a cultivation society where “words become reality”—want a large house tonight, and a large house appears immediately. But the novel iterated 2 times. After the first version failed, the virtual world added a housing-price system: “Once people lose their motivation, how does human society develop? That’s a very large question.”
• The company’s roadmap is more practical. AI for AI is the first major phase. If the system can generate new architectures, theories, and optimization algorithms, “can we use this system to do other things?” That would mean moving into other scientific fields.

14. First Steps: 3 Benchmarks Reach SOTA on the Same General-Purpose System

• The first results, released June 11 under First Steps Toward Automated AI Research, cover 3 areas: optimizing NanoChat’s 5-minute training run until its BPB beat the community figure; speeding up the NanoChat speed run; and finding a new operator that performed “considerably better than the current best SOTA.” All 3 were open-sourced.
• The key detail is that all 3 tests were run by “the same system—it’s general-purpose.” One selection criterion was fast feedback: “You can get feedback quickly, and the system can get up and running quickly.” 曼祺’s observation, which he confirmed, is that the 3 benchmarks cover performance under constrained compute, efficiency, and operators applied to infrastructure—several core layers of a model.
• Why call it a “system” rather than an agent? “The details are still fairly confidential… This way, you can preserve all kinds of possibilities.” Commercialization is “under consideration,” but generality is non-negotiable: the system will not become highly specialized around 1 or 2 specific verticals.

15. Non-Experts Beat a Specialist Team: Operator Optimization Wins Double AI by 10%

• To push back on the claim that the numerical improvements look small, he points to operator optimization. The runner-up, Double AI, is an Israeli team whose CEO Sasha was formerly Mobile AI CEO—a Tesla supplier and semiconductor specialist. Its members “are all engineers who specialize in GPUs,” yet “our result was 10% better than theirs.”
• The contrast is the point: “There aren’t really any GPU experts in our group,” and he himself “had never actually done operator optimization.” The project began with “me alone plus an AI reaching SOTA,” using the company’s own system rather than an external model. His joke carries an edge: “I should be the person who understands operator optimization best.”
• The same applies to the NanoChat speed run. Moving from 79 seconds to 77 seconds looks like a 2-second gain, “but you have to know that 79 seconds was the SOTA the entire community had spent 2 years reaching”—the result of a concerted effort by the technical community to improve efficiency.

16. The Gap from GPT-2 Scale to Frontier Models—and the Honest Admission That There Is No Roadmap Yet

• NanoChat and NanoGPT train models at roughly GPT-2 scale, far below the parameter counts of mainstream large models. Moving to larger scale “will still have some gap,” and the team is continuing to work on it. Related results may come later, but “I can’t disclose the specifics.” The next step is broader optimization: both pretraining and post-training.
• Does the company have an OpenAI-style roadmap toward AGI? “Not at the moment.” His defense is delivered with a joke: OpenAI did not publish a roadmap until it had been established for 10 years. “The timeline will be shorter now,” but “it should not be so fast that it takes only 6 months.”

17. What RSI Still Lacks: No Consensus on Data, Compute Remains a Bottleneck, and Interpretability Is the Bet Beyond Scaling

• On data, “there is still no conclusion”; teams are exploring internally. Some methods are hard and some easy, but “maybe the hard ones work well and the easy ones don’t.” The $650M bankroll “is enough to do some things,” but reaching the level of the “big 3” (likely) remains some distance away.
• The more fundamental view is that RSI requires innovation beyond scaling laws. “If it’s entirely a scaling-law contest, small companies can’t beat big tech.” The only broad direction he is willing to disclose is interpretability: it can uncover insights, accelerate progress, and make models safer. “But I can’t say what exactly.”

18. The Safety View: Under Dog-Training Evolutionary Pressure, Capability and Desire Can Be Separated

• An interesting sociological observation: “The further someone is from the field, the more questions they ask.” Practitioners are less likely to see loss-of-control scenarios as urgent, “because this may not arrive in a single day.”
• His standard answer has 2 layers. First is evolutionary pressure: “When we train AI, it’s basically the same way we trained dogs.” Under that pressure, AI cannot evolve self-awareness like humans. Capability and desire can be separated: many brilliant mathematicians and physicists are extraordinarily talented in their fields but have no interest in power and seek no conflict. Creativity does not require consciousness either: “Those 2 things are not the same.”
• The second layer is interpretability. “Through interpretability, you can truly know what the model is doing,” which can “eliminate people’s fear of this black box.”

19. A Minority Faith: AI Will Ultimately Become Science, and the Next Newton May Not Be Human

• He considers himself a minority among practitioners: “I may be in the minority. Most people may already feel hopeless.” He nevertheless believes “there must be a good principle that makes the model work”; finding it would make models both stronger and safer. His confidence comes from years of studying neural-network training dynamics: since 2020 and 2021, “many problems that I couldn’t figure out suddenly started making sense,” suggesting he may be on the right path.
• His historical analogy is the episode’s most ambitious passage: “I think it will eventually become a science.” Alchemy became chemistry; Tycho’s observations became Kepler’s synthesis and then Newton’s principles. Before Newton, people may have thought the subject was inexplicable and impossibly hard. After Newton, “it would seem fairly simple.” The next Newton? “Not necessarily human. It could be AI, or humans plus AI.” That is why RSI should be built: use all the world’s compute and all its strongest minds, human and artificial.
• 曼祺 asks whether he will live to see a Turing Award given to a system. “If we reach that point, the Turing Award may no longer be very important.” By then, it would be a knowledge-discovery system producing new knowledge every 1 or 2 months. There are few people on the same path: many have given up, while some theoretical researchers have “joined because they couldn’t beat it,” entering OpenAI and joining the scaling push. He names 马毅 as someone still pursuing interpretability.

20. Technique vs. Principle: AlphaEvolve’s Matrix Multiplication Is Interesting, but Not Yet Beautiful

• Asked whether any of AI’s claimed mathematical breakthroughs are beautiful, he cites Google’s AlphaEvolve matrix-multiplication result, which reduces the number of multiplications required. “When you actually look at it, it’s quite interesting,” but the characterization is clear: “A lot of what AI does today is still at the level of ‘technique’; it hasn’t reached the level of ‘principle.’” Technique is a clever concrete transformation, effectively an exhaustive search through a large space to find a good result. Principle is discovering a new theory or chain of logic—something intrinsically beautiful that has never been seen before.
• What would human researchers enjoy once AI discovers new things automatically? “If I could understand how AI actually works, I’d be very happy. Then I could appreciate its beauty.” He still reads mathematical proofs for the pure pleasure of the moment when everything suddenly clicks, even when they have no direct connection to his daily work.

21. The Negative Example of the Self-Driving Era: “It’s Always 5 More Years”

• He joined Google X’s self-driving-car project in 2013, later Waymo. After the initial excitement, he found himself doing peripheral work. He wanted to apply deep learning to autonomous driving, but “there were very few GPUs at the time, and resources went to the favored people.” Level-7 staff had room to maneuver; as a Level-4 engineer, he could work only on small-sample problems such as emergency-vehicle detection—“the sample was so limited that you couldn’t even train a classifier, and might have to write the rules by hand.”
• His organizational observation from that period still applies: many senior people “refused to learn, stopped learning,” instinctively deciding the technology would not work. They needed 2 or 3 years to realize deep learning was transformative; the same cognitive divide persists after large models arrived.
• His judgment on the self-driving industry was correct even then: it was either a success or a complete failure, with no middle state, while fixing corner cases took years. He had already developed immunity to optimistic forecasts: “It’s always 5 more years—people said 5 more years in 2011, and when I joined in 2013, it was still 5 more years.” He adds the fair caveat that autonomous vehicles now operate on the roads in San Francisco and Los Angeles.

22. 2 Types of Researchers: Those Who Set Direction and Those Who Know How to Execute—and 伊利亚’s “Religious Faith”

• During his early years at Meta FAIR, he was still primarily an engineer. He built the Monte Carlo tree search for the Go project and the distributed reinforcement-learning system himself; OpenGo beat Korean professional players 20 games to 0 in 2018 and 2019. The feedback became a personal caption: “When other people opened their computers, they had papers on the screen. When I opened mine, I had code.” From 2019 to 2022, he gradually developed his own taste and has continued to publish at least 1 first-author paper every year. Researchers who are not hands-on “mix up hard problems with extremely hard problems.”
• His taxonomy has 2 categories. The first sets direction and opens up the right problems. 乐坤 (likely) is the archetype: his famous cake diagram put reinforcement learning at the top as the cherry because feedback is scarce, while betting on feedback-rich self-supervised learning. “The entire GPT training process is also partly self-supervised.” His own Coconut (likely, referring to latent reasoning) belongs in this category. The second type “not only knows what to do, but knows how to do it,” extracting the mechanism; that is his path. 伊利亚 and 达瑞 both belong to the first category. “伊利亚 in particular has an almost religious faith in scaling laws—don’t overthink it, just skill, skill, skill, skill.”
• He strongly agrees with 伊利亚’s recent statement that scaling had sucked all the air out of the room: “The dividend from this scaling wave is basically used up… Every research direction has a ceiling.” That leads to his startup rule: “You should start a startup only when you’ve genuinely seen something different. If you strongly believe in scaling laws, you should be in big tech.”

23. The Anti-Big-Tech Organizational Thesis: 150 People Is the Inflection Point

• From the turmoil in Meta’s later period, he draws a sharp conclusion: “This wave of AI is anti-big-tech. All the incentives, job levels, and relationships inside big tech run counter to the direction AI is taking.” As organizations grow bloated, progress slows. Once a manager-executor hierarchy appears, it resembles the advisor-student loop: “Human communication is always the slowest,” while middle management often exists to improve its own position.
• He gives a concrete threshold: “Look at Llama 4 and XCR [as stated]. Both started to break down after exceeding 150 people.” At 150, he argues, the organization crosses a human-systems boundary: people no longer know one another and must rely on structure to obtain information and report upward.
• The counterexample is a large lab whose core teams remain small. OpenAI has more than 7,000 people, “but there actually aren’t that many people working on models.” A team like Crawl Code (likely) is small. His company will stay lean: “We won’t hire too many people.” Some prospective recruits are already interested from Frontier Lab, and discussions or hires are underway.

24. 2 Futures in the Closing Argument: Coding Is the “Xie Gong Clog,” While Original Ideas Remain the Variable

• 曼祺’s closing ties together the 3 episodes: coding ability is Li Bai’s Xie Gong clog—“wearing Xie Gong clogs on his feet, climbing the ladder into the blue clouds.” Kimi K3’s technical report showed that early checkpoints could already optimize kernels. The GPT Speed Run benchmark used by RSI was launched by Neon optimizer developer Jordan Keller (likely), and is now shifting from human-built to AI-built. “RSI is already happening on specific model-training tasks.”
• The fork left for listeners is straightforward. If the curve rises smoothly, “the overwhelming majority of opportunities will remain with the companies that are already strongest.” If there are steps and plateaus, constrained by energy and compute, new methods will be needed: “Original ideas cannot be solved immediately just by making coding ever stronger. They are at least one variable. Which future do you believe in?”