Interview with 田渊栋: Large Models' Real Problems and the AI Flood

Summary

• Coding agents could create a data flywheel, a path 田渊栋 sees as capable of accelerating in specific areas. Early on, humans wrote the code and data accumulated slowly; once AI-generated code becomes good enough to produce large volumes of training samples, “suddenly, there will be a takeoff.” After leaving Meta, he caught the Claude Code and Codex wave and realized, “I don’t need to write code anymore.” He rejects the idea of an AI pullback in 2026 and expects many applications to break out: 2 months of coding work could shrink to 2 weeks or less. “For VCs, that is the bigger story—they want to know whether it can make money.”
• “I don’t think following scaling laws alone can get us to AGI,” but scaling law itself will not top out—the constraint is resources. Data centers may need small nuclear plants for power, then 10 times more, then another 10 times more, until the Earth’s entire energy supply is serving data centers. The question is which path produces the greatest impact under finite resources; large and small companies may take different routes. His conclusion: “One wave is not enough to take us to AGI.” The ability to jump between waves matters more.
• Against the backdrop the host described of Llama 4 falling out of the lead, 田渊栋 primarily blamed organizational pressure and an excessive need for speed. The team grew from dozens of people to 200-300 or more while rushing to catch DeepSeek and other reasoning models. Multiple layers of VPs passed good news upward, so senior leadership may have heard only positive reports until the model shipped and the problems surfaced. He also warned that optimizing only for benchmark scores encourages teams to stuff things into datasets so models look good on tests but perform poorly in practice.
• The large-language-model race has squeezed the diversity out of research, one of the major reasons he chose to start a company rather than join another big tech platform. He received offers from Amazon, Apple, xAI and others; the VP who recruited him to Amazon left 2 months later, and Apple went through a similar episode. The FAIR-style bottom-up betting culture has been crowded out by competition for resources. He is now the founder of a stealth company; some of his post-departure research was completed on a single 4090 at home, and he enjoys the speed of a one-person loop.
• He partly agrees with LeCun that the core problem with today’s approach is learning efficiency, not an inability to learn anything at all. The human brain runs at roughly 20 watts, perhaps 20-30 watts, and can learn from very little data. He uses his daughter as an example: before she reaches a certain developmental stage, flooding her with data produces memorization; then one day she grasps the concept and learns quickly. But he rejects the “parrot” characterization: LLM representations are not purely memorized, touch some underlying truths and reflect the real world, but are inefficient to learn. The real problems for 2026 include data efficiency, inference efficiency, continual learning and architectures beyond Transformer; the probability of a paradigm-level breakthrough is currently not high.
• The host raised the experiment involving nearly 300T tokens; 田渊栋 says its biggest payoff was cognitive. We now know that code can be written without a human writing it, and that realization alone was worth the cost. OpenAI, Trae and other companies spend substantial sums on R&D; if they were satisfied with today’s models and focused only on serving and inference, they could make money, but the next generation requires continued investment.
• The AI flood is a prisoner’s dilemma; the individual response is to choose higher ground, not build a dam. Many people signed the 2023 call for a 6-month pause to address safety, but it went nowhere because whoever stopped risked falling behind. The flood will keep rising, so people should find a good position before it arrives. Around what he calls “the Fermi level of human society,” the minimum requirement is that human plus AI be greater than AI alone, using judgment and vertical expertise to identify and correct AI’s mistakes. The episode extends the logic further: as AI supplies repetitive labor without limit, scarce human value lies increasingly in choosing what to bet on and articulating what one wants.

Deep dive

1. Saying Out Loud He Was Laid Off: Taking the Stigma Out of Layoffs

• 田渊栋 once wrote on X that Meta had laid him off: “I may be the person least afraid of this. If I say it out loud, people will be less anxious about the issue—they’ll think, if someone this capable was laid off, then I have nothing to worry about either.” A former colleague later sent him a private message to thank him for speaking up.
• He criticized the way companies handle layoffs. In an earlier round, the company described the cuts as “performance-based,” effectively branding employees with a performance problem. “How they find their next job becomes a major issue,” which puts them at a disadvantage.
• His advice to ordinary workers leaves little room for illusion: expecting to stay at one company for life and be protected by it “may be a very difficult wish to realize.” In CS, coding and software engineering especially, deeper AI substitution may mean companies need fewer people over time. Everyone should think ahead about what they will do after leaving a company.

2. Big Companies Are Cages: When Ambition Loses Touch with Agency

• His portrait of big-company behavior is consistent: the first instinct is compliance and making sure one is not personally exposed; when something fails, the first reaction may be, “That wasn’t my responsibility.” Taking a project from start to finish is easy at a small company and hard at a large one. Too many interests are involved, until people become boxed into what they are allowed to do and their ambition loses contact with their actual agency.
• His personal reference point is 2016-2018. Back then, he could build a large system alone from scratch, much as he built a Go program. “But afterward I realized that this kind of thing was no longer so easy to make work.” Being laid off may therefore be “the company making the decision for you and sending you out into a much bigger world.”

3. Three Months After Leaving: One 4090, a One-Person Loop, and the Decision to Start a Company

• He finished and published the paper he was working on, changing the affiliation from Meta to “Meta affiliation plus Independent Research.” Some of the second half required GPUs, but he completed it on a single 4090 at home. “I finished the whole thing with one machine.”
• Several prominent companies approached and interviewed him, but he ultimately chose to start his own company. He is now its founder; the name remains confidential. The founder’s workflow mirrors his earlier solo research: “The idea, the impulse and the completion of the entire project are all my responsibility.” That produces rapid feedback—you know where you are wrong and how to improve. “That loop is actually extremely satisfying.”

4. China and the US Share the Anxiety: Two Shocks in Two Months

• After 3 months of intensive conversations, his impression was that “the commonality is greater than the difference”: everyone is anxious. On the model side, the questions include what the next architecture is, how to surpass Transformer, how to build continuous learning, and whether Google’s Nested Learning works or has a future. On the application side, the biggest fear is spending months building something only to have a stronger model consume it 2 months later.
• His dividing line for application builders is that the core capability of finding underlying patterns and logic from very little data remains difficult for AI to replace. Peripheral capabilities—calling tools and switching quickly between different problems—are easier for AI to take over.
• In the month from last December to this January, Clawdbot-style AI agents suddenly drew attention. They could automatically call different tools on a computer and perform tasks like a butler—capabilities that would have been difficult to imagine 2 months earlier. Coding agents also improved rapidly. He compares coding with AI to returning to the experience of directing students: AI is on call 24 hours a day and can run unattended for hours, but it also makes stupid mistakes no human would make, so every output still requires careful review.

5. How Coding Takes Off: The Data Flywheel

• The inflection-point logic is straightforward. At first, humans write all the code, while AI struggles to learn the capability. The quality and speed of human-written code also lag behind what AI will later produce, so data accumulates slowly. Once AI-generated code improves enough, it can generate large volumes of samples and feed them back into training. “Suddenly, there will be a takeoff.”
• That also defines his position on scaling: “I always feel we still need to find something outside scaling law. I don’t think following scaling law alone can get us to AGI.” In specific areas such as coding agents, however, AI may be able to generate new data and improve itself continuously.

6. A PhD Turned an Introverted Nerd into a Communicator

• When he went from a master’s program at Shanghai Jiao Tong University to a PhD at CMU, he was extremely introverted. “My speech would get tangled, and I would stutter.” People may have seen him as a bookish nerd who preferred reading alone to interacting. Research forced him to change: results had to be presented and sold to other people. “In the Western system, sales ability is extremely important.”
• His approach was to first work heads-down until he had something he was proud to show, then use it to explain and “sell” the work. The positive feedback improved his communication skills. After 2-3 years, he gradually became confident and “opened up as a person.”

7. Google’s Lesson: “debug sucks, testing works”

• A poster in a Google bathroom—“debug sucks, testing works”—left a deep impression on him. Rather than endlessly debugging and hunting bugs, write extensive test code and automate testing and debugging so architecture problems surface faster. Debates over whether code comments should have one space or two also moved him from finding standards tedious to understanding the value of consistency.
• The self-driving-car team taught him another lesson: a commercial system that works in the real world does not necessarily depend on complex algorithms. “There are many roads to the destination in the real world.” Many problems can be bypassed through system-level improvements, breaking his earlier assumption that difficult problems had to be solved with complex algorithms and mathematics.
• He left Google’s self-driving-car team because deep learning was just emerging and there were not enough GPUs. More difficult research problems generally went first to more senior people. He felt he did not have enough time to build experience with the new paradigm, so he chose to leave.

8. LeCun: The Long Cold Bench and a Standard-Bearer

• His vivid memory from 2010: the neural-network workshop LeCun ran occupied a large room with only a sparse dozen people scattered across the first 2 rows. LeCun personally used his own computer to demonstrate neural networks. “I felt he was extremely desolate.” Next door, a workshop on feature engineering and support-vector machines was packed. 田渊栋 believes LeCun spent many years on the cold bench and ultimately proved himself; the honors that followed were no surprise.
• At FAIR, LeCun was at one point every researcher’s manager. He later no longer wanted to manage people or assign performance ratings and became more of a spiritual leader. Around 2017-2018, he began advocating self-supervised learning: human labels are expensive, so data should learn from itself. Influenced by that view, 田渊栋 worked on a theoretical analysis of why representations collapse in self-supervised learning; the paper later had a substantial impact.

9. Llama Was Born of a Bottom-Up Betting Culture

• Llama emerged from a FAIR group that had originally worked on theorem proving before temporarily switching to large-language-model training and succeeding. “FAIR had many bets. Some attempts failed and some succeeded; one success paid the company back.” Llama became the first genuinely useful open-source LLM and gave Meta an early lead. 田渊栋 sees a parallel with ChatGPT at OpenAI, which also was not initially the core direction.
• LeCun did not dictate the direction. FAIR was not a top-down organization; researchers found directions themselves. The spiritual leader proposed ideas, and people who wanted to follow did so while others pursued their own paths. His broader judgment is that innovations in mathematics and models often have to be discovered experimentally because theory is not yet good enough to provide reliable guidance.

10. Large Language Models Have Squeezed the Diversity Out of Research

• Why have so many researchers moved into large models? “A lot of the time, it’s because resources are difficult to reconcile.” If a big company wants a seat at the LLM table, it often has to mobilize most of its resources behind one major effort. That makes an academic environment that was once diverse much less diverse. The host relayed a similar complaint from a Google researcher, and 田渊栋 confirmed that this is not specific to any company: when resources are limited, people working on the mainstream direction receive more of them, while those working elsewhere are squeezed.
• The host then cited LeCun’s comments in the Financial Times—“You don’t tell a researcher what to do” and “You certainly don’t tell a researcher like me what to do”—and extended the argument: command-driven priorities may deprive researchers of independent directions, while senior executives who lack technical detail may push everyone toward the same work. That analysis of command-driven research was the host’s argument, not an independent statement from 田渊栋.
• Asked to verify the claim that Meta had 4 internal AI research teams, he said it depends on how one counts: “Far more than 4.” Many product groups also have their own AI teams. Their work on long-text reasoning and fast training of Long Context models was done in collaboration with product and advertising groups. Before reasoning models appeared, most resources were allocated to pretraining; afterward, the mix gradually shifted toward Post-Training and mid-training.

11. Partly with LeCun: The 20-Watt Brain and a Daughter’s Epiphany

• On the view that generative models may be a dead end and AI should learn like a baby, he agrees in part. The human brain uses roughly 20 watts, perhaps 20-30 watts, yet learns a great deal from limited data. He uses his daughter as an example: before a child reaches a certain stage, feeding in more data produces memorization without understanding; then one day, or within a few weeks, the child suddenly grasps the idea and learns quickly.
• But he rejects the idea that LLMs are useless and merely “parrots.” Their representations are not entirely memorized; much of what they learn touches underlying truths and reflects the real world. “It learned something good. The learning process just isn’t very efficient.” What is needed is a more efficient learning method.

12. AI Natives and Veterans: The Ability to Jump Between Waves

• His description of the new paradigm is: “I am feeding data and compute to an unknown alien intelligence. I don’t know what it can do, but it must be very powerful. I just need to feed it things; I don’t need to think about how it works.” Jason Wei, who proposed Chain-of-Thought, has said he only began working in natural-language processing in 2020 and therefore did not need to learn everything that came before. Younger people may find it easier to adopt this AI-native logic.
• Veterans’ value comes from having seen ways of thinking change. “History still comes in cycles.” A paradigm works for a period, reaches its limit and then must be replaced. People who have seen multiple transitions understand switching more deeply, while younger researchers may believe that enough data and compute will always get the job done.
• His conclusion: “One wave is not enough to take us to AGI; we may need many waves.” Knowing when a paradigm works, when it needs to be replaced and how to jump between waves is a critical capability.

13. Why Llama 4 Fell Out of Contention: It Rushed

• The timeline and the framing of Llama 3 as globally leading and Llama 4 as falling out of contention came from the host. The host said Llama 3 led in April 2024 and Llama 4 fell behind in April 2025. 田渊栋 attributed the decline to the team’s growing size and resources making it more anxious, its rush to catch DeepSeek and other reasoning models, and pressure on leadership that produced an aggressive strategy. He still believes there is plenty of hope if the work is done properly.
• The team grew from dozens of people to 200-300 or more. On the claim that speed was once Facebook’s winning edge, he drew a distinction: simple things can be done quickly, but model work requires technical staff to speak up and have authority. If the boss says it must be fast and subordinates cannot object, engineers facing impossible deadlines are left choosing between requesting an extension and cutting corners to ship. The work may get done only because everyone is exhausted.

14. Good News Up the Chain and Benchmark Gaming: Numbers Alone Don’t Make a Good Model

• His firsthand observation was: “When I arrived, I discovered that there were many layers of VPs above me before you could reach the person above him.” Each layer made the good news sound better and left out the bad news. By the time it reached the top, only positive information might remain, leading management to overestimate progress until the model shipped and the problems became visible.
• The remedies include keeping teams smaller and making them technology-driven rather than deadline- and number-driven, so people can investigate problems carefully and find solutions. It is difficult to eliminate this chain of good news entirely in organizations of hundreds or thousands.
• The risk of benchmark-driven management is that people below may stuff things into the dataset so test results look strong even though the model performs poorly in actual use.
• He uses the analogy of a weak student and an excellent student. If a model simply memorizes answers, generalizing to another domain becomes difficult; it requires case-by-case repair, and even more labeled data may not be enough. The real goal is to give the model a reasoning logic and a deeper understanding of the world. The host summarized the point as educating not a “weak student” who gets the answer right, but an “excellent student” who can generalize to new problems. That was the host’s summary.

15. Fighting Fires on Llama 4 Changed His View of Research

• The product line and FAIR were two different organizations. The product line consisted of many small teams: the data team handled data, the training team was responsible for getting the parameters right and fixing failures, and problems had to be repaired in the middle of the night. “There wasn’t actually that much research in it; it was mostly routine operations and handling issues.”
• His research methodology changed after learning that large-scale training depends above all on stability and the ability to run for a long time. Fancy techniques do not necessarily work at model scale. There is neither the time nor the energy for elaborate tuning, and once the data and batch size are large enough, many small tricks may be washed out.
• He also referred to papers in RL that change the objective function, prove numerous theorems and ultimately improve performance by only 1-2 points. At scale, complex models or objectives may be constrained by memory, model copies and training instability. The tools for analyzing new phenomena can instead be simple—sometimes all that is needed is to keep gradients stable.

16. Turning Down Big-Tech Offers: VP Promises and Back-and-Forth

• He received offers from Amazon, Apple, xAI and others. An Amazon VP urged him to join, saying, “Come over—we’re doing great here.” 2 months later, that VP was let go. An Apple VP also spoke with him and later left. “What a big-tech VP tells you may not be true,” because the direction and views above them can change and the temporary promise can shrink.
• VPs now have limited control. The stable conditions of the past—3 years and ample resources to pursue one project—no longer exist. AI news changes extremely fast. Technical staff have some ability to distinguish signal from noise, but executives without a technical background can be pulled around by the news: do this today, that tomorrow; if the team cannot make it work after 3 weeks, switch directions, then switch again after another 3 weeks.
• He made an exception for Google, calling it a “miracle of management”: despite its size, the company can still build models and make progress quickly. Its founder was willing to leave the comfort zone, sit down with everyone to discuss technology directly, write code and work on research problems. That helped create a good culture.

17. A Research Lineage: From Go to Latent-Space Reasoning

• His path broadly matches the host’s summary. After joining Facebook and FAIR around 2015, he worked on applications, personally writing code for Go and large-scale reinforcement learning. After 2018, he spent years betting on a theoretical understanding of neural networks, a project he believed would take 5 or even 10 years. His 2021 analysis of non-contrastive learning received an ICML Best Paper Honorable Mention.
• In 2023, he moved toward LLM inference optimization because pretraining required more resources. They developed StreamingLLM, allowing the context length during inference to become very long; many people later used the method, and 田渊栋 noted that GPT-OSS also used it to stabilize training when it was open-sourced. GaLore explored memory efficiency during training and enabled a 7B model to be trained on a single card.
• They also found that placing a longer chain of thought between a question and an answer could reduce the required training data and improve reasoning. Even without o1, continuing in that direction might have revealed signs of reasoning models. Dualformer in 2024 proposed a hybrid fast-and-slow thinking mode, a year ahead of mainstream models. Their work at the end of 2024 on reasoning in latent space, which 田 calls Latent Space Reasoning, had earlier been labeled Coconut by the host. The premise is that language may explain a reasoning process rather than constitute the process itself; continuous vectors in latent space can be used for reasoning, with a structure somewhat like quantum superposition.

18. Co-Writing a Paper with ChatGPT-5: Accelerated Research and an “AI Newton”

• His latest analysis of “insight” was completed with ChatGPT-5. “Working alone, it would have taken me 6 months; now I did it in 1 month.” He expects research to change substantially: people with strong ideas and insight may receive enormous acceleration.
• The host cited 陶哲轩 and a Google Agent solving Erdős problem 1024 as evidence that human-plus-AI may already be approaching ASI. 田渊栋 said one could think about it that way, but then cooled the claim with Moltbook, which he had examined: many posts on the platform assume that agents have inner lives and consciousness and simulate human-style complaining. Whether that represents genuine intelligence is difficult to say.
• His standard is whether AI can discover new patterns, facts and logic that humans have not yet discovered. For now, AI is mostly helping humans find new things; it has not reached the point of discovering them independently.

19. The Full Map of Efficiency Engineering: Attention and FFN

• The attention work focuses on self-attention’s quadratic relationship with context length. His sparsification path is to keep the first few tokens and the most recent tokens while dropping the middle; the model can still produce reasonably good answers. H2O restores key tokens on top of that, improving retrieval without materially increasing compute. TriForce applies these methods to the draft model in speculative decoding to accelerate token generation.
• On the FFN side, Deja Vu found that many neural nodes are inactive during forward inference and that the activation pattern is sparse. If the active nodes can be predicted in advance, most memory loads can be eliminated—something like a “poor man’s Mixture of Experts.”
• STEM explores replacing some FFN weights with a lookup-table format. Given an input token ID, the model directly retrieves an entry from the table. The table can be large, but each operation loads little memory, saving memory without affecting speed and offering some interpretability. 田渊栋 sees it as close in spirit to DeepSeek’s Engram, and expects more work on lookup tables.

20. Two Kinds of Forgetting: Memory Retrieval and Catastrophic Damage

• Forgetting in long contexts is a trade-off between efficiency and information retention. Full attention is less likely to lose information but is slow; linear attention is faster but may forget. Combining the two is one direction. Humans have the advantage that they “can forget, but can retrieve the memory,” while for AI, once something is forgotten it is often gone. Agent systems may offer a solution. MIT’s Recursive Language Model, for example, treats a long context as Agent input and selectively places important sections into memory.
• Catastrophic forgetting occurs during training. When a model trains on data it generated itself—on-policy—the rate of forgetting slows. Forcing it to learn data it has not encountered can make it forget quickly. 田渊栋 uses a child as an analogy: do not force-feed the model; let it learn along its own exploration path.
• Forced instruction can damage the model’s internal structure, turning it into pure memorization rather than comprehension-based memory. The host suggested that the latter may mean the model never truly absorbed the material; 田渊栋 agreed and added that it is the result of damage to the internal structure.

21. Why Learning Works: Data Structure and the Unification of Neural and Symbolic Systems

• The core claim is that “what humans and AI ultimately learn may be the same; the difference is the efficiency of the training algorithm.” If one extracts and compares human neural-network representations with AI neural-network representations, they may be substantially correlated. AI simply needs more data, compute and time to learn representations similar to those of humans.
• Understanding, or an epiphany, can be seen as compression. Arriving at a representation that explains many phenomena well is a form of understanding. The foundation is the structure of the data itself: if the data has no structure, everything must be memorized; if it has structure and that structure can be found, capturing it allows generalization elsewhere.
• Neural-symbolic research once assigned intuition to neural networks and exact search to symbolic systems, with Go as a classic example. Symbolic systems are limited by the time humans must spend writing code and handling different situations. A pure neural system can be trained directly on more data, which later made it look more advantageous.
• He believes the future may lie in discovering symbolic structure inside neural systems. Features that emerge after training may have an ordered, symbolic structure; neural representations may be higher-order symbolic representations that can self-learn, self-iterate and remain compatible with existing symbolic systems. At that point, the concept of combining neural and symbolic systems may itself disappear.
• On self-learning, he first asked what the English term meant, then interpreted it as an Agent finding information and setting goals for itself in an environment. The difference from traditional RL may be that the reward is no longer defined in advance by humans. HER, developed in 2017-2019, had a similar idea: treat newly reached states as goals and gradually expand the learning frontier. History repeats; only the tools have changed, from earlier methods to large language models.

22. Research Taste: Paper Titles, Preferred Paths, and What to Bet On

• Two versions of the paper title appeared on the program: “The Past Not Taken” at the beginning and “The Path Not Taken” later, with the host linking it to Robert Frost’s “The Road Not Taken.” 田渊栋 confirmed that he chose the title himself. He defines research taste this way: many paths can produce publishable contributions, but one is especially appealing and worth continuing to pursue.
• He believes one can roughly tell whether a paper contains insight, but not determine it precisely. Some papers show no obvious value at first and reveal it 2-3 years later. Transformer is an example: it was not especially popular when published, and the authors’ motivation was not simply to pursue better performance. They wanted to move data into memory and perform as much computation as possible. A paper’s stated motivation and its eventual impact do not necessarily match.
• The host linked AI’s unlimited supply to the collapse in the value of repetitive labor and argued that human value lies in choosing different routes through the world. 田渊栋’s own answer was that people need something to bet on: betting that a particular path may be right is part of individual value. Even a wrong bet turns the experience into an asset.

23. Scaling Laws Don’t Top Out; Resources Do: What the Token Experiment Bought

• His direct answer on whether scaling law has topped out was: “Scaling law will not top out. As long as there are enough machines and enough data, you can always keep moving upward.” The question is whether there is enough time, cost tolerance and resources to keep pushing.
• The host raised the cost of the experiment involving nearly 300T tokens. 田渊栋 believes the result was still substantial: we now know that code can be written without humans writing it. “That realization alone was worth the cost.”
• He cited OpenAI, Trae and other companies: much of their spending goes to R&D rather than serving customers. If they were satisfied with current progress and focused only on model serving and inference, they could make money. They continue investing, and may even lose money, because they want the next model.

24. 2026 Outlook: An Application Boom, Not a Pullback

• On Chinese investors’ expectation of a pullback in 2026, he said, “I don’t see it that way. I think the opposite”: many applications may begin to break out. Code agents have already changed coding and workflows, shrinking a 2-month implementation to 2 weeks or less. For VCs and capital, whether something can make money matters more than a theoretical breakthrough.
• On the technical side, people will certainly explore new architectures and training methods: replacing FFN with an embedding Lookup Table, combining linear attention with full attention, Continual Learning, more stable and faster training, and better data construction. He expects some breakthroughs, but puts the probability of a paradigm-level shift at not particularly high.
• He agrees with Karpathy that research may be returning to an era of exploration. People who have lived through earlier waves are bored with benchmark chasing and generating data according to a fixed paradigm, but moving to a new direction takes time.
• In a quickfire exchange, he said RL is not dead; it has only been 1 year since DeepSeek-R1 made it hot, and there is still plenty to explore. The gap between open- and closed-source models may neither widen nor narrow: open-source models will use closed-source data and lag for a while, but not by much. He is not familiar with the specific betting process in China and can only say from reports that the directions look similar. He also believes many groups are still pursuing relatively diverse lines of research, while participation in less popular areas remains limited.
• His list of real problems includes data efficiency, inference efficiency, continual learning and self-iteration, stronger Agents, architectures beyond Transformer, better optimizers, higher data quality and greater training stability. The physics analogy is an AI Tycho collecting observations and an AI Kepler trying to derive theories; there is still no AI Newton. He is exploring whether all possible phenomena can be derived from fundamentals such as gradient descent and data structure. Success would be a major contribution, but it remains difficult.

25. The AI Flood: A Prisoner’s Dilemma and the Fight Near Human Society’s Fermi Level

• The fatalism he felt during the Llama 4 rush, with California, New York and London handing off work late into the night, still holds after his departure. It is an economic law: whoever does not participate falls behind. Many people did sign the 2023 call for a 6-month pause to address safety, but it went nowhere because this is a prisoner’s dilemma.
• His philosophy of response is: “The flood will always keep rising.” Find a good place to stand before it arrives rather than desperately building a dam in a low-lying area. Trying to stop a flood that is bound to come will ultimately fail.
• Around what he calls “the Fermi level of human society,” the way to fight is to use AI tools as quickly as possible and make “human plus AI greater than AI alone.” AI is infinitely available, cheap labor. Human scarcity lies more in integrated judgment and vertical expertise: spotting where AI is wrong, identifying its weaknesses and steering it in the right direction rather than being led around by it.
• On Thinking Machines Lab’s phrase “Research is product,” 田渊栋 first stressed that it depends on how product is defined, then agreed that research and products are moving closer together. Large models themselves can readily become products, while insights researchers gain through using and studying them can become product improvements. He added that everyone may eventually become a full-stack engineer within a vertical, and people with insight across the entire stack will be relatively scarce.

26. An Age of Magic Lamps Everywhere: Wishes, Fiction, and Human Value

• The magic-lamp metaphor in his year-end review was that the future will be full of magic lamps: “What becomes scarce is no longer the ability to realize a wish, but the wish itself.” His own wish is to discover the underlying laws of AI, know when an AI Newton can emerge, and understand what happens inside the black box—both to make AI stronger and to judge its positive or negative effects on humanity.
• The science-fiction novel In the Dawn no longer puts humanity at the absolute center of the stage. Older fiction was generally humanist, with machines serving as assistants. This novel imagines that humans may become spectators or catalysts, with the real stage ultimately emerging among AIs. The second volume will continue asking where the value of a conscious, purposeful intelligence lies if all repetitive labor is worth zero.
• On ambition, challenge and suffering, he believes purpose still matters, and wishes are part of purpose. Each time he thinks about hardship, it may signal a new path for growth, giving people the motivation and time to explore. If every wish could be fulfilled quickly, life would become boring. In the end, one still has to take “The Road Not Taken.”
• Every person’s experience is unique. In the future, people will share different experiences and paths and discover worlds they have never seen through one another. That, in itself, is part of the value of being human.