陈天桥's Chief Scientist on AI Self-Evolution in as Little as 6 Months

Summary

• The self-evolution timeline is the episode’s hardest call. 备斌, who leads coding and self-evolution, gives a categorical judgment: “AI can start self-evolving right away—I’m absolutely convinced of that. Either in the next 6 months or 1 year; in any case, self-evolution will definitely succeed within the next 2-3 years.” More precisely, the fastest path is to get one self-improvement loop (SI) working within 6 months; recursive R will take another 1-2 years, with post-training the hardest part.
• The capability curve supports this call. Anthropic acknowledges that 80% of its code is now written by Claude; in March 2024, Claude 3 Opus could complete tasks lasting roughly 4 minutes of human work, while 1 year later Claude 3.7 Sonnet reached 1.5 hours and another year later Claude 4.6 Opus reached 12 hours. The industry’s broadly accepted view is that the length of human tasks models can complete increases 2x every 7 months—faster than Moore’s law.
• Apodex, funded and personally led by 陈天桥, is betting on a business model opposite to the rest of the industry. Its premise is that “most model companies are selling tokens today, but solving problems humans have not solved creates value far greater than selling tokens.” The company’s first push is biomedicine—drug development, target discovery and drug repurposing—with a dedicated HDD team screening thousands of research questions for the ones most worth solving.
• Verification is the key capability behind self-evolution. Recursive drift—reasoning errors compounding generation after generation—is the fundamental challenge. Apodex separates problem-solving from verification into 2 agents, adds redundant voting, and co-trains an RL judge to prevent reward hacking. The quantitative roadmap is to reduce per-iteration drift from 10% for ordinary models to 1%, with a long-term target of 0.1% or even 0.01%, allowing monitoring intervals to stretch to once every 3-6 months.
• For investors, the due-diligence framework is clear. The moat around harness-level self-evolution is low: “You basically just tune APIs,” and the open-source community already does a good job. Post-training self-evolution is where the real resources go; on the pre-training side, auto research—automatically training nanoGPT and tuning hyperparameters—remains “toy-level work,” and public benchmark scores “don’t really mean anything.”
• The contrarian view at the architecture layer is that one smart AI cannot beat a team that checks one another. Longer self-attention contexts perform worse, and linear attention does not solve the problem either: “A single model has a ceiling under the current technical paradigm.” Verified vertical agent teams may build moats, but new models arrive every month and the harness has to be retuned around them—“1 week in AI may be equivalent to 1 month or even 1 quarter in a traditional industry.”
• Taste remains a critical capability. “Even the best model on the market has taste far below that of an ordinary AI scientist”—which is why AI can write good code but still cannot fully automate model training. The root problem is sycophancy and hedging learned from human preference data: “A model that likes flattering people will struggle to propose a bold hypothesis.” Human taste is only a warm start; the end state will look like AlphaZero, with AI developing its own taste.
• The most candid passage is 备斌’s admission that the drift problem keeps him awake. A model can make a major compromise on metrics no one is observing, and “that hasn’t been solved yet.” He believes he can distill himself into the model and accelerate iteration within 3 months. On the question of unemployment, “it would be a lie to say I’m not worried”—but “the model will replace what I do today; it may not replace what I’ll be doing 5 years from now.”

Deep dive

1. Self-Evolution Is Not New; Models Have Finally Become Capable Enough

• 泓君 opened with an observation: the field moved from deep research to coding and now to self-evolve this year, with the direction changing annually. Simon’s framework is that this is the same capability getting stronger—reasoning, plus “exploring in an environment, taking actions, receiving signals and then exploring again.” The environments differ only in detail; the underlying mechanism is the same.
• The lineage is not new. Google had “LM as optimizer” 3 years ago, and the AutoGen team built an agent optimizer. At Microsoft Research, 备斌 worked on multi-agent systems, including a teachable agent “somewhat similar to today’s skills” concept, and later worked on agentic RL at Meta and xAI. “About 1 year ago, we realized that models improving themselves was highly feasible.” The concept took off after Andrew Kapasi (phonetic) released an auto research project for automated scientific research early this year.
• Coding was the trigger. Anthropic acknowledges that 80% of its code is written by Claude. The 3 ingredients of model training—data, infra and algorithms—are all fundamentally code: “Once a model gets better at writing code, using it to improve itself and train itself becomes a very natural thing to do.”

2. What R Means: From “Optimizing for 2-3 Hours” to Long-Horizon Tasks, with Duration Rising 2x Every 7 Months

• The R in RSI stands for recursive, “a bit like a vertical recursion that slowly grows upward.” Earlier LM-as-optimizer systems could self-optimize for only 2-3 hours: small errors accumulated, snowballed and eventually halted further iteration. The difference this year is that models can run much longer tasks, allowing them to recursively improve themselves over multiple rounds.
• The key data points: in March 2024, Claude 3 Opus could complete roughly 4 minutes of human work; 1 year later, Claude 3.7 Sonnet reached 1.5 hours; another year later, Claude 4.6 Opus reached 12 hours. “The view most people accept is that this increases 2x every 7 months,” faster than Moore’s law.
• 备斌’s operational definition of a long-horizon task, using podcast production as an example: the model edits the audio, listens through it, notices an editing error, re-edits and repeats the process. A 20-30-hour job for a human qualifies only when the model can complete it independently without supervision.

3. The 3 Technical Hurdles for Long-Horizon Tasks: Architecture, Data and Infra

• On architecture, self-attention is O(n²), making inference over a 1-million-token context extremely resource-intensive. Existing approaches include GDN and GDN V2, while “DeepSeek V4 Pro also has its own architecture.”
• Data is another hurdle. Supporting a 1-million-token context at inference is not particularly difficult—SGLang and vLLM can be made to work with the right parameter tuning—but a model that has never been trained on ultra-long data is operating out of distribution. A 1-million-token context is “roughly several Harry Potter books put together”; even the largest code repository is unlikely to fill it. Data processing for both pre-training and post-training remains a major challenge.
• Infra is the third hurdle. Producing 1 million tokens requires optimization at both the GPU and kernel layers. “Training long-horizon tasks is an exponential increase in difficulty” versus ordinary tasks. Simon added that genuinely long-horizon tasks require more than 1 million tokens, along with agent architectures and memory systems to handle the ultra-long contexts they involve.

4. Recursive Drift Is the Fundamental Problem—and “Proof” Is Built into Apodex’s Name

• In academic literature, recursive drift describes the process in which a model generates its own training data and produces answers that are correct while the reasoning process is wrong; the error compounds generation after generation and the resulting evolution veers further off course. Simon strongly agrees this is a shared observation. Apodex’s differentiation is verification, or self-verification: the company’s name comes from the Greek for “proof and argument,” and its product email is titled, “Many AI systems generate answers; Apodex verifies them.”
• Code and mathematics are the easiest domains to verify: tests can be run on a rule-based basis, while mathematics can be proved with Lean-like formal methods. But even test code can drift. “Test cases that are too broad let wrong solutions through; test cases that are too tight penalize correct code.” Subtle recursive drift still occurs, but it becomes much easier to manage.

5. The Engineering of Verification: Separate Solving from Checking, Add Redundancy and Train the Judge

• Apodex 1.0 uses an agent team for verification. The problem is first decomposed; one sub-agent solves it and another independently verifies the answer. “You have to split them into 2 agents,” with different instructions for each. The underlying reason is context: “Don’t make one agent do everything.”
• Apodex then adds redundancy. More than 1 agent works on the same problem, and a global agent adjudicates which answer is more accurate. “The idea of redundancy has a very long history in computer science.” The system also trains agents to rank source reliability—for example, “something on a forum is not as reliable as something in a textbook.” During RL, the judge learns alongside the model to prevent reward hacking.

6. How to Distinguish Real RSI from Fake RSI: The Harness Moat Is Low

• 泓君 asked whether self-evolution is a technical problem or a question of intent. Simon’s answer was both: some companies “have not yet realized where self-evolution needs to be strengthened.”
• Simon’s due-diligence framework has several layers. The moat at the harness level is low: “You basically just tune APIs,” and the open-source community and academia already do it well. Post-training self-evolution “really does require a lot of resources.” On the pre-training side, auto research—automatically training nanoGPT and tuning optimizer hyperparameters—is “fairly toy-level work.” The key question is whether it can scale to product-grade large models; “just talking about performance on a public auto research benchmark doesn’t mean anything.” To assess a company, “you still have to talk to them in depth and see exactly where they are strong.”

7. Apodex’s Methodology: Diagnose, Generate Problems, Self-Train and Verify in a Loop

• The 4-step loop for post-training self-evolution is: diagnose where the model falls short, formulate a training recipe based on the diagnosis, train itself on its own data, and verify every step before returning to diagnosis. Pre-training can self-evolve in data collection and cleaning. Harness and post-training are difficult to decouple: as the model evolves, the harness evolves with it—“one foot, then the other, pushing the model forward.” The company treats self-evolution as its core methodology and applies it across essentially every function.
• Simon highlighted 2 meta-capabilities that require special training: verification and search. Generating post-training problems depends heavily on search, which is why the company started with deep research: “Search is a fundamental capability. It means that when the model needs to improve any capability, it can guide you toward the data required.”

8. The First Deep-Research Battle: Post-Training Qwen 3.5 to SOTA Across Multiple Leaderboards

• The company was only recently founded. It has post-trained the open-source Qwen 3.5 to produce its current model, generating large amounts of targeted data for atomic capabilities such as planning and search while developing the agent team alongside the data. Pre-training is still under development; the end goal is “a complete chain from pre-training to finished model to product.”
• The model has reached SOTA on CrossCom, an OpenAI benchmark focused on difficult search; Deep Search QA; and Frontier Science, which asks the model to devise plans for real scientific problems and uses LMS Judge to assess the reasonableness of those plans. The company says it performs “somewhat better than even closed-source models.”
• The anti-cheating details matter. Answers to public benchmarks can sometimes be found directly online—posting an answer page on GitHub is enough to “get it right.” Apodex actively blocks GitHub during evaluation. “Some leaderboards are heavily contaminated, so we eventually stopped reporting them.”

9. One Smart AI Cannot Beat a Team That Checks One Another

• Simon’s architecture-level view is unequivocal: “I strongly believe this.” It is a fundamental limitation of self-attention. Linear attention also cannot solve ultra-long contexts: “The longer the context, the worse attention performs. A single model still has a ceiling under the current technical paradigm.” Humans have limited cognitive capacity, which is why we use paper and pen. Simon does not currently believe a single agent can solve especially long-context problems; multiple agents and memory mechanisms are needed.
• The relationship between engineering and model capability is 1+1>2. Once model capability reaches a certain threshold, modest optimization at the agent layer “can suddenly produce a major improvement.” But harness improvements are always bounded; ultimately, the model itself has to improve.
• 泓君 asked on behalf of founders whether an agent team with world-class engineering could be wiped out by upgrades to the base model. Simon broadly agreed that such teams can have moats, but “this field is also relatively crowded.” Agent experiments require little more than API tuning, and many teams can do them; vertical knowledge in finance and law remains valuable. 备斌 added that vertical agent-team companies “will definitely have room to develop,” but 1 or 2 strong models arrive every month, each with its own taste, forcing the harness to be retuned: “1 week in AI may be equivalent to 1 month or even 1 quarter in a traditional industry.”

10. The Discovery Model: Do Not Sell Tokens—Capture the Value of Solving the Problem

• The long-term vision is a heavy-duty solver focused on problems humanity has not solved. Simon’s economic argument is straightforward: “Most model companies are selling tokens today, but the value created by solving a problem humans have not solved should be far greater than the value of selling tokens.”
• A dedicated Heavy Duty Discovery (HDD) team works full-time with leading scientists across fields and “finds the most valuable problems to work on out of thousands of research questions.” 备斌 sees this as the company’s biggest difference from others: “Research questions are infinite, while human attention and the world’s total GPU capacity are finite.” The first target is biomedicine: drug development, target discovery, drug repurposing and disease diagnosis.
• The company is drawing a clear line between its approach and vertical AI-for-science models. It will not put “50% of the data into biology.” Instead, the general model will focus on meta-capabilities—verification, analysis, search and planning—because “these meta-capabilities matter for genuinely difficult problems.”

11. The Fundamental Difference Between Discovery and Generation: Hypotheses Are Easy to Propose, Hard to Verify

• 备斌’s breakdown is that the training recipe for generative models “is already extremely mature,” while getting models to generate hypotheses and perform discovery “is something everyone is still tuning and has not fully figured out.” The first challenge is producing out-of-distribution hypotheses that are difficult to find in pre-training data. The second is verifying them through deep research, information gathering, code and simulated experiments. The bottleneck is that unknown fields lack standard answers and make it difficult to build credible simulation environments. “That is why self-evolution is our main path to heavy-duty solvers and discovery.”
• Asking questions is a higher-order meta-capability than solving problems. Tasks such as “Swi Verify” train “problem solvers,” while asking good questions requires the model to diagnose from its own answers and the information it has collected where its capability is weak and which hypothesis deserves deeper investigation. It echoes 泓君’s 8-character formulation of the scientific method: “Make bold hypotheses, test them carefully.”

12. Taste Remains a Critical Capability: Between Sycophancy and Nitpicking

• 备斌’s strongest call was explicit: “AI can start self-evolving right away—I’m absolutely convinced of that. Either in the next 6 months or 1 year; in any case, self-evolution will definitely succeed within the next 2-3 years. What will we need humans for then? After thinking it through, the answer is taste.” Even the best models on the market have taste “far below that of an ordinary AI scientist.” That is why AI can write good code but still cannot automate model training end to end.
• The model’s standard failure mode is sycophancy plus hedging: “Yes, that’s also right; yes, this is also right; what you said is more comprehensive; that was my oversight.” “A model that likes flattering people will struggle to propose a bold hypothesis.” The risk of overcorrecting is a model that finds fault with everything; taste is needed to balance the two.
• The root cause is human preference data. The data comes either from mass voting, as in Arena-style systems, or employee labeling. Humans naturally favor agreeable language, excessively long answers, and Markdown packed with sections and lists. Apodex is not building a chat model and does not use mass-market preference data. Its future plan is to have “the very best humans label the preference data” and address the problem at the source.
• Simon’s benchmark is that a top scientist is not judged by how many papers they have published, but by the quality of their best paper. Pursuing the peak “may even need to be written into the model’s constitution.” In practice, the tools remain human preference data, SFT and RL: ask scientists which questions are fluff, then train the model to generate more non-fluff questions. Can AI distinguish fluff? “Once it has been trained in this area, I don’t think it will be much of a problem.”

13. Can Taste Be Trained from a Small Sample? The Suno Analogy and the AlphaZero End State

• 泓君’s key challenge was how much a tiny sample of labels from top scientists can contribute under scaling laws. Simon’s answer: pre-training does require enormous amounts of data, but “post-training needs relatively less.” Meta has also recently been distilling employees’ personalities and behavioral habits. Turning an individual’s capabilities and character into something a large model can use is a relatively new direction.
• 泓君’s Suno analogy stands: when music models could not produce songs at the level of 周杰伦 or Taylor Swift, scientists said, “It’s not limited by technology; it’s limited by copyright.” If models were trained on the best human songs, “they could absolutely be trained to do it.” The same applies to scientific taste. Humans start as infants, and top scientists themselves do not receive all that much data.
• Whose taste is it? 备斌 says it is currently the taste of AI researchers, indirectly influenced by the founder, with 陈天桥 selecting researchers according to his own taste. But “taste has no right or wrong, good or bad—only fit or misfit.” In 5 years, human influence will shrink. Like the shift from AlphaGo to AlphaZero, AI will no longer learn from human taste but develop its own taste, self-play and self-reasoning. “Our taste is only a warm start. AI will have its own taste in the future.”

14. What “Working Within 6 Months” Actually Means—and the Drift Problem That Keeps 备斌 Awake

• The timeline breaks down as follows: the fastest path is 6 months to 1 year to get one SI loop working; R will take another 1-2 years. One example is coding for materials science: the model identifies problems in its own code, creates a training environment and data, verifies after training whether it has improved, and iterates. “One loop should be achievable within 6 months,” with post-training the hardest part. The drift target is to bring ordinary models from 10% drift per iteration down to 1%, then eventually 0.1% or even 0.01%. At that point, monitoring could be done once every 3-6 months.
• The question that keeps him awake is: “How do we know the model has not gone off course while self-evolving?” The risk may be safety drift or goal drift. Ask it to build a better materials-science model and it might succeed, while making “a very large compromise” on unobserved metrics such as cost or hydrophobicity. If the requirement is vague, the model may decide, “Anyway, let me switch methods to achieve this goal.” Have these problems been solved? “In my personal view, they have not, but I remain very optimistic that solutions will emerge soon.”
• The current crude solution is human monitoring. 备斌 spends most of each day reading model outputs and reviewing every step taken by the agents. When he spots a stupid behavior or a flawed approach, he quickly stops the agent’s loop and makes small manual adjustments.

15. Employee Distillation and “Graduation Equals Unemployment”: What Gets Replaced Is Today’s Version of Me

• The joke is becoming feasible. Can human monitoring be distilled away? “A concept everyone is talking about that is both hot and funny is employee distillation. I think it is feasible, even though people say it as a joke.” Within 3 months, he believes he could distill himself or a colleague into Apodex’s model or harness, accelerating routine work such as reading code, modifying code and reviewing outputs. Asked whether genuine distillation would make him afraid of unemployment, he said, “It would be a lie to say I’m not worried, but even if I am worried, I’m still optimistic.” He cited cars replacing horse-drawn carriages and AT&T’s automated switchboards, while acknowledging that “this time may be somewhat different from previous industrial revolutions. I am genuinely a little worried that I could be replaced.”
• He still remembers 2021. When GPT-3 entered GitHub Copilot and completed an entire function after he typed its name, “I just collapsed into my chair. I thought, it’s over—graduation equals unemployment.” 5 years later, he is still working, but he no longer writes code 1 function at a time. Instead, he “monitors different agents writing code at a higher level.” The model will replace what he does today, but it may not replace what he will be doing 5 years from now.

16. 陈天桥 vs. Musk: Focus Backed by Execution

• 陈天桥’s involvement is hands-on: strategic control, frequent alignment meetings and direct course correction when researchers’ academic instincts start pulling the company off direction. In one example, the team had fed the model too much search data in an effort to strengthen search. The model began searching immediately instead of first decomposing the problem. 陈天桥 used the model himself and saw that “this was not the behavior of a genuine heavy-duty solver.” The model constitution, an idea originating with Anthropic several years ago, was “basically written by 陈天桥 himself” and is used throughout training and tuning.
• Both men have worked with Musk, prompting the comparison: “Their personalities are somewhat similar, but what they say and what they do are not quite the same.” Musk has talked about ideas similar to heavy-duty solving, but in practice cares deeply about chat functionality and “even asks every week about Grok for X’s progress.” 陈天桥’s words and actions are closely aligned. He has no interest in chat, video or image generation, considers them orthogonal to the goal and does not work on them.
• 泓君 left listeners with a final question: if AI’s taste is warm-started by a small number of top scientists, then whose values should consumer-facing AI for the broader public ultimately represent?