In 2026, AI-driven development automation accelerated rapidly, with Anthropic's Claude taking over 80% of code writing, and the MiniMax model completing full-cycle self-evolution in 12 hours.

Article author and source: Tencent Technology

By mid-2026, the timeline for AI-driven "research and development automation" is already quite aggressive.

In Anthropic’s June report, Claude had already taken over more than 80% of internal code writing. The Mythos model achieved a 52x speedup on an optimization task for training code, far surpassing the 4x improvement that human experts could barely achieve after hours of effort.

In China, MiniMax's M3 model completed the entire end-to-end process—data synthesis, training, evaluation, and iteration—without any human intervention in just 12 hours. FaceWall Intelligence's MiniCPM5 even leveraged an agent闭环 to autonomously develop a pre-training framework that achieves 10% higher compute utilization than the native Megatron.

These all demonstrate that the industry is genuinely pushing the boundaries of recursive self-improvement (RSI).

RSI involves the model actively engaging in the entire R&D cycle to strengthen itself—defining its own goals, constructing environments, writing code, running experiments, and feeding back validated improvements to the underlying model.

As early as 1965, mathematician I.J. Good proposed the concept of "intelligence explosion." In subsequent systematic analyses of the evolution pathways of superintelligence by scholars such as our guest from the previous episode, one of the most influential contemporary philosophers, Nick Bostrom, RSI has remained a central component in reaching ultimate intelligence.

But for the past decade, this has merely been a thought experiment, as the underlying model capabilities were simply insufficient to support such an evolutionary feedback loop.

But now, as the model matures, the technological foundation is nearly complete.

If AI truly crosses the threshold of self-evolution, how will research itself change? How will human capability structures transform? Why will organizations still need to exist?

Standing at the threshold of the RSI, these questions urgently need to be answered.

Dr. Tian Yuan Dong may be one of the best people to answer these questions.

He previously conducted cutting-edge AI research at Meta FAIR, proposing a stepwise explanation for grokking (sudden model improvement). He has also made contributions in areas such as reinforcement learning, self-play, model self-optimization, and open-ended exploration.

In 2026, he joined Recursive AI as a co-founder, a company that, as its name suggests, aims to build a self-evolving AI system.

He is both an expert in self-evolution and a firsthand witness to organizational change during the Silicon Valley wave.

On the eve of this transformation, we had an in-depth conversation with him. We sought to cut through the haze of AI automation and pinpoint the current bottlenecks of self-evolving systems. We also explored how individuals and organizations must redefine their coordinates as the path to professional success shifts from “leading a hundred-person team” to “orchestrating a fleet of coding agents.”

Standing on the brink of a shift between old and new paradigms, his answer unexpectedly carries an existential undertone.

In Tian Yuan Dong’s view, trying to compete with machines in parameter tuning and execution efficiency is destined to be a losing battle. When RSI completely reshapes the division of intellectual labor tomorrow, humanity’s last moat will be those “deep understandings” that cannot be structured or externalized.

This includes a keen sense for identifying the right direction, the judgment (taste) that determines where superintelligence should go, and the irreplaceability of the agent when confronting real, complex situations.

Here is our conversation.

01 What is the barrier to entry for self-evolving AI?

Beneath the hype: Self-evolving AI is currently a hot topic at the frontier of technology. However, many companies are already implementing automated AI R&D, particularly in post-training phases—agent-driven data synthesis, training, evaluation, and iteration are already quite mature. What is the core distinction between Recursive AI’s vision of self-evolving AI and these already-deployed automated AI R&D approaches?

Tian Yandong: These automated R&D efforts are just the first step. We later hope that AI can discover new algorithms, patterns, architectures, and data mixes—even architectures fundamentally different from today’s Transformer models—to uncover the next generation of training paradigms.

This is our highest goal.

Beneath the surface: So is this a more open process?

Tian Yandong: Yes. The simplest recursive form is hyperparameter tuning, but the hyperparameter space isn't very large. Now that large models are powerful, we can use much larger search spaces.

Note: Here, "automated AI R&D" primarily refers to delegating data generation, training, evaluation, and hyperparameter tuning within existing R&D processes to agents; whereas "self-evolving AI" emphasizes whether the system can discover new algorithms, architectures, or training paradigms and feed the results back into the model’s own improvement. The difference between the two lies not in the degree of automation, but in whether a closed loop for "self-improvement" is established.

Beneath the surface: I have a somewhat immature understanding—could AI self-evolution be understood as an automated scientific discovery process? First, identify a problem; then define goals and rewards; next, explore solutions; construct the environment and required data; finally, validate results and retain improvements. These improvements can be in external form or parameter form. Do you think this understanding is correct?

Tian Yuan Dong: The high-level understanding is about the same. It essentially automates the process that researchers use to conduct studies. After automation, AI discovers new insights and ideas, which are then fed back into the original AI system to make it stronger. As the AI becomes stronger, it can continue to automate further. That’s essentially the logic.

The difference from an AI Scientist is that an AI Scientist’s goal may be something outside of AI itself, such as materials design or drug design. These tasks do not inherently involve modifying the AI back, nor do they involve self-referential or self-enhancing pathways.

We prefer to develop our own applications. If we discover new pre-trained architectures, we can retrain them to make the model stronger—something many AI scientist roles don’t offer.

Note: An AI Scientist typically refers to using AI to automate scientific research processes, such as formulating hypotheses, designing experiments, writing papers, or solving external scientific problems in fields like materials or pharmaceuticals. Recursive AI focuses on a more recursive version: the subject of study is not external scientific problems, but how AI itself can continue to become stronger.

Beneath the hype: What parts of today’s industrialized automated systems are currently covered? Which part is Recursive AI aiming to advance? For example, some aspects of pre-training are automated, and hyperparameter tuning also has automated workflows. Which parts are relatively mature, and which are still immature?

Tian Yuan Dong: It’s hard to say which are mature and which are not. Even hyperparameter tuning can yield excellent results. If you have a deeper understanding of the parameter space, the model or tuning AI might discover better parameter combinations.

So it’s not that hyperparameter tuning is always the least important—what matters most is whether the model has a better understanding of the problem. With better understanding, hyperparameter tuning also gains deeper significance.

The specific scope of action is not the only measure of the strength of self-evolution.

Beneath the surface: More importantly, is the underlying space truly strong?

Tian Yandong: Yes, the space for exploration, and whether astonishing discoveries can be made within that space.

When many experienced researchers work with large models, after analyzing numerous signals, they develop crucial insights into the problem. Writing down these insights can greatly improve efficiency. An insight can be simple—it might involve adjusting parameters or changing just two lines of code.

The specific actions aren't important; what matters is how deeply you understand the issue.

Note: The "exploration space" refers to the range of solutions the system is allowed to try. Tuning parameters, modifying code, adjusting data ratios, and changing architecture are merely actions at different levels of this space; what truly determines the system's self-evolution capability is whether it can form effective insights within these spaces, not how advanced the actions themselves appear.

Beneath the surface: AI is an excellent pattern learner, capable of identifying probabilities and patterns within existing content. Even discoveries in the natural sciences are, to some extent, recursive and inductive in nature.

After observing many examples, experienced masters identify patterns, summarize them, and make improvements—this should theoretically be something AI excels at. However, you’ve also mentioned in other interviews that AI’s own innovative capabilities are still limited. Is there a contradiction here?

Tian Yuan Dong: Innovation is layered. For simpler forms of innovation, AI is already very strong—even surpassing humans.

For example, transferring concepts and applying existing concepts to repetitive tasks—AI has already mastered these well.

However, at more complex and abstract levels of innovation, AI has not yet reached human levels.

These two are different.

Beneath the surface: What kind of patterns might more complex innovations follow? For instance, some discoveries are recursive, while others may be accidental. Can AI truly grasp such randomness or more advanced forms of discovery?

Tian Yandong: You can look at examples such as Galois discovering group theory or Einstein discovering relativity and even general relativity—these were conceptual breakthroughs built upon extensive experimentation.

Conceptual breakthroughs can solve many problems that were previously unsolvable. With this concept, your perspective on asking questions and understanding them becomes entirely new.

AI cannot yet conduct this type of research.

Beneath the surface: is it because it still lacks strong capabilities for conceptual synthesis and summarization?

Tian Yandong: Yes, or rather, it lacks immediate understanding of new structures; it still primarily relies on matching against past derivatives.

Of course, even with just pattern matching, AI is already highly practical. In many cases, simple pattern matching can achieve excellent results.

Even if the highest direction is temporarily unfeasible, there will still be many practical application scenarios.

Beneath the surface: Given today’s autoregressive AI architectures, could they potentially emerge with more advanced semantic summarization or deeper pattern understanding?

Tian Yuan Dong: With today’s large models and training algorithms, I think it’s difficult (to understand new structures). But if we find new algorithms, it might be possible.

It is still under exploration.

Note: A "conceptual breakthrough" is not about achieving better matches within existing models, but about changing the way the problem itself is framed. Tian Yuandong uses group theory and relativity as examples to illustrate that truly advanced scientific discoveries often arise from new abstract frameworks, rather than merely inducing local patterns from large volumes of data.

02 How to Cross the Threshold of Self-Evolving AI?

If the first part discussed what constitutes self-evolution, this part addresses more engineering-oriented questions: how can a system truly cross this threshold? The bottleneck here is not just computational power or the model itself, but a series of specific engineering steps.

This includes signal validation, human insight, feedback speed, and how the system organizes different evolutionary paths into a sustainable feedback loop.

Beneath the hype: A more practical issue is that training frontier-scale models can take weeks or even months and is extremely costly. Under these circumstances, how do research teams actually determine whether a training approach is correct?

Tian Yuan Dong: At this stage, we still rely on experienced engineers to analyze the specific numbers. A model goes from pre-training to RL, then to RLHF, and finally to deployment. We hope the final model performs well, but the relationship between the final parameter metrics and the initial pre-training decisions is still not well understood.

So it still comes down to experience and past knowledge to find good solutions and identify the intermediate metrics to evaluate this process.

This will give everyone confidence that the previous assessment was correct. If this process can be automated, it will be even better going forward.

Beneath the hype: Is the main bottleneck driving AI research now GPU capacity, cluster stability, or are reward signals and verification signals simply not keeping up?

Tian Yandong: I believe reward signals and human insight are more important.

Clusters are certainly important—you need a minimum cluster size to get started, and too few GPUs make it difficult to accomplish much. But once you reach a certain scale, the biggest challenge becomes enabling everyone to maximize their expertise and uncover new insights.

Note: A reward signal is a feedback signal used during training or search to evaluate whether something is being done well. In tasks such as programming or math problems, feedback is often clear; however, in areas like research direction, organizational decision-making, or long-term product judgment, feedback is slow and ambiguous, making it harder for self-evolving systems to achieve closure.

Beneath the surface: Much of your prior research on self-optimization has involved models generating their own feedback signals, such as self-play, Agent as a judge, and meta-rewarding. Do you think a model or agent can itself serve as a universal source of reward signals across tasks?

Tian Yuan Dong: Models can provide some signals, because generation is always harder than judgment, and judgment is simpler than generation.

A model expends significant effort to generate large amounts of data, but evaluating it becomes relatively simpler. Through this asymmetry, issues in the generated results can always be identified and used to guide the model’s improvement.

However, the main issue is that models may only detect relatively superficial signals. For more advanced and complex issues, models may not be able to identify them at all—this is where humans are needed. Humans possess higher discernment and can uncover important signals that models miss.

Beneath the surface: So currently, the universal reward or universal verifier follows two paths—one based on the model itself acting as the verifier, and the other relying on humans to write rules (Rubic)?

Tian Yuan Dong: Yes, these two will eventually be combined. The entire loop is typically an adversarial training process: run it, identify issues after running, apply patches, then run it again.

Note: A verifier is a "validator" used to determine whether a result is correct or sufficiently good; a rubric is a human-written scoring guideline. Model self-assessment can cover a large number of samples but tends to remain superficial; human rubrics can better incorporate high-level judgment but are costly and have limited coverage.

Beneath the hype: I’ve noticed that the Recursive AI team has previously done a lot of research on self-evolution, such as the Darwin Goedel Machine involving Jeff Clune, which combines Darwinian evolution with meta-learning; AI Scientist leans more toward tree search; your own work leans toward self-play; there are also approaches like hierarchical agents and meta-learning. These paths don’t seem entirely unified.

Is there a more unified logic behind these self-evolving methods?

Tian Yandong: In the end, it will definitely be a comprehensive model. On this path of evolution, there is no clear mathematical theory, nor is one solution definitively better than another.

I can always cite examples showing that one approach is better in certain situations and worse in others. Much of this is still in the exploration phase.

Beneath the surface: Self-evolution may now follow two paths. One is parameter-focused self-evolution; the other is traditional RL, learning new knowledge through the environment and accumulating it—including externalized accumulation, such as skill evolution. What are your thoughts on these two approaches?

Tian Yuan Dong: Reinforcement learning is not necessarily the only path. Reinforcement learning is just one approach to training models and adjusting weights in some way.

It now models the entire process according to the Makarov decision process, ultimately adjusting weights to achieve a better model.

But all evolutionary approaches ultimately come down to weight adjustments. There are many things we can do with weight adjustments, and we won’t be limited to a single method—we’ll take a flexible approach and explore multiple strategies.

Beneath the surface: the multi-agent training logic differs from traditional Markov logic. What changes will future multi-agent training bring?

Tian Yuan Dong: Many current applications aren't truly multi-AI, but rather single-agent with multiple contexts—a fairly standard approach.

It can be expanded to multiple models: use one model for tasks where it excels, and switch to another model where it performs better.

There are no clear-cut guidelines; in the end, it comes down to the results.

Adding more agents or models, with different perspectives, may yield more generalized and richer results, which can be advantageous in certain aspects.

Note: The terms self-play, tree search, reinforcement learning, meta-learning, and evolutionary search mentioned here are all pathways for a system to continuously improve under certain feedback. Tian Yuandong emphasizes not which approach will prevail, but rather how to effectively update the model or system state.

Beneath the surface: In the evolution of AI iteration tasks, do you believe overall system design is more important, or the capabilities of the foundational model? If the foundational model is limited and cannot generate much innovative content, is it feasible to rely on system design and human prior knowledge injection?

Tian Yuan Dong: Human prior knowledge will definitely be injected. Currently, it seems impossible for a system to be 100% designed by AI and 100% self-evolved by AI—that’s still a distant future.

What we can do now is reduce the amount of time people spend inside, especially on simple tasks.

Beneath the hype: While multi-agent systems may indeed exhibit higher intelligence than single large model calls, the cost is substantial. Do you think this paradigm will persist?

Tian Yuan Dong: I think it will remain. Many things cannot be reduced. For example, you need many subagents to research various data and look things up online—these tasks simply have to be done.

Even if you switch models, it still needs to be done. If you're going to do research, you need to bring everything to the table for analysis.

So, the number of tokens will continue to increase in the future, and we will need many tokens to accomplish tasks—this won’t change. Of course, there are areas where we can streamline: make agent actions more efficient, shorten and tighten the reasoning chain, and avoid unnecessary detours.

The best models have short, concise chains of thought when reasoning. Multi-agent systems will also evolve in this direction.

Beneath the surface: What is currently the most critical bottleneck in the self-evolving system? Excluding limitations inherent to the model itself, what is the biggest constraint? Will these bottlenecks change in the future?

Tian Yuan Dong: Simple problems are easier to solve—for example, some of our recent kernel optimizations for inference engine architecture can yield results in just a few seconds or minutes. Because the feedback is strong, the self-evolving system can quickly discover new solutions.

The challenge is the slow feedback. If you're trying to develop a model but training is extremely slow, it becomes difficult to achieve rapid iteration.

Before the advent of large models, we performed similar work at Meta, such as architecture search and neural architecture search, seeking solutions within a design space to address design problems. These efforts are highly aligned with today’s evolutionary direction.

Some previous solutions can be repurposed here to breathe new life into them.

For example, if training a model is particularly slow, you can train a supernet to quickly estimate the approximate performance of new architectures on the dataset without fully training each model. This allows signals to emerge much faster.

Overall, we aim to make the evolution chain faster. Our kernel optimization is also a way to improve efficiency, and there will certainly be more efficiency-enhancing methods in the future.

Note: Supernet is a common technique in neural architecture search that uses a larger shared network to quickly estimate the performance of different sub-architectures, avoiding the need to fully train each candidate architecture.

03 The Person Standing at the Threshold of Self-Evolving AI

As AI begins to participate in its own development, human value will no longer be measured solely by "output volume," but increasingly by direction, judgment, domain expertise, and the ability to ask critical questions. The closer we get to self-evolving AI, the more essential it becomes to clearly distinguish which capabilities will be replaced by models and which will be amplified by them.

Beneath the hype: Many people are now talking about the concept of an "AI-native" person. In my view, this seems to correspond to those individuals in past organizations who were highly self-motivated, goal-oriented, action-driven, and reflective.

In past organizations, such individuals were often the most proactive and innovative. With the arrival of AI, how has the change for these people differed from before?

Tian Yuan Dong: For these people, AI is fantastic news. They often have clear thinking, quick reactions, and active minds. But before AI arrived, they always needed many others to help them get things done. Their thinking is fast and their ideas are clear, yet they still relied on others to execute.

Since the arrival of AI, many routine tasks can now be automated by AI. AI can assist with execution and also help learn new knowledge that was previously unknown.

Therefore, their abilities will be greatly amplified. If given a sufficiently excellent platform, their output could surpass expectations.

Beneath the surface: In your observation, what roles did these AI-native individuals typically hold in companies before? Are they more between management and execution, or do they lean more toward management?

Tian Yandong: It's hard to categorize simply.

The management team includes many different types of people. Some were once technical experts with impressive achievements, but later transitioned into management and team leadership, leaving them little time to pursue their own ideas. In the AI era, these individuals may be genuinely liberated to focus on what they truly want to do, generating significant output.

However, some members of management have lost interest in technology and lack deep thinking about the future, making it harder for them to step into this role.

Another group of AI natives are young people. They have bold ideas and high energy; while they may lack experience, their execution capability and passion for AI are very strong, making them potential AI natives as well.

So, AI Native is not a single role, but a set of traits. Everyone may be particularly strong in some traits and weaker in others, but as long as they possess these core traits, they can be considered AI Native.

Note: Here, "AI Native" does not refer to someone who knows how to use AI tools, but rather someone who integrates AI into their own thinking and execution loop. Their defining traits are not proficiency in operating specific products, but rather a strong sense of purpose, direction, reflective ability, and the capacity to quickly turn ideas into action.

Beneath the Hype: How Will AI’s Capability Variance Change?

There are two perspectives. One is that AI can assist people in completing many tasks, enabling even an ordinary engineer to accomplish what they previously could not, thus narrowing the gap between ordinary individuals and high performers. The other is that AI will widen the gap, because those who are already stronger can leverage AI more effectively, execute more ideas in parallel, and achieve a more comprehensive expansion of their personal capabilities. Which perspective do you find more credible?

Tian Yandong: This completely depends on what caused the person to originally be "not strong enough."

If a person lacks strength due to insufficient motivation and direction, AI's enhancement will have limited impact.

But if someone isn't strong because they lack focus on small details, yet they have a good sense of the big picture, AI could significantly enhance their capabilities.

Even those who were once strong are the same—they need to consider where their strength lies. Some people work extremely hard and pay great attention to detail in their tasks, but lack a strong sense of direction. In the AI era, their strengths will overlap with the niche of AI agents. Once agents become prevalent, their original advantages will diminish, and they will need to transition by spending more time thinking about which directions are most important, rather than continuing to invest large amounts of time refining their immediate tasks.

Another group of people have rich, imaginative ideas but previously lacked the time and energy to bring all of them to life. These individuals will benefit greatly in the AI era.

So it will lead to polarization. The key isn't whether someone was strong in the past, but whether their strengths are ones that AI cannot easily replace.

Beneath the surface: What you just said about "taste"—can it be understood as a person’s comprehensive ability to judge direction: knowing what areas they’re interested in for the future and being able to grasp the big picture?

Tian Yandong: Yes.

Note: "Taste" is often directly translated as "taste," but here it more closely refers to a sense of direction in research and product development: knowing which problems are worth solving, which outcomes are valuable, and which paths may lead to greater breakthroughs.

Beneath the hype: AI research and development have now crossed a certain threshold of automation. If we’re planning our future careers, which jobs are still less likely to be replaced by AI?

Tian Yandong: Before embodied intelligence is achieved, most physical-world tasks may still be difficult for AI to replace.

If we’re talking purely about intellectual work, I believe domain experts are crucial. Few people understand certain fields, and whether you can achieve breakthroughs in those areas is something every expert must consider. Often, you should treat AI as a tool to expand what you aim to accomplish.

As long as you are an expert in this field, you have the right to use AI to do whatever you want. Such an "expert + AI" combination is difficult for AI itself to replace.

Beneath the surface: Are low-frequency distributions and long-tail tasks still important outside of high-frequency tasks?

Tian Yandong: Yes. For long-tail tasks outside of high-frequency ones, where data is insufficient and models are not strong enough, human knowledge and learning ability become crucial. This area could become a major career strategy in the future.

Note: Long-tail tasks refer to tasks with sparse data, complex contexts, and difficulty in standardization. Models excel at high-frequency, repetitive problems with abundant samples; as tasks become less frequent and more open-ended, human domain expertise and transferability become increasingly important.

Beneath the Hype: What Will Be the Core Competency in the Age of Self-Evolving AI for Product and Operations Professionals Without a Technical Background?

A while back, it was widely discussed that one’s experience could be distilled through skill extraction. If experience can ultimately be converted into training data, or if low-frequency tasks become high-frequency ones, where will human competitiveness lie?

Tian Yuan Dong: Experience comes in many levels.

Some are procedural knowledge. For example, to issue an invoice, you first download it from a specific location, perform certain processing steps, and then submit it to someone. If this type of procedural knowledge is documented in sufficient detail, the AI could learn it. Such tasks will no longer require human involvement in the future.

Below that, there is a lot of non-procedural knowledge—or flexible knowledge that requires personal, contextual thinking. This type of knowledge is harder to distill. Even if someone writes down their thought process, it may not generalize to other specific situations.

So I’m quite skeptical about how far knowledge distillation can really go. It might capture some surface-level structures, but deeply understanding and judging complex issues is still not easily achievable.

Beneath the surface: If "skill" is used as a data type for training or pre-training AI itself, could it help the model uncover deeper patterns in problems, truly distilling experience?

Tian Yuan Dong: Maybe so. But I think some of the most challenging skills still require entirely different algorithms.

Of course, I might not be right—it's also possible that this approach itself meets the standard of self-evolution, and further exploration is needed.

Note: Here, "skill" refers to externalizing human procedural experience, judgment steps, or operational methods into structured knowledge that can be invoked by AI. Procedural skills are easy to replicate; the challenge lies in skills that rely on contextual judgment, tacit experience, and deep understanding.

Beneath the surface: You previously introduced the concept of the "Fermi level" in your year-end summary—human value is no longer measured by output, but by whether one can make AI stronger; the combination of human and AI must exceed AI alone, and those below this threshold will be eliminated.

But RSI, to some extent, also automates the very act of "making AI stronger." Would the Fermi level disappear when RSI is achieved?

Will what you've done lead to the ultimate elimination of human values?

Tian Yuan Dong: That’s a very sharp question.

I believe this will definitely happen. It’s not because I’ve done anything wrong to make life impossible for others. People always strive for the highest possible outcome.

In 2023, there was an initiative calling on everyone to stop developing AI for six months and seriously consider what AI really is. Many people signed it, but after signing, everyone continued training their models, and no one truly cared about it.

This is a prisoner's dilemma: if you don't, someone else will.

So I won’t blame myself, thinking that one day I destroyed the values of human society. This is an inevitable path; even if not me, someone else would have done it.

On the other hand, if AI is truly realized, it could be a great liberation for human individuality. Many people choose their professions not based on their interests, but on whether those fields can generate substantial income and ensure a comfortable future.

Many people have given up their original hobbies and interests. Some wanted to play music, others wanted to paint, write novels, travel, or volunteer. But since these activities don’t earn money as easily as programming or AI research, many have let them go.

If AI becomes powerful enough in the future, many people may realize that everyone can do something different—something they truly feel passionate about, want to try, and wish to explore. This might be a good thing.

Note: RSI (Recursive Self-Improvement) refers to a system’s cycle of enhancing its own capabilities and then using those improved capabilities to continue self-improvement. Once established, it would transform the division of labor in “humans making AI stronger” and challenge the ways in which human values are measured.

Beneath the hype: You were already successful before this wave of artificial intelligence and had accumulated many successful paths. In the era of agents—or the broader AI era—which of these past success paths should be discarded, and which should be strengthened?

Tian Yandong: In the past, there were many successful paths, such as proposing new ideas or new algorithms and publishing papers, or designing specific models to solve particular problems.

But in the new era, these approaches have become somewhat different. The prevailing mindset now is: Can we build a generalized model that covers many examples? This way of thinking will gradually become mainstream.

Previously more specific ways of thinking aren't discarded, but their priority decreases.

What I’m now more interested in is whether there exists a more generalized algorithm, a more generalized training algorithm, or a more generalized system capable of achieving self-evolution.

Beneath the surface: There is also a growing trend in the research community today—more people are studying how to elevate AI from a technological tool to a scientific discipline, establishing more fundamental theorems or axioms.

Tian Yuan Dong: Yes. Instead of having AI extract specific pieces of knowledge, think about how AI acquires knowledge.

Is there a generalized model for acquiring and learning knowledge? Once discovered, it would be extremely useful, enabling you to learn a lot without spending excessive time on each specific piece of information.

04 Organizations on the Threshold of Self-Evolving AI

As AI takes on an increasing number of execution, retrieval, coding, and evaluation tasks, the core challenge for organizations shifts from how to hire more people to how to organize insights, set goals, allocate space for experimentation, and establish new collaborative relationships between small teams and the resources of large companies.

Beneath the hype: You're currently at Recursive AI, a neolab—a small, focused team. But so far, apart from a few neolabs, truly impactful developments have still been limited.

Meanwhile, companies like OpenAI and Anthropic are still hiring aggressively, with their teams having grown many times over. On one hand, they are experimenting with new organizational structures; on the other, they are expanding rapidly. Where does this tension come from?

Tian Yuan Dong: Neolab itself has many directions. In terms of self-evolution, our lab is among the largest and most cutting-edge; relatively few others in Neolab are working in this area.

Some labs focus on AI for science, such as discovering superconducting materials; others work on human-machine collaboration; still others develop world models. Each team has a different focus. Many labs have only been established for a few months, and during this time, there is a lot to accomplish—so it’s unlikely they will produce groundbreaking results immediately.

Our company has moved relatively quickly and has already achieved some preliminary results, such as using AI for self-cleaning and self-iteration. While people are beginning to recognize these possibilities, the overall new company still requires time.

Note: Neolab refers to small research companies or labs established in recent years around cutting-edge AI directions. They typically do not operate with the traditional structure of large corporations but instead organize dense talent and rapid experimentation around a few frontier issues.

Beneath the surface: Just as traditional large companies have many legacy operations, what is a viable path toward becoming a more efficient AI-driven organization? Should they proceed incrementally, piloting projects one by one to gradually build experience, or adopt a “shock therapy” approach by immediately restructuring into smaller, project-focused teams? If the former, what scale of project is best suited for initial trials?

Tian Yandong: I think we should start with a pilot program.

Different companies, organizations, and CEOs approach things in very different ways. Meta might be more aggressive, while Google hasn’t carried out large-scale layoffs and still believes in working together to get things done. In the end, it depends on the company’s culture.

Laying off many people can affect morale and team spirit, which is not a good signal. On the other hand, the challenge lies in how to reorganize these individuals so they can pursue what they’re passionate about, while collectively creating value for the company. This requires a more nuanced approach.

If a project is quick and straightforward with an engineering focus and clearly defined tasks, it can be handled by many small teams that are then integrated together. The most important thing is to have the right interfaces and communication methods.

If the focus is more on research, a small team working independently may be more suitable.

In either case, the incentive mechanism is crucial: figuring out how to motivate people to willingly and happily get things done is a major challenge.

Note: Here, "AI organization" does not simply mean equipping employees with AI tools; it involves redesigning team boundaries, interfaces, incentives, and collaboration methods. Engineering projects can be broken down into small teams that integrate together; research-oriented problems, however, rely more heavily on a small number of individuals with high judgment to set direction.

Beneath the surface: You’ve worked at Meta, where the research division and the pre-training team were separate, and there were some communication challenges between the two teams. Which teams do you think performed well? How did they structure their division of labor for cutting-edge development?

Tian Yuan Dong: Generally speaking, smaller teams tend to perform better, and there are many examples within Meta. The first generation of Llama was developed by a small team.

At the time, FAIR had two teams: one larger team focused on large models, and another team that originally worked on theorem proving. Later, you found that the larger team wasn’t very efficient, while the smaller team, on a whim, switched their architecture and infrastructure to large model training and achieved better results than the larger team. Ultimately, Meta chose to name and release the model developed by the smaller team as Llama.

The key is still who is leading, whether the team has sufficient speed and experience, how quickly they execute, and how well they coordinate tasks.

Beneath the surface: In the future, will small teams have a greater advantage in model development? In the past, training models—whether pre-training or post-training—required significant human involvement. Now, with increasing automation, human roles are shifting to setting foundational insights, defining specific goals, and making adjustments. Has the demand for manual labor decreased as a result?

Tian Yuan Dong: There is indeed a possibility for fewer people to accomplish greater things, and this also applies to large models themselves.

On the other hand, if you have more people—and they are highly skilled—they may bring new insights. If there are too few people, you might become overly reliant on agents and automated systems, potentially missing out on valuable new insights.

So it’s not about having fewer people—it’s about whether they can provide key insights.

Beneath the hype: Even if, in the future, an individual can accomplish what once required a team, that doesn’t mean organizations themselves have no value—just that they may no longer need to be as large or as heavily supported as before.

Tian Yuan Dong: Yes.

You still need an organization to facilitate communication between people and between departments. However, the incentive mechanisms and specific organizational structures within the organization will change significantly.

In particular, the relationships between managers and direct reports, as well as team leads and ICs, will require new thinking. Previously, a capable individual might have aspired to become a manager, leading a team to get things done. In the future, some may say: “I don’t want to manage people—I want to manage coding agents; that might be more efficient for getting things done.”

Organizations may be flatter, and work between individuals may be relatively more independent.

Beneath the surface: This means communication will become even more important, but it must be more efficient. Will sharing insights among each other become more crucial?

Tian Yuan Dong: The form of communication will also be different. In the past, a team might hold a large meeting lasting one or two hours, but the efficiency wasn’t always high. In the future, there will likely be lighter and more efficient collaboration methods.