What exactly is the RSI that VCs are scrambling to invest in? Yao Shunyu, Shi Tianlin and professors from Tsinghua University and Shanghai Jiao Tong University have given the answer.

By | Wang Xinyi

Edited by | Zhang Yuxin

RSI (Recursive Self-Improvement) has become an absolute sensation.

Last Saturday, a group of the brightest minds gathered at West Bund in Shanghai, along with people dedicated to equipping AI with the same level of intelligent capabilities as human brains.

On the other side of the ocean, Yao Shunyu, who had not appeared in public for several months and claimed that Gemini 3.8 Flash was "a major step forward for RSI", was waiting on the other end of the phone; Shi Tianlin, co-founder of Recursive Superintelligence, a Silicon Valley AI self-evolution startup founded half a year ago with a valuation of 4.65 billion US dollars, also joined the discussion.

They gathered for a shared topic: What is happening in the RSI track that is seeing massive capital inflows?

RSI, short for Recursive Self-Improvement, has become the hottest topic since the middle of this year. It refers to the ability of AI to improve itself proactively, ranging from writing code, running experiments, adjusting architectures, optimizing weights, to even setting new goals for itself.

The vast majority of attendees were researchers from major tech companies, universities and laboratories, and many representatives from well-known investment institutions, FAs and incubators were also present. We also participated in this five-hour event in full, where a number of imaginative ideas and controversies kept popping up throughout the venue.

Everyone is envisioning an end state where humans do not need to be in the loop, that is, humans do not need to intervene in the model's self-improvement cycle, and are imagining how major tech companies, laboratories and startups can explore and achieve this end state.

Divergent voices are also emerging: Is self-evolution a pessimistic future? Will it eventually be swallowed up by model vendors? RSI should not be a vertical model without pre-training...

Regarding questions such as how to define, evaluate and verify RSI, and when it will be realized, we have compiled views from Silicon Valley researchers including Yao Shunyu and Shi Tianlin, as well as assistant professors and researchers from universities such as Tsinghua University and Shanghai Jiao Tong University for your reference:

1. RSI will happen when AI companies scale up their own capabilities (such as model capabilities, customer scale, etc.) without needing to scale up the size of their teams at the same time. 

2. RSI is not a single event, but a continuous process. 

3. Judging from some signs at the beginning of this year, RSI is achievable. A major reason is that both coding and Agent Model have matured and reached a breakthrough point. Following the approach of Coding Agent, AI can be allowed to train AI, realize AI's own ideas and improve AI itself, including the architecture, dataset and its own training methods, so as to promote the realization of RSI. 

4. It is hard to tell whether the difference between RSI and the model's fine-tuning on self-generated data is qualitative or quantitative. In a sense, the model generating data by itself is the infant form of RSI. 

5. Data and environment are high-potential business tracks for RSI implementation. 

6. When competing with large manufacturers such as OpenAI and Anthropic, entrepreneurs' opportunities do not lie in general intelligence, but in professional knowledge beyond basic general knowledge. 

7. There is no need to train RSI models from scratch, as their basic performance can hardly exceed that of general large models, nor can they iteratively develop stronger capabilities on their own. 

8. When evaluating intelligence, we should focus on how it acquires a certain capability, rather than what capability it has acquired. 

9. The ultimate form of Verification should not rely on human labor, but in the short term, human participation must be retained, especially in task definition and quality control, where high-quality evaluation by human experts is still required. 

10. The final version of RSI may be a model that can improve itself; but at the current preliminary stage, it can rely on external auxiliary models and collect feedback signals. 

11. The RSI most people picture in their minds may be the final state of RSI - where no human is needed in the entire loop. But in practice, it remains uncertain whether we will reach that day and whether it is necessary to truly pursue that end state. 

The following is the excerpted key content of the co-hosted event themed "When AI Begins Recursive Self-Improvement" by AGI House, HSG and Research AI+:

Before that, let's introduce several guests first —

• Yao Shunyu: Principal Staff Research Scientist at Google DeepMind, focusing on the exploration of Auto-Research direction, graduated from Stanford University with a PhD in theoretical physics, and previously worked at Anthropic.
• Shi Tianlin: Co-founder of Recursive Superintelligence, alumnus of the Yao Class at Tsinghua University, previously worked as a researcher at OpenAI, and dropped out of his PhD program at Stanford University to co-found an AI startup called Cresta.
• Gu Yu: Co-founder of NeoCognition, first author of ApprenticeBench
• Zhou Xuanhe: Assistant Professor at Shanghai Jiao Tong University, Founder of Theseus Labs
• Meng Fanqing: Co-founder of Evolvent AI, first author of RSI Bench Data
• Liu Ziming: Assistant Professor at Tsinghua University, Founder of MetaRing Intelligence, focusing on the AI for AI direction.
• Mu Yao: Assistant Professor at Shanghai Jiao Tong University, Founder of SeeAct AI, working on the RSI direction for embodied intelligence.

In addition, other guests include researchers and PhD students from other universities and laboratories at home and abroad. It is worth mentioning that the organizer AGI House is active in Silicon Valley as an AI startup community and incubator, and this is their first event held in China.

Now let's get to the main topic.

How to Recurse? How to Improve?

Q: How to define RSI?

Yao Shunyu: RSI may not be a single event, but a continuous process. Right now we can only discuss where we are on this continuous timeline.

To give a simple example, models have been able to write code by themselves for several years, and this capability has actually accelerated the research loop, which can be regarded as part of RSI taking place; later models can actively monitor experimental processes and propose new experiments, which can also be counted as part of RSI.

Effectively speaking, the RSI most people picture in their minds may be the final state of RSI, where no human is needed in the entire loop. But it remains uncertain whether we will reach that day and whether it is necessary to truly pursue that end state.

Shi Tianlin: From the perspective of RSI, the model itself must have awareness of its own constraints. It needs to know where its shortcomings lie, and be able to correct itself when it clearly deviates from the expected path. This is an important capability for RSI itself and the model's self-reinforcement.

Of course, the model alone is not enough, and certain harnesses and application scenarios are also needed to constrain the model.

Gu Yu: RSI is a very broad topic, and what several of us present are doing today may be completely different.

Our shared pursuits can be roughly divided into three categories: first, AI finds improvement goals on its own; second, AI keeps getting closer to this goal through continuous iteration; third, the goals correspond to very long-term tasks. For example, when a large model is tasked to build a new large model, its goal may be to obtain a better model, where AI needs to set up budgets on its own, plan the model, summarize methods, and finally get an updated model.

Meng Fanqing: Now we often mention Self-improving, continual learning and RSI together.

From my point of view, RSI means that under a given environment, reward or reward function, the model keeps iterating on a certain target. The target can be various things, such as architecture and weights, and its ultimate goal is to get higher reward scores in this environment.

In the true sense of RSI, the model needs to optimize itself and improve its own performance, rather than optimizing a certain product or another model. The entire system has no external factors other than itself, only itself, the environment and the reward.

Zhou Xuanhe: From the perspective of computer science, the word "Recursive" should be traced back to its most essential level. For an AI system, the most essential level is its parameters and architecture design; one level outwards is various codes and scripts in the Harness; further outwards is the environment it interacts with, including various files and workspaces.

Our definition of RSI is to first apply it to industrial scenarios, find out if there are pain points in the current workflow and where the pain points are. After identifying the pain points, the next step it needs to take is to retrieve solutions.

The final version of RSI may be a model that can improve itself; but at the current preliminary stage, it can rely on external auxiliary models and collect feedback signals.

Has RSI Truly Achieved Self-Evolution?

Q: How to evaluate that RSI has truly achieved self-improvement? Many of you here are working on Benchmark related work, please share the evaluation methods you are developing.

Gu Yu: There are two key points about what Evaluate focuses on and why it is important at this point in time:

The first is ecological validity. Good performance on Benchmark does not necessarily mean that the model can really solve practical problems, and directly correspond to actual production value or economic value. What we want to verify is to turn a general large model into a mature employee that can perform tasks in a real job position. Therefore, what we need to do is to build an environment that is close to real work scenarios.

The second is, what exactly do we want to evaluate about intelligence? I have quoted a point of view on many occasions: Intelligence does not refer to how many things you know, but how you acquire those capabilities.

Zhou Xuanhe: Evaluation is a very important matter, especially for Chinese practitioners in the AI field. After a direction is determined, everyone will quickly push the ceiling of this direction to a very high level through fierce competition.

The current problem is that the existing benchmarks are not good enough, and even some well-known benchmarks have many wrong questions. What should a RSI-oriented benchmark look like? In addition to the three traditional capabilities of Agent: reasoning, exploration and planning, we also need to think about what new capabilities RSI needs. We also believe that in the future, not only AI self-iteration is needed, but high-quality benchmarks still need to combine the efforts of human experts.

Verification is the Bottleneck Restricting the Continuous Evolution of RSI

Q: How to define Verification, or can you share your thoughts on this?

Yao Shunyu: As for which part is harder for RSI to realize: self-improvement, or proving that it has really improved, I think the harder part is verifying that it has really improved.

Many people have done similar things in the past, which was called "cyber criticism" at that time - one AI sets questions and another AI solves them. This approach most easily falls into local optimum: for example, one AI always sets very simple questions that the other can answer correctly; or one always sets extremely difficult questions that the other can never answer correctly. Therefore, the hardest part is that the model must be able to truly understand to what extent it has improved. Compared with solving problems, the hardest part in RSI is definitely defining problems.

Meng Fanqing: Regardless of whether the Verification design itself is correct, even if it is completely correct, it will bring a big problem: reproducibility.

Now the environment is getting more and more complex, involving CPU models, operating systems (Mac or Windows), memory sizes, and different physical platforms will lead to the existence of performance gaps.

In addition, most current benchmarks give an absolute score between 0 and 1 for each question. But these scores do not have strong physical significance, and the absolute scores themselves do not explain much.

One of our company's collaborators proposed another set of scoring standards. It divides the levels into different layers such as Agent-SOTA and Human-SOTA, and each layer corresponds to different Artifacts. On each new platform, running these Artifacts can determine which layer the model's output belongs to. This can avoid the problems caused by the lack of physical significance and hardware gaps mentioned earlier.

Zhou Xuanhe: I saw a work from a Tsinghua University team a long time ago. In the reasoning process, they added perturbation noise to the weights, and verified that different perturbation methods can make the model perform better in reasoning. This actually realizes the capability we are pursuing: learning and adjusting weights while reasoning.

From an internal perspective, we have already designed a module that separates reasoning and memory, there is implicit intermediate state transmission from the memory embedding module, and protection can also be achieved through external means.

RSI Startups Competing for Niche Positions with Tech Giants

Q: How to think about the boundary between RSI startups and model vendors? Where do model vendors have absolute advantages, and where can startups make differentiated breakthroughs?

Meng Fanqing: The concept of RSI is very broad. Scientific research institutions, startups and tech giants are all working on different parts of it, and the structure from top to bottom is as follows:

At the most upstream, companies such as Alphabet, OpenAI, or Kimi, their definition of RSI is basically pure RSI, which means the model iterates on itself including the entire architecture, pursuing higher intelligence.

One level down, for example, some traditional SaaS companies, RSI requires environment and verification. For them, whether they can abstract their original ecosystem into a set of Agent-native services, which is convenient for upstream model vendors to call through Agents and improve the intelligence of Agents in corresponding scenarios, may be the RSI understood by these downstream traditional companies.

Further down, some startups may take the direction of not purely using RSI to iterate the basic model itself, not spending a lot of money to train their own basic model, but using RSI to iterate certain products such as Harness.

For upstream vendors and startups, the goal is relatively clear, which is either to improve the model itself, or to improve downstream performance; while for underlying SaaS companies, their attitude towards RSI may be whether they can provide more environments within the ecosystem.

Zhou Xuanhe: Although everyone's requirements for data are getting higher and higher, I believe that current data and environment are still a promising business track.

It's like trying to plug all the leaks on a leaking boat with your own hands. No matter how many hands a tech giant has, it can't cover all the leaks. Therefore, when models are applied to thousands of industries, a self-evolution paradigm is needed, the two core elements of this paradigm are data and environment.

On the data side, for small companies with insufficient budgets, they need a lightweight data solution that can be converted into users' own assets.

On the environment side, even the best Agent will see a significant drop in task execution scores when performing tasks in a noisy environment. How to fully understand and build a paradigm that co-evolves with your solutions and models is a very important matter.

Gu Yu: I think it can be viewed from two perspectives: the first perspective is that startups serve model vendors, and the second perspective is that what startups do is complementary to model vendors.

As for serving model vendors, for example, selling data and selling environments as Professor Zhou mentioned earlier, these are all very promising businesses.

As for complementing model vendors, as Professor Fanqing mentioned earlier, what model vendors provide is general intelligence, that is, the general capabilities of the model. I think most of the people who choose to start a business should not compete with tech giants on general intelligence, we should focus on specialized knowledge beyond general intelligence.

Some Controversies

Q: What views and definitions in the current industry discussions about RSI do you disagree with or find hard to understand?

Gu Yu: I can't really understand why domestic investors have invested so much money in RSI?

Liu Ziming: Self-evolution represents a very pessimistic future.

The last self-evolution process was the evolution from primates to humans, which was a very long process. Although AI seems to have made a lot of progress now, I want to say that self-evolution is an inefficient path, which means we are stuck in a low dimension.

For example, the current large model is like a whole galaxy, and we are inside the galaxy, so we think the galaxy is everything. But if you slowly expand outward along