From Distillation to RSI: What's Next in the Model Race, with 孟繁青
Summary
SI/RSI and Auto Research are, in 孟繁青’s view, the next focal points in the model race. But he does not expect a standalone “RSI foundation model” to emerge. self-evolving, Auto Search, and RSI can all be formalized as “environment + objective + feedback + continuous iteration”; what still determines RSI capability is the world model, value model, and longer context already internalized by existing foundation models. The opportunity for startups is more likely to be a bridge between model labs and applications than another attempt to build a foundation model.
Distillation is not the decisive factor in China’s catch-up. 孟繁青 believes Kimi K3 may have had access to only one-tenth of the compute available overseas, which in turn forced bottom-up innovations such as Kimi Linear and DeepSeek’s “Multi-Layer Attention” that can scale. Distillation is merely an “accelerator” in an intense competitive environment; even if the channels were cut off, catch-up would become harder but would not stop.
Post-training is increasingly becoming data creation. Once model labs package SFT and related processes as services, ordinary algorithm researchers mainly need to submit data. The truly scarce know-how lies in data correctness, difficulty, trajectories, and distributions; quality can only screen the data initially—“to properly see whether it works, you still have to train on it.”
Benchmarks are shifting from test papers to working environments. Instead of measuring only the “brain” of a model, they now simulate software such as Notion and Feishu, requiring Agents to demonstrate reasoning and operational ability on long-horizon tasks before validating results across multiple scoring points. Benchmarks are therefore becoming closer to product and commercialization platforms.
The high-price window for generic Agent synthetic data is nearing its end. A single long-horizon synthetic task can sell for several thousand, while expert problem-solving data is priced at around 1,000–2,000; as model labs build more of the capability themselves and high-quality benchmarks converge, demand for external purchases will fall. The next wave may be RSI data—for example, having one model post-train a small model while recording the full trajectory, with each sample taking more than 10 hours or even a full day to generate.
There is limited standalone room for a general-purpose Harness, but vertical domains still have an economic case. In general-purpose areas such as coding, models and model labs’ own Harnesses are more likely to be bundled together; in vertical domains, a strong model can guide a smaller open-source model through self-evolution, allowing deployment at lower cost, higher speed, and with stronger privacy. “Writing the Harness” can be delegated to the strongest model, but that does not mean the strongest model must be purchased continuously at runtime.
Over the long run, the moat may shift from chips and talent to organizational throughput. Model releases will probably continue to alternate at the top, while technology circulates across companies; as technical paths converge quickly, the ability to organize pre-training, multimodal work, and Infra determines how fast the company as a whole can move. Model intelligence is also being demonstrated beyond coding, in Auto Research and AI for Science—曲凯’s summary is “from the Olympiad to the Nobel.”
Deep dive
1. Post-training founders’ opportunity lies outside model labs—not in copying them
孟繁青 is a PhD student at NUS who previously worked on Agents and related Infra at Kimi, contributing to releases including Kimi Linear and K2.5. About half a year after founding Evolvent AI, he narrowed the company’s focus to RSI and synthetic data, with a team made up mainly of researchers born in the 2000s.
The biggest change from Kimi to a startup is that foundation-model training inside large companies has become highly engineered, leaving researchers largely focused on producing data within stable processes. At a startup, the platform must be built from scratch, and the company must find an entry point more orthogonal to the main lines of model vendors—something that real applications actually need.
Pre-training researchers start fewer companies not because of a lack of ability, but because their daily work revolves around model architectures, optimizers, and large-scale compute. Without that compute, it is difficult to reproduce the research conditions. Post-training is closer to synthetic data, Agents, and benchmarks, and therefore closer to deployment and revenue.
2. Benchmarks have evolved from “making models solve problems” into simulated work environments
孟繁青 compares old benchmarks to exams: give a model a math problem and check whether the final answer is correct. They essentially measure only the “brain” of the model. Next-generation Agent benchmarks need to build a complete work environment, allowing the model to operate tools such as Feishu and Notion like an office worker—reasoning and taking action at the same time.
The first requirement is proving that the environment is realistic enough. For example, a simulated Notion must produce the same output as the real Notion given the same input. Only when the environment’s behavior is aligned do the Agent’s successes, failures, and training trajectories inside it become meaningful; the environment itself is part of the measurement.
New tasks can no longer be judged through a binary pass/fail check based on a single answer. 孟繁青 says many current benchmarks break tasks into multiple granular scoring points. The grader, task, and environment all require careful design, and in practice humans and Agents still check the results together.
This is also where the commercial value of benchmarks comes from. A sufficiently strong Agent benchmark is not merely a question bank for a paper; it contains a runnable software environment, task system, and validation mechanism. Its engineering implementation is closer to a real product, making it easier to turn into a platform business model.
3. The training pipeline is being commoditized; data judgment is post-training’s core asset
Inside model labs, SFT has been packaged as a service: researchers submit a data request, a remote system runs the training, and the result is returned without requiring them to find GPUs, pull a GitHub repository, or configure an environment. 孟繁青 therefore acknowledges that the barrier to training models has fallen sharply for ordinary algorithm researchers, while the Infra engineers building the services still face substantial challenges.
He rejects the idea that post-training has therefore become nothing more than tedious data labor. The evolution from Transformer to verifiable-reward RL has itself been moving toward simpler, more direct, outcome-oriented forms. As algorithmic tricks converge, the essence of research is laid bare: what data should actually be given to the model?
Data is usually designed around a target benchmark. The most direct test is whether it can raise the score without cheating or hacking, while preserving other capabilities such as writing. Before training, researchers can check correctness, whether the trajectory cheats, and whether the difficulty is appropriate. If it is too easy, there is no upside; if it is too hard, the model cannot learn it. In the end, “you only know after training.”
Evolvent AI’s RSI Bench shows that Agents can already fix mechanical problems—for example, automatically adjusting when they discover that a batch of training data is causing output-format errors. They still cannot generate the kind of data-creation insight a researcher would. 孟繁青’s view is that data know-how remains in human hands, and different researchers’ data can indeed produce different model results.
4. Resource constraints in China have forced innovation in pre-training architectures
孟繁青 observes that the gap between Chinese and overseas models is narrowing. For example, Kimi K3’s release results were already close to the frontier despite possibly having access to only one-tenth of the compute available overseas. Overseas labs may still have unreleased reserves, however, making the true time gap difficult to estimate precisely.
He breaks China’s catch-up into two paths. The first is bottom-up innovation at the pre-training layer by Kimi and DeepSeek, including Kimi Linear and DeepSeek’s “Multi-Layer Attention.” Both aim to achieve better results with less compute without excessive losses in performance or inference speed—a case of resource constraints forcing innovation.
He even believes that globally, there are not many new architectures that have been validated at scale inside companies. Kimi and DeepSeek have at least demonstrated that their approaches can scale, while overseas architectures such as Mamba have not yet achieved the same degree of adoption. This is not a claim of industry consensus, but his own view after discussions with some friends.
By contrast, China’s post-training may be weaker, not because of a shortage of algorithm talent but because of a thinner data accumulation. Overseas vendors such as Claude and OpenAI may have started six months earlier. Data accumulation depends on sustained labor and organizational processes, and a time gap is harder to close than a one-off algorithmic gap. Distillation therefore became a tool for quickly filling the data gap.
5. Models will alternate at the top; real-world feel and leaderboard rank are not the same
孟繁青 expects Chinese open-source models to remain in a long-running rotation for the top one or top two spots: one company releases a major version and leads temporarily, then another company’s new version overtakes it a few months later. A single architectural innovation can create a time gap, but it is unlikely to keep one company ahead indefinitely.
He sees K3’s large parameter count as an important contribution because, at the current scale of available data, continuing to scale parameters still generates returns. Other vendors can catch up, but they need time and the corresponding Infra. The scaling path is replicable, but seeing a parameter table does not mean it can be immediately reproduced.
Outside the leaderboards, 孟繁青 says GLM has clearly led for a long time in actual user experience. He does not know the details of 智谱’s internal work, but based on conversations with peers, he believes its coding edge mainly comes from better post-training data. As for ByteDance, catching up with the current frontier is a matter of time; whether it can continue narrowing the gap with the latest frontier is harder to say.
Talent can be attracted with high compensation, data can be accumulated through investment, and chips still matter. But 孟繁青 ultimately locates the long-term difference in organizational structure. 混元 made visible progress after reorganizing, and choices such as whether to create a standalone multimodal department or merge it with pre-training can change the company’s overall iteration speed.
6. Self-evolving, Auto Search, and RSI are fundamentally the same loop
孟繁青’s unified definition is to give a model a working environment and an objective, let it operate continuously, use the results as feedback to modify its previous behavior, and then try to break through its prior ceiling. Whether it is called self-evolving, Auto Search, or RSI, the formal structure is environment—action—feedback—action again.
The loop can operate at different levels. At the surface, a model self-checks before delivering a result or asks another model to review it. At the middle layer, a Harness automatically organizes multistep work. At the deepest layer, the model generates data and modifies the training process, ultimately training a stronger version of itself. 孟繁青 personally considers all of these forms of ASI, although the public usually uses RSI more narrowly to refer to a model iterating on itself.
Auto Search might repeatedly optimize a GPU kernel, while RSI could cover both pre-training and post-training. 孟繁青 cites one test for AGI: can a model “replicate itself from scratch”? Only if the replicated model goes on to outperform the original does it constitute recursive self-improvement in the true sense.
Synthetic data is not the endpoint of RSI but an intermediate product. Complete self-improvement could also modify the data, algorithms, and architecture simultaneously. Evolvent AI’s experiments have already seen Agents exhibit limited self-evolution at the data layer, but there remains a clear gap between “fixing a format” and “coming up with a training insight.”
7. Harnesses will not disappear, but general-purpose and vertical markets will diverge
Model labs are releasing models alongside their own Harnesses, including Kimi Code and a Harness that 孟繁青 says remains unreleased internally at DeepSeek. Because model outputs pass through this system during post-training, backpropagation makes the model increasingly adapted to it, creating co-adaptation between the model and the Harness.
Harnesses will therefore become simpler, but they will not disappear. In general-purpose areas such as coding, 孟繁青 sees little reason for a third-party Harness to exist; model labs are more likely to provide the model and their own Harness together.
The economics are different in vertical domains. Many tasks do not require a strongest-in-class model at Claude’s level. A strong model can guide a smaller open-source model and a simple workflow through self-evolution in a specific setting, after which only the small model needs to be deployed, delivering lower cost, higher speed, and data privacy.
曲凯 asks whether, if models can already train themselves, they can also write vertical-domain Harnesses—meaning that “the model eats everything” in the end. 孟繁青 distinguishes between the writer and the runtime user: a strong model can write the Harness, but that does not mean it must be used at runtime. 曲凯 also says GPT seems to have just cut prices by 80%; 孟繁青 notes that price cuts are also driven by open-source competition, and prices may not keep falling if a monopoly emerges.
8. RSI does not need a dedicated foundation model; it is closer to intelligent search performed by existing models
In response to the question of why a startup that can conduct self-directed R&D cannot simply build a foundation model better than OpenAI, 孟繁青’s answer is that the so-called RSI foundation model is actually the foundation model model labs are already training. Building a separate RSI-specific foundation model or Harness would serve no purpose in his view.
He describes current RSI methods as “a smarter version of DFS.” The model tries one approach, receives a score, backtracks after a bad signal, and chooses another path. Unlike hard-coded depth-first search, an LLM can decide where to backtrack and prune obviously invalid branches in advance.
Pruning requires the model to predict what future state an operation will produce and the potential value of that state. Technically, this corresponds to a world model plus a value model. 孟繁青 believes both capabilities have already been internalized during foundation-model pre-training. Stronger RSI mainly depends on better internalization of knowledge and continued growth in the context window. He says the current maximum length may reach 1M; a longer context lets the model remember more exploration and backtrack farther.
Evolvent AI therefore wants to build an RSI channel between model labs and application companies, rather than train a foundation model head-on against model labs. 曲凯 extrapolates that future model companies might need only 10 to 20 people. 孟繁青 accepts that this could be true in terms of output, but stresses that real-world organizational and other constraints may not allow headcount to shrink to an extreme.
9. The data industry has moved from ordinary and expert labeling to Agent trajectories
The first phase was ordinary human labeling represented by Scale AI: small data volumes and relatively simple tasks. The second phase shifted to expert data. After o1 appeared, models’ reasoning capabilities jumped, and in early 2025 model labs began searching aggressively for math and physics experts.
The bar for experts was then raised by model capability. Previously, they only had to solve problems; now they also had to write them, and the questions could not overlap with problems already available in the market. 孟繁青 says this has become “too hard.” Expert data is still needed, but demand will not return to its previous scale. Ordinary, high-noise raw data is used more for pre-training, and incremental demand has also declined.
After Claude Code, models evolved from chatbots and super problem-solvers into colleagues, and the data format shifted toward long-horizon Agent trajectories. The synthetic route is to build an environment, set an objective, let the model explore, and save the full trajectory. The human route is to record people interacting with real workplace software.
Human data comes from real work environments but is noisy, requires cleaning, and raises compliance issues. Synthetic trajectories can be used directly for training if the environment, task, and grader are all designed correctly. When asked whether individual users could sell records of their everyday AI use, 孟繁青’s answer is direct: “They won’t sell,” precisely because the data is too dirty and creates compliance obstacles.
10. Current synthetic-data orders are nearing the end; the next wave is RSI trajectories that can train models
The program’s pricing snapshot is that complex, long-horizon synthetic tasks can sell for several thousand per task, while expert problem-solving data is priced at around 1,000–2,000. Human-operated trajectories are typically purchased at scale and do not have a stable per-sample price.
Model labs of course generate synthetic data themselves, but internally generated data can carry the lab’s own bias. They still purchase high-quality external samples to fill out the distribution. As the number of high-quality benchmarks converges, vendors can concentrate internal resources on a small set of core leaderboards, and external orders naturally shift from “the more the better” to selective supplementation.
孟繁青 believes the current wave of ordinary Agent data is nearing its end, and labor-intensive expert-labeling companies will be filtered out. Data companies seeking to keep up with demand must have first-hand, hands-on researchers who understand model training and new benchmarks, rather than relying solely on expanding their labeling workforce.
The next wave could be RSI data—for example, having one model fully post-train another small model and preserving the improvement process as a trajectory. Generating a single sample could take more than 10 hours or even a full day. Evolvent AI is currently using its synthetic-data business to provide cash flow while accumulating know-how, which feeds back into RSI Bench Data and long-term platform capabilities.
11. “Algorithm is data, data is Infra, and Infra is algorithm” describes three sides of the same training bottleneck
孟繁青 agrees with the formulation that “algorithm is data, data is Infra, and Infra is algorithm.” As algorithms such as RL converge toward simpler and more unified forms, the differences between teams increasingly show up in what data they feed the model and whether they can process that data efficiently.
He uses a long-tail example to explain the role of Infra. In a batch of 100 data points, the first 99 can be processed within 1 minute, while the last one takes 100 minutes. If organization and training are synchronous, all the other GPUs wait for the last item. How the Infra handles this long tail directly determines training speed and compute utilization.
As trajectories grow longer, SFT for large models requires more GPUs and more model-parallelism strategies. If a single inference run takes a day or even several days in the future, data requirements will in turn force the algorithms and Infra to be rebuilt. The boundaries between the three are not disappearing; harder data is continually driving innovation in the other two layers.
Synthetic data will not truly be infinite. The volume will increase, but not without limit. The more Agent users there are, the richer the “entropy sources” generated through usage and the better the synthetic tasks that can be fed by them. 孟繁青 expects data volume to grow with model size, while quality gradually improves as well.
12. Distillation is a catch-up tool, not the decisive source of stronger Chinese models
Synthetic data requires an Agent to explore an environment. If the Agent uses an external model such as Claude, that constitutes what is commonly called distillation. 孟繁青 describes distillation from simple math problems as “copying the teacher’s homework”: when the environment is too simple, the teacher’s answer is nearly the entire source of value.
Truly skillful distillation first requires a complex environment, sufficiently good questions, and a correct grader, after which an external model is allowed to explore. The data producer is then contributing more than API calls—it is contributing an entire task and validation system. Lagging models gain more because they have a larger learnable margin.
Leading models can also explore correct trajectories in self-built environments, use those trajectories to train themselves, and improve their scores. 孟繁青 attributes this to RL’s continuous reinforcement process: the teacher and student can be the same model, so synthetic data does not necessarily depend on a stronger external teacher.
He believes the core intelligence of Chinese models comes primarily from pre-training architecture, optimizers, and knowledge acquisition, while post-training shifts the original distribution. Distillation is an accelerator—“use it if you can, but it is not essential.” If all channels were blocked, catch-up would become harder but could continue. He speculates that ByteDance’s caution around distillation may reflect concern that taking a “shortcut” would cause it to miss the insights on the main path.
13. Model intelligence is extending from coding to Auto Research; organizational capability will determine who moves faster
Coding has been the main path for model labs to raise the intelligence ceiling since the beginning of the year, and progress remains rapid, but the training pipeline has become relatively stable. Claude, OpenAI, Qwen, and Kimi are beginning to explore Auto Research, having models repeatedly modify their approaches over long tasks in search of new demonstrations of capability and new training paths.
AI for Science follows the same structure. Having models design molecules, proteins, and other objects is both a relatively direct application and a new measure of intelligence. 曲凯 summarizes the shift as “from the Olympiad to the Nobel.” 孟繁青 also points out that once coding can no longer sufficiently distinguish models, scientific tasks take on the dual role of benchmark and PR.
孟繁青 rejects the categorical view that multimodality is outside the main line of intelligence. Text is highly compressed and therefore learned faster; multimodal information contains more noise and develops more slowly, but it is indispensable for AI to enter the physical world. When resources are limited, companies should not optimize every modality indiscriminately, but multimodality should remain on the main line rather than being completely separated.
Model capability will continue to grow, but that does not mean model labs will always need the same number of highly paid researchers. 孟繁青 favors startups: technology will circulate quickly across companies, and the eventual contest will be over how organizations make the whole system move faster. Evolvent AI has therefore chosen the bridge layer of synthetic data and RSI; in hiring, it prioritizes engineering ability over papers and citation counts.