Topics Pioneers Insight Method Author
Back

Mind Lab’s 骁腾 Talks Chinese Models, Neo Lab, RL Life and RLaaS

·42章经

Rendering the long image…

Topics US-China AI Rivalry and Geopolitical Controls Battle for Frontier AI Moats AI Training Data and Evaluation Benchmarks Enterprise AI ROI and Cost Optimization Enterprise Private AI and Real Compute Demand Specialized SLMs vs General LLMs Long Context and External Memory Bottlenecks
Pioneers 马骁腾

Summary

  • 马骁腾 is short-term bearish and long-term bullish on Chinese models, but sees the gap created by a new paradigm as a more dangerous risk than a compute shortfall. Tighter anti-distillation measures will slow the catch-up, but he still expects domestic compute to reach a usable level “next year” and remains confident in China’s ability to organize data production. Most industrial applications run below 1T parameters, with commonly used models around 300B–500B. But if leaders crack new paradigms such as week-long reasoning and China fails to keep up, the gap in frontier research could widen.
  • Big tech may lose the model-technology race yet still make money from compute and equity stakes. 马骁腾 says large groups will keep weighing training bills against ROI, while startups “die if they can’t build it,” leading to different levels of commitment and data accumulation. 曲凯 points to Tencent’s approach—hiring, funding, compute and independence—as a counterexample. 马骁腾 says that can help a team catch up, but does not guarantee it will be the best: building today’s model in 6 months is not the same as catching the rival 6 months later.
  • Beyond public data, real-world enterprise settings and interactive environments are becoming an important source of post-training data. 马骁腾’s view that “the data is used up” refers to public internet data, not every possible source. The next increment of value lies in enterprises, real life and consumer products. Where there is no standard answer, the deliverable shifts from demonstration data to an environment in which a model can try, verify and learn. He concedes that even acquirers may not know what exclusive data is worth.
  • Overseas, he favors OpenAI; in China, he is strongly bullish on DeepSeek and also credits Zhipu’s data-production system. He thinks OpenAI adjusted its harness and reasoning structure during pre-training, while Anthropic has not yet followed. DeepSeek’s edge is the tight coupling of model and inference: cheaper inference drives more usage and data, and speeds up RL training. Zhipu’s strength is a pragmatic, data-focused approach, with the aspiration to build “the most practical, easiest-to-use model.”
  • By 马骁腾’s standard, Neural Lab must prove a commercial loop, and RSI cannot substitute for real demand. “Only usefulness matters”: start with a product prototype to gather real data, then expand into more use cases and growth after the research works, rather than taking a new architecture out in search of applications. He sees RSI as automated data construction and training loops in verifiable domains, but does not believe model iteration will ultimately need no humans: reward depends on what people want, and private data and real-world experiments do not appear automatically.
  • Post-training services are best suited to companies with PMF, their own benchmarks and a mandate to improve performance or cost. Post-training can improve an existing model’s performance in a business, or transfer similar performance to a cheaper model. In his experience, one training iteration on a frontier model costs tens of thousands of yuan, while a full project can be kept to a few hundred thousand yuan to RMB1M. The bar is “you need your own benchmark”; the main frictions after that are data access, local deployment requirements and willingness to pay.
  • LoRA turns limited compute into an opportunity to customize large models and support a large number of differentiated models. The team started in March 2025 with about 4 H800 servers, then pursued three paths: larger base models, smaller adapters and more LoRAs. He says the team built a system in April that supports inference with 1 million LoRAs on a single base model, while leaving the resulting “agent society” as a possibility rather than a proven outcome.
  • What continual learning may lack is a lightweight unit for storing long-term memory and state, not simply more context. 马骁腾 illustrates the point with his experience of reading 300-plus papers and then “just getting it”: what remains is judgment, not every paper’s details. He is not worried about more capable models and is willing to work toward a future in which different models have different preferences and serve different people: “I think I should fight for the second scenario.”

Deep dive

1. The business form keeps changing; MindLab’s research focus remains post-training

  • 马骁腾 says he entered Tsinghua’s direct-entry PhD program in 2017, graduated in 2023 and stayed on as a postdoc. In 2024, he started a browser company with 谢扬; in October 2025, he founded MindLab with 陈凯杰 and Andrew. The through line is reinforcement learning and post-training, not any one fixed application.
  • The browser was originally a way to collect GUI data and conduct post-training. But he found that an application company has limited resources and must constantly chase the performance of the best models, making sustained model research difficult. Leaving the browser company to build a post-training lab was a way to refocus resources on that work.
  • On the controversy around Macaron and the company, he says startups should not shy away from debate. The substance of both application businesses and RL-as-a-service is using LLM reinforcement learning to improve specific tasks. He sees post-training as a “foundational technology,” like recommender systems: whether it gives rise to Xigua Video, Douyin or Toutiao is not what matters most to him as a researcher right now.

2. Three principles for an RL life: exploration, aligning action with belief and bounded long-termism

  • 马骁腾 uses early Atari training to explain exploration: even 10% random actions, seemingly inefficient, may be necessary to escape a local optimum. In life, that led him to a “90-point rule”: once something is good enough for 90 points, use the remaining energy to explore something new rather than endlessly optimizing the established path.
  • 曲凯 pushes back that an individual can settle for 90 points, but a startup needs to reach 100 or more in at least one area. 马骁腾 accepts the tension: he takes responsibility for exploration, while Andrew and 陈凯杰 fill the commercialization gap, bringing the team to 100. That division is also why he chose to start a company with them.
  • “Aligning knowledge and action” means keeping policy and value aligned: if you believe something matters, act on it, gather experience and use the results to revise your judgment. The point is not to start with an immutable set of values, but to keep updating them and check whether your actual strategy still matches.
  • “Long-termism” also has limits. He uses a 0.99 discount factor and a horizon of about 100 steps: too small and you become short-termist; too large and distant noise prevents convergence. Startups fit this learning loop better because they can get relatively complete real-world feedback. Large companies often have more localized information and room to explore, which is why he says “starting a company is a way of life.”

3. No AI Lab on his résumé does not mean no large-scale training experience

  • Asked how he could train models without working at a major AI Lab, 马骁腾 first admits: “I can’t really know, because I’ve never trained a model inside a big company.” He then describes his own experience: around 2018, he worked on a RMB50M project involving adversarial training on a cluster of 10,000 CPU cores and 100 V100s, when Ray was still at versions 0.2 and 0.3.
  • He sees continuity between today’s challenges and his PhD work: he is still dealing with reward hacking, simulators, reward design and benchmarks every day. The difference is that networks are larger, more expressive and more expensive to experiment with. That experience gave him early exposure to the full RL training stack and infrastructure, rather than just one slice of algorithm research.
  • He pushes back on the story that he stuck with an unfashionable field early on: “How would we have known it wasn’t popular then?” His adviser’s projects set the initial direction; training cars and little characters felt like playing a game and was fun. Getting papers published was the real pain. He also rejects the simple claim that China lacks strong RL researchers, arguing that the bigger gap is a “melting pot” where RL, CV and other fields mix.
  • Breadth across fields is another asset: from traditional RL and theoretical bounds to real-world projects, including lessons from risk-sensitive RL that informed later work on linear RL. His adviser’s explanation was that different experiences lead to different ideas, and even smart people cannot do all the research. That also shapes his hopes for AI: experience should produce different AIs, not just one unified answer.

4. Chinese models are catching up within the old paradigm; the real risk is a paradigm shift

  • 马骁腾 thinks pre-training and post-training for models below 1T parameters have settled into a standard recipe: “Give me the GPUs and I can do it.” Competition is shifting toward data scale, model efficiency and larger-scale training, areas where he sees a clear lead for OpenAI and Anthropic.
  • 曲凯’s sense is that China had been closing the gap over the past few years, but releases such as “5.5X” seemed to widen it again. 马骁腾’s outlook is “short-term bearish, long-term bullish”: tighter anti-distillation measures will slow the catch-up, but there is still room to close the gap as long as the core technical paradigm does not change.
  • His long-term optimism rests on 2 things: he expects domestic compute to remain behind the most advanced Nvidia cards but reach a usable level “next year,” and he believes China is highly capable of organizing data operations. The more dangerous scenario is that leaders adopt a new paradigm and achieve week-long reasoning or Fields Medal-level mathematical proofs, while China fails to break through on that paradigm.
  • He separates industrial use from frontier research. Most real-world applications run below 1T parameters; commonly used Flash-class models are around 300B–500B and already handle the vast majority of office tasks. In frontier areas such as drug discovery, he acknowledges that China will lag the US. That gap should not be conflated with the situation across all industrial applications.

5. Data is not growing out of thin air; the next increment is in private settings and environments

  • 曲凯 asks where the rumored 30T-scale datasets are coming from. 马骁腾’s answer is that public internet data is indeed used up (“数据就是用完了”). Much of the data for today’s pre-training and even mid-training still comes from existing datasets, model distillation and feeding model-generated trajectories back into training.
  • New data can still be organized in specialized domains, but he is more interested in Anthropic’s acquisitions to obtain non-public enterprise data. He is also uncertain about “how much this data is worth”—Anthropic itself may not know. What is clear is the company’s willingness to invest. Data from enterprises, real life and consumer applications may not ultimately flow to pre-training companies.
  • Data is increasingly being reframed as an “environment.” With high-quality standard answers, you can distill and run SFT; without answers, you have to provide a task and room to try, so the model can learn through SFT or RL. Building these environments takes more investment and, in his view, remains relatively overlooked.

6. Big tech’s resource advantage can be offset by group-level bills and organizational inertia

  • 马骁腾 uses Qwen as an example: being in the broadly recognized top tier does not mean the team is clearly behind technically or in resources. But once a huge bill lands in front of the broader group, the team has to answer questions about ROI, conversion and financial optimization. He thinks this happens at every large company; legacy businesses and organizational momentum also constrain breakthrough innovation.
  • His view is emphatic: “The best model will not necessarily come from big tech.” Startups have to go all in because they die if they cannot build it. Long-term data accumulation and cleaning can also leave slower-moving big companies behind. Commercial outcomes are a separate question: big tech can profit from startups’ growth through compute, equity stakes and other means.
  • 曲凯 offers Tencent as a counterexample: once the direction is clear, it can hire the best people, provide funding and compute, and give them independence—enough, in theory, to catch up quickly. 马骁腾 recognizes Tencent’s resolve and Hunyuan’s progress, but says “you can’t say Hunyuan is the best model right now.” Tencent ultimately needs a model suited to future businesses such as WeChat, not simply a winner in the race to AGI.
  • Asked to imagine pulling 80- to 90-point talent from each company and building a model at the same level in 6 months, he says it is “quite possible.” The question is whether the cost would be materially higher than using an existing model, and how far the competition would have moved in those 6 months. GPUs, ecosystem support and operations do not come together automatically just by assembling people.

7. The model companies he favors are betting on structural efficiency and data systems

  • If he could choose only one company overseas, 马骁腾 would favor OpenAI today. He thinks OpenAI adjusted its harness and reasoning structure during pre-training, while Anthropic has not, “as far as we can tell,” made or followed similar changes. That is his technical rationale, not just a ranking of current product performance.
  • In China, he is “strongly bullish on DeepSeek,” particularly its model-inference co-design. He sees the separation of encoder and decoder as an early sign of a design for heterogeneous compute: domestic cards could handle prefill while Nvidia cards handle decoding, using efficiency to compensate for limited compute. Another competitive path is building data teams to solve scientific problems that even the strongest models struggle with.
  • They have also felt DeepSeek’s tight coupling while working with LoRA: an adapter can disrupt the original best balance, though he adds that this may mean the team has not yet pushed things to the limit. He also cites mHC moving from 4 paths to 1 to highlight the system-level impact of inference design: lower prices support more token usage and generate more data, while RL training gets faster.
  • He calls Zhipu pragmatic and easy to use, with the most complete and highest-quality data-production pipeline among the companies they have worked with; Kimi, by contrast, is more innovative. 曲凯 summarizes the distinction as Zhipu doing each thing solidly, DeepSeek excelling at infrastructure and Kimi being more “fancy.” 马骁腾 largely agrees.

8. Product timing is also a choice about where model companies spend attention and compute

  • 马骁腾 criticizes MiniMax for lacking conviction: it had a distinctive linear-attention approach that later saw broad adoption, but abandoned it and returned to conventional techniques. 曲凯 notes that MiniMax was more product- and application-focused from the start, unlike DeepSeek. 马骁腾 says prioritizing revenue and an IPO consumes a lot of attention and affects competition in models.
  • He argues that technology, engineering, commercialization, productization and distribution should form a sequence and feedback loop, rather than stopping at technology. DeepSeek’s decision to hire a product team after building a strong model was “the natural next step,” in his view. MindLab likewise plans to commercialize its LoRA technology through services before moving on to the next round of research.
  • 曲凯 challenges the claim that big tech cannot build good models by pointing to Seedance’s success. 马骁腾 attributes it to data, a clear use case and sufficient compute, and emphasizes opportunity cost: in his telling, ByteDance carved out compute for video, while OpenAI abandoned Sora and redirected compute to language models. He cannot say that definitively explains ByteDance’s position in language models, but believes it is “certainly one factor.”

9. Neural Lab’s test is a research-product loop, not a new label

  • 马骁腾 distinguishes Neural Lab from the traditional paper-and-grant cycle: it is a company taking real money from investors and ultimately expected to deliver commercial value. MindLab’s slogan is “Research Product Co-design”: use real-world settings to inform research, then test whether the research is useful, rather than treating publication as the ultimate measure.
  • 曲凯 proposes a different ordering: OpenAI and Anthropic first make models that are absolutely good, and the money follows. 马骁腾 asks what defines a good model; they eventually land on solving real coding problems and getting into practical use. He emphasizes that “only usefulness matters” (“只有有用是重要的”), warning researchers against “taking a hammer and looking for nails.”
  • They also disagree on whether the model or the product comes first. 马骁腾 initially says research must come first, then product; 曲凯 asks how you can choose a research direction without knowing what users need. 马骁腾 then gives a fuller sequence: start with a prototype and take it to market to gather real data, but avoid scaling growth or paid acquisition. Once the research works, expand into more use cases and identify the next problem.
  • 曲凯 also points out the tension between talking publicly about Neural Lab and not emphasizing the label internally. 马骁腾 says the team cares more about doing solid work, automating training and deployment, and reducing the friction of model training. His message to investors is to ask whether the story “actually has a commercial loop.”

10. RSI can automate data loops, but human needs and data boundaries remain

  • 马骁腾 thinks RSI is largely a “slogan” for now, much like the shift from prompt engineering to context engineering and harness engineering: the emphasis is changing. In practice, AI is taking a larger role in model iteration—finding, constructing and validating data in fast, verifiable environments, then feeding it back into training. “There’s no complicated trick.”
  • He does not believe model iteration will ultimately need no humans, because reward design still has to answer what people actually want. Writing is his example: it is hard for a model to judge whether text is user-friendly, and its writing “may get worse,” so people still need to pull it back. Such capabilities are hard to automate fully through RSI.
  • Even if automated iteration works, it does not solve where the data comes from. A pharmaceutical company’s confidential data is not on the public internet, and a model cannot run experiments on its own to acquire the experience. That changes only if the company actually opens its data and lets AI operate and iterate; an automated training loop cannot cross that boundary by itself.

11. The research value of AI services lies in real problems absent from public corpora

  • 马骁腾 does not see helping companies train models as a business pivot or simple outsourcing. Real data from materials R&D, daily queries in finance and healthcare, and negative user feedback can all provide environments for research. The challenge is continuously turning these streaming tasks into environments and training models on them. If the work cannot deliver commercial value, he considers the research ineffective.
  • He explains the consolidation of companies in data, infrastructure, training and other middle layers with the idea that “the model is the service; the model is the product.” Front ends and harnesses, in his view, differ little: “We’ve never remembered a famous saddle, but we do remember Red Hare.” Someone still has to handle deployment, training, LoRA and benchmarks around the model; the question is what to build in-house and what to outsource.
  • In China, he sees the bigger obstacle for now as understanding and willingness to buy services, not access to GPUs. Post-training uses far less compute than pre-training and may not require the latest cards. That also reflects his view of the China-US SaaS gap: it will take time for the Chinese market to adopt these services.

12. Customers need to know how to judge quality before a post-training project has a clear endpoint

  • The first screening criterion is “you need your own benchmark.” Without an evaluation framework, a company cannot tell whether training worked or build the right environment. Having a benchmark at least means the company knows what it wants to optimize, rather than just starting to experiment with agents, model calls or RAG.
  • The ideal customer is a startup that has reached PMF: frontier models can already serve users, and the next goals are better performance, a closer fit with users and lower costs. These companies often need to compare new models and typically have their own benchmarks. 马骁腾 assesses customers along 2 dimensions: technical understanding and business revenue.
  • Few traditional or vertical-industry companies meet the bar. MindLab prioritizes leading customers with real deployments and clear value, rather than building models for PR. His view is that once a top company validates an approach, others in the industry may follow. If the leading companies do not understand the value of post-training, “the industry probably doesn’t need post-training.”

13. Post-training can improve both performance and cost; project price is not the only hurdle

  • 马骁腾 maps enterprise decisions on 2 axes: the vertical axis is its own benchmark, and the horizontal axis is model cost. A company can improve the performance of its current model or train a cheaper Flash model to reach similar performance. His examples are Gemini 5.3 and Gemini 5.3 Flash; the goal is to move the performance bar up or left.
  • Fees are generally based on tokens and correlate with data volume and compute consumption. His rule of thumb is a few tens of thousands of yuan for one iteration on a frontier model, and a few hundred thousand yuan to RMB1M for a full project; smaller models cost less. 曲凯 says it is well worth spending around $100K to try for a company that has raised a Series B and secured $10M–$20M. 马骁腾 agrees.
  • Once a company has the funding and a benchmark, the main friction is data access. Some fast-growing companies prioritize results; he notes that their data already flows to model providers when they use those providers’ services. Others require local deployment, which MindLab can support, but at a higher cost than using a service.
  • 曲凯 asks why customers do not build their own teams if the method is standardized and the cost is not high. 马骁腾 points to the underlying engineering cost of large-model training: coordinating multiple forms of parallelism, training correctly, quickly and cheaply, and continually adapting to new models. He acknowledges that some companies will build in-house; if the business only needs a very small model, it does not need to hire them.

14. LoRA started as a response to compute constraints and is moving toward million-scale differentiated inference

  • 马骁腾 says they started working on LoRA “entirely because we had too few GPUs”: around March 2025, they began testing DeepSeek LoRA training on about 4 H800 servers. Later, around September last year, they saw Thinking Machines publish “LoRA Without Regret,” which argued that LoRA can approach full fine-tuning in many cases. That matched their observations and energized the team.
  • They used Tinker as a reference to build Mint, their service, then pushed in 3 directions: skill up by training on larger base models for greater intelligence; skill down by making each LoRA smaller to cut costs and speed up iteration; and skill out by training more LoRAs on that basis.
  • He uses human genetic differences as an analogy: even when the difference is “less than 0.5%”—or, jokingly, perhaps “less than 0.2%” between him and 曲凯—people can still differ in behavior, voice and appearance. A fixed base model carries most shared capabilities, while a small adapter changes behavior. That is his intuitive way of understanding model differentiation.
  • He says the team built a system in April that can run inference with 1 million LoRAs attached to one base model. The idea that follows is still a hypothesis: a system with many LoRAs and different behaviors could become an “agent society” and generate new forms of interaction. He does not claim that such a society has already been realized.

15. Continual learning needs to preserve judgment, not every detail

  • 马骁腾 has recently been focused on what could break through beyond the standard recipe. AI can read 300-plus papers a day, but can it develop insight? His own experience is that after reading enough, “it just clicks”: when he sees a new paper, he can quickly spot the key points and judge how it might contribute to the next project, without remembering every detail of what came before.
  • He thinks this compressed state needs somewhere to live, and today’s models lack that mechanism. Continual learning may need “a lightweight, iterative unit,” perhaps something like LoRA but with a new way to update it, capable of carrying longer-term memory and improving learning efficiency. That is a research question, not a settled solution.

16. General models may get stronger, but he chooses to work toward a more diverse future

  • 曲凯 ends by asking whether a stronger general model could generalize its way through embodied AI, personal agents and even MindLab’s work. 马骁腾 replies that even if some people dismiss personal agents, the people at Anthropic responsible for data procurement still buy a lot of this data. He sees products as the best way to get real-world data. His example is Claude Code: Anthropic led OpenAI by about 6 months, built Claude Code and accumulated a lot of data, giving it an edge in coding during that period; OpenAI took a long time to truly catch up. 曲凯 says that models eventually generalizing to solve problems does not contradict buying data today; they do not fully agree.
  • On the physical world, he acknowledges that a larger general model will likely understand it better than a specialized embodied model. But “GPT-6” understanding the world is different from deploying it in an embodied system: it still has to deal with causality, latency and interaction with the body. One possible path, in his view, is to use a powerful model for its capabilities and then distill them into a VLA suited to controlling the body, rather than skipping the engineering steps.
  • Asked whether current work could become worthless, he offers 2 scenarios. If an all-knowing, all-powerful AI really emerges, work, careers and even social order could all be reshuffled, reducing today’s values to zero. If not, models will flourish in many forms, each with its own taste, preferences and areas of strength, serving different people and industries.
  • He does not say the first scenario cannot happen, but says, “I think I should fight for the second scenario” (“我认为我应该为第二种情况而奋斗”). Rather than compressing society into a single answer, he wants AI to make the world more diverse. Some people may lose their jobs in the short term, while others use new technology to find more interesting work they actually want to do.