锐评国模与 Neo Lab，畅谈 RL 人生与 RLaaS | 对谈 Mind Lab 首席科学家骁腾

Summary

This 42章经 episode interviews 马肖腾 of MineLab about reinforcement learning, enterprise model post-training, AI-company competition, and the research-to-market discipline of a New Lab. Its central operating thesis is enterprise post-training as a feedback loop: customer environments supply proprietary data, tasks, failures, and benchmarks that can support repeated model improvement rather than a one-off fine-tuning job. The conversation also connects LoRA-based model diversity, Continual Learning, Research Taste, and RL-inspired decision heuristics while keeping costs, data access, evaluation, latency, and human-defined rewards visible.

Key Claims

• 马肖腾 describes a path from Tsinghua automation and large-scale multi-agent RL research into MineLab, whose business shifted from consumer applications toward enterprise reinforcement-learning and post-training services.
• Reinforcement-Learning Decision Heuristics turns exploration versus exploitation, policy-value consistency, and discounting into personal rules: preserve some randomness, use action to revise beliefs, and balance long-term direction against executable near-term steps.
• The source treats post-training as a reusable foundational capability similar to recommendation systems: value comes from repeated exposure to real behavior, feedback, and vertical tasks rather than from a single generic model recipe.
• Public web data is described as increasingly exhausted, so durable advantage moves toward proprietary enterprise data and interactive environments that models can act in, receive feedback from, and learn within.
• Enterprise Post-Training Loop requires a customer-owned benchmark before optimization; without a baseline and stable evaluation, neither capability improvement nor cost reduction can be demonstrated.
• Post-training can improve performance on a private benchmark or distill useful behavior into a cheaper model, but local deployment, privacy requirements, distributed training, and serving infrastructure raise implementation cost.
• The guest estimates that one frontier-model training iteration may cost tens of thousands of yuan and that an end-to-end engagement may range from hundreds of thousands to roughly one million yuan; these are experience-based ranges, not audited market prices.
• New Lab is defined here by a research-product-design loop: real use cases should shape research questions, prototypes should collect evidence before large-scale promotion, and usefulness matters more than publication alone.
• DeepSeek is praised for model-infrastructure co-design and Zhipu AI for practical usability and data production, while OpenAI and Anthropic are treated as leading overseas labs; these are the guest’s comparative judgments rather than benchmark-backed rankings.
• RSI is narrowed from an open-ended autonomy slogan to greater AI participation in finding data, creating verifiable feedback, and improving training pipelines; human reward design, taste, and private data remain necessary.
• Mint uses LoRA-based specialization to explore larger base models, cheaper adapters, and many coexisting behavioral variants on shared foundations.
• Continual Learning is framed as the unresolved problem of compressing long experience into durable judgment or “taste” without retaining every detail, possibly through lightweight and repeatedly updated memory units.

Key Quotes

“90 分主义” - the guest’s shorthand for leaving capacity for exploration rather than optimizing every current task to completion.

“企业要做后训练，首先得有自己的 benchmark” - the source’s evaluation prerequisite for enterprise post-training.

“真正的 New Lab” - the source’s label for research organizations that connect technical work to real use and commercial feedback.

Connections

• 马肖腾, MineLab, and Mint - guest, company, and LoRA training-service context.
• Enterprise Post-Training Loop, Model Post-Training Bottleneck, Agent Post-Training, and AI Verification - data, environment, benchmark, and delivery loop.
• Low-Rank Adapter Model Diversity, QLoRA, Model Distillation / 模型蒸馏, and Specialized Intelligence - low-cost specialization and heterogeneous-model branch.
• New Lab Organization, Research Taste, and AI Organization Design - research, product, design, and commercialization relationship.
• Reinforcement-Learning Decision Heuristics, Agent RL, and Organizational Long-Termism / 组织长期主义 - RL concepts applied to action and time horizons.
• Continual Learning, Persistent Agent Memory, and Human Taste as AI Training Signal / 人的品味作为AI训练信号 - compressed experience, durable judgment, and human preference boundary.
• DeepSeek, Zhipu AI, OpenAI, Anthropic, and Model-Infra Co-Design - model-company and infrastructure comparison set.

Contradictions

• No settled factual contradiction is adopted.
• The title says “Mind Lab” and names “骁腾,” while the supplied body consistently uses “MineLab” and “马肖腾”; the wiki consolidates the body forms under MineLab and Ma Xiaoteng / 马肖腾 while preserving the mismatch as source-level uncertainty.
• The source’s claim that sub-1T model training has become a relatively standard recipe is qualified by its own emphasis on tacit infrastructure, architecture-hardware co-design, data access, environments, and evaluation.
• Company rankings, cost ranges, LoRA/full-training comparability, domestic-compute forecasts, and the future balance between general and specialized models remain guest judgments without independent validation in the supplied document.