头雁
头雁|Nov 16, 2026 11:20
Qu Kai's 42-chapter lecture, "A Sharp Commentary on Chinese Models and Neo Lab, A Discussion of RLaaS," is definitely worth listening to; I finished it over dinner tonight. It mainly discussed their personal reflections on Neo Lab and RSI, continuous learning, and competition among Chinese models. Personal Reflections: - Analogizing post-training to recommendation systems in the internet age: the underlying technology, and what products are built on top (TikTok, Toutiao, etc.), is not the focus at this stage. - Learning philosophy: strong generalization ability, cross-domain expertise (achieving 90 points in one field, achieving 20 points in other fields can have a multiplicative effect); unique insights gained from different experiences. Competition between Chinese and American Models and Outlook for Chinese Models: - The quote from "The Cost of Staying" illustrates the cost of large companies "staying" (maintaining existing business models and profit structures)—especially when new technologies may impact core businesses (such as search advertising), organizations may forgo groundbreaking innovation to protect existing revenue. This is essentially a concrete manifestation of the innovator's dilemma in the AI era. - US Lead: OpenAI and Anthropic are significantly ahead (earlier data processing and large-scale RL, more GPUs). Mature standard recipes already exist for models below 1T; the main gap lies in data organization and scaling efficiency. - Attitude towards domestic models: Short-term pessimism (distillation becomes more difficult, catching up becomes slower), long-term optimism (domestic computing power will be "usable" next year; Chinese teams have stronger data organization capabilities, and even large US companies have significant Chinese involvement in data work). What should domestic model companies focus on next: - The core remains Infrastructure + Data. - Infrastructure: Co-design model architecture and infrastructure to adapt to the latest GPUs and even domestic GPUs (DeepSeek-V4.1-Flash's causal encoder-decoder architecture is an example; in the future, encoders may use domestic GPUs for prefilling, and decoders may use NVIDIA GPUs for decoding). - Data: Concentrate resources on breakthroughs in cutting-edge fields such as scientific discovery. The data has shifted to an "environmental" context—if there's a standard answer, perform SFT/distillation; if not, a problem and environment are needed for the model to trial and error. Neo Labs vs. RSI: - A true Neo Lab must focus on potentially commercially viable projects, pursuing a closed loop, rather than simply publishing papers and applying for grants. Mind Labs advocate research-product co-design: research is driven by real-world scenarios, judged by "usefulness" rather than publication potential. - Sequence: Research/modeling first, then product development. The ideal closed loop involves first collecting real-world data for product prototypes, conducting experiments, validating the results, and then expanding. - Domestic Neo Labs often emphasize global models and RSI (the main research direction), while overseas focuses more on vertical domains and new architectures. - RSI is currently largely a marketing slogan, essentially similar to prompt/context/harness engineering, but emphasizing a higher AI weight in training iterations (automated data pipeline, model finding data/creating verifiers). However, it's difficult to completely eliminate the need for humans: reward design requires human definition of requirements; proprietary data (such as pharmaceutical secrets) cannot be acquired autonomously by AI. RLaaS (Reinforcement Learning as a Service) and Post-Training Business: - Providing RLaaS (training models for companies) is not contradictory to Neo Lab's narrative; it's a continuation: we've always hoped AI would learn from real-world environments. - Cost: Pricing is token-based; for models like DeepSeek-V4-Pro and GLM-5.3, a single iteration costs tens of thousands of RMB, and a project typically costs hundreds of thousands to a million RMB. Building an in-house team is costly and difficult to execute quickly; larger models are better suited for external services. Current Focus and Future Outlook: - Most concerned with continuous learning: Models lack a mechanism for storing long-term memory and "taste" (compressed insights gained from reading numerous papers). We need lightweight units similar to LoRA but supporting longer-term, more efficient updates (Ilya is also considering similar issues). https://www.(xiaoyuzhoufm.com)/episode/6aec6bc06381f0f964c45068
Share To

HotFlash

APP

X

Telegram

Facebook

Reddit

CopyLink

Hot Reads