In this episode, 骁腾 (Pony), co-founder and chief scientist of Mind Lab, discusses the current state of Chinese AI models, the role of Reinforcement Learning (RL), and the concept of RL-as-a-Service (RLaaS). He shares insights on how RL principles like exploration vs. exploitation, alignment of policy and value, and long-term thinking can guide both AI research and personal life. The conversation also covers the competitive landscape of Chinese model labs, the importance of data and infrastructure, and the future of model training.
Summarized by Podsumo
Reinforcement Learning (RL) is likened to a life philosophy, emphasizing exploration vs. exploitation, alignment of policy and value, and long-term thinking as key principles.
The cost of fine-tuning top models like GPT-4, Claude, or Gemini is relatively low (a few million RMB per project), making it accessible for companies with a clear benchmark.
Chinese labs like DeepSeek are seen as strong contenders due to their efficient infrastructure design, while others like Zhipu are praised for practical data handling.
The concept of RLaaS (RL-as-a-Service) is emerging, where companies like Mind Lab help others train models, focusing on real-world data from specific industries like finance or healthcare.
The future of AI may involve multiple specialized models rather than one all-powerful AGI, with each model having its own 'taste' and serving different needs.
"Life itself is a reinforcement learning process. You need exploration to escape local optima, and you must keep updating your values to align with your actions."
— 骁腾 (Pony)
"The most important thing for models is to have a place to store long-term memory and taste. Without that, continuous learning is impossible."
— 骁腾 (Pony)
"If there is no clear benchmark, you cannot judge which model is good. Training a model without a benchmark is like shooting in the dark."
— 骁腾 (Pony)