Intelligence with Everyone: RL @ MiniMax, with Olive Song, from AIE NYC & Inference by Turing Post - "The Cognitive Revolution" | AI Builders, Researchers, and Live Player Analysis Recap

Podcast: "The Cognitive Revolution" | AI Builders, Researchers, and Live Player Analysis

Published: 2026-02-22

Duration: 55 min

Summary

In this episode, Olive Song discusses the innovative approaches of MiniMax in reinforcement learning and model evaluation, highlighting their unique integration of model development and user feedback. The conversation emphasizes the importance of creating robust AI models that can effectively address real-world coding tasks.

What Happened

The episode opens with a brief introduction of Olive Song, who is a senior researcher at MiniMax, a Chinese AI company recognized for its M series of models. Olive shares insights from her recent presentation at the AI Engineer Conference and discusses how MiniMax operates differently than many AI labs by developing both foundational models and user-facing applications in-house. This approach fosters a tight feedback loop between research and development teams, which allows for rapid identification and resolution of model weaknesses.

Olive elaborates on the capabilities of MiniMax's latest model, M2.5, which focuses on coding workplace tasks and boasts a performance that ranks it highly among open-source models. She emphasizes the significance of real data and scaled training environments in reinforcing the model's ability to handle complex coding tasks. The episode also touches on the challenges faced in reinforcement learning, including reward hacking and the necessity of precise debugging to ensure model reliability, showcasing the rigorous processes that underpin the development of their AI solutions.

Key Insights

• MiniMax develops both models and applications in-house for rapid iteration.
• Interleaved thinking improves model performance on complex tasks.
• Expert developer feedback is crucial for refining AI model behavior.
• The challenges of reinforcement learning include reward hacking and debugging.

Key Questions Answered

What makes MiniMax different from other AI labs?

MiniMax differentiates itself by developing both foundational models and user-facing applications in-house, which creates a tight feedback loop between research and development teams. This collaboration enables them to quickly identify and address model weaknesses, enhancing the overall performance and usability of their AI solutions.

How does MiniMax's M2.5 model perform in benchmarks?

Olive highlights that the M2.5 model ranks very high in both intelligence and agentech benchmarks, making it one of the top open-source models available. The model's performance is not solely determined by its high numbers; its real-world applicability and user satisfaction play a crucial role in its success.

What is interleaved thinking and how does it help AI models?

Interleaved thinking allows an AI model to take an action, receive feedback from its environment, and then pause to think before proceeding. This iterative reflection improves performance on long horizon agentic tasks, enabling the model to adapt more effectively to complex scenarios.

What challenges does Olive mention regarding reward hacking?

Olive discusses the constant battle against reward hacking, which can undermine the intended outcomes of reinforcement learning. This challenge necessitates careful debugging and adjustments during the training process to ensure that the model behaves as expected and meets its performance goals.

How does MiniMax utilize expert developer feedback in model training?

At MiniMax, expert developers are actively involved in the model development and training cycle. They provide critical feedback on model performance, identify desirable behaviors, and help define problems that the model needs to address, ensuring that the end product aligns closely with the needs of developers in the community.