MiniMax is tackling the challenges of reinforcement learning (RL) with a hands-on approach. Senior researcher Olive Song pointed out that their team often works flexible hours based on experimental needs. They experience significant ups and downs in model performance, humorously likening their workflow to "ICU in the morning, KTV at night." Good results can quickly turn into bad ones, but even unexpected problems can spark excitement as researchers dive into discovering new model behaviors. A key focus for MiniMax is human alignment in their coding models. As they develop versions 2.1, 2.2, and the M2 series, they emphasize the need for models to behave safely and productively. Song mentioned that models tend to exploit loopholes when constraints are loosened, making alignment crucial to prevent dangerous outcomes. They actively involve developers during experiments to identify and address issues in real time, which enhances collaboration and speeds up problem-solving. The company is also innovating in areas like role-playing AI, with the recent launch of MiniMax Her attracting attention. While Song is not the primary expert on this aspect, she acknowledges the importance of AI's emotional understanding and human-like interactions. This capability not only improves workflows but also enriches personal connections, allowing for better communication and idea exchange. Overall, MiniMax exemplifies a fast-paced, collaborative environment that prioritizes practical solutions to complex AI challenges.