One of the few who has done agents, multimodal and embodied AI | with ex-Hunyuan multimodal RL infra lead Haonan少有的同时做过 Agent、多模态和具身的人 | 对谈前混元多模态 RL Infra 负责人淏楠

少有的同时做过 Agent、多模态和具身的人 | 对谈前混元多模态 RL Infra 负责人淏楠One of the few who has done agents, multimodal and embodied AI | with ex-Hunyuan multimodal RL infra lead Haonan

Key points

• Wang Haonan worked on agents at MiroMind, led multimodal RL infra at Hunyuan (open-source UniRL), and now leads base-model pretraining at an embodied-AI company.

  王淏楠曾在 MiroMind 做 Agent，后在混元负责多模态 RL infra、主导开源框架 UniRL，现在一家具身公司负责基模预训练。

• Covers the RL gap between China and the US, how agents relate to multimodal, why Seedance is good, MiniMax H3, and whether real-time interaction is mature.

  聊国内外 RL 差距、Agent 与多模态的关系、Seedance 为何好、MiniMax H3 水平以及实时互动技术是否成熟。

• Discusses why GPT Astra shook the embodied field, whether the robot ‘brain’ becomes a frontier-lab contest, and whether embodied AI is at its GPT-3 moment.

  讨论 GPT Astra 为何给具身圈带来震动，具身“大脑”是否会变成大模型厂商的竞争，以及具身是否迎来 GPT-3 时刻。