This episode features an in-depth interview with Luo Fuli, head of Xiaomi's large model team, discussing the profound paradigm shift in AI from pre-training to post-training, heavily influenced by Agent-based frameworks like OpenClaw. She highlights the critical role of post-training, efficient model architectures, and collective intelligence in driving AI's accelerated productivity transformation and narrowing the gap with top international models.
Summarized by Podsumo
The AI paradigm has shifted from pre-training to *post-training (PoseTrain)*, with Agent frameworks like *OpenClaw* enabling models to compensate for weaknesses and achieve higher performance in complex tasks.
Luo Fuli experienced a personal 'aha moment' with *OpenClaw*, realizing its autonomous, empathetic, and research-assisting capabilities, marking it as a 'fossil-era' breakthrough.
Effective Agent development requires a significant reallocation of computational resources, with a suggested *3:1:1 ratio for research:PreTrain:PoseTrain* cards, emphasizing the growing importance of research and post-training.
The future of AI development hinges on *collective intelligence* and *egalitarian organizational structures* to foster rapid iteration and innovation, especially by embracing diverse talent like undergraduates.
Xiaomi's *Mimovir Flash and Pro* models prioritize 'Non-contextual Efficiency' (能康效率) through architectures like Hyper-sparse attention and MTP to achieve *high inference speed and cost-effectiveness*, making them highly suitable for Agent applications.
"环境反而比经验更重要。" (Environment is more important than experience.)"
"我第一天跟他对话的时候从凌晨两点时去到了六点天亮...你可能第一个感受是OK 他是他非常有自主性然后他非常有灵魂。" (My first conversation with it went from 2 AM to 6 AM... your first feeling is that it's very autonomous and has a soul.) — Luo Fuli on OpenClaw"
"我觉得它是一个非常好的放大器。" (I think it's a very good amplifier.) — Luo Fuli on Agent frameworks for mid-tier models"