Problem Definition In Research
150. 对英伟达研究副总裁刘洺堉的4小时访谈：Cosmos 3、世界模型、武术、黄仁勋影响我的，和你不需要击败所有对手 adds Liu Ming-Yu / 刘洺堉’s practical version. Liu says the core of research is asking the right question; his own path from image generation to Cosmos 3 shows problem definition changing when scale, customer use, and Physical AI impact become more important than a clean paper problem.
Problem definition in research is the ability to make a question precise, tractable, and consequential enough for a field to move. 133. 对谢赛宁的7小时马拉松访谈：世界模型、逃出硅谷、AMI Labs、两次拒绝Ilya、杨立昆、李飞飞和42 treats this as one of the core research abilities.
Source View
Xie Saining uses Fei-Fei Li and ImageNet as the central example: the contribution was not only collecting data, but defining image classification so that deep learning systems could be compared, improved, and understood. He generalizes the lesson to research itself: losing the ability to define problems means losing much of the ability to innovate.
140. 对姚顺宇的4小时访谈：请允许我小疯一下！在Anthropic和Gemini训模型、技术预测、英雄主义已过去 adds Yao Shunyu / 姚顺宇’s frontier-model version. He argues that top labs are less worried about whether AI can improve in general and more worried about whether the next task is well defined enough to train, evaluate, and scale. In this source, a good AI research problem needs a usable environment, data, feedback signal, and evaluation standard.
AI4S 需要狂人与野心家｜对话英灵殿 Odin："如果神存在，我怎能容忍自己不是神？"【公路播客】 adds Haotian Odin / 浩天’s biological-startup version. His All-Modal Molecular World Model claim is a problem-definition move before it is a proven architecture: if the real biological problem crosses small molecules, proteins, RNA, and DNA, then separate single-modality tools may define the problem too narrowly.
150. 对英伟达研究副总裁刘洺堉的4小时访谈：Cosmos 3、世界模型、武术、黄仁勋影响我的，和你不需要击败所有对手 adds the infrastructure-company version. Liu Ming-Yu / 刘洺堉 defines the problem not as “make the strongest video model,” but as helping Physical AI developers through better data, better starting points, and better environments; that problem definition makes World Foundation Models more precise than the broader World Models label.
Connections
- Fei-Fei Li and ImageNet — core example in the source.
- Xie Saining — speaker who draws the methodological lesson.
- Research Taste — adjacent judgment about whether a problem and result are meaningful.
- Representation Learning and Multimodal Intelligence — technical areas where problem framing shapes progress.
- World Models — field where definition remains contested between video generation, prediction, causality, action, and product interface.
- Yao Shunyu / 姚顺宇, Frontier Model Scaling, Long-Horizon AI, and ML Coding — frontier-model training version added by episode 140.
- Haotian Odin / 浩天, Yinglingdian AI / 英灵殿, All-Modal Molecular World Model, and AI Drug Discovery Platform — molecular-biology problem-definition branch added by the Shizilukou Crossing source.
- Liu Ming-Yu / 刘洺堉, Cosmos 3, World Foundation Models, and Physical AI — infrastructure and world-model problem-definition branch added by episode 150.