Basically he states the big data approach is fundamentally flawed because the LLM’s cannot reason at all. Everything it generates is probabilistic. So the systems and robots he builds are built to be “aware” of their surroundings and context, not relying on a dataset, but on ad-hoc input. An example cited in the article is the Tongtong 2.0 robot which uses a pillow to give it some extra height so that it can retrieve a book from a bookshelf otherwise out of reach. It sees the book, realizes it can’t reach it and then finds a solution so that it can.
This article doesn’t go into great detail about his work (it’s more about his life) but does touch on the notions of it: https://www.theguardian.com/news/ng-interactive/2025/sep/16/song-chun-zhu-why-one-of-the-worlds-most-brilliant-ai-scientists-left-the-us-for-china
Basically he states the big data approach is fundamentally flawed because the LLM’s cannot reason at all. Everything it generates is probabilistic. So the systems and robots he builds are built to be “aware” of their surroundings and context, not relying on a dataset, but on ad-hoc input. An example cited in the article is the Tongtong 2.0 robot which uses a pillow to give it some extra height so that it can retrieve a book from a bookshelf otherwise out of reach. It sees the book, realizes it can’t reach it and then finds a solution so that it can.