What Actually Matters in AI? - with Zhuang Liu (Princeton)

Zhuang Liu, assistant professor at Princeton, discusses which parts of modern AI matter and which are historical accidents, drawing on work such as ConvNeXt, which showed convolutional networks can match Transformers with careful design, Transformers without normalisation using dynamic tanh, ImageBind, studies of CLIP's blind spots and evidence that large 'diverse' datasets remain distinguishable from one another. The conversation asks whether architecture research is still worthwhile, what good data means, and why vision bridges modalities while language drove adoption.