Jack Morris discusses his research journey and key contributions to understanding large language models from an information-theoretic perspective. He explores how information is stored and compressed within models, highlighting that GPT-style architectures store around 3.6 bits per parameter. The conversation also covers embedding inversion, where text can be reconstructed from vector representations with high accuracy, revealing the surprising richness of embedded information. Jack introduces the idea that major AI breakthroughs often stem from new datasets rather than novel methods, citing examples like BERT and transformers. He also touches on model universality, drawing parallels to computer vision techniques like CycleGAN, and emphasizes the importance of adapting to evolving AI education and engineering practices.