• The article argues that the use of tokens in today's generative AI models is a significant limitation, as they restrict the models' ability to generate truly novel and creative content. Tokens are pre-defined units of text that the models are trained on, which can lead to a lack of originality and a tendency to regurgitate existing information.

• The author suggests that the reliance on tokens forces generative AI models to operate within the confines of the data they were trained on, rather than being able to truly understand and reason about the world in a more holistic way. This can result in outputs that lack coherence, depth, and nuance, as the models struggle to make meaningful connections beyond the surface-level associations they've learned.

• The article proposes that the next generation of generative AI models should move beyond the token-based approach and explore alternative architectures and training methods that allow for more flexible and open-ended generation. This could involve incorporating more advanced reasoning capabilities, better understanding of context and semantics, and the ability to draw upon a wider range of knowledge and experiences to create truly innovative and meaningful content.