“Vector-based memory in AI models could have 1,000 times more information bandwidth than token-based communication, which is limited to about 16 bits per token for a 100,000-token vocabulary.”