Fooling around with encrypted reasoning blobs

The author explores the concept of encrypted reasoning in LLMs, where the model's internal monologue is encrypted and sent to the client, and discusses potential security implications and side channels.