Optimizing LLM Interaction through Vocal 'Rambling'

Andrej Karpathy proposes a counter-intuitive method to improve interaction with large language models (LLMs): prolonged vocal 'rambling'. Rather than formulating concise queries, he suggests speaking freely and disjointedly for several minutes, using speech recognition.

This approach provides LLMs with a larger volume of contextual data, even if it's messy. Surprisingly, LLMs excel at restructuring these incoherent streams of thought into clearer and more coherent responses than the user's initial intention. The result is better mutual understanding between the user and the LLM, thus reducing the need for subsequent corrections.