Andrej Karpathy has some advice for AI users trying to write beautifully crafted prompts.
Stop. Instead, you should just ramble. On, and on, and on.
In a Tuesday post on X, Karpathy — a leading AI researcher at Anthropic who coined the term “vibe-coding” — said he sometimes turns on voice mode and prattles at a large language model for about 10 minutes.
He called the sessions a “total mess, anything goes, full stream of consciousness.”
“Sometimes the LLM needs more bits to understand what you’re trying to achieve,” he wrote.
The approach helps properly divide responsibilities between the human and the bot, he said: Speaking lets the human unload every half-formed thought about a project, while the chatbot reconstructs the monologue into something cleaner and more organized.
Karpathy said he often starts by issuing a warning to the chatbot — that his thought process might be jumbled and the text might contain typos.
“The result is that you improve the mind meld and have to correct things less from that point on,” he said.
Other X users weren’t fully sold on Karpathy’s pitch to blather.
Karpathy’s pitch also comes as Silicon Valley is sharpening its AI voice tools — and trying to get users to speak up.
Earlier this month, OpenAI unveiled GPT-Live, the new model that’s powering ChatGPT Voice. The company’s $230 mini-keyboard also has a voice-recording key that sends audio prompts directly to an agent. Anthropic also launched updates to its voice model on Thursday, allowing users to choose between its Opus, Sonnet, and Haiku models.
