Anthropic AI researcher Andrej Karpathy (안드레이 카파시) suggested long voice input, rather than meticulously written prompts, as a way to use generative AI effectively. He said it is more efficient to speak at length in a stream of consciousness and let AI structure it than to try to organise thoughts in advance.

On July 24, Business Insider reported that Karpathy said he often uses a method in which he turns on voice mode and speaks freely for about 10 minutes about project-related thoughts, rather than polishing sentences when talking with AI.

In a post on X, formerly Twitter, on July 22, Karpathy described the approach as a "stream of consciousness" that is "completely messy and anything goes."

He said that if users deliver fragmented ideas and concerns all at once, rather than spending time trying to craft a finished prompt from the start, a large language model can organise them more systematically.

"Sometimes a large language model needs more information to understand what the user is trying to achieve," Karpathy said. "It is effective for a person to speak half-formed ideas as they are, and for a chatbot to reconstruct that monologue into a clean output."

He said the method is also simple in practice. He starts by telling the AI his thoughts may be mixed up and that typos or wrong expressions may be included during voice input. He then speaks at length about the project and has the AI summarise and structure the main points. Karpathy said the process helps align context and reduces what needs to be revised later.

His approach has not persuaded all users. Some X users responded sceptically, saying a long voice monologue does not always produce better results. Others shared examples of voice input and tried testing its effect.

The remarks also come as big tech companies step up competition over voice-based AI features. OpenAI earlier this month unveiled a new model, GPT-Live, that runs ChatGPT’s voice function. It also introduced a $230 mini keyboard with a dedicated voice recording key, presenting an environment in which voice prompts can be delivered directly to an AI agent.

Anthropic on July 24 also announced a voice function update. Users can choose their preferred model, including Claude’s Opus, Sonnet and Haiku, to use voice conversations.

In the industry, generative AI input is seen as gradually expanding from text-centric to voice-centric. Karpathy’s suggestion is also interpreted as an approach to use voice not just as an input method but as a new interface for organising human thinking and dividing roles with AI.

Karpathy stressed that AI users do not need to feel burdened to write perfect prompts from the start. He said letting AI structure incomplete ideas and long monologues may help achieve desired results with fewer revisions.

As voice-based AI functions rapidly become more advanced, attention is focused on whether this kind of conversational approach can take hold among users as a new way to use them.