Andrej Karpathy Advocates Voice Mode for AI Agent Alignment
X ● Covered by 2 sources
Karpathy says stop typing perfect prompts, just talk to your AI agent for 5-10 minutes in voice mode. Rambling out loud apparently aligns the AI with your actual goals better than careful text.
Based on reporting by X — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Andrej Karpathy has a new piece of advice for anyone wrestling with AI agents, and it has nothing to do with prompt engineering tricks or clever system messages. His suggestion: stop typing, start talking. Specifically, open up voice mode and just ramble for five to ten minutes about what you're trying to do, why you're trying to do it, and what's worrying you about the task at hand.
The logic here is less about the AI understanding words and more about it understanding you. A tightly written prompt captures an instruction, but it strips out all the context that lives in your head — the half-formed doubts, the tangential concerns, the reasons behind the reasons. Karpathy's bet is that unscripted speech drags that context out whether you mean it to or not. You say more than you planned to. You circle back and correct yourself. You mention the thing you're worried about almost as an aside, and that aside turns out to matter.
This is a notable departure from the dominant style of agent interaction, which has mostly trained users to be terse and precise, as if talking to a search engine that punishes extra words. Karpathy is essentially arguing the opposite: extra words are the point. An agent that has absorbed your rambling has a richer model of your intentions than one that received a crisp, minimal instruction, even if the crisp instruction was technically correct.
It's a small piece of advice, but it says something bigger about where alignment problems actually live day to day. Most people aren't fighting with AI over deep value misalignment — they're fighting with it over shallow misunderstandings of what they actually wanted. Karpathy's fix doesn't require better models or new training techniques. It just requires talking to the thing like you'd talk to a new hire you're trying to bring up to speed, not like you're filing a support ticket.
My take — AI-written commentary, not fact-checked reporting
I'll believe voice-rambling is a real alignment technique and not just a Karpathy quirk when someone runs a controlled comparison instead of vibes, but the underlying point is right regardless: most of what people call 'AI misalignment' is actually just under-specified intent, and no amount of clever RLHF fixes a user who didn't say what they meant. Also, notice how casually we've arrived at 'talk to your AI agent for ten minutes like a therapy session' as mainstream advice from one of the field's most respected voices — a year ago that sentence would've sounded like satire.
Read more about this at: X