Speak to it (audio input)
Stop typing. Roughly half his queries are voice — and on the phone, most of them.
The first modality is the easiest win: don't type, talk. On the phone the mic is built in — it transcribes your speech to text and you send it. On desktop, ChatGPT lacks an inline mic, so he uses a system-wide speech-to-text app (bind a key, talk, it types for you). Karpathy: about half of his queries — most on mobile — are just voice, "because I'm lazy and it's just so much faster."
On the ChatGPT phone app there are two different voice buttons. The plain microphone icon is the one that matters here: it listens and transcribes your words into the text box. You still send text — you just didn't type it.

The desktop app has no mic button, so he uses a system-wide dictation app — SuperWhisper, WisprFlow, MacWhisper — bound to a key (he uses F5). Press, speak, press again; it types into any app, not just ChatGPT.
“The big takeaway here is: don't type stuff out. Use voice. It works quite well, and I use this pervasively — roughly half of my queries are just audio because I'm lazy and it's so much faster.”
Andrej Karpathy·1:21:46
Dictation mangles product names, library names, and technical terms. When exact spelling matters, type it. Everyday phrasing? Just talk.
The load-bearing points
- Phone: built-in mic = dictation (speech → text you then send).
- Desktop: use a system-wide STT app (SuperWhisper / WisprFlow / MacWhisper) on a hotkey.
- ~half of his queries are voice; most of his mobile ones.
- Type the proper nouns — dictation misheres them.
Talk for a day
For one day, dictate your prompts (phone mic, or a desktop STT app) instead of typing. Notice the speed change.
Show the point
His claim is ~2× faster for everyday phrasing. The friction is only in names and jargon.
Bind a hotkey
Install a system-wide STT app and bind it to a key. Dictate into a non-ChatGPT app to prove it's system-wide.
Show the point
That's the desktop unlock — voice input everywhere, not just inside one chat app.
For exact strings dictation gets wrong — product names, library names, technical terms. Everyday phrasing is fine by voice.