Why Everyone's Talking to AI (With Their Voice)

Think about the last prompt you typed to AI.

You probably trimmed it before you sent it. You skimped on the backstory that felt like too much, dropped the tangent that was only half relevant, and tightened the rest into a couple of clean lines.

You wrote it the way you'd write a Slack message to a coworker who is slammed: polite, efficient, respectful of their time.

That instinct is a good one, and it's a big part of why people like working with you. But it’s also working against you every time you chat with AI.

AI doesn't need the courtesy. It reads your backstory and your tangents as fast as you can hand them over, and it's good at ignoring the parts that don't matter. The messy, unedited version of what's in your head can often be the better prompt.

So today, I want to break down:

• Why the compression you learned from humans backfires with AI

• Why talking gives more context than typing (and takes way less time)

• The newly updated voice feature you should try today (instructions below)

—@dharmesh

Compression Is for Coworkers

Think for a second about why you compress a message to a human.

Their attention is finite and you want to be respectful of their time. Nobody wants to be hit with a novel-length email they have to parse for 10 focused minutes before replying. So you learn to strip it down. Say the thing, skip the preamble, respect the clock. It's a useful skill, and most of us have been practicing it our entire working lives.

Almost none of it applies to AI.

The model doesn't get impatient at paragraph four. It reads the whole mess as fast as you can hand it over, and it is very good at ignoring the parts that don't matter.

Every writing habit you've picked up in your career was built for an audience with limited attention. Short emails, give the bottom line up front, keep it under a page -- all of it exists because human attention is scarce and you don't want to spend more of it than you have to.

AI is the first audience where that constraint just isn't there. When attention stops being scarce, the goal flips: stop trying to say as little as possible, and start handing over as much as you've got. Context is queen, after all.

Why Talking Beats Typing

Once you accept that more context beats less, the next question is a practical one: what's the fastest way to get what's in your head into the model?

For most folks reading this, the answer isn’t your keyboard. It’s by speaking.

You talk much faster than you type, and speed is only half of the advantage. A second-order consequence of speaking to AI is that you don't self-edit out loud the way you do inside a text box.

Now, plenty of you have probably tried talking to your computer or smartphone before and decided it wasn't for you. I understand why. Speech-to-text has been around for decades, and for most of that time, it forced you to talk like a robot.

You simply couldn't speak naturally. If you wanted "Can you pick up milk today?" you had to actually say the words "question mark" out loud. The software transcribed what it heard, but it didn't truly understand anything. It was not an enjoyable experience.

Thanks to AI, that's now changed.

Speech-to-text powered by AI is much more powerful today. You can talk the way you'd talk to a person -- pauses, false starts, tangents and all -- and the punctuation and formatting will be taken care of automatically.

There are dedicated apps for this that you can use across any application on your computer and smartphone. Willow and Wispr Flow are two of the big ones that you’ll like run into online (note: I’m an indie investor in Willow). Both have a free trial -- but I should note you don’t need a third-party app to test speaking to your AI apps if you’re not ready to take that step just yet.

Similar functionality (without a few bells and whistles the paid apps offer) is already sitting in the apps you're using: the dictation button in ChatGPT or Claude runs through AI as well. Tap the mic and just start talking.

Last week OpenAI released an upgraded voice model that’s supposed to make talking with AI feel much more like having a real conversation. It’s very cool and I’d recommend you try it.

The blue arrow points to the speech-to-text feature that allows your speech to be converted to text. The orange arrow points to the Voice Mode option that was recently upgraded, which drops you into a one-on-one conversation with AI.

What makes this new model (GPT‑Live) powering the upgraded Voice Mode special is that it continuously processes input while generating output. If you ask it to research something, like the best dad joke of all time, it will do so mid-conversation without skipping a beat. It’s really quite something, and I recommend trying it yourself.

Try It Yourself

Speech-to-text isn’t for everyone, but if you haven’t tried it recently, I’d recommend revisiting it to see how much it’s improved today:

• For work: OpenAI’s ChatGPT, Anthropic’s Claude, and Google’s Gemini all have a similar microphone icon you can press to try turning your stream of consciousness into a prompt.

• For fun: Open ChatGPT on your phone, tap into voice mode, and start a conversation.

Remember -- don’t put much effort into compressing your thoughts. If you feel like you’ve rambled on too long or gone too off-track, you can always end your voice note with something like, "Okay I’m done. Pull out what actually matters here, organize it, and tell me what to do next."

If you had tried this a year ago, you would have had a completely different experience. Which is why I like to bring up my own rule about this: in a world where capability expands exponentially, retest your assumptions every six months.

Hit reply after you’ve tried speaking to AI -- I’d love to hear how it went. I try to reply to a few people every week, but I always read every response 🙂 

—Dharmesh (@dharmesh)