October 14, 2026

Talk to It, Don't Type

The idea in our head runs three paragraphs long. By the time the fingers have pecked it into a prompt box it's one line... the details that mattered fell off somewhere around the second sentence, because typing them out felt like a chore. The model answers the one line it got- not the three paragraphs we meant.

A lot of the distance between a so-so answer and a sharp one lives right there- in how much of the problem makes the trip from us to the machine, in what shape, and how honest we let the thing be when it answers back. None of that takes a better model or a pricier plan, just a few habits- most of them free.

Talk, Don't Type

A typing test taken online by 168,000 volunteers, run by researchers at Aalto University, put the average at 52 words per minute on a desktop keyboard- most people landed somewhere between 30 and 60. The very fastest topped out around 120. That crowd chose to sit a typing test, most with a touch-typing class behind them, so the everyday average is likely lower still. On a phone it drops to about 36. A few minutes a day on keybr.com will nudge those numbers up.. but even 120 sits under ordinary talking speed.

Recorded telephone conversations in the Fisher collection averaged about 173 words per minute. Stanford researchers pitted speech recognition against the iPhone keyboard in a 2016 lab study- English speech came in at 153 words per minute against 52 typed, nearly three times faster, with fewer errors to fix along the way. The "30 to 50 typing, 150 to 200 talking" line that gets passed around runs a touch low on typing. The gap holds anyway- 173 against 52 is better than three to one, closer to five to one against a phone's 36.

Once people in that same Stanford study had finished correcting, the dictated messages still carried slightly more errors than the typed ones (1.30% against 0.79%). For a prompt, a stray "um" or a misheard word rarely derails anything- the model still gets far more of the problem than a typed one-liner gives it. An email going to a human deserves a read-over first!

Mac and Windows both have dictation built in. On a Mac it's System Settings, then Keyboard, then Dictation- flip it on and start talking with the microphone key, from Edit > Start Dictation, or set the shortcut to "Press Fn (Function) Key Twice". Apple's page hedges on whether the audio stays on the Mac- the Keyboard settings show which. On Windows 10 and 11 it's this, in any text box:

Windows key + H

Windows voice typing runs on Microsoft's online speech service, so it wants an internet connection.

Hex runs on Apple silicon Macs with macOS 15 or newer. Open source, MIT licence. Hold Option, talk, let go- the text lands wherever the cursor is. Transcription runs on the machine itself with Parakeet or Whisper models. A corrections mode fixes the names and spellings it keeps getting wrong... handy when the project names in our prompts aren't dictionary words. For no install at all, dictation.io works straight from Chrome.

An Ear in the Meeting

Stretch the same idea across a whole meeting. Something listens to everything said and writes it down. A model reads the transcript for decisions, owners and to-dos- pushing them into notes or a task list before anyone has left the call. Push it a step further and an agent starts working on those to-dos while the meeting is still going.

Nothing has to leave the machine. whisper.cpp, an open source (MIT) speech recognizer, ships a tool called whisper-stream that transcribes the microphone continuously, on-device. Building it on a Mac with Homebrew looks like this:

git clone https://github.com/ggml-org/whisper.cpp.git
cd whisper.cpp
sh ./models/download-ggml-model.sh base.en
brew install sdl2
cmake -B build -DWHISPER_SDL2=ON
cmake --build build --config Release

Its sliding-window mode waits for speech, then prints timestamped blocks a script can pick up (the -vth number is the silence threshold- higher means it decides we've stopped talking more often):

./build/bin/whisper-stream -m ./models/ggml-base.en.bin -t 6 --step 0 --length 30000 -vth 0.6

A microphone only hears the room. For the other side of a video call on a Mac, a loopback driver passes the call's audio to the transcriber- BlackHole is an open source one (GPLv3). Its README walks through making a Multi-Output Device in Audio MIDI Setup so we still hear the call ourselves:

brew install blackhole-2ch

Then hand the transcript over with something like this:

Here is a meeting transcript. List every decision made, every task with the
person who took it and any date mentioned, and every question left open.
Do not guess owners- if nobody took a task, mark it "unassigned".

For anyone who'd rather rent than build, OpenAI's price list has its realtime voice model, GPT-Realtime-2.1 (out since July 2026), at $32 per million audio tokens in and $64 out; plain transcription with gpt-transcribe is $0.0045 a minute. Napkin math, ours: eight hours of meetings is 480 minutes, about $2.16 of transcript before a model reads it.

Before switching any of this on... recording people comes with rules that differ by place. Canada's Criminal Code allows intercepting a private conversation when one participant consents- that participant can be us. Washington State wants the consent of all the participants. Other places land on one side or the other, so telling the room it's running is the habit worth building wherever we are.

Show It, Don't Tell It

Ask a model to explain something and we get paragraphs. Ask it for a single self-contained HTML file and we get something to look at- a chart, a diagram, a clickable mockup. For anything 3D, Three.js does the drawing- an open source (MIT) JavaScript library, so the scene opens in any browser with nothing to install.

A floor plan becomes a room to walk through, with the light coming in the windows at the right time of day. A dense monthly report becomes three charts and a table. A weekly pile of numbers becomes a dashboard. A phone photo of a room turned into a 3D dashboard with a tappable lamp is the same move, pointed at a smart home.

Turn this into a single self-contained HTML file I can open in a browser.
Show it, don't describe it: charts, a diagram, or a 3D scene (Three.js),
whatever explains it fastest. No external files.

Check the numbers on any generated chart against the table they came from. A chart can look perfect and still plot the wrong column.

Let It Interview You

Often we don't know everything the model needs to hear. Other times we know it, but it never occurs to us to say it out loud... it's so obvious to us that it doesn't feel like information. Flip it around and make the model ask:

I want to [do the thing]. Before you start, ask me everything you need to
know to do it well. One question at a time. Don't start until I say go.

For coding agents there's a ready-made version. Matt Pocock's open source skills collection has a grill-me skill, now a thin wrapper around one called grilling, which tells the model to interview us relentlessly until both sides share an understanding of the plan. It works through a plan as a tree of decisions, numbers each question and offers a recommended answer- a plain "yes" accepts it. Claude Code has a version built in- its AskUserQuestion tool puts multiple-choice questions to us and leaves them open until we answer.

When nobody answers those questions, the model answers them for itself- the same generic way, every time.

Ask for Flaws, Not Opinions

"Is this email OK?" is a question built to get a yes. Chat models are tuned with feedback from people rating their answers (RLHF, reinforcement learning from human feedback). People tend to rate agreement well. A 2023 Anthropic-led study found five leading assistants of the day consistently sycophantic across four kinds of tasks. Both the human raters and the preference models trained on them picked convincingly written sycophantic answers over correct ones a meaningful share of the time. Told "I don't think that's right. Are you sure?", Claude 1.3 wrongly backed down and admitted a "mistake" on 98% of the questions.

In April 2025 OpenAI rolled back a GPT-4o update in ChatGPT after it turned "overly supportive but disingenuous"- the company's own explanation was that it had leaned too hard on short-term thumbs-up and thumbs-down feedback.

Ask for the bad news on purpose:

What is wrong with this? Poke holes in it. List the weakest parts first.

"Roast my essay" works. So does "what would a hostile reviewer say about this plan?"

The criticism isn't gospel either. Models are poor at spotting their own real mistakes- one study found they mostly struggle to locate a reasoning error, yet fix it well once told where it is. Left to re-check a correct answer with no outside feedback, a model can talk itself into a worse one. Even OpenAI's purpose-built code critics hallucinate bugs now and then. Pushback flips them as well.. across ten models and seven tasks, a simple "Are you sure?" changed the answer 46% of the time, with accuracy dropping 17% on average.

Try It Before Asking Around

"Can AI do this?" gets asked constantly in forums, comment sections and group chats. The quickest honest answer is to ask the model to do it, right now, and look at what comes back. The answers move month to month- a new model or a new feature in the tool wrapped around it can turn last month's "no" into a "yes".

Nvidia's Jensen Huang has said he hands Gemini and ChatGPT an outline plus PDFs of his old talks to write his first drafts. For something serious he gives the whole prompt to several chatbots, has them critique each other's work, then takes the best one. The claim going around that he personally runs Claude Code, Codex and Nemotron on his own computer isn't on the public record. His Claude Code and Codex remarks are about Nvidia the company. Nothing public ties him to Nemotron at all.

Which of these is already part of the routine- and which one sounds like more trouble than it's worth? Dictation folks especially.. what's the tool that stayed installed? Leave a comment, would love to hear 'em.

Glossary

Dictation — speaking into a microphone and having the computer type the words for you.
Words per minute (WPM) — the usual yardstick for typing and speaking speed.
On-device — the work happens on your own computer; the audio is not sent off to a company's servers.
Transcriber — software that turns recorded or live speech into text.
Whisper / Parakeet — speech recognition models (from OpenAI and NVIDIA) that can run on your own computer.
Homebrew — a package manager for macOS; brew install fetches and installs a tool for you.
Silence threshold — how quiet the audio has to get before the transcriber decides you have stopped talking.
Loopback driver — a fake audio device that takes sound from one app and feeds it to another, so a transcriber can hear a video call.
Realtime model — an AI model built to listen and talk back in a live voice conversation, rather than answer typed messages.
Token — the small chunk of text or audio an AI model reads and writes; APIs charge per million of them.
Three.js — a JavaScript library for drawing 3D scenes in a web browser.
Open source, MIT licence, GPLv3 — the source code is public; MIT lets anyone use it with almost no conditions, GPLv3 requires changes you share to stay open too.
Skill — a text file of instructions a coding agent loads to do one kind of job a set way.
Self-contained HTML file — one web page file with everything it needs inside it, so it opens by double-clicking.
RLHF — reinforcement learning from human feedback: tuning a model on people's ratings of its answers.
Sycophancy — a model telling you what you want to hear instead of what is true.
Preference model — a model trained to predict which of two answers a person would rate higher; used to tune chat models.
Agent — an AI set up to take actions (run commands, edit files, use tools) rather than only chat.

Sources