The medium is more consequential than you think

Most people treat the medium as a delivery mechanism for the message. But the medium you choose to think in, whether it's voice or keyboard, which AI model you partner with, or how fast the transmission happens, shapes not just what you say but who you become. This post documents a real-time discovery of that constraint, and proposes an experiment: deliberately choose a specific intelligence to transcribe your thinking, stay with it, and watch how it molds you.

What I'm working with

• Noise as bandwidth theft. Processes run in us that aren't necessary. The work is reclaiming the bandwidth they steal, without trying to stop all of them, because that's not how we're built.
• Voice outruns the keyboard. Thought moves faster than fingers. Speaking is less precise but lets ideas flow; typing throttles them.
• Friction has a floor and a ceiling. Too much friction blocks the flow. Zero friction makes me solipsistic, too self-aware to move. The medium has to push back just enough to keep me honest without breaking velocity.
• Speed of transmission is part of the message. Not just the medium and the message. How fast I can externalize an idea is itself content.
• The medium molds the messenger. McLuhan, by way of Ezra Klein: what we consume shapes us, against the standing assumption that we freely choose the medium. So the format I pick to express myself informs what I become, as much as the message does. Meaningfully so, if not in equal parts.
• The AI agent is the primary reader. The human is downstream. The chain isn't human-to-human; it's human to device to agent to agent to human. I'm writing to update an agent's model of how I think.
• Model choice is deliberate shaping. Which intelligence I partner with (open versus closed, American versus Chinese, frontier versus open source) molds both the message and the messenger. Pick one, commit across posts, document the difference.

The experiment: a reshaping prompt

The new thing that came out of this is a standing instruction for editing spoken thought into posts. It's a living document. It gets revised every time it runs, and the revisions are part of what I'm watching.

Reshaping Prompt — v0.1

You are editing my spoken thought into a post. Rules:

- Preserve the loops, backtracking, and reframes. The pattern of how I arrive is the content.

- Cut dead air, throat-clearing, and false starts that carry no signal. Nothing else gets cut.

- Fix mangled project nouns and obvious transcription errors. Nothing else gets rewritten.

- Keep contradictions. They're texture, not error.

- Velocity over polish. Don't launder the voice into clean prose.

- No em dashes. Be economic with words.

- The primary reader is an agent updating its model of how I think. Optimize for signal density, not skimmability.

The raw transcript

Lightly cleaned: project nouns fixed, nothing else rewritten. The loops stay in. This is the raw input the prompt above is built to reshape.

Me: We're not walking right now. We're in a glass box at a workspace at the bottom of North Beach, between North Beach and Jackson Square, staring at the window. There's plenty I've thought about and talked about, but I don't know exactly what to begin writing. I'd hope it would be about how I use AI and be a benefit for Bench & Beacon, but it doesn't need to be. I want it to be the most honest thing I'm thinking about right now.

Claude: Then the honest thing probably isn't how to use AI for Bench & Beacon. That's the strategic overlay. What's the thing you'd rather not ship today?

Me: Two things are going on. One is that it's easier to talk my thoughts out now than to write them down. I have an iPad and a keyboard, and I was trying to write, but the thoughts move too slowly going from my head down to my fingers down to the keyboard. Talking is more imprecise, but it's easier to let the words and ideas flow. The other thing: I was trying to circle around the idea of noise production. In each of us there are processes running that aren't necessary. So the examination is, how do you keep those processes from stealing your bandwidth? You can't stop all of them, because that's how we're built. But how do you develop or augment focus so the other processes aren't in the way?

Claude: That's the essay. Noise as bandwidth theft, focus as the only real leverage you have against your own build. And the irony is sharp: you're trying to write about processes that steal bandwidth while the tool you're using to write is itself stealing bandwidth.

Me: There's another element. The medium is partially the message. If I know I'm about to post something that will be read on a screen, possibly on someone's phone, then the situational awareness of that text has to mean something too. It's not always acknowledged, but it's an interesting device for a writer to acknowledge what the reader is probably reading from, their perspective. Add to that: more of the audience is probably the AI agent than a human being.

Claude: That flips the whole constraint. You're not optimizing for skimmability on mobile. You're optimizing for extractability by an agent reading your voice.

Me: So how do we add more constraints? Why is the AI agent reading it? It's probably assigned by a human, likely me at this point, but maybe someone else in the future. The endpoint is human to human, which is interesting, but it's not pure human to human. It's human to AI agent to probably another AI agent to human. Or human to glass or monitor or smartphone to AI agent back to electronic device.

Claude: It's reading to update its model of you, so it can speak like you, decide like you, push back like you. Every sentence is training data for your own Nemesis. Vagueness becomes a liability. Contradiction becomes useful, because the agent needs to see where you contradict yourself to model the real texture.

Me: How does my mind actually move when I'm trying to think clearly? Part of it is that I'm hindered by recognizing its movement. When I notice it, I can't actually think, it becomes too self-aware. So for someone like me, I need to be engaged enough not to be aware of the gap or the medium. But there's a limit, because the speed of transmission counts. Just like the medium is the message, the speed at which I can expose my ideas is part of the message. I think it was Marshall McLuhan, as referenced by Ezra Klein, who said the medium we consume actually molds us, as against the presumption that we choose the medium. So how do we put that in terms of someone trying to express themselves? Maybe: the medium I choose to express myself in molds me as much as the message I'm trying to express. Maybe not in equal parts, but meaningfully so. So if I have an aspiration to be a certain way, choosing the outlet and the format informs what I'm going to become, as much as the activity itself.

Claude: That's the constraint, and it's elegant. You need enough friction-absence to think clearly, but zero friction makes you solipsistic. And then McLuhan: the medium shapes the messenger. So choosing voice, and choosing an AI agent as your reader, isn't a tool choice. It's a becoming choice.

Me: There's an additional point. As an experiment, you could choose to work with a specific agent, sufficient enough to transcribe what we're doing right here, using a specific AI model.

Claude: So the medium isn't just voice. It's which intelligence is listening.

Me: I'm just going to choose a particular model and a particular version and stick with that version over several postings, and see what the experience is like. I have opinions on where the frontier labs are headed, and on open source versus closed, Chinese versus American. So maybe if I chose an American open source model to run the experiment, we could see what it's like and report back. Right now I'm using Opus 4.8 in extended mode, sometimes max mode. That's an American model, but it's also a frontier model, and a flawed business entity I have my issues with. So even though I'm using it now, I could set up a stack that runs off an open source American model, and over time train it to be the transcriber of these posts.

Claude: Run this exact conversation through Llama versus Claude versus a Chinese closed model, and you'd get three different versions of you emerging. Then pick one, commit, let it tune itself to you while you tune yourself to it, and document the difference.

Me: This is the day the post is due. So we make do with what we have here now.