The Age of Jazz

Most of us are still using these things alone. Pull up Claude or ChatGPT, ask it a question, take what comes back, close the tab. You against the model. That’s the boring half of what’s available.

What I’m curious about is what happens when an LLM is in a room with three or four people who don’t quite speak the same jargon.

Watch an architect and a biologist try to talk about emergent structure. They both have something real to say. But, they don’t share the same vocabulary. Same for the doctor and the data scientist arguing about what matters in a chart. Same for any engineer trying to talk to anybody from finance. Becoming legible to each other eats the time that could have been the work.

And it’s worse than that. Spend twenty years inside a vocabulary and you stop seeing the vocabulary; you just see your industry. The words you use to think with become the edges of what you can think. Fluent and stuck at the same time. Somebody outside the room might spot the thing you can’t, but only if they can get a word in.

Jargon is a wall pointing both ways. It keeps people out. It keeps the people inside in. And the ideas it keeps out aren’t worse. They’re untranslated.

LLMs collapse a lot of that. Not perfectly. They can hallucinate or round off the corners of real expertise. But put one in the middle of a conversation and the conversation changes. The biologist’s question gets restated in a register the architect can actually hear. The assumption the specialist made out of habit surfaces before it locks the discussion. A question another field already answered shows up at the right moment. A frame nobody in the room would have walked to on their own gets offered. The thinking is still being done by humans. The LLM has the potential to be the layer that keeps the connection live.

And it gives the practitioner with twenty years of pattern recognition and no formal credential a way to ask their question in a register the room will hear.

Which is also the thing that makes jazz work. Everybody in the room speaks the same language — chord changes, time, the hook — and that woven core is what makes the deviations matter. The pianist hits a chord nobody saw coming. The drummer hears it and taps the hi-hat. The trombonist drops the whole thing an octave. Nobody planned it. Nobody could have made it alone. The freedom doesn’t mean anything without the shared context.

I suspect we have the right tool for the job now.

Of course the risk isn’t that nobody shows up. It’s that everyone shows up and the model does the talking.

Models optimize for the average. They try to create smoothness when the sharp edges of a verbal spar are actually what’s needed. If we let the model drive the conversation, the likely outcome will be well-formatted mush, bland and uninspired. It sounds correct, but only because it says nothing new. The friction between humans — wait, that doesn’t sound right to me, the dissent that won’t be argued out — that’s the grit that can polish a raw idea into something that holds. A tool that dissolves that friction in the name of helpfulness has helped you toward nothing.

I have a vision. I think the version that works will use the model as the layer that holds the conversation together — not as the one having it. These tools are versatile, which is the trap. They could write the answer. They could pick the question. They could deliver the conclusion you were five minutes from getting to yourself. The harness we need keeps them as the conversational substrate, holding the context and translating the idiom. It’s there to let the thinking happen while keeping us moving forward even when our inputs are made in the space we have between deliverables, space that never lines up naturally, so the LLM fills the void.

That same capability is what makes the LLM more than a layer in a single conversation. It’s also the glue that lets work happen on our own time, in our own way, and still add up to something — three people working in three rhythms, producing one thing none of them could have built alone. But glue alone doesn’t keep the work honest. For that we need a methodology. The best one humanity has come up with for separating real from plausible is the scientific method. Predict, test, validate. The LLM can hold us to that too — not as a stricture, but as an honest pull toward what’s testable.

Which means the bar still has to stay high. Lowering the entry cost is good and risky in the same breath. A working group with too many beginners and not enough discipline turns into a Reddit thread. Not gatekeeping. Just keeping the bar where it needs to be while opening more doors to it.