Machines Don't Have Taste

I was building a feature for a screen companion app — a tool that watches your screen and helps you learn software by voice. Think of it as a tutor that can see what you see and point at the thing you need to click next.

The question came up: what if the user wants to replicate something they saw in a video? A CapCut tutorial, a design showcase, a motion reel. How does the tool ingest that reference and hold it across a session?

Simple enough for tutorials. Fetch the transcript, pin it as context, let the AI ground every answer in the specific steps the user is following. Straightforward engineering.

Then I thought about a specific video I’d seen. A motion designer who also happened to be learning chess. He used his CapCut skills to narrate his learning — not with words, but with motion. The way a piece name flew in was timed to a beat drop. The way a capture animation synced to the tension of the position. The audio wasn’t the content. The motion was the content.

And I realised: no transcript captures that. No single frame captures it either. The thing that made the video remarkable existed between the frames — in timing, in easing curves, in the relationship between sound and movement. It was temporal. It was felt.

The AI could not watch that video and “get it.”

I could.

The taste layer

There is a layer in every domain that machines cannot reach. It is the layer where you watch something and feel that it works — not because you can articulate every technical decision, but because you have spent enough time in the domain that your nervous system recognises quality before your conscious mind catches up.

That is taste. And taste is not transferable to a prompt.

You can tell an AI “make the text fly in from the left.” You cannot tell it “make it feel the way that video felt.” The feel is a compression of hundreds of micro-decisions — the easing curve, the hold duration, the relationship to the audio, the negative space before the next element enters. You learned those micro-decisions by watching, making, failing, and watching again. The AI learned a distribution over what text-fly-in parameters tend to co-occur in training data.

These are not the same thing.

Two types of operator

Every tool creates two types of user.

The first type understands the domain. They have taste. They know what good looks like before they can build it. The machine closes the gap between their vision and their execution. They describe the motion they saw, the AI helps them build it in CapCut — keyframe by keyframe, parameter by parameter. The human is the taste layer. The machine is the execution layer. This works.

The second type does not understand the domain. They want the machine to have taste on their behalf. “Make it look like that video.” The machine produces something. It is technically correct. It is also dead. It has no feel, no timing instinct, no reason behind any specific choice. It is a distribution average wearing the skin of a creative decision.

The first operator is amplified by the machine. The second is limited by it.

What machines actually do

LLMs do not generate new knowledge. They search. At extraordinary scale, across extraordinary breadth, with extraordinary speed — but they search. They find patterns in what humans have already explored. They surface connections humans missed within existing domains. They recombine.

This is genuinely powerful. The space of unexplored combinations within known domains is effectively infinite. Most of what we call new ideas are recombinations anyway. The machine accelerates this by orders of magnitude.

But recombination is not the same as origination. Nobody prompted an LLM and got general relativity. Nobody will. The architecture finds patterns. It does not look at the universe and think “what if space itself bends?” That requires a kind of contact with reality that statistical models do not have — because they have never lived in the reality they model.

The motion designer who narrated his chess learning through animation — that idea came from living in both domains simultaneously. Feeling the tension of a chess position and knowing, from years of motion work, that a specific easing curve could express that tension. That is domain fusion through lived experience. No model has lived experience. It has training data.

The implication for builders

If you are building with AI — and you should be — the question is not “what can the AI do?” The question is “what do I know that the AI doesn’t?”

Your moat is not the tool. Your moat is the taste you bring to the tool.

The operator who understands their domain deeply and uses machines to expand their reach within it — that is the right model. The operator who expects the machine to have taste on their behalf will always be disappointed.

So: what domains do you have genuine taste in? Not domains you have read about. Domains where you can feel when something is wrong. Where you have opinions you cannot fully articulate. Where you notice things others walk past.

The machine will expand your reach in those domains faster than anything in history. But it will not give you the domain. That part is yours to earn.

Build as a thinker. The machine is waiting for instructions from someone who actually knows what to say.