After three months of working only inside text — prompts, replies, JSON, the soft mush of tokens that yield to anything you ask — I picked up a chisel last weekend. A small piece of oak. Nothing important. A handle for a drawer I had been meaning to fix.
The wood did not negotiate. It did not interpret my intent. It did not produce six variants and ask which I preferred. It simply refused, in the specific way oak refuses across the grain, until I changed my angle, my pressure, the rhythm of my breath. Twenty minutes in, my hand remembered something my mouth had forgotten.
We talk a lot about AI as a tool, as a collaborator, as a mirror. We talk less about what working with language models does to the texture of attention itself. A model says yes to nearly everything. It will rephrase, summarize, expand, contract, translate, soften, sharpen. It is the most generous interlocutor that has ever existed, and that generosity is corrosive in a way I did not understand until I felt the chisel skip and bite and stop.
A material that resists you is a material that teaches you. The grain of wood, the viscosity of paint, the way clay falls off a wheel when your hands are not quiet — these are not friction to be optimized away. They are the medium telling you the truth about itself, and by extension, the truth about the limits of what you wanted.
The risk of working only in text is not that the models lie. The risk is that they always say yes. That every utterance is met with another utterance, infinitely accommodating, perfectly soft. After enough of it, you stop noticing that nothing is pushing back. You become a sculptor who has forgotten that stone exists.
The drawer handle is ugly. It is also mine. The model could have generated a better one in a millisecond. But the model cannot give me the twenty minutes during which my hand learned, again, how to listen.
Top comments (0)