The Baker's Thumb: Allowing AI to See What It is Missing
When human deviation is detected but may be missing data.
A baker discards a crust that looks perfect. It didn't yield to her thumb.
To an AI system optimizing bakery operations, this looks like waste--a deviation to correct. To me, it looks like intelligence. Knowledge residing in her hands that the recipe cannot capture: the humidity in the air, or the specific texture of the flour.
When we optimize with AI, we capture temperature, timing, measurements, and process. The knowledge of the thumb is missing.
The Incomplete Map
We are currently building systems that optimize for the map while being unaware of the terrain.
We see it when a delivery driver navigates gridlock. He diverges from the GPS not because he is inefficient, but because he has a "shadow resume" of experience. He knows the loading dock is actually around the block, or that a specific street floods in the rain.
To the system, these are errors. To us, they are agency.
The question isn't whether AI should correct errors. It should--it corrects mine all the time. The question is whether AI can learn to ask: Is there expertise here I'm not seeing?
The Experiment
I am not a coder; I am an observer of these quiet human moments. But I wanted to know if we could translate this "shadow resume" into something a machine could respect.
I gathered pieces of my work that narrate human agency--from pastry chefs to tradespeople--and ran a simple experiment. I did not change the model's code; I changed its context. I simply asked the model to inquire before optimizing.
The result? When the model was allowed to consider human intent, it stopped treating divergence as failure. In preliminary tests, it became significantly more accurate at recognizing expertise by showing the model that human agency exists.
A Mosaic of Human Intelligence
I have created this work not as a software product, but as a map of intangible intelligence.
The Repository: A collection of "Portraits of Human Intelligence" (PHI). This is a body of written work I used to translate what the model could not see.
The Evidence: For those interested in the methodology, I have a preprint available.
I am sharing this because we don't need AI that can only follow the map; we need AI that understands why the pilot pulled back on the stick.
If you are interested in the technical details, please contact me for a draft arXiv Preprint: phi@joesterly.com
Acknowledging the photographers who share images on Pixabay under CC licensing, and especially Shirley Hoy for the cover image.