Application: Agency in the Age of AI

Why I Built This

Most mornings I open a conversation with an AI and work on something I don't yet understand well. The goal is to come out knowing more and able to do more than I did going in. If the thinking happens in the machine and stays there, I have gained nothing, so I build my working habits around keeping it.

Two leaks

The first leak is meaning that is made and never taken up. By taken up I mean internalized, so that it becomes something I know. The gap looks like reading a few pages and not remembering a single word. An hour of work produces an insight, a cleaner framing, a decision about what to do next, and a week later it is gone, or I am acting as if it never happened. I call this the uptake gap.

The second leak is quieter. AI output is the wording most likely to be believed. It arrives shaped like a conclusion, and a reader can take it up without ever seeing what it was a result of. It is like receiving the end result of an experiment without knowing the hypothesis. Two mistakes follow, and they will look familiar to anyone who has met Type I and Type II error. I can accept something that nothing supports, or I can discard something sound because I can't see why it works. Nothing in the prose helps me tell the two apart, since believable and grounded read the same on the page.

Why it happens to me

I am monotropic. Monotropism is a way of attending in which attention narrows into one interest at a time and goes very deep. Deep focus makes insight easy. The mundane work of carrying an insight forward, writing it down, connecting it to the last thing, coming back to it, is what falls away, and without it nothing is internalized. That is a predictable pattern, and a predictable pattern can be designed around. I expect other people lose meaning the same way for other reasons: a full calendar, a noisy week, a tool that forgets.

Tools as corrections

So I build tools that correct for it. The approach matters more than the list. I design around the way I actually work and treat each failure as information about what to build next. The first is Connective Tissue, my name for a system of shared anchor files and update protocols. It links chats and projects so they move together. Each session starts from the current state of the work, with the reasoning kept alongside the conclusion, which is the answer to the second leak. The AI I use most is Claude, and nothing here depends on that choice. Continuity of that kind belongs to the files. Making the material mine takes a practice, and that is the part I am still building, something closer to a pedagogy for working with AI. How the system works is laid out on my work page.

What I think AI is for

Belief Space is the sum of what human minds have so far perceived or conceived. Each of us holds a region of it, and a region has edges, the places where what we can see runs out. I see three possible roles for AI at those edges, in ascending order.

The first is finding the edge. That is a diagnosis: where my understanding stops, and what is actually constraining me, which often differs from the problem I walked in with. The second is helping me cross it, by supplying what I need for the next step and letting me test that step. The third is leaping to a belief space no one has occupied before. That is a possibility I am not yet claiming.

The uptake gap sits between the first two roles. An edge that is found and not internalized is only a note. So I try to end each session with a next step chosen for what it will teach me whatever happens, which gives the result a hypothesis to belong to. Where things stand: edge-finding works. Crossing is partial. Leaping is an aspiration.

Where this goes

Meaning that gets taken up compounds in the person who took it up. A transcript just sits there. The tools exist to produce the first kind.