Cognitive Debt.

A conversation between Max Schoening and Geoffrey Litt, on software factories, software labs, and what AI does to understanding. The idea that stayed with me was cognitive debt: what happens when the work moves faster than your understanding.

The frame is factory versus lab. A factory is for work that is already understood, where the goal is throughput. A lab is for work that is not yet understood, where the goal is not the artefact but the understanding produced by making it. AI blurs the two. A prompt becomes code, a loose thought becomes a document, and the artefact arrives before the understanding has formed. That gap is the debt.

The objection is that we never understood the whole stack anyway. True, but a compiler hides machine code, whereas AI can hide intention, rationale, and trade-offs. The question is not whether you understand every layer. It is whether you still understand the layer where your judgement matters.

A test for the debt. Can I follow this, change it, explain it, and recover if it is wrong? If not, the output may still be useful, but debt has been created. The danger is not that machines do more. It is that humans contribute less while believing they are contributing more.

The conversation also sent me down a Geoffrey Litt rabbit hole. His is the territory I have been circling, and his homepage puts it more cleanly than I have. The phrase that stuck was "designing and prototyping environments for thinking." His inspirations page maps the lineage he works from.

It also pointed me at Bret Victor's Magic Ink, which I have not read yet. It splits software into three kinds and treats them as fundamentally different problems:

• Information
• Manipulation
• Communication

This continues the thread from Universal Machine. The best tools do not produce on our behalf. They extend what we can see, try, understand, and change. The question I keep returning to is not where to add agents, but what new senses a tool can give the person using it.