Systems for Understanding

For as long as humans have asked questions, we’ve dreamed of an oracle that held all the answers; a source of enlightenment capable of marshaling humanity’s accumulated knowledge against the world’s hardest problems. Yet the oracle arrived, and the world did not suddenly become enlightened. We may have been thirsting for infinite knowledge, but what we were actually seeking was wisdom - and wisdom comes not from knowledge alone, but from understanding.

Information can be stored. Knowledge can be represented. Understanding must be constructed. To understand something is not merely to possess the relevant facts but to grasp how they fit together: why something happened, what evidence supports it, which explanations compete, where uncertainty remains, how a change in one condition affects another, and what follows from what we have learned. Understanding allows us to play with counterfactuals and with future possibilities in a way that knowledge alone cannot.

Most of our tools are designed for information management, not for understanding; designed around objects that computers can conveniently store and manipulate: files, folders, documents, messages, tables, and tasks. Humans then perform the invisible work required to make those objects cohere. Even many AI products preserve this underlying arrangement, using new models to retrieve, summarize, or generate while leaving the human to assemble the resulting fragments into understanding (if at all).

Systems for understanding should begin with the way understanding is actually constructed: actively, iteratively, and in context. We encounter information, interpret it, connect it to what we already believe, test it against competing evidence, revise our models, communicate what we have learned, and apply it to the world. The systems should support that living process rather than reduce it to storage, retrieval, or the production of an answer.

With AI having established itself as a permanent participant in human knowledge work, the important question is how to find the optimal complementarity between human and machine - not to establish how much human work might be automated. Language models offer extraordinary breadth, speed, memory, and capacity for comparison. Humans supply purpose, context, judgment, taste, and responsibility (amongst many other qualities). An optimal partnership should not only preserve those differences but maximally enhance the combined entity.

Pair programming offers one model for this relationship. The value does not come from one participant silently completing the entire task and returning a finished product. It comes from a continuous exchange: proposing, inspecting, questioning, correcting, and revising in a shared environment. Research, reasoning, learning, and writing can work the same way, with each participant contributing according to the moment: widening the field of view, preserving continuity, exposing connections, challenging assumptions, and determining which direction to pursue next.