Understanding Is the New Bottleneck
As AI automates execution and synthesis, the critical bottleneck for knowledge workers shifts from producing output to understanding what to do—requiring new skills like asking good questions, evaluating AI suggestions, and navigating ambiguity.
Background
- Geoffrey Litt is a researcher and engineer who previously worked on dynamic personal software tools at MIT and Ink & Switch, often focusing on bringing fluid, spreadsheet-like interactivity to unstructured data.
- This essay argues that as AI code generation (e.g., GitHub Copilot, Cursor, Claude Artifacts) makes *writing* code cheap and fast, the slowest step in building software becomes *understanding* existing codebases — what they do, how they fit together, and whether a change is safe.
- The "new bottleneck" idea contrasts with earlier eras where writing code was hard (typing) or where reading code was hard (no good tools). Litt claims we now have excellent tools for writing and decent tools for reading, but the *understanding* required to make architectural decisions hasn't been automated.
- The piece proposes that future programming tools should shift from helping us *write* programs to helping us *understand* them — e.g., interactive queries about code structure, simulating hypothetical changes, or visualizing dependencies — rather than just generating more code faster.