The compiler was supposed to end programming too

The article says programming has been “ending” since FORTRAN, because each wave of automation removes some manual coding while pushing the work up a level. It argues that AI follows the same pattern, but faster, and that the hard part remains deciding what should exist and checking that the result is right.

Reading notes

#

• A made-up 2031 Tuesday shows agents finishing some tickets overnight, but stopping to ask about a refund rule that requires a human decision.
• FORTRAN is presented as an earlier promise to eliminate coding and debugging, but it mostly removed tedious work and raised the level of programming.
• The pattern repeats across assemblers, C, Stack Overflow, npm, Copilot, and coding agents, with each step reducing some manual work.
• Stack Overflow survey data suggests rapid AI adoption, but also falling favorability and distrust of AI output.
• A METR study found developers expected AI to save time, but the measured result was a 19% slowdown.
• METR’s time-horizon trend suggests agents can finish longer tasks over time, and the article uses that line to motivate the 2031 scenario.
• More AI-generated code can move the bottleneck to reading and review, which is why pull requests wait for people even when code arrives faster.
• Faros AI data is used to show higher task completion and merge volume, but also much slower review times and more incidents.
• The article treats the specification, design, and testing of the conceptual construct as the part AI still does not solve well.
• Hiring data suggests younger software developers are being affected more than older ones, and the article worries that junior apprenticeship could weaken.
• The closing guesses for 2031 are smaller teams, specs and tests as the source, more software built for one person, reading code as a career, and clearer labels on every change.