Jeff Shrager, Kent M. Pitman and Mia Golec revived parts of two early AI programming systems. Simon’s 1961 Heuristic Compiler modeled programming as problem solving; MIT’s KBEmacs combined reusable patterns under a programmer’s direction.
- The restored compiler reproduced published examples. One required a repair: removing a description that made the program choose a failing path.
- The headless KBEmacs core reproduced the square-root sequence from a 1984 self-test. That sequence built two programs using 14 commands.
- The team reports using ChatGPT for much of the transcription and Claude Code for restoration and debugging, under human direction.
The authors compare the old systems’ explicit plans and problem-solving loops with today’s coding agents, while cautioning that they are not direct technical ancestors.
Caveat: These are partial restorations. Only one of six KBEmacs test sequences was replayed, and reproducing old outputs does not prove the systems’ theories of human programming.