A blog post analyzing a recent MIT/Wharton paper on AI coding tool productivity, examining how gains from autocomplete, sync agents, and async agents decay as work moves up the software delivery pipeline. The author translates the paper's economic production functions into Amdahl's Law terms, finding that the global parallelizable fraction (P) stays stubbornly around 35% across all three generations of AI tools. This means the hard ceiling for total shipping improvement is roughly 53%, because the remaining 65% of developer time — planning, meetings, alignment, and reasoning — cannot be parallelized by current AI. The analysis explains why macroeconomic and team-level productivity gains from AI coding tools remain modest despite impressive task-level speedups.
10.8K Impressions1 Comment