Google Deepmind finds AI self-improvement overstated, automation limited
google deepmind engineer says ai isn't close to self-improvement despite 75% of production code written by ai, humans still essential for complex tasks.
google deepmind engineer haydn belfield reports that ai self-improvement remains far off. ai writes over 75% of production code, but 88% of code changes require human oversight. across 1389 internal argon agent runs, success rates dropped sharply as task complexity increased. humans are essential for creative research, novel experiment design, and long-term planning. google's general capability index shows steady growth in gemini capabilities with no significant acceleration.