Google says self-improving AI isn't anywhere close
decoder ·
Google engineer Haydn Belfield states that self-improving artificial intelligence remains far from reality, despite significant automation in other areas. While AI now generates over 75 percent of production code, human oversight is still required for approximately 88 percent of changes to ensure quality and safety. Internal data from Deepmind's Argon agent shows that success rates drop sharply as task complexity increases, with humans remaining essential for creative research and complex problem-solving. On the RE-Bench benchmark, Argon solves only half of the tasks, a notable improvement but insufficient for full autonomy. Google’s General Capability Index indicates that model capabilities are growing at a steady, linear pace without any measurable acceleration driven by AI itself. Although the risk of AI-driven research automation is currently rated as low, Google is building infrastructure to monitor potential future shifts. These findings contrast with warnings from prominent researchers like Geoffrey Hinton, who have predicted a rapid "intelligence explosion" where AI could automate all R&D within years.
decoder