Google engineer Haydn Belfield stated that artificial intelligence research at DeepMind is not experiencing a measurable acceleration driven by AI itself. According to reports from The Decoder, Belfield noted that while AI now writes over 75 percent of production code at the company, a significant majority of those changes—88 percent—still require human oversight. Data from the company’s internal General Capability Index reportedly shows steady, linear growth rather than the sudden exponential speedup often discussed in industry projections, leading to the conclusion that true self-improving AI remains distant.
The Reality of Coding and Capability Metrics
The distinction between AI assisting with software engineering and AI independently advancing its own capabilities is a central debate in technology development. While large language models have become deeply integrated into daily software workflows—writing three-quarters of production code at a major lab demonstrates high utility—the reliance on human review for nearly nine out of ten code changes highlights current limitations in autonomy and reliability. The General Capability Index serves as an internal metric for tracking progress, though specific details on how this index is weighted or calculated were not provided in the announcement.
Unanswered Questions and What Lies Ahead
Despite these insights into DeepMind’s internal operations, the announcement did not specify the exact parameters of the General Capability Index, nor did it outline specific timelines for when researchers expect self-improving capabilities to become feasible. Observers will need to watch for future disclosures from Google regarding its internal productivity metrics, as well as whether other leading AI laboratories report similar plateaus or accelerations in recursive self-improvement.
As the industry continues to debate the trajectory of machine intelligence, current data suggests that human engineers remain firmly in the loop for the vast majority of software development tasks.
Lead image: ThisIsEngineering / Pexels
