Google engineer: AI writes more than 75% of production code, and there is no clear evidence that independent research and development is accelerating. Google engineer Haydn Belfield said that although AI’s automation in model development continues to increase, there is currently no measurable evidence that AI autonomously driven research is accelerating.

Google engineer: AI writes more than 75% of production code, and there is no clear evidence that independent research and development is accelerating. Google engineer Haydn Belfield said that although AI’s automation in model development continues to increase, there is currently no measurable evidence that AI autonomously driven research is accelerating. Belfield pointed out that currently more than 75% of production code is written by AI, but about 88% of code changes still require manual supervision or review, and each change must also undergo manual review before submission. The proportion of code changes handled entirely by AI has increased from 7.5% to 11.9%. In 1,389 Argon internal agent running tests, the higher the task complexity, the lower the success rate of AI unattended intervention. A survey of 12 senior DeepMind researchers also showed that AI has a high degree of automation in well-defined tasks with verifiable results, but it still relies on humans in creative research, experimental design, training anomaly diagnosis, and long-term planning. In the RE-Bench benchmark test, which contains 74 real machine learning engineering tasks, Argon's task solution rate is about 50%, which is higher than Gemini 3.7 Flash's 27%, but there is still a gap between achieving full automation. In order to track changes in AI capabilities, the research team constructed a General Capability Index (GCI) covering 123 models and 92 benchmark tests. The data shows that the Gemini model capability is growing steadily at a rate of about 0. 1 GCI points per month, and there is no statistically significant sign of acceleration. The team believes that the performance of Gemini 4 Argon that is slightly higher than the growth trend line is mainly due to the increase in computing power and organizational investment, rather than the acceleration of research and development driven by AI independently. At present, the research team believes that the risks brought by AI-driven research automation are still relatively limited, but they are continuing to improve the monitoring system. At the same time, more than 20 researchers, including Jeffrey Hinton, Joshua Bengio and OpenAI chief scientist Jakub Pachocki, warned that AI may achieve a high degree of automation in research and development work in the next few years, compressing years of progress into months. Different judgments highlight that there is still great uncertainty in AI's self-improvement capabilities and its potential risks. via AI News (author: AI Base)