AI systems now write 80-90% of code for model development
AI systems are increasingly automating the development of next-generation models, writing 80-90% of the code for training environments. This acceleration is seen as a pathway to superintelligence, though safety research lags significantly behind capability research.
AI systems currently perform up to 26% of AI research tasks, according to Anthropic. The development of next-generation models is increasingly automated, with AI writing 80-90% of the code for training environments and evolutionary algorithms. This automation is seen as the pathway to superintelligence, potentially skipping the traditional AGI stage. OpenAI recently solved hundreds of open mathematical problems, some of which would have warranted a Fields Medal if solved by humans. Balde notes that while AI excels in specific domains like coding and math, it still fails on simple 'jagged frontier' tasks. He warns that the pace of capability doubling is under 90 days, with some experts predicting recursive self-improvement within one to two years. Balde argues that the rapid automation of AI research by AI itself creates a control problem. He states that safety research lags five to ten years behind capability research and receives 200 times less funding.