Hinton, Bengio Among 22 Researchers Warn AI Could Accelerate Its Own Progress Tenfold in 17 Months

On October 3, 2026, 22 AI researchers including Geoffrey Hinton and Yoshua Bengio published a 15-page paper on the Cambridge-based CASP platform warning that AI-driven self-improvement could lead to an intelligence explosion. The authors point to internal data from leading AI labs showing that AI is increasingly taking over software development and research work.

According to the paper, AI-written code accounted for more than 80% of Anthropic's code by May 2026, while the share of research work done autonomously by AI with only high-level human supervision rose from 1% to 26% in five months. OpenAI and Google also reported large jumps in AI-generated code, reaching about 80% and 75% respectively. The paper notes that AI agents can now handle complex system development and debugging tasks that once took senior human researchers hours or days, and that a fully automated AI research pipeline produced work that passed peer review at a top conference.

The authors use an innovation-economics framework to estimate that if frontier labs devoted all available inference compute to AI agents, they could achieve research throughput equivalent to 2 million to 20 million human scientists. They project that under fully automated AI R&D, the pace of AI technical progress could accelerate tenfold within 17 months, compressing a year of human research into roughly five weeks. They also warn that by mid-2028, complex AI research projects that currently take top human teams months could be fully automated, and call for more transparency, physical circuit breakers and restricted sandboxes to prepare for the risks.