AI Researchers' Views on Automating AI R&D and Intelligence Explosions

Many leading AI researchers expect AI development to exceed the transformative impact of all previous technological revolutions. This belief is based on the idea that AI will be able to automate the process of AI research itself, leading to a positive feedback loop[1]. In August and September of 2025, we interviewed 25 leading researchers from frontier AI labs and academia, including participants from Google DeepMind, OpenAI, Anthropic, Meta, UC Berkeley, Princeton, and Stanford to understand researcher perspectives on these scenarios. Though AI systems have not yet been able to recursively improve, 20 of the 25 researchers interviewed identified automating AI research as one of the most severe and urgent AI risks. Participants converged on predictions that AI agents will become more capable at coding, math and eventually AI development, gradually transitioning from ‘assistants’ or ‘tools’ to ‘autonomous AI developers,’ after which point, predictions diverge. While researchers agreed upon the possibility of recursive improvement, they disagreed on basic questions of timelines or appropriate governance mechanisms.

Introduction. An ‘intelligence explosion’ is a hypothetical scenario introduced by I.J. Good in 1966[2], in which AI systems become capable of recursively improving their own capabilities (designing AI successors that can design AI even better). Yudkowsky [3] defines an intelligence explosion in terms of the return on cognitive investment, when investing cognitive resources in improving cognition accelerates the returns from those investments. Recursive improvement need not look like conventional AI R&D, future AI systems might be able to directly plan and program recursive improvements, or improve by improving their physical hardware. However, a critical capability milestone may occur when AI systems can do the work of today’s AI researchers, or AI R&D, at which point, such an AI could improve the development of it’s successor. In August of 2025, a version of OpenAI’s GPT-5 model won a gold medal at the International Math Olympiad[4].

Discussion / Conclusion. An epistemic divide emerged between frontier AI companies and academic institutions regarding ASARA development, reflecting deeper differences in organizational culture, incentives, and proximity to cutting-edge capabilities. All participants at frontier companies demonstrated active engagement with ASARA scenarios and reported regular internal discussions about recursive improvement dynamics. They note that these conversations are encouraged. For instance, Participant 6 said, “I’ve been impressed so far, at least internally, with the discussions people have had that especially that the leaders are aware of these worries and talk openly about them, and bring them up on their own and encourage us to think about them." By contrast, academic participants often expressed limited consideration of these possibilities, with one noting: “I don’t even think the majority of AI researchers even think about this problem” (P17, PhD Student). Participant 7 also brought this up when they said, “When I ask questions like this, even if people in the labs are not believers in AGI or something, these are still discussions that we have at work. They’re encouraged by the top...