Toward White-Boxing the Internal Circuits of LLMs

Last updated: October 11, 2026

This is research into the safety of superintelligence, a topic whose risks have been widely pointed out in recent years. We aim to white-box the internal mechanisms of generative AI, establishing technology for directly monitoring concept manipulation during inference and identifying dangerous reasoning.

At this stage, to establish our research methodology, we are focusing on our other project, "Metaphor Circuit Research," through which we are developing the analysis tools and methodologies needed to white-box LLMs.

White-boxing the internal circuits of LLMs is a very important theme in AI safety, and we believe it is also an essential technology for making generative AI, which has become so important and indispensable to our society, safer and more valuable. It also has implications for clarifying the theoretical foundations of LLMs. We place AI safety first and will continue to advance research and development of LLMs' foundational technology.