Roman Leventov comments on AI interpretability could be harmful?

[Roman Leventov]

I wonder whether GFlowNets are somehow better suited for self-destruction/​non-finetunability than LLMs.