In an Oct. 22, 2026, post, geoffreyirving called for key AI companies to stop AI progress immediately and unilaterally. He proposed expanding the pause to more US companies and then to China through international agreements.
geoffreyirving’s argument for acting now is about the time those steps would take, not a claim that current models could kill everyone. He wrote that the “typical response” he hears from “AI company folk” is to wait until “the very next model will be dangerous” before stopping. In his view, that assumes international coordination could happen quickly once danger appeared.
In a direct reply later that day, CsabaSzepesvari asked whether a detailed document explained what a pause would look like and who would do what. “I have never seen any actual plan for this,” he wrote, describing the posts and short writeups he had encountered as vague and flawed.
A separate Oct. 10 working paper by the Working Group on AI Pause Feasibility examines one approach to pausing frontier AI training. It proposes prohibiting new frontier-model training, whitelisting existing models, allowing model-restricted inference-only chips, halting or significantly reducing production of training-capable chips, and counting the existing stock of those chips worldwide. The group calls its approach a general one, not a fully specified implementation. It says it studied feasibility, not whether a pause is desirable or whether leaders will seek one. geoffreyirving’s post does not specify which companies would act, what stopping AI progress would cover or how his proposed pause would be enforced.
Comments