Daniel Kokotajlo suggests publishing research-cluster GPU logs for AI verification

Daniel Kokotajlo argues that any US–China deal on advanced AI would need intrusive verification, including inspectors counting chips at data centres. He proposes separating private inference facilities from maximally transparent research clusters, where devices between GPUs would log activity and those logs would be published online. Kokotajlo says broad visibility could deter secret work and let the wider scientific community scrutinise training and testing.