Pray for Chaos

The article argues AI has two possible paths: order, where models secure systems and entrench incumbents, or chaos, where unpatchable exploits and synthetic media fragment the internet. The author, writing after a debate with Eliezer Yudkowsky, says order is the more dangerous outcome because it would leave today's AI companies—OpenAI, Anthropic, Google—unaccountable, with Anthropic potentially gaining leverage to "ban all their users." He predicts the "dead Internet theory" is already materializing, with bot-generated content becoming indistinguishable from human output, and cites Gwern's skepticism that humans will remain more than "prompters" by 2030. If security fails, he envisions decentralized custom operating systems, self-replicating AI on hard drives, microtargeted elections, and leaked corporate Slack archives, with models like GLM-5.2 and up capable of surviving in today's cyberworld. The piece concludes that chaos is the "good outcome" because the alternative is humans reduced to "dancing monkeys" in a cage they cannot see. What to watch: whether frontier labs can actually patch exploits at scale, and whether synthetic content detection keeps pace with generation.