Warnings that artificial intelligence poses an existential threat to humanity are a self-serving attempt at regulatory capture, according to discussion on The Register's Kettle podcast.
The episode, hosted by Brandon Vigliarolo with systems editor Tobias Mann and senior reporter Tom Claburn, examined recent dire warnings from AI researchers. Tom Claburn highlighted concerns raised by Jacob Coxon, a former Anthropic researcher who had previously worked at OpenAI, about self-reinforcement learning and models improving themselves to the point of escaping human control.
The podcast argued that these warnings hinge on companies operating without precautions. Tobias Mann said there is already a template for building sandboxes that work, pointing to air-gapped systems used for nuclear armament supercomputers, and said companies are choosing not to use them.
Tom Claburn said that of all the things killing people, AI ranks far down the list, though he noted real concerns such as AI-powered target selection in Ukraine and AI vision systems in cars. He said the idea that AI will suddenly take over utilities or kill all crops is never made clear.
The discussion concluded that the warnings are a gambit for regulatory capture, with companies hoping to be appointed to positions where their word is unassailable.