What else happened on Nov 18th
Today's topic is recursive self-improvement (RSI), meaning the process in which AI builds the next models itself. The Future of Life Institute (FLI) put out a press release about a statement by scientists, writers and company founders. The signatories say Anthropic is not a safe AI company. A Forethought researcher goes through METR's report on the acceleration of AI R&D inside Anthropic and asks what the report implies for an intelligence explosion. That is the name for self-accelerating growth in model capabilities, which could start once AI fully automates AI R&D. The Lucid Labs post is about something else: how to stop a party that reports to a verifier from claiming that its model inference takes up more AI accelerator memory than it really does.