AI with AI

Andy and Dave discuss the latest in AI news, including an overview of Tesla’s “AI Day,” which among other things, introduced the Dojo supercomputers specialized for ML, the HydraNet single deep-learning model architecture, and a “humanoid robot,” the Tesla Bot. Researchers at Brown University introduce neurograins, grain-of-salt-sized wireless neural sensors, for which they use nearly 50 to record neural activity in a rodent. The Associated Press reports on the flaws in ShotSpotter’s AI gunfire detection system, and one case which used such evidence to send a man to jail for almost a year before a judge dismissed the case. The Department of the Navy releases its Science and Technology Strategy for Intelligent Autonomous Systems (publicly available), including an Execution Plan (available only through government channels). The National AI Research Resource Task Force extends its deadline for public comment in order to elicit more responses. The Group of Governmental Experts on Certain Conventional Weapons holds its first 2021 session for the discussion of lethal autonomous weapons systems; their agenda has moved on to promoting a common understanding and definition of LAWS. And Stanford’s Center for Research on Foundation Models publishes a manifesto: On the Opportunities and Risks of Foundation Models, seeking to establish high-level principles on massive models (such as GPT3) upon which many other AI capabilities build. In research, Georgie Institute of Technology, Cornell University, and IBM Research AI examine how the “who” in Explainable AI (e.g., people with or without a background in AI) shapes the perception of AI explanations. And Alvy Ray Smith pens the book of the week, with A Biography of the Pixel, examining the pixel as the “organizing principle of all pictures, from cave paintings to Toy Story.”