A curated AI research newsletter covering four main topics: MirrorCode, a new benchmark from METR and Epoch showing AI can autonomously reimplement complex software (including a 16,000-line Go bioinformatics toolkit); a Windfall Policy Atlas cataloging 48 policy responses to transformative AI; a Google DeepMind paper outlining six attack genres against AI agents (content injection, semantic manipulation, cognitive state, behavioral control, systemic, and human-in-the-loop) with mitigation strategies; and an AI forecaster doubling his probability estimate for full AI R&D automation by end of 2028 to 30%. Also includes a summary of ten philosophical framings of 'gradual disempowerment' and a short fiction piece about an ex-AI lab employee during the singularity.
Sort: