Import AI issue 451 covers several AI research topics: a Stanford professor's framework for 'political superintelligence' with three layers (information, representation, governance); a robotics paper (DexDrummer) showing robot hands are still poor drummers; Google research arguing future AI alignment requires building hybrid human-AI social institutions rather than perfecting single models; Meta-affiliated researchers' 'Hyperagents' (Darwin Godel Machine) that self-improve via recursive meta-agent loops with strong benchmark results; and HorizonMath, a new benchmark of 100 unsolved math problems where top models score only 7%. The issue closes with a speculative fiction piece about AI lab sabotage.
Sort: