Frontier AI companies criticized for lacking safe self-improvement plans

A critic compares talk of “recursive self-improvement” with an unspecified role for humans to “Full Self-Driving (Supervised),” arguing that the wording can satisfy some critics while misleading them about the goal. They say the criticism applies to all frontier AI companies, none of which has even a concept of a plan for safe self-improvement. They add that DeepMind’s control and alignment agendas may be the closest to such plans, though much remains to be done.