5 pointsby matt_d8 hours ago1 comment
  • promptsaredead2 hours ago
    I think a lot of SOTA models are already going this way. Long autonomous tasks in claude code / codex already do this to stay on track and avoid multiplying errors over many steps.