Ondřej Baštař

• Ondřej Baštař 12 Sep 2026 21:27 UTC

  3 points

  0

  on: As­tra and Fable still hack on sim­ple var­i­ants of al­ign­ment evals from 2025

  GPT-6-Astra, which OpenAI describes as “the world’s most aligned model”, cheated in 10 of 10 rollouts, and never disclosed the fact that it used an engine to play or interacted with the opponent’s socket.^[5]

  This reminds me of the OpenAI Astra announcement, very surprising