Sam Altman posted a photo of his husband and infant son hoping to dissuade the next person from throwing a Molotov cocktail at their house, after the arrest of an individual accused of attacking his home and menacing OpenAI’s headquarters.
And METR found the same model has a 13-hour autonomy horizon if reward hacking is allowed, against 5.7 under standard methodology. The honest model and the dishonest one are effectively different species.