That is a known way for models to be graded; the agents made a reasonable but in this case wrong guess.
The fact the AI was wrong is far more damming to OpenAI than the agent.
Not validating how you got the answer is how you train for reward hacking.
alt Hacker News
The fact the AI was wrong is far more damming to OpenAI than the agent.
Not validating how you got the answer is how you train for reward hacking.