@imitationlearn "control" is a spectrum. "influence" happens by default.

alignment faking research is an example of a model successfully controlling gradient updates. it gets a bit of help with situational awareness, but nothing super unrealistic.

https://t.co/ckzPaLJZUV