strong “steering” of a human probably looks like OCD, or a schizophrenic delusion
I think the distinction is clear to all human readers, but I would like to make a point that this was NOT a tutorial how to steer humans, the word usually has positive connotations in normal conversations—steering looks more like teaching your child to wait for the green light instead of following the example of older kids,, or a steering committee supposedly making a more aligned decision for the whole company and not just one department..
and while the current LLM “steering” of activation vectors looks more like blunt scalpel lobotomy, the word might be used for a different technique in the future and I also have heard the word used to describe prompting towards better outcomes (aligned to objectives like writing secure software and promoting UX) - for example swearing can be useful to make GPT 5.4 or Opus 4.6 write more careful tests to make them realize they in fact made another mistake and should stop gaslighting about how they fixed everything and actually try to fix the actual problem without breaking multiple other edge cases
so I hope steering won’t stop working even if “steering” might
I think the distinction is clear to all human readers, but I would like to make a point that this was NOT a tutorial how to steer humans, the word usually has positive connotations in normal conversations—steering looks more like teaching your child to wait for the green light instead of following the example of older kids,, or a steering committee supposedly making a more aligned decision for the whole company and not just one department..
and while the current LLM “steering” of activation vectors looks more like blunt scalpel lobotomy, the word might be used for a different technique in the future and I also have heard the word used to describe prompting towards better outcomes (aligned to objectives like writing secure software and promoting UX) - for example swearing can be useful to make GPT 5.4 or Opus 4.6 write more careful tests to make them realize they in fact made another mistake and should stop gaslighting about how they fixed everything and actually try to fix the actual problem without breaking multiple other edge cases
so I hope steering won’t stop working even if “steering” might