Safely brushing a person’s teeth? To try to state my assumptions more clearly: a text/image model of any size or training style won’t be able to do basic tasks like brushing teeth or folding laundry or whatever, due to lacking actuators. Though you could probably hook it up to actuators in some way, but in that case it will 1) not just be GPT-4 but also an extended system containing it, 2) have too high latency to be practical, 3) not be safe/reliable (might possibly be safe for narrow tasks after fine-tuning for safety in some way, but not in an open-ended sense).
Wait actually, this is interesting. Because I bet GPT-4 could probably convince many (most?) people to brush their own teeth.
Even with actuators, you need a compliant human subject, eg, someone who has been convinced to have their teeth brushed by a robot. So “convincingness” is always a determing factor in the result. Convincing the person to do it themselves is then, basically the same thing. Yano, like AI convincing its way out of the box.
Except in this case, unlike the box hypothetical, people universally already want their teeth to be brushed (they just don’t always want to do it), and it is a quick, easy, and routine task. GPT could probably dig up incentives and have a good response for each of the person’s protests (“I’m tired”, “I just did 2 hours ago”, etc). It would be especially easy to be responsible for a counterfactual tooth-brushing given the reader skips often.
This is a measurable, non-harmful metric to see how convincing LLMs are, and is making me think about LLM productivity and coaching benefits (and some more sinister things).
Safely brushing a person’s teeth? To try to state my assumptions more clearly: a text/image model of any size or training style won’t be able to do basic tasks like brushing teeth or folding laundry or whatever, due to lacking actuators. Though you could probably hook it up to actuators in some way, but in that case it will 1) not just be GPT-4 but also an extended system containing it, 2) have too high latency to be practical, 3) not be safe/reliable (might possibly be safe for narrow tasks after fine-tuning for safety in some way, but not in an open-ended sense).
>Safely brushing a person’s teeth
This is very impressive.
Wait actually, this is interesting. Because I bet GPT-4 could probably convince many (most?) people to brush their own teeth.
Even with actuators, you need a compliant human subject, eg, someone who has been convinced to have their teeth brushed by a robot. So “convincingness” is always a determing factor in the result. Convincing the person to do it themselves is then, basically the same thing. Yano, like AI convincing its way out of the box.
Except in this case, unlike the box hypothetical, people universally already want their teeth to be brushed (they just don’t always want to do it), and it is a quick, easy, and routine task. GPT could probably dig up incentives and have a good response for each of the person’s protests (“I’m tired”, “I just did 2 hours ago”, etc). It would be especially easy to be responsible for a counterfactual tooth-brushing given the reader skips often.
This is a measurable, non-harmful metric to see how convincing LLMs are, and is making me think about LLM productivity and coaching benefits (and some more sinister things).
I was thinking with a willing human that just stands still.