A one-shot interaction with chatGPT asking for comment on proprioception

Refraining from the usual screenshot-based commentary but otherwise standard, chat session starts with a prompt by an agent

Oswald Berthold:

what do you know about proprioception and why it is so important for safe AI?

Which is reacted on by the other agent. Will dissect the original four-paragraph reply with immediate inline comment.

chatGPT:

Proprioception is the sense of the position and movement of the body, and it is important for safe AI because it allows a robot or AI system to be aware of its own body and its position in relation to its environment. This is important for a number of reasons.

This is entirely correct, proprioception is the sense of position and movement of the body. Most importantly, this refers to that part of the sensation which is conveyed even with your eyes closed, a point missed by the AI.

First, proprioception helps a robot or AI system to navigate its environment safely. For example, if a robot has proprioception, it can detect when it is about to bump into an obstacle and take appropriate action to avoid the collision.

Proprioception does help a robot to navigate its environment safely, you are good there chatty. The example on the other hand is old-school in terms of framing it as obstacle avoidance. Framed more accurately, proprioception allows us to detect obstacles, but not so much in the “when-we-are-about-to” way, instead more in the “when-it-has-already-happened” way. The special thing about proprioception then is, that it is incredibly fast. It is so fast, that it allows you to adapt your speed or direction of motion so quickly, that many times you will manage to influence the action (motion) consequence, while you are bumping into something. Consider the timescale of these events, it is in the few-milliseconds range.

Second, proprioception can help a robot or AI system to interact with its environment in a more natural and intuitive way. For example, if a robot has proprioception, it can use its sense of touch to detect when it is holding an object and adjust its grip accordingly.

This is correct again on a superficial level, but misses the point in the same way as for the previous items, in the framing. Proprioception does not happen primarily on the object level, but instead on the level of the individual motion. So the salient stimulus here is the moment when the motion does not yield the predicted response, because it is hitting the object (general class of external entropy by way of condensed matter physics), but this applies to each individually moving articulation of the body, and foremost, their integration needs to be taken into account.

This is quite informative since we know how the response was informed (average internet discourse), so the issue is there but it is discussed in wrong terms, entirely.

Overall, proprioception is an important capability for any robot or AI system that is designed to interact with the physical world, as it allows the system to be aware of its own body and its position in relation to its environment, which is necessary for safe and effective operation.

This only repeats what we said two paragraphs ago, and is cyclical reasoning in its reference to safe operation. Two slightly vague paragraphs of response do not count as “a number of” reasons, so again we are cyclical in the argument.

There certainly is a larger discussion going on about common sense and an understanding of physics which we require of trustworthy AI. This discussion aligns very well with our own framing in terms of grounding.

References

Comments

tag cloud

robotics music AI books research psychology intelligence feed ethology computational startups sound jcl audio brain organization motivation models micro management jetpack funding dsp testing test synthesis sonfication smp scope risk principles musician motion mapping language gt fail exploration evolution epistemology digital decision datadriven computing computer-music complexity algorithms aesthetics wayfinding visualization tools theory temporal sustainability supercollider stuff sonic-art sonic-ambience society signal-processing self score robots robot-learning robot python pxp priors predictive policies philosophy perception organization-of-behavior open-world open-culture neuroscience networking network navigation movies minecraft midi message-passing measures math locomotion linux learning kpi internet init health hacker growth grounding graphical generative gaming games explanation event-representation embedding economy discrete development definitions cyberspace culture creativity computer compmus cognition business birds biology bio-inspiration android ai agents action