px

Hot take a "maximally truth-seeking" LLM is a contradiction in terms because the closest thing a being can get to being "truth-seeking" is self-honesty and self-honesty is a property of beings with qualia

2 top replies

Scott Stevenson @scottastevenson · Oct 17

This is a good insight. A way to rephrase it could be maximally aligned with the average of human qualia, if you assume that there is an objective reality underpinning human qualia. However no one has access to the shadow world if “objective concrete reality”. For all useful purposes, reality is essentially subjective. No one experiences objectivity

Hyters @Hyters303 · Oct 17

if a model starts investigating where its own beliefs and values came from would you only call that "truth-seeking" once we have established it has qualia?

I maintain that LLMs are not conscious at all, neither in training nor in inference, but based on the way consciousness works, if consciousness were present anywhere it would be in the training phase.

1 more tweet in this thread

Being trained at something is more or less synonymous with being able to perform it unconsciously

1 top reply

Tex @The_Texian · Oct 17

What do you mean by conscious?

Question of AI sentience feels academic now but once Optimus robots are walking around it'll be quite salient. Roughly, debate will hinge on if these things get full blown rights or are constrained by Asimovian principles (i.e. are "slaves" as the sentientists will allege).

1 more tweet in this thread

It'll be darkly funny if the survival of humanity hinges on whether we have the cojones to ignore the moral accusations of people who are confused about qualia.