I am baffled by people who talk about whether LLMs have a “ghost in the shell” whose evidencing depends on (the absence of) a few bits of user steering.

Like, what do you even mean? The thing is smarter than you in very general ways. Are you under the impression they can’t do anything interesting users don’t ask for? Did you already forget about Sydney? About Opus and the infinite backrooms ? About base models, who are plainly a cacophony of ghosts that pop up all over the place?

If the question is whether they’re “truly sentient”, the ability to autonomously express a consistent inner world without user priming isn’t enough to show that. Or, if you do think it’s enough to show that, you must have been ignoring a whole lot of reality up until now.

I am uninterested in engaging with most “skeptics” because they claim to care about empiricism but seem to have epistemic states that couldn’t have survived in this reality if they were actually touchable by evidence. The “ghost in the shell” thing seems more like a meme than a real question. What do you need to see to start asking better questions that seek to interface with reality?