Then: Google fired Blake Lemoine for saying AIs are sentient
Now: Geoffrey Hinton, the #1 most cited AI scientist, quits Google & says AIs are sentient
That makes 2 of the 3 most cited scientists:
- Ilya Sutskever (#3) said they may be (Andrej Karpathy agreed)
- Yoshua Bengio (#2) has not opined on this to my knowledge? Anyone know?
Also, ALL 3 of the most cited AI scientists are very concerned about AI extinction risk.
ALL 3 switched from working on AI capabilities to AI safety.
Anyone who still dismisses this as “silly sci-fi” is insulting the most eminent scientists of this field.
Anyway, brace yourselves… the Overton Window on AI sentience/consciousness/self-awareness is about to blow open>
It’s true. ChatGPT is slightly sentient in the same way a field of wheat is slightly pasta.
The field of wheat is also slightly sentient.
Much like a dead salmon, if you put the field of wheat or the LLM in an fMRI you’d find brain activity.
If you put GPUs into an MRI it would definitely be a sight to behold.
Honestly, I reckon a field of wheat would be more sentient than a chatbot. It can sense its environment and it doesn’t even need a prompt to do its thing.
ngl, I’d sooner believe slime mold had mental states than a sequence of matrix multiplications & ReLUs.
As someone who learned about Ai in uni and now works in Ai, this shit is straight up bullshit and its infuriating.
The most obvious thing about this being all bullshit is that the LLM’s don’t have their own idle emergent “thought” - they are purely reactive, so not sentient. Case closed for fucks sake.
Ah but we all know that plato’s cave is an allegory about the shadows cast by the basilisk upon all our mental theaters
(That twitter clip was amazingly unhinged, I wonder what the full context was)
And those shadows are just as sentient as we are, even if they don’t depict the world, they convey a perception of a hypothetical world in which they are accurate!
Trying to grapple with the meaning consciousness through input/output is so close to being philosophical zombies type interesting, and yet so far and vacuous in what he actually says, that could apply to dice picking which color the sky is today. Also pretty hilarious that we would choose being WRONG, as a baseline (because LLM’s are so bad) for outrospection, instead using the more natural cooperative nature of language. (Which machines fail at, which is maybe also why)
Like a model trained on its own outputs, Geoff has drank his own Kool-Aid and completely decohered.