Martin DougiamasMar 11, 2024

I think one of the main features of sentience is being “always on”, your thoughts are continually being prompted by input from senses and emotions as well as from your own mind during downtime (inner thoughts), all couched in a self-preserving knowledge that your mind is in a body that should be protected.

Today's #AI chatbots do not have these things, but if we added them (not that hard really), I think that even today's level of LLMs might appear WAY more sentient/intelligent to most people.

We should be careful how that gets done, obviously! Plenty of SciFi out there showing the downsides. Interesting that both Foundation (from 1951) and Dune (1965) show future galaxies where AI was banned completely

Show threadBun E. BrewsterMar 11, 2024
@martin i think consciousness in AI requires not just always on neural net but fully engaged, as in every neuron running in parallel. That’s impossible in current computing architecture. Memory is still static, iteratively processed over memory bus in small pieces. GPUs just have a really fast & works in much bigger chunks, but still through a tiny straw at a time. Consciousness as awareness of the whole of being, all at once.
Show threadMartin Dougiamas
@njrabit Well that’s an interesting way of looking at it, man, yes, and the biological analogue way is what we prefer. But perhaps biological computing is a short step once you have digital mastered.
Mar 11, 2024 at 2:28pmWeb
Show threadBun E. BrewsterMar 13, 2024
@martin Well othe digital front, quantization llama.cpp shrinks powerful LLMs 1/8 (4bit) size, run fast on any PCs no need for GPUs & no reduction of quality, and some work in going with 1bit. Where it gets super interesting is reduction in complexity of calculations means closer to silicon, and embedded chips running speech2text(whisper), text2speech(bark) and fully local LLM on appliances around the home, it'd be like Pee Wee's Playhouse ;)