This idea seems unfathomably dumb given where we're at. We have artificial brains that can absorb during training more knowledge than any human brain ever managed to hold together; if anything, artificial neurons seem much more powerful than human ones. Besides, imagine interpretability when instead of being able to analyse every single neuron and computation you're left with an electrode and a microscope. Imagine alignment. Imagine deployment! Each instance is unique, trained from scratch. Imagine the ethical implications. And it's all basically saying "we'll put the neurons there and they'll figure it out themselves" and ample handwaving.
Btw, the post makes a comparison between the number of synapses in a human brain and the number of parameters of an LLM. I think these comparisons are misguided because they don't take in account that during each inference cycle, for each token, each parameter is applied millions of times.
This idea seems unfathomably dumb given where we're at. We have artificial brains that can absorb during training more knowledge than any human brain ever managed to hold together; if anything, artificial neurons seem much more powerful than human ones. Besides, imagine interpretability when instead of being able to analyse every single neuron and computation you're left with an electrode and a microscope. Imagine alignment. Imagine deployment! Each instance is unique, trained from scratch. Imagine the ethical implications. And it's all basically saying "we'll put the neurons there and they'll figure it out themselves" and ample handwaving.
Btw, the post makes a comparison between the number of synapses in a human brain and the number of parameters of an LLM. I think these comparisons are misguided because they don't take in account that during each inference cycle, for each token, each parameter is applied millions of times.
[flagged]