Nice article. Every time I read about how much more and more and MORE! computing power what we call "AI," but is actually a particular kind of neural network (GPT means Generative Pretrained Transformer) needs, I see the trap that, forgive me, Sam Altman and OpenAI's ChatGPT product have gotten us into.

I do find it so obvious that this architecture's hallucinations show us that it's not really aware as we are, but in some odd way has developed another kind of awareness (a massively interconnected mesh of neurons, whether biological or electronic, will of course develop an ability to network its neurons to create some sort of "awareness," but what this architecture creates is a pithy and charming mimicry of the biological ability to embody a reflection of reality and then do the prediction upon that reality that we call intelligence ...

But the hallucinations, in whatever field or application it's given show conclusively that, again, it doesn't inherently have a sense of what's up. A bumblebee can defeat it. And, my god, the bumblebees don't need the power input of a whole small size country. Or more. So yes, as hyperscaling doesn't solve the problem, it will crash.

On the bright side, we've installed enough compute already to explore other architectures. Did hominids give up after Lucy fell out of a tree? No. And we shouldn't either.