I uploaded my own book manuscript to chatGPT bc I wanted to do some summarization work. I was surprised (and somewhat dismayed) to find it still hallucinated repeatedly when asked to confine itself to material from a directly-uploaded PDF.
yup. Maybe none of these models do what it says on the tin. Maybe back-propagating errors is nothing like what our brains do. What if actual AI is still decades away?
Yeah, I think the basic NN is maybe not going to make this work and just adding inputs have a scaling issue. The brain is organized into circuits, not layers. Neurons lose potential when they fire, and it doesn't do backpropagation. All probably important aspects of why it woks.
People: “AI is going to replace us!”
Also People: “Ah this AI thing kinda sucks.”
Still People: “Ok let’s keep refining AI so it’s more ready to replace us.”
People Again: “AI is going to replace us!”
FWIW, the freebie ChatGPT version still seems to be giving you 3.5 by default; if you didn't explicitly change it, try doing that (or try Gemini or Claude). Output from all of them has to be checked and flyspecked, but GPT 3.5 is notably worse than the others.