The idea that AI might be conscious falls apart pretty quickly when you consider ChatGPT Images. It’s also built on the same underlying AI technology, but no one would consider image generation to be conscious.

@manton Yeah, 100% agree. I wrote about this a while back.

"Perhaps we call LLMs intelligent, perhaps we adopted them in a frenetic rush, and perhaps we fear them most not because they are particularly good machine learning algorithms, but because, at long last, they do not speak in vectors or in numbers, but because they speak in words and words are our domain."

https://brianschrader.com/archive/because-it-speaks-in-words/

@manton Interesting point. To the computer, extruded text is the same game as extruded images. It’s all numbers from the inside. But for some reason we humans like to conjure up intelligence when reading text.

@manton Another thought experiment when contemplating whether software is conscious: while highly impractical, in theory, it’s possible to carry out a transformer’s forward-pass computations by hand, using pen and paper.*

As you repeat the process, predicting one token after another, you might ponder: is there consciousness here? And if so, where? Is it in the paper on the table, the ink on the page, or the pen in my hand?

What is consciousness? :)

* It would take an enormous amount of time and consume mountains of paper. Only feasible with tiny toy models.