Hacker News Reader

Comment by Al-Khwarizmi | Hacker News Reader

That's part of the reason to be wary about the possibility that our AI models may become conscious.

Most people happily dismiss the idea that a GPT-like model can be conscious, but the truth is that we don't know practically anything about how consciousness arises in living beings that are more similar to us and that we have been coexisting with for thousands of years. We don't even know how to tell if such a being is conscious or not (most would agree that a dog is conscious and a bacterium is not, but what about a mouse? A mosquito? A tardigrade? Where is the limit?). It takes a lot of hubris to make claims about if and when AI models can become conscious, with what we know right now.

View on HN

Your post touches two points worth to think about:

The test for consciousness: if we do have zero criterion, how we are gonna find out if something is conscious?

That is already a good indicator how much we do not know about the „mechanism” of consciousness.

Point two is that we in fact know so little, that we cannot even nail a probability for some complex mechanism if is conscious.

A car: probably not, too far away from things that are conscious.

An ant?

A microphone that listens to its output?

A k-means run on some data?

An llm?

Some electricity in a biological neural network?

If only we had a proven criterion to decide!

Ned Block explores this difficulty in his The Harder Problem of Consciousness paper. https://www.jstor.org/stable/3655621

He uses Commander Data from Star Trek as the example. Since we don't know whether it is something unique about our biology that leads to consciousness or consciousness is functional, then we can't determine whether Data is conscious. It also gets into the weeds about what sort of functional equivalence might or might not count. Data is obviously not the exact same functionally as we are, neither are LLMs. There are important differences. We know neurons don't use backpropagation, for example.

So what is the basis for being conscious?

The reason I'm skeptical that GPT-like models can ever reach consciousness is because they're centred around language, whereas consciousness isn't. While it's entirely possible to reach the same result with different methods, my intuition tells me that language-based models are just not going to cut it ever, which is admittedly very unscientific.

I see it as more there are degrees of "consciousness" (whatever it is). Human consciousness varies in a single human over time (sleeping, tired, alert, drunk etc) and the range appears to vary (let alone across species, I'd grant dogs a possible overlap, having been outstarted by them more than once).

Reply on news.ycombinator.com