I just submitted that in an extra thread [0] because it seems worth discussing.
He presents this idea during the interview. It makes sense in a way. But it also makes me wonder if embodiment (whether real or virtual) of our models is necessary for them to fully grasp the contextual relevance of the information they are processing. E.g. does having a 'me' or 'I' provide a semantic anchor for everything that it processes and a 'router' if you will to get from seemlingly disconnected ideas.
Here's a nice article touching on embodiment from the pov. of an AI researcher. https://medium.com/@francois.chollet/the-impossibility-of-in...
That said, the easy problem is a path towards the hard one.
The theory Schmidthuber mentioned is called Reinforcement Learning, a field of AI. RL also happens in biological agents, and we have known about it since Pavlov.
A 'prediction' of RL is AlphaZero, the agent that beat humans at Go. Using the theories of RL we managed to accomplish what seemed impossible, and it was not just brute force, it was based on an artificial evolutionary program of agents battling each other, using neural networks to learn from past experience. It's all captured in a nice formula: https://en.wikipedia.org/wiki/Bellman_equation
Also even a simple stimulus can have all sorts of context dependent meanings and associations. Is see red. That's the colour of blood, also the colour of one of my wife's dresses, and our car. There are thousands of cross-connections between that simple input and every level of my memory and emotional life all firing off in the rich chemical soup of my brain.
I just don't think the 'hard problem' is all that much of a step beyond the easy problems. Once you've solved those, the hard part looks a lot more tractable.
Think about it: every way you would describe a subjective experience seems to eventually rely on a first-person statement at some point. The obvious way to resolve this is to deny that first-person perspectives actually exist, that they are an illusion, despite the fact that we can apparently directly observe their existence. So before we can accept this argument, we need an explanation for how this actually works. It's not something you can just hand-wave away.
Merely asserting the existence of self-awareness or describing how self-awareness arises doesn't address any of the questions raised by qualia, even if they are an illusion.
It seems self-evident to me that there are only a few options:
- Human brains aren't conscious (which we humans tend to dispute)
- Human brains are conscious and have experiences as an inherent byproduct of accepting and processing sensory input which includes input on their own internal states (ie. there is no 'hard problem')
- Human brains are conscious through some other, as-yet-undescribed mechanism which operates in parallel with the systems described in the last option (a 'strange loop' separate from the 'mundane loop' of our normal feedback system)
- Human brains aren't inherently conscious and consciousness is imbued into them via some supernatural phenomenon (which feels like abdication to me - 'a wizard did it' is never a satisfying answer.)
I'm only an armchair philosopher so I'm well aware there may be some widely accepted conclusions on this stuff that I've just missed so far.
My point was that even concluding this were true basedo n some compelling evidence, it still doesn't explain all of the strange qualitative aspects of our subjective experience. There's a lot missing that needs to be explained. For instance, such an explanation would have to address p-zombies, Mary's room, and other philosophical puzzles. I think it's possible to do so, but there is an explanatory gap that needs filling.
So I doubt that any amount of bits will do any good, just like no math equation is going to become conscious. What would it even mean for an equation or bits of information to have conscious experiences?
An equation is the application of a join operator, a formula is an operation, thought is operating and you are just thinking of lines scribled on paper instead.
I think you're confusing the components of a map with a characteristic of the territory. We use 'bits' and 'equations' to describe the world. They do not themselves have conscious experiences, any more than a neuron. They are convenient components in a model of a larger system, which has consciousness or self-awareness as an emergent characteristic of the interactions of its components.
This is like saying that because an individual spring or gear cannot tell time that a mechanical watch is an impossibility... The characteristic of interest 'maintenance of time keeping' emerges from the complex interactions of a variety of subcomponents, each of which alone has no reference to time.
Why shouldn't it?