MIRI's Approach – Machine Intelligence Research Institute

I am highly dubious that the the qualitative nature of awareness when removed from symbols is of any value.

I have long ago accepted that I "observe" my brain—that is, my brain registers its own actions. At this point, understanding in the sense of the Chinese Room Experiment is mostly a question of "how do you want to serialize these symbols?". The meaning is in the relations between symbols in that person (or computer's) mind—mind-bogglingly complex, sure, but hardly non-serializable.

I find it very interesting that consciousness itself has had such a difficult time being hammered out. I have read quite a bit of philosophy on the subject (Searle's publications among them), and it seems that there are fundamental disagreements over diction. I have had a very, very difficult time wrapping my head around the arguments for both "free will" as some kind of quantum effect (see: Penrose's The Emperor's New Mind and its "sequel") and arguments for some specially defined consciousness with requirements for "awareness". Think about a dolphin: it's fairly easy to imagine what it might be aware of, although the particulars are obviously unreachable because our brains are not wired to be aware of the same sensors of which a dolphin brain is aware. That's the practical limit, though, in terms of having difficulty grasping what a dolphin might experience—you might have the same difficulty understanding how a blind-from-birth person is aware of the world because you can't cancel out your own visual wiring. This isn't really a barrier in terms of having the same ability as a blind person (or vice versa) except where that sensory awarenesses is critical.

Now, in terms of meaning, talk to people with abnormal thought patterns—schizophrenics, bi-polar people, borderline personality people, OCD people. For instance, many with severe personality disorders have a tendency to dichotomize everything with difficulty integrating "shades of grey" into every day thinking. Others—e.g. some compulsive liars—have a very difficult time pinning down specific meanings from an objective standpoint. Compartmentalization is a mechanism allowing multiple truths in compartments while allowing contradictions in a general sense. Meaning is evidently subjective. Which is more "meaningful" to you, attempting to understand the mapping of incomprehensible numbers of physical neurons, or attempting to understand the mapping of incomprehensible numbers of non-linear equations? They both come out to about the same level of "meaningless symbol" processing with no "magic" to me.

The structure of models of current neural networks are very, very rudimentary now, and are probably dozens of orders of magnitudes less complex than those of the human brain. But current research—including analyzing the brain, understanding how to describe it in terms of current understanding of neural networks, serializing it to a model, and simulating the model—are even now, within reach of being able to simulate a nematode brain. At that point, the arguments over consciousness merge for "how big/smart/aware/whatever does a brain need to be to be conscious" and "do we have strong AI"? It becomes a game of "is this particular instance of AI closer topologically to something we have qualitatively shown to be useful as a strong AI contender via something like the chinese room experiment (I'm pretty sure a lot of humans would fail a chinese room experiment because we can be really dumb, so we can compare/contrast against a human success rate), or to something we modeled after biological research.

Consciousness is nothing special anymore. The magic is modeling the facets of awareness you find fascinating or unique, in any language you want. If you don't think you can model it, try to articulate what quality you would have difficulty modeling. I suspect I would have difficulty understanding the quality in my own experience.

    I see the future as being one in which strong AI
    emerges from whole brain emulation, starting around
    2030 as processing power becomes cheap enough for an
    entire industry to be competing in emulating brains
    the brute force way, building on the academic efforts
    of the 2020s.

For example, we've had fast enough computers to emulate nematode brains for 20+ years but we are still not able to emulate one and have it learn.

http://www.degruyter.com/view/j/jagi