Ask HN: Artificial General Intelligence: emergent consciousness impossible?

[gsanghera]

What is wrong in this?

Let's assume emergent consciousness is possible (whatever consciousness might mean), so that a program can eventually say "OK, I exist / feel / think".

How would this program understand time though? You and I (even animals) understand time by being present here and now. We can rely on our memory to know what happened before, but we don't confuse right now with the past. Not even a second back.

But for a program, with (say) a memory defined as a traversable memory, and cpu cycles defining the moments, all it can "see" is a database with states. It could traverse to any such state independently and declare that to be "now". It has no way of really knowing when is the real now.

Interaction with an agent solves some part of this problem. The newest feed from its sensors could define the now, or the last line of the database. But with perfect recall, why would the conscious program choose to stay there? For it, every state is as real as any other. Any interaction with its sensors might just be interpreted as noise in one dimension, something that can be tuned out with the appropriate response. With every passing tick, the traversal scope keeps increasing.

We don't have such a choice. I can't choose to live in the past (barring mental illness). But I guess what I'm saying is that such a program could choose to "live" in any state in its past.

Clearly this is absurd. Permutation City discusses this problem a little, but still thinks emergent consciousness is possible. I didn't find that satisfactory for the reasons above.

Any thoughts?

[tgflynn]

I am going to make two claims:

1) Consciousness is not a computation.

2) Science will never be able to disprove claim (1).

If you doubt claim (2) then explain how you could prove to me that you are a conscious being as opposed to a figment of my imagination invented by my subconscious mind in a similar way to the "beings" one experiences in dreams.

So obviously I don't accept your assumptions and further I think that your confusion follows from those assumptions.

You seem to be imagining an AI as some kind of independent being that can make "free" choices. But an AI, at least to the extent that anyone working in the field is able to conceive of it, is nothing like that whatsoever. An AI is little more than an optimization algorithm that attempts to optimize some utility function provided by its programmer. So the answer to your question is that if it is useful for the purposes of optimizing its utility function for a (sufficiently advanced) AI to process data in temporal order then it will do so, otherwise it will not.

[gsanghera]

I think I was making the same point as you - that consciousness is indeed not a computation.

But I actually wanted to hear from someone who thought otherwise, rather than you - dream brother :)

[anoniuyiu33412]

your answer somewhat reminded me about a current trend in advanced neurological stuff, could be some paper (I'm not googling it sorry, no time now), and a recent interview with some Ace AI guy at Facebook.

Something like this "human could be no general intelligence at all"

So we apparently are not much than some kind of pattern prediction engine, the whole societies we build is just a subproduct of some base algorithm, probably we are just bigger - talkative - ants.

[anoniuyiu33412]

> I guess what I'm saying is that such a program could choose to "live" in any state in its past

It could, but if an algorithm is somewhat conscius, could also probably have concurrent access to the stored information (the past), and just like us, it won't be able to make much of it after some time, there is just so much information you can generate from a given set of data.

Specially if the algorithm has a purpose, or 20.000 purposes .

At that at point the AI could choose to advance its relative point of view of the time to the current events, in order to try to add new stuff to the whole almost unvariable dataset it already has generated.

No guarantees can exist if the algorithm has no purpose at all.

[anoniuyiu33412]

> How would this program understand time though? You and I (even animals) understand time by being present here and now. We can rely on our memory to know what happened before, but we don't confuse right now with the past. Not even a second back.

What ever question you can make about anything related to time, and get answer of information already stored, that is the past.

Anything about you can make a question and you have no stored information, you determine there is no existing information about the subject, that's the present.

Anything after the absent of information is the future. Empty patterns.

[gus_massa]

> But with perfect recall,

This assume an infinite hard disk (or at least big enough). And even in that case, some info can be simplified to get faster access, and keep the details in a cold storage.

[gsanghera]

What I meant was - the ability to revisit the past perfectly, rather than a "dimmer" version of it.

[gus_massa]

To revisit the past perfectly, you need a huge hard disk.

Video uses about 1GB/hour, that's almost 9TB/year. And now add something like 360° video, or a quality, or more fps. And now add all the other senses. To revisit the past perfectly, you need a huge hard disk.