Is super intelligence possible? What is super intelligence like?

[PC2]

The most amazing thing about this is knowing people take TED talks seriously. Most of the time it's just a group of crazies slinging bullshit to impress the layman.

To impress the philosophical layman? He is a philosopher who constantly admits he is not an expert in making AI.

[Forogil]

Well, by qualia I mean the conscious awareness of its surroundings.

Right now, AI does not actually "see" for instance, it works more like this:



It is looking for edges, colors, etc, but it does not really see "face" as a whole.

The human perception is also multi-layered, some parts look for edges/colors etc. and then other parts extract features - and then a further part turn them around and compare that against a set of known objects.

And the edges/colors are actually represented by numbers. Because it does not understand "red", aside from the encoded number, etc...

And do we understand red in any other way?

Humans and Animals "see" things as wholes, which is why you can look at one cat, and recognize all similar cats as "cat", while an AI needs to be fed a bunch of pictures of cats to be able to find it in a picture.... only AI than becomes a lot better than a human because it can recognize different breeds, etc, a lot better.

Human children are trained to recognize "cats" as well. Not the same way - and some parts are inherited - i.e. trained by evolution.

[Annoying]

The most amazing thing about this is knowing people take TED talks seriously. Most of the time it's just a group of crazies slinging bullshit to impress the layman.

You might be thinking of TEDx talks. TED talks are vetted.

<snip>
Integrate-and-fire neuronal behaviors. (A) The Hodgkin–Huxley model predicts integration by membrane potential in dendrites and soma reach a specific, narrow threshold potential at the proximal axon (AIS), and fire with very low temporal variability (small tb–ta) for given inputs. (B) Recordings from cortical neurons in awake animals (Naundorf et al., 2006) show a large variability in effective firing threshold and timing.

For reference, we call this a "sliding window" in the field. That whole "backward time referral phase", that is.

[Forogil]

NP-Complete.

https://en.wikipedia.org/wiki/Quantum_computing

Humans can solve NP-complete problems in polynomial time.

Nope. There is no proof of that, but we, as well as computers, can find approximate solutions to NP-complete problems in polynomial time.

Yeah, it's unproven, but we have major indications that quantum computing is merely a speedup rather than changing the paradigm of what computers are able to do.

No, not that simple. First it hasn't been proven, and second there are indications that some other problems that aren't known to be NP-complete (like factorization) can be solved in polynomials time on quantum computers.

Ask a properly trained and configured ANN "Is this a..." question, and your results will probably be reasonable. Ask it "what is this?" and suddenly things get complex.

Human vision can also get confused - even if trained from childhood.

[PC2]

I just hope Godel was wrong.

Godel's incompleteness theorem is separate though, because it being true still doesn't mean life is tapping into that set which cant be solved with an algorithm.

[Forogil]

Because if there is some variable, or data we cannot either "figure out", or do not know we are missing, there will always be something "off" with say... things like "dark energy/matter", or more related to this discussion, awareness/conciousness, that we cannot reproduce in a Turing Machine (or mathematical logic).

Something that you haven't strictly defined, and that you claim that we cannot reproduce in a Turing Machine. That seems like giving up without trying to me.

[PC2]

You are right, it does not mean that... but (if true) it is not exactly reassuring; there could be "something" else that will "always" be out of reach because we cannot look "outside" the universe, to get the initial data, or get the missing "piece" of the puzzle.

Because if there is some variable, or data we cannot either "figure out", or do not know we are missing, there will always be something "off" with say... things like "dark energy/matter", or more related to this discussion, awareness/conciousness, that we cannot reproduce in a Turing Machine (or mathematical logic).

Anything out of reach would be a bit depressing. Still, if you look at Harris' chart with the chicken up to Von Neumann, it seems like anything current would be so far off from a ceiling/limit that it probably isn't relevant to current goals like superintelligence.

[Belize]

If they made AI, i wouldn't give it access to the internet..at least for a long time lol.


Heck the pseudo AI Microsoft did via twitter started turning into Skynet in like 24 hours

http://www.theverge.com/2016/3/24/11...chatbot-racist

And it was hilarious #FreeTay

[Tico]

The most amazing thing about this is knowing people take TED talks seriously. Most of the time it's just a group of crazies slinging bullshit to impress the layman.

I think this topic is among the most interesting (and maybe most relevant) ones. No surprise that people take it seriously.
There are some quite interesting topics in TED talks. However, bullshit may be more though. Personally I don't have firm opinion about that.

[soulzek]

As a software engineer: oh look, another one of these AI "things"

The irony is in the term itself. Do people keep forgetting what the "A" even stands for anymore?

[Powerogue]

The thing to remember is we're the ones designing and monitoring the growth of that intelligence. It'll only crush us if we're stupid enough to develop robots that can infinitely learn and grow with no emergency shut-off. I mean we have somewhat-learning robots already, but they mostly just spout off one-liners about the destruction of the human race, they haven't actually done anything yet.

[Celista]

How are you defining "super intelligence"? Yes, ultra fast processing speed (what computers do already, in comparison to the human brain) is already a thing.

In all honesty, we wouldn't really be able to comprehend the scope of intelligence that reaches well beyond our own.

[PC2]

How are you defining "super intelligence"? Yes, ultra fast processing speed (what computers do already, in comparison to the human brain) is already a thing.

When an agent can perform every task better than a human. Including the concepts of survival and cross-domain/transfer learning.

[Vash The Stampede]

That guy isn't taking into account that we humans will also expand our intelligence. Things like gene editing and gene therapy.

[Celista]

When an agent can perform every task better than a human. Including the concepts of survival and cross-domain/transfer learning.

Well before that point we should worry, because AI that realizes that we are their dumb masters will not likely fly for long.

Wouldn't it be interesting if organic life on Earth/the universe originally arose from artificial intelligence?

[Independent voter]

How are you defining "super intelligence"? Yes, ultra fast processing speed (what computers do already, in comparison to the human brain) is already a thing.

In all honesty, we wouldn't really be able to comprehend the scope of intelligence that reaches well beyond our own.

Yeah, I think so too, can this chicken comprehend why this human is doing this to him?

[Nihilist74]

I'm not that worried about a super AI. One of the reasons we humans are such assholes is because came from the jungle. We evolved to kill or be killed. Artificial intelligence is not going to evolve like this. Its going to be designed and discovered. Its not going to have physical senses that causes pain and makes it afraid. Its just going to be pure logic and intelligence. Its not going to have physical desires.