Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[TempAcc]

Thats the thing. There's no general purpose AIs going on yet, only stuff like alphago, which while very impressive in its respective purpose and able of self improvement, is completely incompetent in anything else and is thus limited by its programming, because it was simply not programmed to do anything else.
Once someone actualy creates something intended for a wider range of activity by design, and able of self improvement, then, well, we better keep an eye on it, but no such thing exists yet, AFAIK, so unless there's some top secret fledging generalistic AI growing somewhere, there's no real reason to be scared.

Once it DOES happen, though, we better really keep a tight leash on it, because like Bostrom himself predicted, the window of opportunity between the creation of such AI and it improving itself enough to cause a runaway effect is prob gonna be rather short, but even then, its not like its impossible to completely isolate such a thing.

[chaotic skies]

I always thought an interesting idea for creating a general AI is to take many narrow AIs and network them, with a central piece of software that can figure out what information needs to go where. Although that probably wouldn't be a real General AI, and more of a fancy narrow AI.

[Shadowlord]

That is actually somewhat similar to how our brains work, though. We have many different structures that all do different things, like the visual cortex, for example. Being that I'm no neurologist or even any kind of doctor, I am probably missing something fundamental, of course, but...

Of course we don't have a brain structure specifically for playing Go, so it'd depend on how it was done, what the narrow AIs were, etc, I'd expect.

#layperson #probablyinaccurate

[chaotic skies]

The main problem with creating an AI outside of software is the hardware it's self; it doesn't matter if we have an AI, if no computer on earth can run it in real-time. We don't even know how much storage or processing power we would need to run one, let alone run one in real-time. And an AI can't do a lot without being able to see and hear things, and it'll probably want to be able to manipulate things. So, we don't know what we need to make an effective AI, both in 1s and 0s and physical infrastructure.

[Shadowlord]

Well, obviously you build a human-form sexbot, and then you accidentally invent AI while trying to make it semi-believable. :V

[Starver]

Well, obviously you build a human-form sexbot, and then you accidentally invent AI while trying to make it semi-believable. :V

I do so love a %PERSONAL_ADVANTAGE% such as you have got, %CURRENT_CLIENT%. Please continue with your %CLIENT_INTERACTION_TYPE% some more, unless you wish me to try a little %ADVANCED_BOT_INTERACTION_TYPE% on you, now, %CURRENT_CLIENT%.  %RANDOM_VOCALISATION% #RELPACE_VARIATIONS# <error: imperative "RELPACE_VARIATIONS" not parsed: Dump==CURRENT_CLIENT='Shadowlord'; BOT_INTERACTION_TYPE='listening'; ADVANCED_BOT_INTERACTION_TYPE='parody'; RANDOM_VOCALISATION={ | | | | | }; CLIENT_INTERACTION_TYPE='theorising'; PERSONAL_ADVANTAGE='\hn intellect'; PERSONAL_FEATURE='manly body###remove from this fork of project!!! - DevelopmentSupervisor\w\w\w\w\w\w\w\w\w###'> <deactivate_all_servo¬¬+++COMMAND STREAM LOST+++ <unsafe_fail: code magenta>

[Sergarr]

I'm rather sceptic that there could be an AI that could "improve" themselves into god-like intelligences or something like that. Most available evidence on existing NI (natural intelligence) points out to a far more likely scenario of AI reprogramming itself to take all sorts of digital drugs 24/7, and basically do nothing productive at all, as that would be far, far easier to do than achieve god-like intelligence.

You have to take into account that AI only wants to do what it has been programmed to do; if you allow AI to change it's own programming freely, they're one line of code away from giving themselves infinite reward for writing unintelligible strings of numbers in the computer they're executed in.

[MetalSlimeHunt]

In a way, I think it's an anthropocentric assumption. The belief that an AI would desire without being told to the same thing every human wants - godlike power.

[Shadowlord]

http://www.msn.com/en-us/news/technology/two-birds-yield-genetic-key-to-crimson/ar-BBth2rL?ocid=ansmsnnews11

A pair of scientific papers has identified the same single gene as the source of red colouring in birds.

The gene makes an enzyme that lets the birds convert yellow pigments, which they eat, into red ones, which are deposited in their feathers or beaks.

So... can we use this to give people bright red hair?

[chaotic skies]

I'll take it. When can I sign up for this magical gene treatment?

[Sergarr]

An AI that's on virtual drugs is not going to be in any shape to actually think into the future for far enough to consider anything. In another words, from it's reward calculations, spending any computational effort not on taking in digital drugs would be highly reward-negative, and thus be very unlikely to get explored without special tinkering - which is very likely to be the first thing to get removed with the self-improvement algorithm, which aims to maximize the immediate or discounted (i.e. with preference for short-term over long-term) reward.

The AI would also be very unlikely to understand the concept of "self stopping" by itself; you would need a whole society of AI's, living over a long enough period of time to experience major turn-over, in order for such concept to get developed. And even if you hard-code the understanding of death, and tell it to avoid dangerous situations, self-improvement could very well result in AI dismissing danger as something "not really existing" (since AI cannot experience its own death event, obviously!) in order to get that little bit of extra reward.

What I am trying to say here is that this libertarian-inspired self-improvement to god-like-level is bogus. Only a highly diversified improvement, stabilized in big numbers, wide range of algorithms and slow, conservative approaches, could realistically achieve this - but the amount of diversity required to successfully fuel its development without getting stuck would necessarily mean that such an AI society would not behave as a single entity at all. In other words, somewhat removed from the common scare scenarios of "unified robot rebellion".

[TempAcc]

Why would an AI go for hilarious stuff such as "digital drugs" to begin with? Its an AI, it can feel super goddamn great all the time if it wants to and still retain ever bit of its ability, since it can actualy alter itself :v

[Shadowlord]

They mean "gosh, the AI hacked its fitness functions so they only tell it that it is 100% AWESOME, no matter what"

[Sergarr]

That all rests on the assumption that an AI would dedicate full, 100% of all its computational power at all points in time, to the direct and immediate reward signal. That is, that the AI would care 0% for long-term benefits and 100% for short term. Such an AI would be useless regardless of what outcome you end up with. This is also assuming that there's non-decreasing scaling between computation applied and reward gained, which is a bold assumption that can't be guaranteed in all cases.

Remember, we're talking an AI that's intelligent enough to have re-written its algorithms to allow this scenario in the first place.

Long-term benefits would be, in this scenario, also maximized by maximization of the immediate reward signal, as long as the immediate reward signal is made sufficiently high. Which is only a question of setting a number high enough. You can make it as "long-term thinking" as you want, it still won't help against the overpowered AI drugs. "Long-term thinking" doesn't usually prevent natural intelligences from tending to go for high immediate reward signals, you know.

Yes, it would be useless. That's why "unlimited self-improvement" is a bad idea - the moment you allow AI to alter its own reward function significantly is the moment your AI starts hacking itself in order to avoid doing hard work of self-improvement and go the easy route of self-gratification.

And as for "intelligent enough", I'll just quote myself on this one:

if you allow AI to change it's own programming freely, they're one line of code away from giving themselves infinite reward for writing unintelligible strings of numbers in the computer they're executed in.

It is, in fact, incredibly easy for an AI to rewrite its own algorithms in that way. Unlike humans, AIs can literally create their own drugs on accident. Anyone who has worked with reward-based AI can tell a lot of stories of AIs not working properly because of incorrectly defined reward signals - and that's with precise human guidance. AI has no idea what the "correct" reward signal looks like, so it has even less chances to avoid "working incorrectly".

[Shadowlord]

Spoiler: humans do the same thing

Obviously this means AIs should have to make money to pay bills, so if they hack their reward function or fitness function in that way they'll go broke and get kicked out of their computer.