Print

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

[Reelya]

Well there are uses. e.g. you could make a decent news aggregator out if it, especially if you trained in on getting the right type of people to "like" your shares, you could make an industry-specific aggregator. Plus, if the bot always fed people stuff they're likely to like, then people would think the bot was handy to know. The bot could in fact be feeding different groups very different shares: e.g. sharing alt-right anti-hillary memes one moment and the next moment sharing anti-trump memes to another crowd.

[penguinofhonor]

You could also just slip some advertisements and product placement in between the rest of the posts. Then you're making money, and that's the most useful thing of all.

[Reelya]

Spoiler (click to show/hide)

[Liberonscien]

Can you elaborate on that?

Why do you think there was a thread here where people created chatbots based on forumites, actually?

Do you feel happy talking about every post by everyone over a day or two being imported into the thing and seeing what happens?

Tell me more about the markov chain generator, which is not terribly advanced as an AI.

Why do you like, for example, getting maximum number of likes on its facebook posts?

Heh, this reads like a chat bot said this.

[Reelya]

I kinda wish you'd renamed this the "AI Chatbots" thread. Because when I clicked it I expected a discussion on actual AI. It was disappointing when it's only about chatbots, because they're basically the garbage toy level of AI.

[Zanzetkuken The Great]

I kinda wish you'd renamed this the "AI Chatbots" thread. Because when I clicked it I expected a discussion on actual AI. It was disappointing when it's only about chatbots, because they're basically the garbage toy level of AI.

Only reason why it is currently on that path is due to my misunderstanding of how advanced the thing in the OP was.  No reason the discussion cannot change.

[McTraveller]

So I've been thinking a bit recently about AI and all the stuff in the almost-mainstream media about it. I kind of follow it with more than idle curiosity, but not much more than that, but today I came to what I think is a realization on a few things:

1. What most people today consider "Artificial Intelligence" is actually really "Artificial Skill". Even the most famous of these, Google's Go playing program, is just very skilled.  Generally speaking, most things we have today are very skilled pattern recognition (and even forecasting) machines.  These are "highly skilled" activities, but I don't think they are intelligence.

2. I believe that there are three characteristics that should be used to identify an AI which are vast improvements on, say, the Turing test.  These two characteristics are: 1) the ability to know if a concept is correct (or incorrect) without someone telling you and 2) the ability to teach / explain a concept to another intelligent being; and 3) the ability to determine one's own goals / decide to perform tasks other than what was "programmed" and explain why those choices were made.  This is a broad classification, and also generally includes "the ability to learn", which is why, say, a supermarket calculator which is "skilled" at arithmetic won't ever be (has never been?) considered intelligent. It also excludes those highly-skilled single-purpose systems.

Most expert systems we have do not really have a way to know, without outside confirmation, if they have correctly or incorrectly categorized a problem.  Consider doing your homework - you will often know with pretty high probability if you have an understanding of a concept and if an answer is correct without having to check.  I'm not aware of many expert systems that "know" they are correct.

And then consider the ability to educate - we have evidences of other animal life that is highly skilled, but doesn't really teach or pass on that information, and even other animals that do pass that information on.  But I'm fairly certain we don't know if those species have any kind of implicit understanding of correctness or not...

Also, I fully understand that my criteria exclude some humans from being classified as "highly intelligent" - there are many people that can't teach others, and many who cannot explain why they choose certain actions, and some people who can only do things they are instructed to do. I note, though, that historically such individuals have been labeled with some form of the word 'impaired', or classified as "less intelligent" - this did factor into my use of this as a classification.

Anyway, those are my ramblings, further discussion encouraged.

[wierd]

I hate to be abstruse, but a human brain is just the same concepts, in much higher stacking and proportion.

Biological cells are surprisingly robust, and efficient, given what they are.  They are very hard to beat in terms of general practicality. They are sufficiently cheap, that nature produced systems capable of general intelligence, BY CHANCE.

Should similar levels of practicality be reached with artificial automata, artificial general intelligence is definitely possible.

The problem is that artificial automata DO NOT come anywhere close to the robustness and practicality of organic automata. As such, we are stuck with "Highly domain optimized, stripped down, analogs of significantly lower complexity and capabilty."

[McTraveller]

Oh I wasn't at all saying that we couldn't get there - I was just proposing something to be a bit more robust in terms of classifying "intelligence" - especially because so many people say "we don't even know how to define intelligence!"

I was proposing a definition there that doesn't implicitly exclude computers and doesn't necessarily just assume that all humans have it, and also doesn't assume that "it's got to be a neural net."

Kind of like the Turing test, it is based on observation - does it know it is correct/incorrect? Can it teach? Is it general purpose?

[wierd]

Any program that implements a cache, "knows" things.

This is essentially what any human who "knows the answer" is doing. They have either previously had the answer given to them, and stored it (in the cache)-- or they have found the answer, and stored it (in the cache.)  Later, the question is asked again-- the human retrieves the answer from their memory (cache), and already knows it is the correct answer.

Computers do the same basic thing all the time now.

[McTraveller]

It's not the cache - it's not "knowing the answer".  It's being able to know that the answer is correct, when it hasn't been previously stored in a cache.  Or better yet - being able to know that what someone told you to put in the cache is good or not.

So something like seeing a new problem, coming up with a solution, and then running it through a "is this solution correct" heuristic would satisfy the criteria I proposed.  So call it "cache plus heuristic" maybe.  For math problems this is trivial, but for something like "is that thing I just saw through the trees a bus, or something else?" doesn't have as clear-cut a heuristic to "know" if it is correct or not (interestingly, not even for humans) - but you can be sure a person who sees a bus through trees.

That's the kind of thought experiment I was proposing, even if my details are slightly ethanol-influenced at the moment.

[Reelya]

Oh I wasn't at all saying that we couldn't get there - I was just proposing something to be a bit more robust in terms of classifying "intelligence" - especially because so many people say "we don't even know how to define intelligence!"

Maybe it's over-specified however. The majority of humans might not be able to meet your criteria in a strict sense.

Are all humans really able to just deduce that a concept is correct without any outside confirmation? That's sort of questionable, because the knowledge of humanity didn't arise like that, it arose through people throwing out both true and erroneous statements, and seeing what stuck.

The idea of a perfectly logical entity just determining universal truth and falsehood of concepts from first principles is in fact much more fanciful than the idea of making an artificial brain that has basic consciousness. We know that if you chunk enough (trillions of) neurons together with some appropriate stimuli then the mass self-organizes into intelligence, because it happens in all animal species. We have no evidence that a universal truth machine is possible, and in fact there are mathematical proofs saying that it is not.

Are all humans able to teach something to another being? Many humans are able to do intelligent things, but teaching is an entirely different skill, and it's questionable whether learning comes from the teacher or the students. e.g. teaching by example is just doing, and the student then learns by copying.

the ability to determine one's own goals / decide to perform tasks other than what was "programmed" and explain why those choices were made.

but the fact is that machine learning systems are only "programmed" by us, to about the same extent that a neuron is "programmed" by the DNA that went into it. So just because there was an underlying design system which shaped how a neutral network grows doesn't tell you anything about whether it's possible to produce intelligence with it.

Anyway, humans can't even explain why we make the decisions we make. We're easily screwed over by simple cognitive tricks. Sure, a better formulation of intelligence is needed, but your model has very little to do with how human intelligence actually works, so it's in danger of becoming a type of strawman.

[Antioch]

AI doesn't need to be a general AI of human capabilities to be extremely beneficial to humanity.

I think smart algorithms and neural nets designed for specific tasks will be some of the most major technological progress we will see in the coming decades.


Take for example the Watson AI that is now already being consulted to assist in diagnosing medical patients.

Google Deepmind is also rapidly improving their learning algorithms, the most known being their AlphaGo program that a while ago beat the top ranked Go players in the world 60-0 in online Go matches.

[McTraveller]

AI doesn't need to be a general AI of human capabilities to be extremely beneficial to humanity.

I think smart algorithms and neural nets designed for specific tasks will be some of the most major technological progress we will see in the coming decades.


Take for example the Watson AI that is now already being consulted to assist in diagnosing medical patients.

Google Deepmind is also rapidly improving their learning algorithms, the most known being their AlphaGo program that a while ago beat the top ranked Go players in the world 60-0 in online Go matches.

Yeah, the medical stuff is very useful.  But is it "intelligence" or is it something else? I tend to lean toward it being as Ipsil said - just massive statistical analysis machines.  There is that strange 'qualia' that is missing with what we have today in terms of AI that makes those things great pattern matchers or good game-players (even very complex games!).

The majority of humans might not be able to meet your criteria in a strict sense...

The idea of a perfectly logical entity just determining universal truth and falsehood of concepts from first principles is in fact much more fanciful than the idea of making an artificial brain that has basic consciousness. We know that if you chunk enough (trillions of) neurons together with some appropriate stimuli then the mass self-organizes into intelligence, because it happens in all animal species....

Well yes - I don't actually presuppose that all humans are actually "intelligent".  I think that's something that reeks of some kind of bias.  It doesn't help either that "intelligence" is actually a spectrum, not a "yes or no" situation.

Put another way - if you had to propose a definition for intelligence, what would it be? I mean, the dictionary says "the ability to acquire and apply knowledge and skills" - so maybe that's sufficient?  And it's pretty broad.

[Antioch]

AI doesn't need to be a general AI of human capabilities to be extremely beneficial to humanity.

I think smart algorithms and neural nets designed for specific tasks will be some of the most major technological progress we will see in the coming decades.


Take for example the Watson AI that is now already being consulted to assist in diagnosing medical patients.

Google Deepmind is also rapidly improving their learning algorithms, the most known being their AlphaGo program that a while ago beat the top ranked Go players in the world 60-0 in online Go matches.

Yeah, the medical stuff is very useful.  But is it "intelligence" or is it something else? I tend to lean toward it being as Ipsil said - just massive statistical analysis machines.  There is that strange 'qualia' that is missing with what we have today in terms of AI that makes those things great pattern matchers or good game-players (even very complex games!).

(...)

I would certainly call these machines a form of intelligence. In part because they are actually learning machines that train themselves to perform tasks.

And their classification in no way diminishes their merits, they WILL start to perform a lot of tasks we consider to require intelligence.

There is a large difference between how computers play chess and how AlphaGo plays Go. The chess computers use fixed algorithms and brute force to solve the problem, Some parameters are tweaked over time to make them stronger.

AlphaGo however is a learning machine. It started like a beginner and by looking at human Go matches and playing millions of matches against itself it became insanely good. AND IT IS STILL IMPROVING ITSELF.