Anyone else feel we are on the cusp of some really exciting and scary AI?

[Locut0s]

Let me be clear I'm not talking about general purpose human consciousness level AI. It's hard to know if we will ever get there. However the pase of recent AI development from things like Alpha Go and many many other applications of neural networks is... I have to admit a little unsettling to me. No I'm not some doomsday conspiracy theorist. Skynet isn't here soon, is it ;-)?

The kinds of things we are seeing now in AI though are definitely extremely exciting. I think we are now seeing what we can argue to be "thinking" on a limited level. Remember these don't operate by brute force searches. Take a look at the recent breakthroughs with Alpha Go Zero.

https://en.m.wikipedia.org/wiki/AlphaGo_Zero

It managed to beat the previous Alphago 100 games to 0 and that is the one that beat Lee Sedol. More impressive still it reached that level in just 3 days of training Vs months for the previous. AND it does so on much less hardware. It's also coming up with new go moves that human players have never seen that the master players are now looking to copy.

You may argue that this is not "thinking" because it's just a program. And it's so limited in scope. All it can do is play Go at a superhuman level. Well sure but that's the start. And it's clear to me that lots of animals engage in limited thinking. Remember that computers are able to leverage their own advances and improve upon themselves much faster than biology, at times exponentially.

We are already seeing the use of neural networks in our daily lives in things like steadily improving voice recognition. You don't realise how scarily good it is until you get a Google Home. My father has one and we were playing around with it the other day. Yes I know home itself is just a cheap piece of plastic and the neural networks are in tons of server farms but that's beside the point.

We are in for some very interesting times. I really do think though that we need to start seriously thinking about AI safety. You don't need Skynet or general consciousness for some scary unintended consequence. Especially since these are black boxes to everyone most of the time.

 

[IronWing]

It is kind of interesting. How does one program AI to recognize that a solution, while correct, is also stupid?

 

[Locut0s]

It is kind of interesting. How does one program AI to recognize that a solution, while correct, is also stupid?

What?

 

[Locut0s]

It is kind of interesting. How does one program AI to recognize that a solution, while correct, is also stupid?

Ohh you mean how would an AI recognise a correct yet still stupid solution?! Yes that's kind of at the heart of what I'm saying we need to be careful of. Because we don't know how AIs work entirely we don't understand entirely how they reach the solutions they find. So we are left with our own human level intelligence as the controlling factor. And as smart as we are, we are also very stupid. And our own stupidity has been our own undoing a lot historically.

 

[Yakk]

There are so many resources worldwide being consumed by AI it is impressive. Any potential university grad associated with AI gets bought up by the private sector, nevermind professors. It's actually a big problem for universities in the last couple years.

It is kind of interesting. How does one program AI to recognize that a solution, while correct, is also stupid?

A self learning AI does exactly that. It learns one way to do things that works, until it doesn't, then it tries variations until it finds another way that works for the new problem, and the previous one(s) .

Case in point, the AI Google is developing to beat DOTA 2 (Google's new project). It is nearly impossible to play against as a human player (tested against arguably the best professional DOTA2 player worldwide) because of the bizarre way the AI has thought itself to win. We could say the solution the AI developed on its own is stupid, but it works.

Subsequent study allowed human players to beat the AI, and the AI will work through those new variables also. The AI is playing games against itself 24/7 and systematically learning every time.

 

[BurnItDwn]

I'm waiting for the car AI that sacrifices the driver of the car in order to save a larger number of lives.

 

[Red Storm]

The Go stuff is definitely really interesting. I'd love to see the progress they make with Starcraft 2 in a few years. If they can develop a Starcraft AI that can actually compete with and/or beat the best in the world... that will be incredible.

 

[Locut0s]

I'm waiting for the car AI that sacrifices the driver of the car in order to save a larger number of lives.

That's actually an interesting example. Machine learning is used for self driving cars currently but they don't make any of those kinds of decisions. However once self driving cars prove themselves able to handle just about any normal situation, they are close now, then it makes sense that insurance companies might want them to introduce some kind of cost function for the life of the driver. A computer can react far far faster than a person. In the advent of an accident this means that it might be able to "choose" how the driver gets injured if it can't avoid a collision. The insurance companies would want whatever way that injury occurs to be cheap to cover. Whether that involves sacrificing the driver is an interesting argument to have. But if we don't think about these things at all they will be upon us before we are ready.

 

[Locut0s]

The Go stuff is definitely really interesting. I'd love to see the progress they make with Starcraft 2 in a few years. If they can develop a Starcraft AI that can actually compete with and/or beat the best in the world... that will be incredible.

How does Star Craft compare to Dota 2? As Yakk mentioned Google already has an AI that beats the best at that game and that's not a really simple game.

 

[Red Storm]

I'm waiting for the car AI that sacrifices the driver of the car in order to save a larger number of lives.

I'm trying to think of a situation like this in the real world but am coming up empty. The only way I can see it is if the car is already driving way faster than it needs or something, which is what a human would do, not an AI. Considering the car will already be driving safe, I don't see how it will happen.

For example, you're on a bridge and there's a bunch of people in front of the road, so the car determines it has to drive off the road to not hit those people, which may kill the driver. I don't see how this works because AI can already "see" better than we can, and in all directions at the same time. Add to that the car won't be speeding because it's AI run not human controlled, and I don't see how that scenario would realistically come about.

How does Star Craft compare to Dota 2? As Yakk mentioned Google already has an AI that beats the best at that game and that's not a really simple game.

Pretty sure I read that it was a more simple AI that was programmed to run a select (few?) builds, that after a while the human players were able to easily beat once they learned the quirks of its play style. To me that's no different than the hardest level AI you already see in RTS games. Once you learn the quirks, you can beat it. A truly progressive RTS AI would not have these quirks, it would keep adapting and formulate new moves.

 

[BurnItDwn]

That's actually an interesting example. Machine learning is used for self driving cars currently but they don't make any of those kinds of decisions. However once self driving cars prove themselves able to handle just about any normal situation, they are close now, then it makes sense that insurance companies might want them to introduce some kind of cost function for the life of the driver. A computer can react far far faster than a person. In the advent of an accident this means that it might be able to "choose" how the driver gets injured if it can't avoid a collision. The insurance companies would want whatever way that injury occurs to be cheap to cover. Whether that involves sacrificing the driver is an interesting argument to have. But if we don't think about these things at all they will be upon us before we are ready.

I can't take credit for the idea. It's been floating around the interwebs for a couple of years now. Essentially, automakers have asked the federal government to set the guidelines, they clearly don't want the liability.



https://gizmodo.com/should-your-driverless-car-kill-you-to-save-two-other-p-1575246184
https://www.technologyreview.com/s/542626/why-self-driving-cars-must-be-programmed-to-kill/

 

[BurnItDwn]

I'm trying to think of a situation like this in the real world but am coming up empty. The only way I can see it is if the car is already driving way faster than it needs or something, which is what a human would do, not an AI. Considering the car will already be driving safe, I don't see how it will happen.

Thankfully those scenarios are very rare.
Generally, most scenarios involve human drivers in a different vehicle breaking rules and putting the self driving car and it's occupant(s) in danger. (Or, very rarely, a disaster or mechanical failure setting off events)

For example, you are on a city road, there is a traffic jam to the left of you, there is a crowd of pedestrians on the sidewalk to the right of you, and there is a truck coming at you head on at high speed. The self driving car may allow the truck to smash you to bits since otherwise, you're escape route could kill many pedestrians.



EDIT:
Also need to consider, foggy weather, bad visibility from heavy snow, or optical/sensor being comprimised somehow.

 

[Yakk]

How does Star Craft compare to Dota 2? As Yakk mentioned Google already has an AI that beats the best at that game and that's not a really simple game.

Pretty sure I read that it was a more simple AI that was programmed to run a select (few?) builds, that after a while the human players were able to easily beat once they learned the quirks of its play style. To me that's no different than the hardest level AI you already see in RTS games. Once you learn the quirks, you can beat it. A truly progressive RTS AI would not have these quirks, it would keep adapting and formulate new moves.

The best non-cheating RTS AI currently is AotSE, it is adaptive, but self learning. Meaning the AI in AotSE may not react the same way to similar situations, but won't permanently add this to its knowledge base.

The AI used in DOTA 2 by Google, and the new entry in StarCraft by Facebook are self learning AI's.

In Google's case, so far they are working on 1 champion in 1 position. A MOBA game is many orders of magnitude more complex than boardgames, which is why they are attempting it.

 

[Locut0s]

Thankfully those scenarios are very rare.
Generally, most scenarios involve human drivers in a different vehicle breaking rules and putting the self driving car and it's occupant(s) in danger. (Or, very rarely, a disaster or mechanical failure setting off events)

For example, you are on a city road, there is a traffic jam to the left of you, there is a crowd of pedestrians on the sidewalk to the right of you, and there is a truck coming at you head on at high speed. The self driving car may allow the truck to smash you to bits since otherwise, you're escape route could kill many pedestrians.



EDIT:
Also need to consider, foggy weather, bad visibility from heavy snow, or optical/sensor being comprimised somehow.

As soon as self driving cars start entering the market they will take over very quickly. It doesn't matter IMHO that people will push back. A lot of people me included love driving. But accident and fatality rates are going to drop like a stone as soon as they start becoming common. Insurance companies, auto makers, and law makers will take note. Not too long after people will think you are crazy to want to drive your own car. I could be wrong but just my 2 cents.

 

[Red Storm]

As soon as self driving cars start entering the market they will take over very quickly. It doesn't matter IMHO that people will push back. A lot of people me included love driving. But accident and fatality rates are going to drop like a stone as soon as they start becoming common. Insurance companies, auto makers, and law makers will take note. Not too long after people will think you are crazy to want to drive your own car. I could be wrong but just my 2 cents.

I can't wait for self driving cars. I can get in my car and resume my Netflix binge while I'm taken to work/home.

The best non-cheating RTS AI currently is AotSE, it is adaptive, but self learning. Meaning the AI in AotSE may not react the same way to similar situations, but won't permanently add this to its knowledge base.

The AI used in DOTA 2 by Google, and the new entry in StarCraft by Facebook are self learning AI's.

In Google's case, so far they are working on 1 champion in 1 position. A MOBA game is many orders of magnitude more complex than boardgames, which is why they are attempting it.

They are doing it with Starcraft 2 as well. Excited for progress in that game.

 

[BurnItDwn]

As soon as self driving cars start entering the market they will take over very quickly. It doesn't matter IMHO that people will push back. A lot of people me included love driving. But accident and fatality rates are going to drop like a stone as soon as they start becoming common. Insurance companies, auto makers, and law makers will take note. Not too long after people will think you are crazy to want to drive your own car. I could be wrong but just my 2 cents.

I think it'll be like 20-30 years of "self driving cars which have an override for manual driven mode for private property, trails, offroading, etc ", and even then, people will want to drive their classic or collector cars.
So essentially, new self driving cars will have to co-exist with Model T's until the end of time.

 

[Red Squirrel]

It's already happening, look at Youtube Content ID. And it's bad, because AI (at least that level) cannot reliably differentiate between fair use and other exceptions of copyright law. It just knows that the content is the same or similar as other content, and then removes/strikes/demonetizes etc the video. But because of Youtube being so big, whatever their AI says is right/wrong may as well be taken as being law because you will get punished based on what they say not what the law says. They could go around and delist websites from Google that their AI determines should not be listed, for example too (if they don't already). AI in this sense can be very dangerous, and really this is how it will take over. Imagine AI being used to determine if you are guilty of a crime for example or to determine what kind of medication you should take instead of a doctor, or other day to day things. So everything would be binary with zero exceptions or considerations for each particular situation. Actually the movie iRobot goes over this when the robot saves the person in a car under water. The AI determined the best person to save based on the chance of survival (which kind of make sense when you think of it logically) and not the fact that the other person was a child. AI has no emotions. This can be very dangerous and unfair in a lot of situations where AI will be used to make decisions.

People think of killer robots when they think of AI taking over, but reality is, it will be AI that is sponsored by major corporations and/or government that take over. Killer robots that take over would actually be easy to defeat, because chances are even the government would want to defeat them too. But corporate AI is going to be legal and backed by the government so even if you know a way to defeat it you won't be allowed to. Chances are also good that any kind of robot will always have a kill/override switch. I guess there is the idea that the AI becomes self aware and disables that switch because it wants to ensure it's own safety, but honestly I really don't imagine AI will get to that point, as it's not really the goal to make a self aware AI. The goal will be to streamline and automate decision making based on input data.

 

[Hayabusa Rider]

Soon truck drivers will be on unemployment. Factory workers as well. Office workers, middle management. Improved mobile machines which can pave a driveway or do plumbing, most menial jobs. The people in power remain controlling government and law. Professional jobs go- nurses, physicians, surgeons. No jobs and no incomes.

Hardly as nice a time as TNG.

 

[Red Storm]

Soon truck drivers will be on unemployment. Factory workers as well. Office workers, middle management. Improved mobile machines which can pave a driveway or do plumbing, most menial jobs. The people in power remain controlling government and law. Professional jobs go- nurses, physicians, surgeons. No jobs and no incomes.

Hardly as nice a time as TNG.

That's when UBI (Universal Basic Income) kicks in. It has to, or yes, a lot of people will become jobless then homeless very quickly.

 

[Locut0s]

That's when UBI (Universal Basic Income) kicks in. It has to, or yes, a lot of people will become jobless then homeless very quickly.

Agreed we need to plan for this with something like UBI. Because it's coming and fast. Like said above a HUGE number of jobs are within reach of automation. From menial up to things like surgeons. And you can't really fault either the companies or the people for wanting it it's cheaper, more efficient, and usually safer. If planned for this could actually be a great thing as a lot of those jobs are not happy jobs. If we have something like a UBI it potentially frees a LOT of people up to explore other aspects of their lives. This really is just a natural progression that started with the industrial revolution. Our lives are a LOT better for such automation. BUT whenever large scale automation has taken over people have been thrown out of work and suffered a great deal. Turmoil and violence and even wars can result. But if planned for properly a lot of good can as well.

 

[GobBluth]

If you marry TrackingPoint to DARPA's petman, you get the actual Terminator.

https://www.youtube.com/watch?v=5S4ZPvr6ry4
https://www.youtube.com/watch?v=Pmteh_NChOQ

 

[Kaido]

Anyone else feel we are on the cusp of some really exciting and scary AI?

Nah.

It's all a sham.

 

[Hayabusa Rider]

Here's a little thought on possibilities assuming the proper technologies develop concurrently.

So you have no job. You have a universal income but as human societies go we are useless. We had no choice in all of this.

So..

As computing power grows people are paid with VR time. Medicine keeps your body healthy and fit even while sedentary. You go offline to sleep or if you want to get out in "the real world". Everyone gets their own Matrix they can control or have designed for them so there's mystery left. Not glasses and a tank but direct injection into the sensory areas of the brain to override reality, if there really is such a thing.

If you want a nice steak and you have something which cannot be distinguished between the best steakhouse and electronic impulses it makes no difference to the one doing the dining. Everyone gets to be a god or devil or king or serf. A knight, an astronaut, anything at all the the imagination can create.


Now is that all good or bad and why?

 

[ultimatebob]

I don't know, man... I still can't get my Google Home to send text messages yet. I think that we're pretty far away from it becoming self aware and murdering us all.

 

[Hayabusa Rider]

I don't know, man... I still can't get my Google Home to send text messages yet. I think that we're pretty far away from it becoming self aware and murdering us all.

"Hey Google, order my usual pizza form the usual place and have it delivered in an hour. Also find the fastest route with the least traffic to home. Oh, I have an appointment with my dentist at 9:00 AM with my dentist on Friday. Remind me an hour before it".

When will that happen?