Comparing computers today to "four year olds" is lunacy... the four year old is capable of many tasks no computer can easily accomplish. That said, eventually it won't be a joke, and it would be nice if we had this conversation as a species before then.
> A few years ago, Giannandrea compared artificial intelligence to a 4-year-old child. Today, he revised his statement and said that it’s even worse than that. “They’re not nearly as general purpose as a 4-year-old child,” he said.
It seems like right now "superhuman AI" is a buzzword that people like to use when they want to be covered by the press.
I'm surprised OpenAI didn't chime in. Physicists seem to use Aliens or multiple dimensions for this purpose (But some also use AI for the same effect).
It sort of distracts people from actually asking real questions like how to use AI / ML responsibly because the former doesn't require much to speculate about.
Journalists don't feel like they are qualified to report on the actual technology (which is a good thing), don't bother learning anything in order to become qualified (which isn't), don't bother speaking to qualified people on the front line of this technology (which is horrible).
So what they have resorted to is reporting on these "philosophical" topics, because all you need for that is a fucking opinion, right. It's a great Faustian bargain because you then get all those companies and people, who similarly have no clue but are fishing for PR, to pile on.
See "should the autonomous car hit the pedestrian or save its passengers" or "this artist drew a lane marker around his beater car".
Until it's conscious, AI is just a tool and the same ethics apply when using it as when using any other tool. If you use it to hurt people, to deceive people, to steal from people, etc, you go to jail. Well, ideally. But like insurance companies that are forbidden from charging different rates to people based upon their being a member of a protected class or any proxy which becomes essentially equivalent. So if their ML system starts jacking up premiums on one group of people because its found an indicator it likes, they're still breaking the law. Even if they can't explain why it keyed on that indicator beyond "look... here's a list of numbers. Those are weights in the neural net. We don't know what they mean."
I don't fear the AI vs humankind scenario, the one that might get scary in the next decade or two is the humans with AI-powered weapons vs humans with AI-powered weapons.
And it is not about destroying things, it is about destroying trust, society and government. Right now it's about online ratings, ads, sports articles and similar things, but I'm sure there are people working on targets that are more suitable as weapons.
So he compared modern computers to 4 year old and then said worries of superhuman AI are overblown? That's kind of worrying. 4 year olds typically get a lot smarter.
Agree. How far open does the barn door have to be before it goes from 'irresponsible' to 'responsible'? When does it become a matter of public interest?
What time scale is this guy assessing risk on? Does he have or expect grandchildren? Does he mind if the rest of us do?
I'd appreciate it if industry people like this could behave a little more... responsibly.
One concern is the prospect of AI overtaking human intelligence. Another one is the extent to which companies can employ these technologies on consumers? The latter have some much more present dilemmas that are both commercial and ethical. Companies naturally prioritizes the commercial aspect.
There's a simple explanation for Musk's views on the dangers of AI: it's too dangerous to develop on Earth, it must be regulated. OTOH, you can go to Mars where you don't have to worry about such dangers and would be free to pursue all sorts of research (AI, genetics etc) that's too dangerous for Earth.
Everybody knows getting to Mars is his life's work, but he doesn't want to just get there, he wants to colonize it. The first several thousand people will absolutely risk their lives for the sense of adventure and novelty alone (I could be amongst them), but you're not going to get masses of hundreds of thousands of people moving for this reasons, which will wear off quickly anyway.
All previous mass movements had a strong economic incentive of some kind - getting land, getting gold, trading natural resources for gold, etc. Mars isn't going to be sending raw material back like the fur traders did, and even after 2C global warming Earth will still be way more hospitable than Mars.
If Mars can become the only place where certain classes of activity are allowed, if it's the only place where you can go an experiment and test with little or no oversight, then exchange your findings for money from Earth, then I can't see a more powerful economic driver for getting masses of people there.
I expect some of the first people to arrive on Mars will be military or police or both. When was the last time you even saw scifi where the civilians got there first and had the opportunity to build something new? No, the governments on Earth will absolutely see Mars as a new colony they own, and their laws apply there. It takes too much infrastructure to get there to figure they'll let people do it and face the almost inevitable threat of them declaring 'independence' once they're out of the range of ballistic weaponry. So send the weaponry with them, along with authority-projectors. The idea of a fresh start to human society, of being able to jettison some of the things that only stick around because everyone has forgotten why they were started in the first place, is one of those things that's just too good to be true. Possible, certainly, but too potentially dangerous for those trying to hold onto the status quo back on the homeworld.
The great risk of AI to survival is in military applications, I believe. Not because of super-intelligent Skynets - in fact their lack of general intelligence may make them more dangerous - but because of superhuman response times. While banning military AI is as useful as banning nuclear weapons (impossible on a practical level), I feel that Google and Facebook and others should be leaders and address these issues, rather than conveniently bury their heads because the problem, though unavoidable, is still on the horizon, difficult, and disturbing and inconvenient to their happy visions.
Imagine a human stock trader trying to compete with a 'flash' trading program, which operates on the scale of (micro?)-seconds. Now imagine they have guns and are trying to kill each other. Imagine that the survival of millions, of your nation, of democracy and liberty depend on winning. The military has no choice but to build robots that autonomously decide and act to kill people and to destroy the things that people build; it's that or surrender and learn to speak Russian (or Chinese or the language of whoever does build those robots).
The U.S. military already has systems that do this, such as the guns that defend ships from incoming missiles - no human could identify, aim at and shoot down a supersonic missile quickly enough. They plan to implement it in other situations such as, interestingly for HN readers, defending and attacking IT systems: Imagine an AI that can launch and adjust attacks in micro- or even just milli-seconds, while the human sysadmin's brain is just having the first inklings of thought about the response procedure to the first attack. Again, either surrender or use autonomous AIs yourself.
AIs and associated military robots raise other serious issues: Throughout human history military power was tied closely to the magnitude of wealth and population; this restricted the major threats to just a few great powers. But perhaps you will only need to build a robot army, not a human one; one good AI development team and an underground manufacturing facility (anywhere in the world) might be sufficient; your AI doesn't have to be precise or make good decisions, it only needs to kill so rabidly that others surrender. Nuclear weapons at least require very specialized materials, limiting potential manufacturers (though even NK can do it now). Could Singapore do it and become the Rome of the future? Saudi Arabia? Japan? Finland? In fact, who needs a nation - wealthy individuals, groups, or corporations might pull it off. What about GE, Boeing, Google - and those are just a few American companies.
Putin recently said that "the one who becomes the leader in this sphere will be the ruler of the world."
Well first off any time someone wants to try making their technology the centerpiece of their war effort, you just ignore the technology. You attack the people controlling it, the people building it, the infrastructure, etc. And yeah, all the 'civilized' countries have invested very heavily into ballistic weaponry and agreed that's how they will fight their wars... but when it comes to survival, to wondering whether what is left after the war will even be worth preserving, those biological and chemical weapons aren't going to be left on the table.
The US' view of drone technology really disturbs me. They see drone strikes as 'not warfare' and have permitted civilian agencies with no military involvement to conduct them (the CIA). I'm not sure if they realize that a drone could fly quite easily in the air above Manhattan or not. Most likely they will simply do the slimey thing of waiting until that happens and THEN deciding it's an act of war when somebody else does it. If they can even attribute it. I'd expect such uses of drones against a capable adversary to transform the control rooms in Kansas into the primary target. You can shuffle war around all you want, but no adversary you are trying to kill is going to just sit back and let themselves be picked off while you and yours stay safe at home, expecting the robots to take the fall. It's war. They're going to show up wherever your blood-filled bodies are and they're going to destroy them. If it's truly overwhelming, they might just decide to salt the earth and leave you nothing to rule, destroying whatever it was you were killing them for just to spite you. The only winning move is not to play.
But aside from that, the trading example you mentioned is a good one. So if you build a system that is splendidly fast at reacting... you won't have gotten much of anywhere. The arbitrage markets, where that kind of thing is valuable, are already owned by people rich enough to make sure you never get a piece. In the typical market, those systems have a weakness. They can't account for their own impact upon the market. This is a fundamental problem, and not one which can be solved with more computing power or even better algorithms. In order to accurately predict a system that is impacted by the actions you take in it you need a system bigger than what you're simulating. So until the whole of the global human economy can be simulated at a speed greater than realtime, it probably won't be a problem there. The systems will also face problems where they encounter unexpected feedback loops when interacting in the same market as other automated systems.
There are dangers, though. My primary worry is turning machine learning systems loose on fundamentally unpredictable data, and then reading the systems outputs like the meaningless tea-leaves they are but giving them total trust. We know you can't predict 'black swan' events, but that doesn't stop people from claiming they can. And nothing about machine learning will change the fundamentals like 'you can't predict a chaotic systems whose initial condition you can't even know.' A high school in Dubai bought a 'crime prediction' ML system and presumably is using it to target students before they do anything wrong. I'll be surprised if that's not debuted in US schools within 5 years. And probably on the street in the UK on the same timeframe. The sad thing is, if the systems actually work (in the sense of actually calling out bad things before they're done), no one will be satisfied. I graduated high school right before Columbine happened. I watched as schools adopted these lists of "warning signs" and used them to target students and persecute them under the guise of 'preventing school shootings'. There was one problem with those lists that nobody mentioned. A list of warning signs that was accurate would identify 1 or 2 kids...
In Musk's defense, he was responding to questions about "general purpose" AI ( aka "real" artificial intelligence ) and how if that is achieved then we'd reach a moment ( singularity ) which will fundamentally change our place in the world. That's what he called "summoning the demon" and that's a legitimate point.
However, the odds of achieving true AI anytime soon is remote at best. So this guy complaint about musk has some validity too.
I used to think AGI was a long-way off myself. Until I saw this video: https://youtu.be/Aj-zNjff7wY?t=1621 He's not just simulating a neural network, which is useless if you want to create an actual mind. He's involving the body. He's actually got the recipe for the approach to AGI that has, in my opinion, by far the best chance of succeeding.
I have only an amateur-interest level knowledge of neuroscience research, but even I know that the body isn't just contained in your skull. Its neurons spread throughout your body. And dualism is flatly wrong. There is no body/mind separation. They are the same thing. Changes to the body are reflected by changes in the mind. People who experience total facial paralysis, for one random example, lose the ability to express emotion. And then lose the ability to feel it. And then lose the ability to even remember the subjective sensation of feeling it. It changes their person on the most fundamental level. How are you going to get that from a neural net with zero inputs based upon the biofeedback mechanisms of our meat-based body? You won't. If you get anything conscious at all, it would end up being profoundly different from a human mind. Give it a body, however, and enable the biofeedback even virtually... then you're talking.
Napkin sketch prediction that before we get to actual AI that will benefit us / enhance free will and lives / not just enslave us for their own / rich peoples / military benefit, we will run into some fairly shitty versions.
25 comments
[ 3.8 ms ] story [ 35.1 ms ] thread> A few years ago, Giannandrea compared artificial intelligence to a 4-year-old child. Today, he revised his statement and said that it’s even worse than that. “They’re not nearly as general purpose as a 4-year-old child,” he said.
(quoted from https://techcrunch.com/2017/09/19/googles-ai-chief-thinks-re..., which also has video of the entire thing and is IMHO a better resource than the thin article above)
It sort of distracts people from actually asking real questions like how to use AI / ML responsibly because the former doesn't require much to speculate about.
So what they have resorted to is reporting on these "philosophical" topics, because all you need for that is a fucking opinion, right. It's a great Faustian bargain because you then get all those companies and people, who similarly have no clue but are fishing for PR, to pile on.
See "should the autonomous car hit the pedestrian or save its passengers" or "this artist drew a lane marker around his beater car".
And it is not about destroying things, it is about destroying trust, society and government. Right now it's about online ratings, ads, sports articles and similar things, but I'm sure there are people working on targets that are more suitable as weapons.
What time scale is this guy assessing risk on? Does he have or expect grandchildren? Does he mind if the rest of us do?
I'd appreciate it if industry people like this could behave a little more... responsibly.
Everybody knows getting to Mars is his life's work, but he doesn't want to just get there, he wants to colonize it. The first several thousand people will absolutely risk their lives for the sense of adventure and novelty alone (I could be amongst them), but you're not going to get masses of hundreds of thousands of people moving for this reasons, which will wear off quickly anyway.
All previous mass movements had a strong economic incentive of some kind - getting land, getting gold, trading natural resources for gold, etc. Mars isn't going to be sending raw material back like the fur traders did, and even after 2C global warming Earth will still be way more hospitable than Mars.
If Mars can become the only place where certain classes of activity are allowed, if it's the only place where you can go an experiment and test with little or no oversight, then exchange your findings for money from Earth, then I can't see a more powerful economic driver for getting masses of people there.
Imagine a human stock trader trying to compete with a 'flash' trading program, which operates on the scale of (micro?)-seconds. Now imagine they have guns and are trying to kill each other. Imagine that the survival of millions, of your nation, of democracy and liberty depend on winning. The military has no choice but to build robots that autonomously decide and act to kill people and to destroy the things that people build; it's that or surrender and learn to speak Russian (or Chinese or the language of whoever does build those robots).
The U.S. military already has systems that do this, such as the guns that defend ships from incoming missiles - no human could identify, aim at and shoot down a supersonic missile quickly enough. They plan to implement it in other situations such as, interestingly for HN readers, defending and attacking IT systems: Imagine an AI that can launch and adjust attacks in micro- or even just milli-seconds, while the human sysadmin's brain is just having the first inklings of thought about the response procedure to the first attack. Again, either surrender or use autonomous AIs yourself.
AIs and associated military robots raise other serious issues: Throughout human history military power was tied closely to the magnitude of wealth and population; this restricted the major threats to just a few great powers. But perhaps you will only need to build a robot army, not a human one; one good AI development team and an underground manufacturing facility (anywhere in the world) might be sufficient; your AI doesn't have to be precise or make good decisions, it only needs to kill so rabidly that others surrender. Nuclear weapons at least require very specialized materials, limiting potential manufacturers (though even NK can do it now). Could Singapore do it and become the Rome of the future? Saudi Arabia? Japan? Finland? In fact, who needs a nation - wealthy individuals, groups, or corporations might pull it off. What about GE, Boeing, Google - and those are just a few American companies.
Putin recently said that "the one who becomes the leader in this sphere will be the ruler of the world."
The US' view of drone technology really disturbs me. They see drone strikes as 'not warfare' and have permitted civilian agencies with no military involvement to conduct them (the CIA). I'm not sure if they realize that a drone could fly quite easily in the air above Manhattan or not. Most likely they will simply do the slimey thing of waiting until that happens and THEN deciding it's an act of war when somebody else does it. If they can even attribute it. I'd expect such uses of drones against a capable adversary to transform the control rooms in Kansas into the primary target. You can shuffle war around all you want, but no adversary you are trying to kill is going to just sit back and let themselves be picked off while you and yours stay safe at home, expecting the robots to take the fall. It's war. They're going to show up wherever your blood-filled bodies are and they're going to destroy them. If it's truly overwhelming, they might just decide to salt the earth and leave you nothing to rule, destroying whatever it was you were killing them for just to spite you. The only winning move is not to play.
But aside from that, the trading example you mentioned is a good one. So if you build a system that is splendidly fast at reacting... you won't have gotten much of anywhere. The arbitrage markets, where that kind of thing is valuable, are already owned by people rich enough to make sure you never get a piece. In the typical market, those systems have a weakness. They can't account for their own impact upon the market. This is a fundamental problem, and not one which can be solved with more computing power or even better algorithms. In order to accurately predict a system that is impacted by the actions you take in it you need a system bigger than what you're simulating. So until the whole of the global human economy can be simulated at a speed greater than realtime, it probably won't be a problem there. The systems will also face problems where they encounter unexpected feedback loops when interacting in the same market as other automated systems.
There are dangers, though. My primary worry is turning machine learning systems loose on fundamentally unpredictable data, and then reading the systems outputs like the meaningless tea-leaves they are but giving them total trust. We know you can't predict 'black swan' events, but that doesn't stop people from claiming they can. And nothing about machine learning will change the fundamentals like 'you can't predict a chaotic systems whose initial condition you can't even know.' A high school in Dubai bought a 'crime prediction' ML system and presumably is using it to target students before they do anything wrong. I'll be surprised if that's not debuted in US schools within 5 years. And probably on the street in the UK on the same timeframe. The sad thing is, if the systems actually work (in the sense of actually calling out bad things before they're done), no one will be satisfied. I graduated high school right before Columbine happened. I watched as schools adopted these lists of "warning signs" and used them to target students and persecute them under the guise of 'preventing school shootings'. There was one problem with those lists that nobody mentioned. A list of warning signs that was accurate would identify 1 or 2 kids...
However, the odds of achieving true AI anytime soon is remote at best. So this guy complaint about musk has some validity too.
Musk knows nothing substantial about AI except what he's read in SF. Not a great resume for a Cassandra.
I have only an amateur-interest level knowledge of neuroscience research, but even I know that the body isn't just contained in your skull. Its neurons spread throughout your body. And dualism is flatly wrong. There is no body/mind separation. They are the same thing. Changes to the body are reflected by changes in the mind. People who experience total facial paralysis, for one random example, lose the ability to express emotion. And then lose the ability to feel it. And then lose the ability to even remember the subjective sensation of feeling it. It changes their person on the most fundamental level. How are you going to get that from a neural net with zero inputs based upon the biofeedback mechanisms of our meat-based body? You won't. If you get anything conscious at all, it would end up being profoundly different from a human mind. Give it a body, however, and enable the biofeedback even virtually... then you're talking.