Transformacja cyfrowa

#TheFutureIsYours Kształtowanie cyfrowej przyszłości Europy

Uwaga: tłumaczenie automatyczne może nie być w 100% dokładne. Pokaż tłumaczenie automatyczne

A global cooperation on AI

19/04/2021 19:54 30 komentarze

Artificial intelligence (AI) seems to be considered by many countries only as a strategic economic industry, where the goal is to develop AI faster than the others.
But some people* in the field are worried that AIs could, in the next decades, become more intelligent than humans, and, if not carefully designed, could become an existential risk; that is, a threat to humanity in general.

The problem is : we don't know how to "carefully design" an AI that is more intelligent than us, nor when (or even if) we will know it, nor what amount of work will be necessary.

For this reason, it is important for major powers (including the EU) to meet and cooperate extensively to make sure that, if a smarter-than-human AI appears, it shall be for the good of humanity.
Additionally, the EU should support work on safe AI, that is currently almost only made in the US.

Humanity has shown, through climate change, its ability to create threats to itself; but unsafe AIs could be way more difficult (or literally impossible) to stop. Let's make sure we solve the problem before it is too late.

*If interested, see, for example :
– For a short outline :
Stuart Armstrong, Smarter than us (free on the internet)
– For more details :
Stuart Russell, Human Compatible: Artificial Intelligence and the Problem of Control.
Nick Bostrom, Superintelligence.

Powiązane wydarzenia

Il 16 Dicembre dalle ore 16:00 alle ore 18:00 per tutte le classi quinte via meet si terrà la prima conferenza legata alla consultazione europea “Quale futuro per il progetto europeo?”. Questa prima conferenza sarà tenuta dal dott. Massimo Contri (dirigente di azienda nel settore ITC) e avrà come tema la Transizione digitale (Intelligenza artificiale, big-data, internet: concentrazione tra poche aziende e conseguenze per la concorrenza e diritti dei consumatori. Organizzazione del lavoro e nuovi sistemi produttivi: polarizzazione del lavoro tra professioni di alto livello e lavori precari. Nuove professionalità: impatto sulla formazione ed opportunità per la crescita occupazionale)

Zatwierdzony przez

i jeszcze 23 osoby / osób (rozwiń) (ukryj)

i jeszcze 24 osoby / osób (rozwiń) (ukryj)

26 Wyraź poparcie

Nr ref.

Numer wersji 1 (z 1) zobacz inne wersje

Sprawdź odcisk palca

Link udostępniania

Liczba komentarzy: 29

Sortuj według:

Konwersacje

Jan-Marten Spit

20/04/2021 11:23

Hi,

The human brain is the most complex structure in the know universe. It comprises about 86 billion neurons.

The largest artificial neural networks comprise about 20 million simplifications of neurons. We are nowhere near having AI at 'human levels'. We are at the level of a simplified frog. Still a capable animal, but we don't fear we will be taken over by frogs. Or elephants.

I can understand that some are in awe about what contemporary neural networks can achieve. But neural network that can recognize a magpie in a picture cannot open the door. It does not even know what a door is. in fact, it does not even know what a magpie is. In fact, the neural network does not 'think' at all. it is a mathematical approximation method. A very capable one, but its ability should not be overrated.

Even if we had a quantum computer that could approach the complexity of the human brain, it will still be confined to the existence of computer software. Which relies on, for one, a power supply.

Marie Saulnier

20/04/2021 18:41

Hi :) !

Indeed, as you say, AI is very far from having a (general) human level; but it is very difficult to know how technology will look like in, say, 50 years (think of AI in 1971). Due to the risks, it is natural not to wait until the last minute to make sure it is safe.

Many AI researchers actually think it is quite probable for human-level AI to appear in the next 50 years (though they have wildly different opinions and have been mistaken in the past, in both ways; it just shows how unpredictable AI progress is); see, for example : https://arxiv.org/abs/1705.08807v1

(By the way, something that cannot think can still be an existential risk; atomic bombs don't think, but, once launched, kill; climate change doesn't think, but, once started, kills. That said, I'm not afraid of today's AIs.)

1500 characters are too short to justify everything, and that's why I added references at the end of my idea. If you don't have time, please consider at least reading https://bit.ly/3egCtXU

Jan-Marten Spit

22/04/2021 09:31

First paragraph, i agree.

Second paragraph, human level AI. In specific domains or equal to the human brain? We already have human level AI in playing Go, that is not future. AlphaGo cannot cook dinner.

having good code and good hardware is but a fraction to the deployment of useful AI. Most of what goes in it, is human intellect, even though it is called AI. It lives in a habitat of good training data, proper trainign parameters, hardware, operating systems, electricity, including remote power stations and power switches. The relation between AI and humans is that of Almighty Creator and its unwitting subject. AI will always be below humans in the chain of command, as we can survive in environments where AI (and its mechanized subjects) perish.

I'd say a more realistic and pressing danger is that AI - which is indeed very potent technology - will enhance the power of some humans over others - as is already the case.

I enjoyed reading that last link though, good summary.

Marie Saulnier

22/04/2021 12:52

I speak about artificial general intelligence (AGI); in the survey I linked to, it is defined thus : "High-level machine intelligence” (HLMI) is achieved when unaided machines can accomplish every task better and more cheaply than human workers."

You describe the present state of AI, , but it doesn't need to work the same way in the future, and there are incentives to make possible for AIs to adapt to their environment or to change their own parameters to achieve their goal. AIs only do what their source code says they must do; their source code doesn't tell them they must remain below humans in the chain of command. If their goal is to maximize paperclips, they'll try to make it impossible for humans to stop them, and, then, maximize paperclips (see Bostrom).

"we can survive in environments where AI (and its mechanized subjects) perish"
Like ? I'm pretty sure you'd die on Mars. Anyway, that's irrelevant (not enough space to develop, tell me if you want, I'll add a comment).

Jan-Marten Spit

27/04/2021 16:13

"we can survive in environments where AI (and its mechanized subjects) perish"

is what i said. Do i have to explain humans can do without, say, electricity?

The habitats in which humans can survive are much broader.

"AIs only do what their source code says they must do; their source code doesn't tell them"

This is bad logic. You propose that AI can surpass humans "because" we did not tell it it shouldn't.

"If their goal is to maximize paperclips, they'll try to make it impossible for humans to stop them"

If i would ask you to defend that statement you would require dozens of pretty unlikely assumptions.

You cannot defend unverifiable statements by Bolstrom to defend the same statements. Such practice is called religion. Obviously, you can believe what you want, but what Bolstrom fantasized is not science. Or even coherent.

Mario Cannistrà

21/04/2021 11:50

This is a very flawed argument. Saying there is nothing to worry about because "we're not there yet", or that "you can just pull the plug" are very short-sighted, and dangerous ways to think about this.

Please do more research on the Alignment problem, the commonly (and debunked) proposed "solutions" to it, and the challenges it presents. You will find that any solution that you might propose is very superficial, has been debunked and deemed completely ineffective or insufficient. So in short, so far the alignment problem is unsolved.

Regarding the "we're not there yet". Yes, and be glad that we aren't, because if we were close to AGI without having solved the alignment problem, it would already be too late.

Progress in AI and AGI research is accelerating at unprecedented levels, and with companies like DeepMind and OpenAI fully dedicated to making AGI, it might not be long until we achieve it. Just look at the latest research done.

Jan-Marten Spit

21/04/2021 19:21

I do not present an argument, so it cannot be flawed either. I present reality.

I write neural networks. I understand the mathematics behind it. It therefore also understand the capabilities and the limitations. And therefore to what extent the technology is hyped. The research, i already did.

It is a flawed argument to use the fact that the threat is, at least currently or in the foreseeable future, not real, as evidence we are just lucky. That is circular reasoning.

You say we have a problem that we fundamentally cannot solve and it is upon us pretty quickly, you don't like me pointing out reality by labeling it a flawed argument, and you appeal to authority (a propaganda technique) by claiming i need to do more research for having to audacity to disagree with you.

Now what does that sound like. That sounds like a cult.

AI is potent technology, but it is not magic.

Mario Cannistrà

21/04/2021 21:48

I don't think I'll change your mind, but for anyone else reading this:

Even a monkey can write neural networks, I'm just a web developer and I have written some, knowing how to implement a NN (which is narrow AI/ANI) doesn't mean you know about AGI, and vice-versa.

And yes, neural networks are limited, that's not the point. I'm not saying that AGI will emerge from
ANIs just by throwing computing power at them, you might need different algorithms. As a counter-example OpenAI's GPT3 has demonstrated linear growth (it did not plateau), and "intelligence" by adding more "parameters" which takes more computing power. So adding more parameters will likely continue this trend, we don't know if indefinitely or not, or if it will lead to AGI, but that's not impossible.

And I never said we can't solve this problem, but it's hard, and we better start soon.

I would continue to educate you, but characters here are insufficient, so you must do it yourself, that's not "appeal to authority".

Jan-Marten Spit

22/04/2021 10:31

"Even a monkey can write neural networks"

nonsense.

"knowing how to implement a NN (which is narrow AI/ANI) doesn't mean you know about AGI,"

Yes, it does, as an ANN is actual technology and AGI an acronym / language element to communicate an idea currently without implementation. GPT3 is a NN. And it is not going to "become AGI by adding more parameters" because it is not 'general' but very domain.

"And I never said we can't solve this problem"

Yes, you did: "You will find that any solution that you might propose is very superficial, has been debunked and deemed completely ineffective or insufficient."

A statement that also reveals your zealot approach to people who have the audacity to question your proclamations.

"I would continue to educate you"

As you know you are bluffing and you know i know you are, this statement is not directed at me.

Mario Cannistrà

22/04/2021 13:54

because it is not 'general' but very domain.

How do you know a language model couldn't be general? Do you have any evidence? If it was able to reason about anything, you wouldn't call that an AGI?

Yes, you did

No, that doesn't mean what you understood it to mean. I said you might propose. Meaning that you are proposing things that have already been proposed, and are not aware of them, so you will continue doing that, since you refuse to educate yourself.

audacity to question your proclamations

Feel free to question everything, but at least provide evidence to back up your claims, and don't spread misinformation if you clearly don't know anything about the subject

you are bluffing

Feel free to ask me anything, as ignorant as your questions might be, and as limited as these comments are... I will point you to sources to read, or watch, but I know you won't use them, so...

Jan-Marten Spit

23/04/2021 08:07

@Mario

"How do you know a language model couldn't be general?"

A? You mean 'this'. I was responding to something particular, that YOU brought up. AIG systems do not exist at this time.

" I said you might propose. Meaning that you are proposing things that have already been proposed"

Might is future tense, proposing is present tense. Word weaseling instead of admitting.

"but at least provide evidence to back up your claims"

You are claiming very extraordinary things, and you have NO proof any of it will ever come to pass. You speak about AIG as if it exists. You need to supply backing proof, and you have none.

You failed to admit to mistakes or apologize for your appalling behavior. You refused rational debate from the first reaction to me. You deny writing what is visible as evidence centimetres above it, and you engage in ever more hysterical claims the more you get cornered in your lack of in-depth knowledge on the matter.

Yes, i am done with you.

Hans Sarpei

21/04/2021 15:51

not sure whether i'd call the human brain is the most complex structure in the universe and pretty sure it doesn't really matter, but other than that, i have to agree.
The amount of pseudo-scientific "AI research" academic chairs with little relation to engineering created within the last few years is astonishing. If someone wants money for research today, he's got to add anything AI in the project title.
"Artificial Intelligence" today is almost 100% basic machine learning methods applied to a good model. It's awesome what can be done with it, and, like any software, if used in critical infrastructure, you have to make extra-sure that there are no bugs. But the notion that any of the methods we use today might be the base for some "I, Robot" stuff; i'm sorry but unless someone finally describes on a technical level how exactly that would play out, i call bull.

Jan-Marten Spit

21/04/2021 19:37

The complexity of the human brain matters because it is a prerequisite for the supposed threat. At 80 billion neurons each having 10.000 connections to other neurons, the human brain is the most complex structure in the known universe. And as that complexity delivers human intelligence (which we do not fully understand), yes, it matters in a discussion about AI possibly surpassing human intelligence.

So whilst a neural network may be able to recognize a face, it is quite another matter to discover it has to shoot everybody that threatens to turn of the 3 power stations 50 miles away. For that it would have to 'understand' that electricity to the chips is existential and not a constant. I wonder how it would learn that without being shut down. And how would it learn that when it is shut down.

If given a few minutes of critical thought, the problem is quickly discovered to be good stuff for whiskey and campfires. And bestsellers.

Mario Cannistrà

21/04/2021 21:58

You clearly have not read any of the related literature, such as Nick Bostrom, or even thought about it critically, like you encourage others to do.
If you did, you would know that all your "counter arguments" have been addressed several times in this field, and do not hold up.

Again, if you want to have a serious discussion, research the stuff, it's pointless to repeat it here in these short comments. If you don't want to educate yourself, at least stop spreading misinformation, pretending you know stuff, I do not appreciate such behavior.

Hans Sarpei

21/04/2021 22:50

gotta disagree, in my view the issue is not the computational complexity of the brain that can't be reached (aren't we there already?), but rather that it is too hard (and pretty much a waste of effort and power) to formulate a vast and unspecific problem like "be like a human" and to create a model and feedback loop that show senseful results.

not sure why you are so fixated on that "the human brain is the most complex structure in the known universe" thing, it's obviously wrong. for one, everything that encompasses the human brain is more complex, no?

Jan-Marten Spit

22/04/2021 11:03

@Mario

"You clearly have not read any of the related literature, such as Nick Bostrom, or even thought about it critically, like you encourage others to do."

Nick Bolstrom is one philosopher that wrote a popular book - doomsday stories sell. Nick Bolstrom is NOT an AI specialist. You also fail to mention that Bolstrom does NOT represent the field of AI research. Quite a few AI researches estimate the probability of super intelligence beating humans as 'never', whilst most take the safer bet of saying 'not in the foreseeable future'.

"at least stop spreading misinformation, pretending you know stuff, I do not appreciate such behavior."

Yes, you are spreading misinformation, and yes, you are pretending to know stuff, and as you get ever more cornered in your own quackery, you start to get fairly unpleasant.

Stop spreading misinformation.

Jan-Marten Spit

22/04/2021 11:23

@hans_sarpei

"in my view the issue is not the computational complexity of the brain that can't be reached (aren't we there already?)"

no, we are not. look at the numbers again. 80 billion neurons with 7.000 - 10.000 connections each.

moreover, the artificial neurons as they are modeled by the mathematics of AI are simplifications. Human neurons are suspected to be much more complex, so even if we would be able to match the number of neurons and synapses of the human brain, we probably will not have matched the complexity of the human brain.

"the human brain is the most complex structure in the known universe" thing, it's obviously wrong. for one, everything that encompasses the human brain is more complex, no?

A human in a box is slightly more complex than a human, yes. But the box is not more complex than the human.

Marie Saulnier

22/04/2021 13:36

Bostrom isn't an AI researcher, but he has still answered many of your arguments; given how few characters we can use, it is impossible to answer everything here (if you prefer, try Russell instead).
As the survey I linked to shows, there's no consensus, but the median is around 50 % chance before the 2060s (and even 1 % chance of an existential risk is still huge enough to think about it).

Why should AGI have the same "complexity" as the human brain to achieve the same results ? We achieve smarter-than-elephant intelligence with 3 times less neurons, according to : https://bit.ly/3vcgwA4. Humans are not optimized for intelligence, even less for general intelligence, and AGI doesn't have to mimic human neurons. For example (I repeat I don't say NN would be sufficient, I don't know), today's articial neurons are much faster than biological ones. Human brains are simply the stupid gradient-descent-like solution of a fitness-maximizer (aka evolution); engineers can surely do better.

Hans Sarpei

22/04/2021 21:10

@Jan-Marten Spit
disregarding the discussion of whether those numbers matter, 8*10^14 seems achievable to me.

as i have said, i would put the real problem somewhere else: stuff would have to be modelled, and that's just too hard to get sufficiently precise. i guess we agree on that, but imv this is the main thing.

I don't even think it is 100% impossible in theory, but that it'd cost incredible amounts of time, effort and money. And then you'd maybe have something that might resemble human intelligence and wastes a lot of energy, but what the heck do you want with it? You don't want another human, you want to make useful stuff that takes reasonable effort to make and ideally improves things a bit. If you do want another human, that's easier to make, too.

I don't see this on its way, i agree with Jan-Marten Spit on the pain, and i think one should worry about a few of the millions of more likely and desastrous risks first.

and how the heck are we at maximum comments depth already?

Jan-Marten Spit

23/04/2021 00:14

@marie_saulnier

"Bostrom isn't an AI researcher, but he has still answered many of your arguments"

The matter before is if the extraordinary claims are properly argued. I do not have to defend a lack of belief.

"Why should AGI have the same "complexity" as the human brain to achieve the same results ? We achieve smarter-than-elephant intelligence with 3 times less neurons"

Because perceptions are simplifications of human neurons, and the 'intellect' of AI is a function of the number of synapses - there is a minimum for a NN to learn an XOR gate for example. The complexity matters.

Although an elephant has 3x the neurons, most of them are in the cerebellum, whilst the human cerebral cortex is 0.5 the weight but contains only a 1/3 of the neurons compared to humans.

"Humans are not optimized for intelligence"

I would not know what that "means". Humans created AI.

"even less for general intelligence, and AGI doesn't"

exist at this time.

Jan-Marten Spit

23/04/2021 00:36

@marie_saulnier

"today's articial neurons are much faster than biological ones."

today's artificial neurons are -simplifications-. there is lots about neurons and the human brain we do not understand. so how can you claim neurons are inferior if you do not know what to measure let alone have actual measurements?

"Human brains are simply the stupid gradient-descent-like solution of a fitness-maximizer (aka evolution); engineers can surely do better."

I am just baffled hearing an argument 'for' the potency of AI by denouncing one of the most important ideas and techniques behind machine learning as 'stupid'.

Moreover, did you not smell the paradox in claiming that 'humans neurons are inferior, engineers can do better'? What would those engineers be using when they are 'doing better'?

Jan-Marten Spit

23/04/2021 01:06

@hans_sarpei

"disregarding the discussion of whether those numbers matter, 8*10^14 seems achievable to me"

in case a synapse would require a 10 byte double to track its weight, that is 80 petabyte. And 10^14 connections to calculate through, per iteration. I would not dare say that is impossible in the future.

It seems we agree there are more pressing, and actual problems with AI.

Mario Cannistrà

21/04/2021 11:07

I'm happy to see that there is already a proposal about AGI, and AI safety specifically.

I think it's the most under-discussed and potentially most important problem that we will ever face, even above wars, climate change, poverty, etc...

It really can't be understated how important this is, and the fact that most people don't know about it makes me deeply worried for the future of humanity.

The AI alignment problem, or control problem, needs to be solved as soon as possible, because several research groups are working on AGI right now, and good progress is being made, so we don't know when AGI will emerge, and if it does before we solve the alignment problem, it will be too late.

The potential of AGI could be immense, either in a positive, or negative way, it will probably be the last invention of humanity. Solving the alignment problem will make a positive scenario more likely after AGI emerges.

This comment has limited characters, so exploring a complex topic like this is hard.

Konwersacje

Comment moderated on 03/05/2021 17:30

Mario Cannistrà

21/04/2021 22:00

You are talking about narrow AI, and not AGI. While I do think that narrow AI is a very real and worthwhile problem to talk about, I think this post is mainly focused on general AI, as the OP says "The problem is : we don't know how to "carefully design" an AI that is more intelligent than us", and quotes Bostrom's "Superintelligence" in the sources.

Marie Saulnier

21/04/2021 22:42

Indeed, this idea is mainly about AGI (I should have been more precise in, like, the title ? in retrospect, I could probably have picked a better one), though I also think that today's AIs should be discussed.
Why not post another idea about it (and post the link in comments here afterward) ? I could do it, but I don't know anything about AIs involved in nuclear "second strike", and I'm worried that, were I to write it, I would forget the obvious. So let's share what we think are the most important problems about today's AIs and …
Now, I notice how inadequate this platform is for constructing ideas together … well, I will still mention that recommendation algorithms (capable of influencing a large part of the public opinion in many ways, and not aiming to be beneficial to society) should probably be mentioned in this new idea.

Jan-Marten Spit

22/04/2021 14:25

A real problem with AI, right now, is that it enhances the power of some over many. Humanity is currently providing huge datasets to learn from to a handful of over-sized conglomerates, mostly operating in different countries so outside democratic control.

This problem already exists and will therefore get worse with advances in AI. No 'super-intelligence' required, just a bit of intelligence suffices to do evil.

So why are we focusing on a hypothetical problem not considered realistic by many in the field, whilst not addressing a problem that is actually occurring now, and considered a problem by almost everyone in the field?

Dávid Lakatos

16/05/2021 08:50

I agree that AI development should be planned out and regulated. Progress will not stop, but we can and must direct it.
For starters, don't try to make humans. Computers and machines do a lot of things a lot better than humans, but fail in other areas. Build on the strengths of everything that already is. Let computers compute, and let humans make art or whatever each individual is most talented/capable at.
If you somehow still do end up making humans, have a specified set of tests (a Turing test is just not enough anymore) and values that qualify something as human, and give the thing human rights.

Kyle Hargraves

10/10/2021 03:18

I am referring to Marie's substantive post of 19 April.

It is all very well (and laudable) for human beings to cooperate for the greater good (JS Mill) but the desire or
inclination, on the evidence for doing so, isn't compelling. The map of Europe has changed significantly during each century from the Treaty of Westphalia (and indeed prior to that treaty). Even the EU possesses internal tensions of nationalist origin and given the possibility of a very great strategic advantage (military, trade or whatever) arising from AI cooperation will be at best : slight.

Normative appeals to our better natures may possess some influence but only up to a point; ultimately nationalism, as the Seven Years War and both World Wars refer, triumphs. The same will prove true for cyber warfare and AI with the two intertwined. As it stands, about 1.8 billion people have control over circa 6 billion (7.8 - 1.8) people now and I, for one, foresee no material changes long term.

Nicolò Boggian

05/01/2022 12:02

What about a single employement contract for Digital Work in Europe?
A single employement contract for Digital Work in Europe

Konferencja w sprawie przyszłości Europy

Warunki korzystania z platformy Najczęściej zadawane pytania (FAQ) Jak mogę zorganizować wydarzenie? Karta Konferencji Oświadczenie o ochronie prywatności Plenarne grupy robocze Przydatne dokumenty Document Directory

Kontakt

Zadzwoń do nas pod numer 00 800 6 7 8 9 10 11
Skorzystaj z innych form kontaktu telefonicznego
Napisz do nas, korzystając z formularza
Spotkaj się z nami w lokalnym biurze UE

Znajdź UE w serwisach społecznościowych

UE w mediach społecznościowych

Instytucje UE

Search for EU institutions

Możesz uzyskać dostęp do platformy z konta zewnętrznego

Wybierz swój język

Języki urzędowe UE

Transformacja cyfrowa

#TheFutureIsYours Kształtowanie cyfrowej przyszłości Europy

A global cooperation on AI

Powiązane wydarzenia

Zatwierdzony przez

Odcisk palca

Udostępnij:

Link udostępniania:

Prosimy o wklejenie tego kodu na swoją stronę:

Liczba komentarzy: 29

Konwersacje

Konwersacje

Proszę się zalogować

Wybierz swój język

Języki urzędowe UE

Transformacja cyfrowa

#TheFutureIsYours Kształtowanie cyfrowej przyszłości Europy

A global cooperation on AI

Powiązane wydarzenia

La Transizione digitale

Zatwierdzony przez

Udostępnij:

Link udostępniania:

Prosimy o wklejenie tego kodu na swoją stronę:

Liczba komentarzy: 29

Konwersacje

Konwersacje

Proszę się zalogować