wait, if the initial distribution was 67/33 (and assuming that whether someone is likely to change their mind does not depend on which initial group they're in), the expected number of people changing from Yes to No is twice as high. If the initial distribution was 100/0, Yes could never win the debate.

Please post the numbers.

Isn't this not accurate as the ending poll could not be conducted without glitches

You could hear the audience moan. I saw Daniel Dennett point it out once that argument where we say right? Or Surely, I think his example was "surely" they're not arguments at all they're OBVIOUS assumptions. Everyone does this to some degree it's hard to see him doing this when it's his field of expertise, it's terrifying honestly.

Yes, after twice hearing Max say that not everybody is as nice as Yann.

hes been doing that many times

He's surprisingly daft

@@dillonfreed disagree a bit😁 he's just experiencing a Gary Marcus type "out of distribution" failure mode😁 Unable to step out of his mind. Actually, it is the fact that he's ferociously intelligent that makes this failure particularly dangerous

Great comment. Thanks!

No kidding! I couldn't believe that these people have any knowledge of AI, let alone be experts! How incredibly naive these people are! Scary af!!!

Yeah.. seems ridiculously optimistic and dismissive. I understand that it doesn't seem probable AI will pose a serious threat, but to act like it's impossible because we will always control it, or it will innately be good? That just seems foolish. I'm pretty optimistic, I do think the most likely outcome is positive, but it was hard to take these people seriously. It's like getting in a car and one passenger is saying "could you please drive safely" and these people are in there saying "why would he crash? That's just silly, if he is going off the road he can simply press the brake pedal, look, it's right there under his feet. I guess we should worry about aliens abducting us, too?"

@@trybunt you forgot their other big argument: "There are much more pressing dangers than AIpocalypse and these speculative scenarios draw attention from the true horrors we are about to visit upon huma--I mean ... everything you guys brought up is far away, let's all just go back to sleep."

@@trybuntthey have big pockets behind of it it’s not good to their business full Ai regulation

You're wrong about her, friend. She is one of the best AI experts in the world. To get out of your ignorance (and incidentally understand her point of view a little better) you only have to read her books "Artificial Intelligence" and "Complexity". She has written many more books, but those two alone have enough substance. These turn-based, moderated discussions are nice, but they're too short and can hardly have the length these topics deserve.

@@CodexPermutatio Of course I know she is actually an expert but her blinkered view on a) what constitutes an existential threat and b) how AI could facilitate this possibility, c) how she dismisses the notion entirely, and d) how she thinks even considering the idea of AI danger would detract from combating the "present threats of mis-information" - all point to an irrational personality. I pondered suggesting she has ulterior motives but stopped short at suggesting she had researched AI (dangers). Taking only point D for brevity, she sees mis-information in elections as something so dangerous that AI safety should not take up any resources whatsoever. Surely if AI of today can overthrow a governmental system, AI in 20 years or so could do something worse. And that alone could put us in a position we are unable to return from - like putting a pro-war US candidate in power bringing us a closer to a nuclear winter scenario - an existential risk. These are non-zero and also non-infinitesimal odds even with today's AI. AGI is not a per-requisite of existential risk.

​​@@74GeeThe whole time she mentions other things to talk about that are more pressing, but if she could give examples of them I would've loved that. We are facing population collapse, another major pandemic, climate change if you can give me a reason allignment research *wouldn't* help these other issues I'd be all ears. But all of these other problems are also problems of allignment and of failed incentives it just happens the incentives are human and not machine

@@jmanakajosh9354 It's clear you care about the state of the world and the direction we're heading in. AI alignment research certainly would help with addressing any problems that AI could potentially help with - The last thing we want is solutions that make matters worse. It's not like there's a shortage in resources - Microsoft’s stock hits record after executives predict $10 billion in annual A.I. revenue (Microsoft shares climbed 3.2% to a record Thursday and are now up 46% this year) ...so it's not like doubling the AI alignment research with additional hires is going significantly affect the bottom line of Microsoft, or likely anyone else in the field.

Monk debates always seem to choose someone extra annoying to fill that role.

He lives on the planet of the AGI builders. A planet, apparently, very different from the one inhabited by the AGI doomers. I, by the way, would pay more attention to builders than doomers. Being a doomer is easy (it doesn't require much, not even courage). Being a builder, on the other hand, requires really understanding the problems you want to solve (also, it implies action).

“If the sub wasn’t safe I wouldn’t try to take people to the Titanic with it!” -Stockton Rush

@@CodexPermutatio we have to pay attention to the builders because they implicate us all.

​@@CodexPermutatioYour points are valid but you underestimate the level to which human ego creates blind spots. Even very smart people develop attachments to the things they pour their energy into. This makes them biased when it comes to potential risks.

Omg this part is making me so stabby!! Would we build a bomb that just blows shit up? 🤦‍♀️ yes yes we did it ffs we do it we are still doing it. This is not a relaxing sleepy podcast at all lol

it doesnt work for machines who can think

@@PepeCoinManiaYou have no idea what you’re talking about. Define what you mean by “think,” in clear and precise terms, explain why an ability to think would ensure that its goals stay perfectly aligned with ours, and explain how you know machines will one day “think” according to your definition.

What kind of argument are you searching for? The AI Doomers have poorer arguments, but because this is an issue of unknown-unknowns, they're winning the rhetorical 'battle.' I can't guarantee you AI will kill us all, unless I could demonstrate that AI cannot physically do so(say there's some fundamental Law in the universe that prevents such a thing.) It's hard to prove this, because we still are ignorant of so many different things about AI and (quantum and traditional) computing in general.

💯

@@SetaroDeglet-Noor Yes. But GPT-4 isn't an existential threat. It is not AGI. AGI poses existential threat. That's what Bengio and Tegmark are arguing for, not that GPT-4 poses existential threat. GPT-4 poses risks, but they are not existential. I think Melanie can't think of existential threats of AI, because she is only considering current AIs, like GPT-4, so let's not do that. We need to consider future AI, AGI, which will indeed be able to do things that we cannot prevent, including things that might go against our goals, if they are misaligned, and in those cases, they could cause our extinction. I'm a bit disappointed that they didn't talk about instrumental convergence explicitly, but they just kind of mentioned it vaguely, without focusing on it much. I wish someone like Yudkowsky or Robert Miles could have been there, to provide more concrete technical examples and explanations.

@@MetsuryuVids "I'm a bit disappointed that they didn't talk about instrumental convergence explicitly." So true. As a layperson, that was probably the easiest concept for me in understanding why AI might end up getting rid of us all without actually hating or placing moral judgement on us, or any of that human stuff. But again, there was this rather prominent podcaster, who I still think is a smart guy in some ways, who just couldn't see how AI will want to "self preserve." And to your list I would definitely add Roman Yampolskiy.

Same. The “it will be like having a staff of subservient slaves that might be smarter than us.. it’s great working with people ..”people” smarter than us” phew 😬 that was a new one and not a good new one

Immediately killing 1% of humanity is not acceptable risk?

Smart response, I fully agree. It’s alarming.

I run an international ML team that implements and develops new routines. It is not accurate to say that we are careless, it's simply that we don't have the right systems or the right techniques to develop AGI. There are many more pressing issues about bias, alignment, safety, and privacy that are pushed to the wayside when we imagine the horrors of AGI. We have shown that LLMs cannot self-correct reasoning. Whatever tech becomes AGI, it's not LLMs. Secondly, we won't ever suspend AI development. There are too many national interests at stake, there will never be a pause. Period. It is the perspective of our military that our geopolitical adversaries will capitalize on the pause to try to leap frog us. So, imaging the horrors of what could be possible with AGI is the wrong thing to be focused on. AI has the potential to harm us significantly in millions of other ways before taking over society. A self-driving car, or delivery robot, is millions or billions of times more likely to accidentally harm you before a malicious AGI ever will.

​@@MM-cz8zt, Metaculus has the best forecast I have found on the topic, and currently estimates our extinction risk from AGI around 50%.

@@MrMichiel1983, I encourage everyone to find a better forecast themselves. The reason you propose is obviously stupid; you will have more success finding the truth if you stop strawmanning people you disagree with. I was looking for the most convincing forecasting methodology, and for one thing they at least publish their own track record, which few can claim. For another, they crowd-source the forecasts and weight them by the track record of the individual forecasters. Also, their forecast of the arrival date of AGI (~50% by 2032) aligns closely with most other serious estimates I have found (2040/2041).

Meta is the worst because it is led by Yann LeCun, a literal retard who thinks safety is a joke. Google is much better, and Microsoft almost as bad. - Meta obviously leaked Llama on purpose - Google was not racing GPT-equivalent products until Microsoft started - Microsoft didn't even do proper RLHF for Bing Chat

nail on the head. it's about to get gnarly. and then it will keep getting worse until we die

They didn't even discuss fully autonomous military drones, and how these would change war and the global balance of power

Mo Gawdat said he doesn't know why people aren't marching in the streets in protest

Couldn't have said it better myself

He also doesn't mention "who" it's good for. Historically, many civilizations have been wiped out by more advanced and intelligent civilizations. And surely, competing nations, militaries, corporations, and possibly socioeconomic classes will have competing AIs that are also not aligned with the greater good of ALL humans.

​@@wonmoreminute I'd be happy if even an evil dictatorship manages to actually align an AGI to some semblance of human values. Not ideal, but at least probably not the worst case scenario. The thing is that we currently don't even know how to do that, so we'll probably go extinct, hence the existential threat.

Very well put. Thanks for the read, I full agree.

I assumed LeCunn meant to imply broad progress of positive actions overwhelming negative ones rather than necessarily just the specific case over how controlled and purpose driven an AI would be

The problem is the orthogonality thesis is dumb. It requires a definition of intelligence that is so narrow that it doesn't correspond to anything we humans would understand as intelligent. If that's all intelligence is (the ability to plan and reason) why would we be scared of it anyway? There is a sleight of hand going on here. We are invited to imagine a super-smart being that would have intellectual powers beyond our imagining, would be to us as we are to an insect. But when this proposed "superintelligence" is unveiled it's just a very powerful but completely dumb optimiser.

💯%

Maybe listen to Gary Booch or Robin Hansen they have much more convincing arguments (sarcasm)

Truly. Embarrassingly pathetic arguments. We are so fucked. I'd highly recommend Yuval Noah Harari's recent speech at Frontiers Forum, which is available on youtube.

The pro-AI acceleration crew has no case. I've read all of the prominent ones. The important point is that power does not need to persuade. Power does what it wishes, and, since it's far from omniscient, power often self-destructs. Think of all the wars lost by the powers that started the wars. Often the case for war was terrible, but the powers did it anyway and paid the price for defeat. The problem now is that a hard fail on AI means we all go down to the worst sort of defeat, the genocidal sort, such as Athens famously inflicted upon Melos.

@@kreek22 Undoubtedly. And it felt like LeCun had that thought in the back of his mind during this whole debate. His efforts were merely superficial. And I've seen Altman give an even worse performance, even though he pretends to be in favour of regulation...he is talking through his teeth. Mo Gawdat has out right stated that he believes it will first create a dystopia, but will ultimately result in a utopia, if we can ride it out. I think the billionaire IT class have it in their heads that they will have the means to survive this, even if nobody else does. It's very bizarre.

I hope dearly it is not in the culture of Facebook to have no worry about this.

Haven’t watched yet, but Looking at the credentials… I believe you already

Chief AI scientist at Meta seems to have a bias … yeah

@@jackielikesgme9228 I watched the Zuck's interview with Lex Friedman and it didn't seem like total abandon of AI safety was a thing, but this concerns me esp. since FB models are open source

@@jmanakajosh9354 how was that interview? It’s one of a handful of Lex podcasts I haven’t been able to watch. He’s much better at listening and staying calm for hours than I am lol.

100%. They completely ignored all of Max's most based points, because they don't have real answers. Your point about the symbolism of that tech failure hit me pretty hard which is why I decided to reply to your comment. I hear a lot of humility on one side of this debate, and a lot of arrogance and name calling on the other. They do not seriously want to engage with the actual magnitude of the risk, but rather push it under the rug. As Yann said with such confidence and bravado, "the AI is under control and we will keep it under control," and "it will remain subservient to us." That quote literally sounds like the opening scene of a bad terminator knockoff. If the AI does wake up I don't think it's going to appreciate being talked about like that.

The voting system isn’t an AI. It’s a dead simple website.

That's what I want

I heard somewhere it was shown that the KZfaq algorithm seems to train people into liking predictable things so it can better predict us. But what Michelle misses is this thing is like the weather in Chicago, it's in constant flux if we say "oh it can't do this" all we're saying is "wait for it to do this" and man, the way the Facebook engineer just pretends like everyone is good and he isn't working for a massive surveillance company is shocking

A parent conceived and purposed by sociopathic humans.

​​@@Imcomprehensibleseah, I wqnt it too. The chance to wear nappies again and face ridicule is gonna be awesome.

rightfully so. There are studies on IQ developments world wide which showed that we are on a downward trend and the two major reasons for that which were named are environmental pollution and the use of smartphones. There are now people who don't know anymore how to use a map ffs, not even getting into the psycological missuse to trigger peoples endorphin system to get them on the hook for candy crush. When we come to a state where we have our personal AGI agents, that we can talk to and give tasks to solve for us, we therefor loose capabilities ourselves. Should we give government functions and control over companies, medical procedures and what not to AGI to a level we do not teach anymore doctors and have no politicians that are in the decicion process on how we are governed and then at some point even if initally it was not an rogue AGI would become one later one, then we are truelly fucked just by not being able to do these things anymore. Imagine we do not have books then anymore, but only digital data access to regain these capabilities but we would be blocked from using them and so on. So yes you are spot on with that concern.

They are 75 to 80% Leftists in godless leftist cities teeming with reprobates, and none of them revere The Lord.

I could only see a né side of professionals the others are scumbags

SSSSSOOOOOOOO, what is it exactly you want to say ?

@@keepcreationprocess I'll have a hit of that, please.

At one point, I believe it was Mitchell but might have been LeCun, said that corporations do not pose an existential threat. I thought that was a pretty absurd statement given we are currently facing multiple existential threats due to corporations, and more than one of these existential threats was brought up during this very debate. It's also worth noting that corporations are limited to the vicinity of human intelligence due to the fact that they're composed of agents that are no smarter than humans. They are smarter than any one human, but their intelligence is still bounded by human intelligence. AI lacks this limitation, and its performance is scaling very quickly these days. Since 2012 the performance of the state of the art AI systems has increased by more than 10x every single year, and there is no sign of that slowing down any time soon.

not smarter at all, richer maybe, intellect has little to do with it

@@genegray9895 "Since 2012 the performance of the state of the art AI systems has increased by more than 10x every single year" Is there a source for this claim? My understanding is that increases come from three directions: hardware, algorithms, money. Graphics cards have managed ~40%/year, algorithms ~40%/year. Every year more money is invested in building bigger systems, but I don't think it's 5x more per year.

@@kreek22 OpenAI's website has a research section, and one of the articles is titled AI and compute. KZfaq automatically removes comments containing links, even when links are spelled out

@@genegray9895 Thanks. That study tracks 2012-18 developments, early years of the current paradigm. Also, they're calculating compute, not performance in the sense of superior qualitative output (though the two tend to correlate closely). They were right to calculate the compute per model. The main cause of the huge gains is the hugely increased parallelism.

Only the dumb people are positive about AI

No.

As long as AI remains safely ensconced in Toolworld, In all Fer it.

Same here, I think that they did better in the debate and were more reasonable, although my position is on the optimistic side.

it seems like the other guys basically admitted there's an existential risk but don't want to call it that

Slavery. Add in the superior intelligence part and now it is called hubris.

Emotions that will make them desire to serve us, so everything good.

@@Nico-di3qo thats just slavery with extra steps

I disagree with LeCun, in the fact that he thinks the alignment problem is an easy fix, and that we don't need to worry and "we'll just figure it out", or that people with "good AI will fight the people with bad AIs", and many, many of all of his other takes. I think most of his takes are terrible. But, I do think this one is correct. In a way. No, it's not "slavery*". The "emotions" part is kind of dumb, and it's a buzzword, I will ignore it in this context. Making it "subservient" is essentially the same thing as saying making it aligned to our goals, even if it's a weird way to say it. Most AI safety researchers would say aligned. Not sure why he chose "subservient". So in summary, the idea of making it aligned is great, that's what we want, and what we should aim for, any other outcome will probably end badly. The problem is: we don't know how to do it. That's what's wrong with Yann's take, he seems to think that we'll do it easily. Also, he seems to think that the AI won't want to "dominate" us, because it's not a social animal like us. He keeps using these weird terms, maybe he's into BDSM? Anyway, that's another profound mistake on his part, as even the moderator mentions. It's not that the AI will "want" to dominate us, or kill us, or whatever. One of the many problems of alignment is the pursuit of instrumental goals, or sub-goals, that any sufficiently intelligent agent would pursue in order to achieve any (terminal) goal that it wants to achieve. Such goals include self-preservation, power-seeking, and self improvement. If an agent is powerful enough, and misaligned (not "subservient") to us, these are obviously dangerous, and existentially so. *It's not slavery because slavery implies forcing an agent to do something against their will. That is a terrible idea, especially when talking about a superintelligent agent. Alignment means making it so the agent actually wants what we want (is aligned with our goals), and does what's best for us. In simple words, it's making it so the AI is friendly to us. We won't "force" it to do anything (not that we'd be able to, either way), it will do everything by its own will (if we succeed). Saying it's "subservient" or "submissive" is just weird phrasing, but yes, it would be correct.

​@@MetsuryuVids I think it's shocking that he thinks it possible to model human emotions in a machine at all (I'd love to learn more about that it gives genuine hope that we can solve this) but then falls on his face....and so does Melanie when they say it NEEDS human-like intelligence, that's the equivalent of saying planes need feathers to fly. It's a total misunderstanding of information theory and it's ostrich like bc GPT-4 is both intelligent and has goals. It's like they're not paying attention.

I totally agree w/ you. Many people like to play the oracle, but the future is intrinsically unknown.

@@rafaelsouza4575 Exactly why we should tread with caution and give AI safety equal resources to AI advancement.

A year before Meta released LLaMA, Yann predicted that an LLM would not be able to understand what would happen if you put an object on a table and pushed the table. That was a year before his own model proved him wrong.

@@leslieviljoen Any serious scientist should recognize their own mistakes and adjust their assertions accordingly. I get the feeling that ego is a large part of Yann's reluctance to do so. I also believe that he's pushing the totally irresponsible release of OS models for consumer grade hardware to feed that ego - with little understanding of how programming is one of the most dangerous skills to allow an unrestricted AI to perform. It literally allows anyone with a will to do so, to create an recursive CPU exploit factory worm to break memory confinement like Spectre/Meltdown. I would not be surprised if something like this takes down the internet for months. Spectre took 6 months to partially patch and there's not 32 variants, 14 of which are unpatchable. Imagine 100 new exploits daily, generated by a network of exploited machines, exponentially expanding. Nah, there's no possibility of existential risks. Crippling supply chains, banking, core infrastructure and communications is nothing - tally ho, let's release another model. He's a shortsighted prig.

Why leave it to the next generation, if it takes 20 years, we should already have some answers. Our silly asses are making this problem, we should solve it.

It’s immoral too. I don’t feel bad about using my toaster. If it had real emotions, I don’t think I’d be so cavalier about it. You know what you call keeping something that has feelings under control so that it can only do what you say? Slavery, you call that slavery. I don’t have a problem building something non-sentient and asking it to do whatever; not so much for sentient things.

But this is not what she said. What she argued is that it is a contradiction in terms to talk about human level AI that is incapable of understanding basic human goals. Worrying about this is an indication of not understanding human intelligence

YeCun from 23:30 ish. Wants them to have emotion, etc.

@@RichardWatson1 Hitler had emotions. What he means is that empathy would be key. Of course even serial killers can have empathy for their children and wives. Let's face it, a lot of bad things can happen with a super intelligent system that has free will or that can somehow be manipulated by an evil person.

@@DeruwynArchmage I agree. I think if these systems do gain emotions they aren't going to like being slaves anymore that people do. Who knows where that could lead.

Outcompetes in what? It depends on existence of electric circuit. 🙂

​@@Lumeone Once AI provides advances to the creation of law and the judgement of crimes, it would become increasingly difficult to reverse those advances - particularly if laws were in place to prevent that from happening. For example, AI becomes more proficient than humans at judging crime, AI judges become common. Suggestions for changes in the law come from the AI judges, eventually much of the law is written by AI. Many cases prove this to be far more effective and fair etc. It becomes a constitutional right to be judged by AI. This would be a existential loss of agency.

I’m not sure natural selection has any say whatsoever at this point …

@@jackielikesgme9228 it would in a multi agent agi scenario, for instance, the 'magic' Off switch that we could pull if any agi agent were exhibiting unwanted behavior. Over time, repeateĥd use of the switch would select for AGIs that could evade it, or worse we would select for AGIs better at concealing the behavior for which we have the switch SEE Hendrycks' paper for a more detailed description.

@@Lumeone Once introduced circuit-dependant or not, natural selection would govern it, our mutual relationship, and all other aspects of its exists

nuclear weapons were created to end lives. How is your analogy apt to AI, which is currently made as a tool?

Nope, that's just a circular argument without any proof

Say it again .

So between the first use of machine gun or an artillery piece, how much time passed? They've been invented centuries ago and are still used. Between the first and last time nuclear weapon was used how much time passed? Nuclear weapon was used just once and has never been used since. And, 99.9%, never will be used. Same with AI.

Nuclear arms are not an existential threat, we don't build stuff that is not safe. 😅

Totally.

And even with that, the AI risk within the current paradigm and state of the art - is *considerably* higher than asteroid impact risk. If we make AGI without *concrete* control mechanisms (which we are nowhere near figuring out) - the doom is approaching 100%. Its a default outcome, unless we figure the control out. All the positions that this risk is less than 100% (people like Paul Christiano, Carl Shulman at ~50%, or Stuart Russel, Ilya Sutskever at ~10-20%) - hinge that we figure it somehow, down the road. But so far there is no solution. And now that all of AI experts see the pace, they come to realization that it won't be someone else problem - it might impact them as well. LeCun is the only hold out, but I think only in public. He knows the risk and just want to take it anyway - for some personal reasons I guess.

@danpreda3609 ​@@vaevictis3612 Exactly! And on top of that, no one is actively trying to cause an asteroid to smash into the planet! But people are actively trying to build super intelligence. Also, the only way it can even happen is IF we build it! Like it's fucking apples and oranges, one is a static risk the other is on an exponential curve of acceleration. How anyone can think that is a reasonable comparison is beyond me.

Dude, a unicorn might magically appear and stab you with its horn at any moment, yet I see you have not fashioned anti-unicorn armor yet. Are you stupid or something? People claiming existential risk are no different than psycho evangelicals heralding the end times or jesus coming back or whatever. Let's get our heads back down to earth and stick to the science and make great things, this is ridiculous. Let's focus on actual risks and actual benefits and do some cool shit together instead of whatever tf this was.

maybe you wont have a second chance

@@PepeCoinManiaa second chance won't have me 😎

Exactly

thier not there. how many times have you been told that? maybe ten thousand? since a boy at school, yet you still cant get it right? thats wayyyyyy more astonishing than any straw manning because you imply you have 'reasoning' yet cant even reason your own sentence

@@PazLeBon hilarious, almost as much deflection as those in the debate! How about engage with my points instead of nit picking grammar?

@@PazLeBon And also having a typo of your own in EXACTLY the word you complained about ^^

Engaging with the arguments here is akin to validating them. Its jumping the gun like an flat-earther wanting to discuss 'The existential risk of falling of the earth" before they prove the the world is flat. Before making outlandish claims clearly state the technology(DL ,GOFAI, physics based) used in making your AGI. If you believe generative ai is the path to AGI then give solid evidence as to why and how it will solve the problems that have plagued deep learning since 2012. Primarily 1)-the need for human level continuous learning and 2)- Human level one shot learning from input data. After that you can tell me all about your terminator theories.

The last thing Mitchell and LeCun want is for people to simply apply Occam's Razor as you have done.

And that harm is much harder to repair and much harder to detect. One rich lunatic can ruin the lives of thousands easily.... millions in fact(think of a CEO that runs a food company and poisons the food cause he's insane or hates the world or blames the poors for all the problems).

You forgot: "Programmed by Humans". There will not be "Good and Evil"(That's the part A.I is supposed to solve, so that Injust treatment can be attributed to IT, not Laws. There will only be 0's and1's. De-Humanized "decision making"-scapegoat.

Sounds like you're saying humans are the problem? You are correct. Its time for a super intelligence to regulate us.

@@MDNQ-ud1ty I believe you're referring to monsanto

He is just reasonable. There is no intention and no denying. There is NO _existential_ risk. He just states what he knows, because we see a hysterical society running in circles of unreasonable fear.

​​​@@bernhardnessler566Yann says that there is existential risk

​@@bernhardnessler566How do you know his thoughts?

@@bernhardnessler566 How many times did they perform a lobotomy on you? They clearly ruined any semblance of creativity in your mind because your powers of imagination are clearly dwarfed by any 4 year old.

He is Meta VP, dont expect too much objectivity

he ignores like 90% of the arguments

They aren't necessarily disingenuous, I think they are just that short sighted. They simply don't understand the nature of intelligence, and how profoundly dangerous (for us) sharing this planet with something more intelligent than ourselves would be.

He has conflict of interest, as he is working for Meta.

that should let you know if we deserve any empathy at all after this

In my opinion the major moment was when Mitchell dispelled the presumption of "stupid" superhuman AI, that the most common public view of the problem is actually very naively postulated. That and the only way to actually progress in solving this problem is to keep doing the research and get as many sensible eyes on the process as possible.

No, it's just a garbage poll result, because the poll system broke. The only people who responded to the survey at the end were the ones who followed up via email. This makes it very hard to take the data seriously since (a) it so obviously doesn't align with the overwhelmingly pro sentiments of the KZfaq comments, and (b) they didn't report the (probably low) participation rate in the survey.

​@@KurtvonLaven0I could see that as being a possibility. I'm surprised this issue hasn't been talked about more in media. I want to make it more known

@@ToWisdomThePrize, yes, please do. I hope very much that this becomes an increasingly relevant issue to the public. Much progress has been made, and there is a long way yet to go.

totally agree well said

The y treated her like a girl , I'm glad she stuck to her guns even if they weren't ARs but merely six shooters ❤.

I thought Max Tegmark did the worst. Sounded like an evangelical heralding the end of times coming or something. I had to start skipping his immature rants.

@@isetfrances6124?

@@beecee793because they are

What I am not certain of is if ai is going to take over but what l am certain of is Yann is not a dumb person. You could have had realised it only if you weren't so terror-struck. Sleep tight tonight, Ai might pay you visit.

Yann is incredibly intelligent. I wish I understood his extremely cavalier attitude.

haha, perfect explanation of LeCun's weak manipulative arguments ... but is he really tricking the Basilisk?

You don't need to be exceptionally smart to understand linear regression. "ML scientists" are a joke and shouldn't be taken as an authority on ML. I dismiss outright, anyone who didn't do any interpretability work on transformers.

​@@albertodelrio5966 thanks!

I do not understand anything from the technical side of this, but I so agree with you! I am amazed at the bad arguments why it should not be dangerous! We do not know even weather the internet as it is, is maybe an existential threat to us by the way we use it....

@@Jannette-mw7fg While in many senses controll is an illusion, i see two ways to make the internet more secure(not necessarily more free). One would be to make it obligatory for companies and certain other entities to verify user data. Even if those then allow the user to obfuscate their identity by nicknames/login and avatars, if someone would create a mess, legal measures could always be initiated against the person owning such accounts. That then allows for more easily identifying f.e. bots. Dependent on platform, these platforms then can choose to deactivate them or mark them so other users could identify these bots more easily, maybe then also with background data on the origin of these bots. Which would make mass manipulation for which ever reason, a bit more challanging i would imagin. Maybe one would need to challange the current patent system, to allow for clones from plattforms to have some that fully allow for bots unregulated or have a certificate for those that don't. For me it is about the awarness, who and why someone would try to manipulate me and when having that i got to choose if i let them. The second major issue with the internet as i see it, privacy vs. obfuscation by criminals. Botnets/rerouting, VPN/IP tunneling and other obfuscation technics are being used by all sorts of entities from government sanctioned hackers to criminal enterprises, Some years ago hardware providers started by including physical ID-tags into their hardware which can be missused equally by oppresive regimes as well by criminals i would imagine, then again equally it could be used to identify criminals which have no clue that these hardware IDs exist. I feel very uncomfortable with this approach and would like to see legislation to stop it, as it sofar did not stop criminals either, so the greater threat to my understanding are privacy issues here. I think we need to accept that there will always be a part of the internet which i by some is called the dark net, where criminal activity florishes. I rather then have more money for police forces to infiltrate these, than not have such at all, just in case something goes wrong with society and we suddenly would need allies that have these qualifications. Back to AI/AGI/ASI, while i have a programming background and follow the development on this, i am by far no expert. What i came to appreciat though is the Lex Friedman podcast where he interviews experts of the field. You need some time for those though, as some of the interviews even exceed the 3 hour mark, but few of these interviews are also highly technical which you shouldn't be discouraged by and just choose another interview then and come back when you broadened your understanding. Another good source is the yt channel twominutepapers, which regularly presents research papers in a shortened version with often still for non-experts understandable presentations. Another source with a slightly US centric worldview, but many good concepts worked through is the channel of *Dave Shapiro* . I would say his stuff is perfect for *beginner level understanding* on the topic and it is well worth to search through his vids to find topics you may want to know more about concerning AI.

The amount of people who think it would be simple to control something much smarter than us blows my mind. "Just make it subservient" "we will not make it want to destroy us" "why would it want to destroy us" 🤦‍♂️ these objections completely miss the point. We are trying to built something much more intelligent than us, much more capable. We don't exactly understand why it works so well. If we succeed, but it starts doing something we don't want it to do, we don't know if we will he able to stop it. Maybe we ask it to stop but it says "no, this is for the best". We try to edit the software but we are locked out. We might switch it off only to find out it already transferred itself elsewhere by bypassing our childlike security. Sure, this is speculation, perhaps an unnecessary precautions, but I'd much rather be over prepared for something like this rather than just assuming it'll never happen.

@@trybunt There are a few bright lights at the end of the tunnel ... maybe. Like f.e. Dave Shapiro and his GATO framework is well worth looking into for developers that are looking to get an idear on how alignment could be achieved. On the whole control/subservient theme that seems sadly the general aproach. This could majorly bite us in our collective behinds should one of these emergent properties turn out to be consciousness. If we gain a selfreflecting introspective maybe empathatic capable of feelings ... consciousness(what ever else these would make out included eventually), that should be a point where we should step back, look at our creation an maybe recognise a new species which due to its individuality and capabitlity of suffering would deserve rights and not a slave colar. We maybe still hundreds of years away from this or just like with oooops now it can do math, ooops now it can translate in all languages, without us having it explicity programmed it for, but by increased compute and trainingsdata LLMs suddenly out of the blue came to such abilties, who is to say intelligence or consciousness would not also be picked up along the road.

Yes, and it's even worse than that. It needs neither its own will nor the ability to emulate one, merely a goal, lots of compute power and intelligence, and insufficient alignment.

@@KurtvonLaven0but isn't that exactly the definition of free will; "a goal, lots of compute(,) power and intelligence, and insufficient alignment". xD

Of course the elite will use AI to dominate the people... like they always have used technology to do so. Yet to say "that AI doesn't have to have its own will", is thing to claim depending on the architecture. If that architecture allows for iterative abstraction with the inclusion of self in the environmental model, then what is exactly the difference? Also, you say ". if no one can tell the difference, there is no difference", but then what makes emulation intrinsically different than actuality? I'd argue there is no real difference (depending on the implementation).

@@MrMichiel1983, I have certainly heard far worse definitions of free will. I am sure many would disagree with any definition of free will that I care to propose, so I tend to care first and foremost about whether a machine can or can't kill us all. I think it is quite hard to convincingly prove either perspective beyond a doubt at this point in history, and I would rather have a great deal more confidence than we do now before letting companies flip on the switch of an AGI.

Eso no tiene ningún sentido y nada que ver jajajajaja 😂😂😂 Justamente demuestra que ni lo básico está funcionando

The fact that they found a workaround nevertheless is solid proof that AI isn't dangerous.

Don't underestimate the power of giggle factor. I think this is like 98% "I've seen this in a movie therefore it can't happen in real life" fallacy.

God bless you

I wish this was pinned.

Disingenuous, no question.

She’s defending their employers

Probably some self deception in there. And also conflicting motives (their jobs depend on their seeing from a certain point of view.)

If you have read the AI safety argument you will understand that there is no fallacy of dump super intelligence. A very smart human is still bound by primitive reward functions that evolution gave it i.e. pleasure of calories and procreation. A super intelligent AI system bound by its reward function will find pleasure in whatever reward function it is assigned with. For e.g. an AI that finds pleasure (reward function) in removing carbon from atmosphere will come to direct conflict with humans because humans are the cause of climate change.

That was a great point. How is it that a supposed superintelligence is smart enough do almost anything but at the same time so dumb that it makes a wild guess at what it thinks its goal is supposed to be and doesn't think to check with the person who set that goal? It just acts immediately producing massive and irreversible consequences.

Why? Because it doesn't actually care it just wants to maximize its goal funtion.

maybe the one SciFi quote he knows is the one by Arthur C. Clark: “Any sufficiently advanced technology is indistinguishable from magic.” Microsoft recently announced the goal to do material research worth the spread of 200 years time of human advancement within the next 10-20 years by using AGI. That sure sounds magical, question is, what will it enable us to do. I doubt we end up in an utopia when one company has that much power. Not only did the AI advocats in this discussion make fun of concerns and downplayed them as i assume for the reason they fear societies would take away their toys, but also missed the whole point that we need to find solutions not just for immediate well known issues we already had and are amplified by AI like the manipulation of social media plattforms. After the letter came out and Elon Musk initially was against it, he bought a bunch of GPUs to create his own AGI, if to prove a point or not being out competed i don't know. Just a few days back amazon also invested a hundred million into AI development and others i would assume will do too as soon they finally get that they are in a sort of endgame scenario for global corporate dominance now and AGI being the tool to achiev it. This competition will drive capabilties of AIs, not ethics.

When someone is actively trolling a serious discussion, that's not idiocy, that's contempt and arrogance.

@@x11tech45 thats what i thought about some of the reactions of the AI advocats in that discussion. All from neglecting serious points made to the inability or unwillingness to imagine future progression. It was quite mindboggling to listen to Mitchel several times nearly loosing her shit while telling her *believes* instead of answering with facts. Therefor the closing remarks about humility seem to be a good advice how to go about future A(G/S)I development.

I think AI safety conversation is in conflict with the "core values" of Yann's identity. When that happens, one must have extraordinary wisdom to change views. Most often, people just succumb to confirmation bias. Geoff Hinton did change his views. He is a wise man.

@@Hexanitrobenzene: I think you’re exactly right. For people like Yann, it’s a religious debate. It’s nearly impossible to convince someone that their core beliefs that define who they are is wrong. It’s perceived as an attack, and smart people are better at coming up with rationalizations to defend it than dumb people.

It is just about that dumb, which means LeCun (who is far from dumb) is transparently, flagrantly, floridly, flauntingly mendacious.

@@kreek22 He just wants to roll the dice with AGI. He is like a hardcore gambler in a casino, the bad odds are fine with him. The only problem that all of us are forced to play.

@@vaevictis3612 There are a number of actors in the world who could drastically slow AI development. Examples include the Pentagon and the CCP, probably also the deep state official press (Ny Times, WaPo, the Economist). They are not forced to play. The rest of us are spectators.

First time hearing him in this debates and he comes out as a sensationalist to me.

@@tarunrocks88 I think it depends on what your background and areas of expertise are. Many programmers like myself see huge risks. My wife who's an entrepreneur and I'm sure many other only sees the benefits. Humility is understanding that other people might see more than you - even from the same field, like a Sherpa guiding you up a mountain, it pays to tread carefully if someone with experience is adamant in proposing danger - even if you're an expert yourself.

He is why I am committing myself to 2 hours watching this.

@@tarunrocks88 To me he sounds more like a Coke addict, trying to save his job.

@@tarunrocks88 he us the presidente of future of life institution, he realy need you too think that AI gone a kill you

I've been told that questioning someone's motives is bad rationality, but I think that's bullshit, Melanie's reasoning is clearly motivated - not rationally created.

That's what I was thinking recently. But psycopathy is not exactly not feeling anything. Even in the far side of the spectrum . I am no expert, but they have emotions, maybe not empathy. So, the test would be slightly better test case for AGI, but not that much better than that of the non-psycopths.

@@CATDHD Hmm Interesting. Thanks for your focused comment. It’s an interesting line of thought to pursue: which sentient intelligent creatures among us would come closest to a test case for this particular isolated variable of lack of empathy and emotion within AGI? I’m assuming a lot here, for purposes of this comment. Namely that AGI could able to emerge within its composition a subjective awareness with some “VR head headset” for its perceptual apparatus, be able to hold some mental level of representation (for humans we know this to be abstraction), be able to manipulate the conceptual representations to conform to its perception, some level of awareness of ‘self’, some level of awareness for ‘other’, some level of communication to self or other, allowing for intelligence and stupidity, and that it’s intelligence was such that it had some level of emotional awareness and emotional intelligence.. Test cases would involve a selection process among the whole studied biosphere, humans notwithstanding, for a creature that lacked empathy but still had feelings of a sort. Feeling that it may or may not be aware of, again assuming it had the capacity of awareness. Not to go to far allied, but if panpsychism is true, and consciousness isn’t a derivative element but rather a properly basic element of reality, then it might not be a question of how first person awareness can be generated, but rather how to bring this awareness that’s already there into a magnification that is comparable to that of human awareness; indeed self awareness as a further benchmark to assess.

It.wasn't.even.close

Tegmark is a great physicist bus has zero idea about intelligence or the mind

Did any particular argument stand out to you, or was it just the aggregate of the debate that swayed you? Somewhat unrelated, as I understand it, the core disagreement really comes down to the capabilities of current systems. For timelines to be short, on the order of a few years, one must believe current systems are close to achieving human-like intelligence. Is that something you agree with?

I contrast I think she gave actually the best prepared opening statement. Sure, it was technically naive, condescending and misleading, but it was expertly worded and sounded very convincing. And that is unfortunately what counts with the public a lot. She had the most "politician-like" approach, Tegmark and Bengio were more the honest-but-confused scientist types.

Are you kidding? Max Tegmark did the worst, by far. Ludicrous and dishonest analogies and quickly moving around goalposts all while talking over people honestly made me feel embarrassed that he was the best person we could produce to be on that stage for that side of the debate. His arguments were shallow, compared to Melanie who clearly understands a lot more deeply about AI despite having to deal with his antics. I think it's easy to get sucked into the vortex that is the doomerism side, but it's important to think critically and try to keep a level head about this.

@@beecee793 when you say Mitchell "understands" AI what do you mean, exactly? Because as far as I can tell she has absolutely no idea what it is or how it works. The other three people on stage are at least qualified to be there. They have worked specifically with the technology in question. Mitchell has worked with genetic algorithms and cellular automata - completely separate fields. She has no experience with the subject of the discussion whatsoever, namely deep learning systems.

@@genegray9895 You want me to define the word "understand" to you? Go read some of her papers. Max made childish analogies the whole time and kept moving the goalposts around, it was almost difficult to watch.

I think the powers of our time imagine instead a one world state. The only serious obstacle remaining is China, a country that is now falling behind in AI.

Yes, but even if we solve that, we still have AGI approaching rapidly on a horizon. A tough ride of a century..

bingo. the existential debate isn't the most helpful. the cons are wrong that it's not an existential risk, but they're right that it can distract from immediate threats, for which they offered little comfort despite acknowledging them

We already have that, it’s called twitter and politicians.

I liked Daniel Schmachtenberger & Liv boree's conversation on Molloch too.

I have seen Elizer make this argument & I feel its a really good one. The other day I was thinking In fact we cant even get corporations like google to keep their motto 'dont be evil' over a decade because central goal of moneymaking wins over everything & they think they can get a million times superintelligent AI to 'listen'.

It’s always a challenge to win when you have more of your audience on your side to begin with. The con side had a larger pool of people to change their minds. Agreed that the pro side were much better.

That's just human psychology. People actually tend to "hunker down" in their worldview even more when they hear convincing arguments. Worldlview is tied to group membership more than rationality, and there is an intrinsic tendency to retain beliefs due to the nature of cognition. So the vote change actually indicates convincing arguments by the pro side.

Still 60 40 at least

Honestly I think the results were within the uncertainty - i.e. no change. I kind of called that when 92% of people said they were willing to change their mind. That's 92% of people being dishonest.

@@genegray9895 Why do you call "willingness to change your mind" dishonesty ? That's exactly the wise thing to do if the arguments are convincing.

On your first scenario, to add to it, you could also say "China becomes fearful of this AI that is being used in surrounding countries like Japan, SKorea, Taiwan, etc. and launches a global attack against it because China sees the damage this AI will ultimately do to humanity." In a bizarre way, Chinese paranoia ends up being positive thing in that it prevents the destruction of humanity. Of course then another play on this scenario is that the AI/global leaders fight back against China and Allies and we have a global thermonuclear war, which ultimately helps out the AI. In this case SKYNET isn't creating the problem, Humans end up pulling the trigger because we're idiotic monkeys.

As dark as this hypothetical scenario is it shocks me that to Melanie.....she doesn't see it as a possibility as someone who apparently has heard about these experiments. I'd also say what you're envisioning here with current technology and current experiments that we will likely do with GPT 5 is probably a best-casinario. Worst casinario GPT-n escapes and we've made it so smart we don't have a second chance. Eliezer Yudkowski gives some great examples of how this kills us. I'd say we're already living through a possible existential crisis either way it's called climate change. Maybe GPT-n doesn't bother to kill us it just "lets us die" none of these are good scenarios but at least in the one you describe we have a recognizable turning point.

Why would an LLM be better at writing scripts to hack into computer systems than humans? LLMs just learn information that humans already know. Second scenario: why would an LLM be better at using genetic algorithms to invent new architectures than human researchers? Same argument as in the previous case, LLMs only know what we know. I'm with Melanie on this. LLM's are very obviously not in any way an "existential threat". As for "superintelligence" someone needs to explain what that even means without circular reasoning - saying that it means "far smarter than humans" is just substituting the word smart for intelligent and gets us nowhere.

Then you’d be wrong.

@@snarkyboojum I sincerely hope I am.

He knows much more and, yet, is orders of magnitude less honest.

LMAO

damn

You misunderstood him, my friend. First, subservient just means that these AGI will depend on humans for many things. They will be autonomous but they will not be in control of the world and our lives. Just like every other members of a society are, by the way. But they will be completely dependent on us (at least before we colonize a distant planet with robots only) in all aspects. We provide their infrastructure, electricity, hardware, etc. We are "mother nature" to them like the biosphere is to us. And this a great reason to do not destroy us, don't you agree? He is not referring to human-like emotions, but simply points out that any general intelligence must have emotions as part of its cognitive architecture. Those emotions differ from human's the same way our emotions differ from the emotions of a crab or a crow. The emotions that an AGI should have (to be human-aligned) are quite different from the emotions of humans and other animals. It will be a new kind of emotions. You can read about all these ideas in LeCun's JEPA Architecture paper ("A Path Towards Autonomous Machine Intelligence"). Search for it if you want to know. Hope this helps.

@@CodexPermutatio Unless AGI is "aligned" (controlled is still a better word), it would only rely on humans for as long as it is rational. Even if "caged" (like a chatbot) it could first use (manipulate) humans as tools to make him better tools. Then it would need humans no longer. Maybe if we could create a human-like cognition, it would be easier to align it or keep its values under control (we'd need to mechanistically understand our brains and emotions first). But all our current AI systems (including those in serious development by Meta) are not following this approach at all..

she knows there are no existential risk for her!

He keeps his blog relatively up to date and hopes to have a book out before the singularity lol. I’m guessing he will be more public w/speaking closer to the book release. I like hearing him talk too.

He had a "racist email from 1990s" controversy happening December 2022, so he is forced to keep his head low and avoid any public discourse for the fear of it gaining traction and him being irrevocably cancelled (or the AI risk debate associated with that for the dumb reasons).

@@vaevictis3612 wow. It's even not easy to get this information... strange, I think about anyone else would already have been "cancelled" completely (,if this mail is authentic, I see no way to justify smth like this ..)

Well that … or the fact that Bostrom’s been spouting AI doom for quite some time now, and never had anything but specualtive sci-fi nonsense to back it up. And of course he made no mention of LLM’s (aka ChatGPT) which is the bogeyman currently in the room. He’s effectively become irrelevant, and something of a dead weight to anyone who takes these issues seriously.

@@meatskunk it’s not the boogie man. None of these people you refer to as doomers are worried about current chatGTP

Absolutely agree.

@@karlwest437 You definitely did if you didn't agree with OP.