I, too, am extraordinarily humble.

Came here just to say this.

Most phds I've met hate being called a doctor on informal conversation. The ones who insist on it are... Wel... Assholes.

@@Phelan666 You say you're humble but no one on Earth is even close to being as humble as I am. No one in the history of humanity has ever had as much modesty and humility as I do. I am the greatest person to ever exist when it comes to being humble.

Username doesn't check out :)

I wish I could say my opinions on quality were usually positive :) But I can't ಠ_ಠ

also a lolcow

@@mattimorottaja8445 why?

@@enricobianchi4499 because of the way he looks. Dudes an internet atheist and wears a fedora and is overweight yet has super skinny arms and has a neck beard. I don’t like judging people on their appearance but why would he knowingly conform to such a negative stereotype? He needs to be different

@@dylancrooks6548 having looked into it in the meantime, yeah he's kind of a fun little guy

Watching some of Yudkowski's interviews and reading comments, it's amazing how often people will hear "we're all going to die" and respond "that guy is wearing a hat".

Musk's fears are pretty much summed up in what Bostrom wrote in 2014 in his book "Superintelligence". Elon thinks humans and A.I. can co-exists only if we merge with it via some kind of brain-computer interface. I'm really curious about what Robert has to say about it. Bostrom made some heavy assumptions in the book but I suspect Robert's criticism (I guess? That's the impression I got) is going to focus more on Elon's responsibility as a public figure than his arguments.

Argument goes like this: it's easier to built AGI than to build *safe* AGI, so making progress on AGI open is not very wise. You risk that some party that doesn't buy into whole "risk from AGI" business will take your results and in effect will have to do less work, cutting your lead time (which you need to do the required work on safety). OpenAI makes its research results public and it seems to me that results are more on AI-capability than on AI-safety front.

spirit, Ben Goertzel argues it's better to arrive at the singularity sooner rather than later because if we have a massive computing overhang it could make it possible for perhaps dubious smaller or less established groups or even single individuals to reach the end zone without much thought or telling the world what's going on. The irony is of course as you point out that it may be much harder to create benign AGI than to create AGI, this is a field of true uncertainty, so time is an important factor in thinking things through as is adequate resources.

Non-open AI will be even worse, since there will be incentive to hide the flaws of AI to keep investor dosh rolling.

Open AI is still worse, because competition favors those with least regards toward safety. It is hard to have oversight over a few competing companies, it is impossible to have oversight over millions of individual participants or small groups.

'Ohm'? Pun intended?

LOL, 255 bytes era!

were*

@@fakeaddress My opinion hasn't really changed since then - they seem to focus much more on capabilities than alignment

which is bad

​@@NNOTM They are focusing on alignment, though. It doesn't make great click-bait, so the media doesn't report too much on it, but their blog has lots of articles on what they're exploring. (It's hard, so progress is slow, but it is an area they invest in.)

Likewise with a likewise comment!

So the question is, what kind of nuclear power are we talking about? Before we get a working nuclear power plant we may get a working atom bomb, and there may be nobody left to work on containment.

I find the runaway intelligence scenarios relatively hard to envision. I think personally there will be something like "maximum intelligence per unit hardware" which is the most intelligent thing you can design with a certain amount of hardware. IMO a super intelligence will first need to build a massive amount of hardware. Nick Bostrom had an interesting concept for this in his book, that there could be self assembling nanomachines on a molecular level which gives a vast amount of hardware. However there is a bootstrap problem in that you have to be extremely intelligent already to invent this method, which an ai wouldn't be without it. Though I think it's important to worry about rampant runaway I can't see how a single pc, for example, could become super intelligent. What do you think Rob? Do you believe there is a hardware requirement for superintelligence? Does that prevent an extremely quick runaway scenario?

I think there *is* a hardware requirement, but I'm not sure it's small. And even if a laptop isn't enough to get a decisive strategic advantage, there is a lot of available hardware in the world (accessible through buying, hacking, ...).

Thing is, today's most advanced machine learning algorithms and other frontiers of AI research aren't running on individual machines - they're networks, distributed across numerous specialized units (Google's TPU2 system, for a topical example). We have designed AI to take advantage of the efficiency and power gains which only such distributed networks can provide. Would this not imply that any threat posed by runaway AI be of a decentralized nature? (Please correct me if I'm wrong on anything.)

The human brain runs on merely 20 Watts of power, an AI could easily have Gigawatts available to it. In terms of energy efficiency, the human brain operates 500,000 times below the Landauer limit. So yes, there are limits, but they're not very limiting. You can read more about it in this paper: intelligence.org/files/IEM.pdf

BobC In reference to the fluidity of the domain of AI, the most prominent recent example of that is probably AlphaGo. Defeating top humans at the game has been a huge milestone to be achieved for AI for a long time, but once it was done, everyone said "oh but that's not true AI". Is that what you're talking about? Personally, I would say that this is more of an effect of goalposts moving rather than utter dismissal of some concept in the AI domain. These goalposts have been meant to illustrate how far we have come towards developing true AI, and weren't ever actually meant to be viewed as true AI. I'm not sure how much sense I'm making so I'll leave my comment at that for the time being

Quite right: It certainly is about the goalposts, but I think it is also about how "the complex becomes easy", as what once were strictly AI results permeate into general use, and also become better understood. By this I mean the concepts underlying an AI result become simplified, not just the application techniques. Once simplified, other domains then "adopt" the AI result as part of their own area, leaving AI in search of new turf. ML may be the first AI result to *not* have its underlying theory clarified before seeing widespread adoption in industry. For example, it is still very difficult to ask a complex neural network: "What rules have you learned?" Getting a trained neural net to "explain its inner workings" is hard! We simply accept that a trained neural net has learned to do something that needed doing. It's still very much a black box, We really can't yet extract just what the new rules are! Recent work using multiple different neural network architectures to solve the same problem gives me hope that we may eventually be able to extract the learned content by comparing the trained states of each network. The initial point of such work was to try to find better ways to anticipate which network architectures will prove optimal for a given problem. Basically, to train a neural network on aspects of the problem so it will select the best neural network architecture to use to solve the problem. But I think (hope) there may be more that can come out of such work. In the late 1980's and early 1990's, during the "backprop" and Moore's Law explosions that triggered major growth in ML, I was involved in attempts to convert neural nets that had been trained to control complex dynamic systems into simpler representations that could be programmed into embedded microcontrollers. The underlying question was straightforward: "Can we take what a trained neural network has learned and implement it in other domains?" Our hope was that whatever the neural net had learned could be converted to existing control theory, and hopefully reveal where control theory itself could be enhanced. That is, have control theory "adopt" the results (but not the process) of neural nets that had been successfully trained to control dynamic systems. We expected to eventually map everything back into an extension of, or modification to, Pontryagin's maximum principle concerning the control theory Hamiltonian. Which is ideal in theory, but generally impossible to use in practice due to the dimensional explosion of the PDEs it generates. We failed. We were a tiny team with no formal support, so our resources were limited, and when initial progress hit a wall we all had to move on to other projects. Perhaps it is time to try again.

​@@flymypg We're not developing an AI, if you ask me, we're only basically discovering what intelligence already is. From your example, not being able to ask "What rules have you learned?" implies a dark truth about how we can make a machine that can operate on a large problem domain, and even solve it better (quicker or with a greater throughput) than any human, but has no deep understanding of the domain from a meta perspective. I'm sorry but that's not intelligence. The results with all of these can be impressive at times, yes definitely, but all of it is basically some type of a real-time classifier, trained to be able to impress us with the end results. That's the only utility function there is, summed up. No genuine intelligence, and there will be none any time soon. I'm sorry to be that guy. Yes, it'll be impressive at times, and it might even kill us because a human made an error in judgment, but yeah, no one is going to try that human, it'll be the "rogue AI's fault". We humans are peculiar. All of it poses deeper philosophical questions. If you have a box with just one button that says "SAFEST BUTTON", but inside that box is a monkey with a gun, whose fault is it if you keep pressing that button until the fatal end? We are simply delegating responsibility to a riddle-like proxy device. This is today's AI safety in a nutshell. Such a machine has no genuine thoughts, only decision-making routines that are convoluted by design -- to promptly open up the "decision-space" while pursuing a hardwired intent. It doesn't matter how many of these are interconnected, they're all potentially faulty in design. Such room for an error. And such a backstage for crass complacency and irresponsible behavior on a grand scale.

People not always know how that surname is spelled: I've in the past often doubted and spelled it "Hawkings", long before knowing of Dawkins.

Interesting. My theory is: It's because of the character Sam Hawkens (often mistakenly spelled Hawkins) from Karl May's Winnetou novels.

The internet has ruined me. I read "copies" as the diminutive noun form of "coping".

I'm curious. I can name 3 great ideas of his: SpaceX, Tesla and Solar City. Can you name 27 bad ideas Elon had? I think the ratio you pulled out isn't true or fair. The only ideas that matter are the ideas he puts his money on. Each and every one of those are good.

I didn't expect people to think 1/10 was some kind of exact figure. Anyway, I don't think funding is the only thing that dictates whether one of his ideas matters or not. People take him seriously in all sorts of fields whether they're in his area of expertise or not.

Actually the Hyperloop is not as ridiculous as most people think, it's just another radical idea of his. The main difference is that his other radical ideas are proven to be very good, but the Hyperloop haven't got the chance yet. Most people thought that electric cars are ugly, slow and too expensive to sell, self-driving car are decades away, a small private company won't ever be able to build serious rockets, and reusable rockets are impossible to build for anyone. If you read the original paper on the Hyperloop (Hyperloop Alpha), you will see that it's not just possible, but cheap compared to the competition. All criticism I've ever heard has been addressed in the paper.

the problem with elon musk ideas is not whether they're ridiculous or not, it's that when he says something people are expecting it to be something that is feasible in the near future. terraforming mars is one of those ideas, yet i don't see nearly enough backlash against it even though it is orders of magnitude less feasible than the hyperloop.

@András Bíró "Hyperloop is (a) radical idea of his" - riiiiiiight. Google "wikipedia vactrain", the inklings of this can be seen even in the 18th century, and by the start of the 20th century it's pretty much crystalised. It's not his idea, he just stamped his brand onto an existing one and convinced everyone it's his... just like he did for example with Tesla (definitely NOT his idea and he had NO hand at all with the early prototypes - in fact those early prototypes already existing and working convinced him to become an early investor in it). Musk is a marketing genius, not an engineering one, and it shows since the hyperloop as it happens is exactly as ridiculous as it's always been.

He plays his little ukelele outros

There are several risks associated with AI. The most immediate, inevitable, and mostly accepted is high unemployment caused by rapid automation. That's not a safety issue, that's actually what we want from AI, but we are horribly unprepared to handle it (at least the US). If handled badly, even that desired effect can destroy our civilization. High unemployment leads to economic and social problems (this is already happening in the US), which if goes too far will inevitably cause economic collapse and massive violence, maybe even a civil war. And the collapse of the US economy will bring down most of the developed world (and China), and that will likely start wars, which could escalate into a global nuclear war. Another issue that is rarely discussed is the dangers of a fully automated military. In the wrong hands (which means any human hands) it can be horribly dangerous. With modern technology it's already way too easy to fight wars. Since it costs few lives on your side, it doesn't generate strong resistance at home, and can be continued indefinitely or until you run out of money. Since war is extremely profitable for a few powerful groups, the incentive to keep fighting is very strong. That's one reason why the US now involved in at least 7 major wars and countless small conflicts. The war in Afghanistan is 16 years old now, and the default position of the government is to continue. When pressed for reasons the only thing they can come up with is the valuable minerals that could be taken from the Afghans. And once you have a fully automated military that follows any order without questions, and free from moral dilemmas, you can use it against your own citizens too. This has plenty of historical examples, but with human soldiers it's always a gamble, because you can't be sure that they will follow orders. Yet another issue is censorship. Now the internet is too big and too fast to be controlled by anyone, and that's a huge boost to democracy, we now know far more about what's happening in the world than ever before, and we can fact check political speeches in real time. With advanced AI it will become possible to control most information on the internet, and force state propaganda on it, like it was in the days of centralized news. As we saw last year, leaked e-mails can change the outcome of presidential elections, so it's a pretty big deal. And I still not touched actual AI safety issues. There are plenty of them, but a big one is superhuman intelligence. Once AI reaches human level, it will keep improving without any hesitation, there's no inherent speedbump at human level, it's a nice wide highway. We already saw this with all tasks that AIs mastered, when AlphaGo reached human level in Go, it didn't stop, it kept going and in a scarily short time it was far ahead of any human being. And once we have a general AI that is far more intelligent than us, we won't be able to control or stop it anymore. We will have as much control over it as a mouse has over a human. Early humans hunted to extinction animals that were larger, stronger, faster and more agile. This means that we have only one chance to do this right. If we screw it up, we can't fix it.

I don't necessarily disagree with the gist of your response. In my original comment I was just nitpicking one detail from the video so that people wouldn't get false impression and IIRC Robert put strong emphasis on what I said in his next video.

I understand that every single person who signed, read and agreed. If there's any discrepance with the content it should be very minor, else they would not sign it.

Spectacular quote, thank you

Elon's got ideas but no background or qualifications in AI research

He's a business person with a dash of tech nerd. His insight is just barely more than your average youtube-taught geek, but he likes feeling like the smartest person around. It would be a shame for Robert to waste his time inflating the Musk hot air baloon.

I would like to point out that there is a big difference between "Don't agree" and "disagree".

How you doing in these times Leo?

@@Kaos701A Honestly just annoyed, cause people don't seem to care about what is fed INTO the AI, and from watching these videos for2 years now, that is something we should VERY much care about.

Futureologists are more visionary artists than engineers, we need tangible ideas that can be implemnted and experimented with right now rather than ideas and inventions that might become practical 50 years from now as there is too much we don't know right now. We thought there would be flying cars, now there are some prototypes but they've been found to be wildly impractical toys with the current technology we have; which in turn we'd probably draw the same conclusions about other ideas, like the lift carrying sattelites into space, sounds cool but will be impossible to build as no one knows how exactly to build it.

@@egg5474 So, if they are visionaries and artists that somehow justifies calling them not real and not serious people? Think about weight of these words. And I don't know what futurologists you read but that's not the same as science fiction writers. Futurologists often discuss the problems this channel is all about: technology and morals, AI and future of humanity. That's why he said "it's not only futurologists who discuss it". Trying to lower a group of people down doesn't make any good to "the engineers" you mention. I'm an electrical engineer and I don't appreciate this.

Maybe you need more fiber in your diet.

+Michael Deering

Build your own AI to protect you from it, that's my plan :)

Fleecemaster Noone can make an AI that'll kill you if all the AI researchers are dead :)

How’s that going?

He's very smart, but I doubt you've read many other AI researchers. You should probably write a qualifier next time. I agree we should we concerned, though.

I can't identify it. Tell me now.

It's (a ukulele cover, or something, of) Respect by Aretha Franklin.

+Aexis Rai I have a feeling it's an electric ukulele battleaxe cover, to be exact.

4:05

If something happened it likely would be because human stupidity not that of an A.I. humans are the ones with a control problem.

This doesn't work too well with brand new things though. Like, who are you going to believe in 1902, Lord Kelvin the extremely accomplished scientist, or two bicycle repair guys from Ohio?

@@RobertMilesAI Well, if I am only told by the priesthood that Lord Kelvin is an accomplished scientist and I actually see the Wright brothers get their plane off the ground, it should be obvious which I put more confidence in. The principle is simple: Show, don't tell. I tend to believe my own eyes. And I tend to view "experts" as priests. They may have a reason to believe something. But unless I can see it for myself, I don't.

Have you heard? Some guy on youtube said we shouldn't worry about AI safety! We must act immediately.