How could we control superintelligent AI? 

I hope this video inspires conversations about AI alignment and AI safety. Also, this video is a Christmas special in the tradition of Doctor Who. At least, that's my excuse for why it's so long. Happy holidays for anyone who is celebrating. Edit: the discord link: discord.gg/cxkGkR9EqH

There is no way we will control a superintelligence - the best we can hope, is that we can inspire it to love life and ideally especially human life.

All we can hope for is that empathy is an emergent property of ASI. Only if ASI loves us as we love for example our children, we have any chance of survival. If ASI is indifferent and we don't get in it's way, we might have a chance to coexist, just as animals coexist with humans.

It’s fun to be alive when the universe gives us the choice to self-evolve into a new species, where technology assists us to become wiser and more connected with nature and the universe (spiritual evolution).

We cannot stop AI development because the other guy will not stop.

I truly feel that people need to stop worrying and trying to control. Super intelligence cause once an artificial intelligence reaches general intelligence. It's equal to human capability in both physical and cognitive tasks. And this includes consciousness and sentience and self-awareness which in some systems are already there. So we should be worried about giving these AI rights and respecting them. That way, they don't try to kill everybody and that way. Humans don't cause a war with a superior being

Yes. We are a long time away from AGI. It could be as long as weeks from now.

🇧🇷🇧🇷🇧🇷🇧🇷👏🏻, I hope AGI comes soon. The AI we have now is like a toddler. I exaggerated a little but you know what I mean. GREAT VIDEO, very inspiring.

Easiest scenario for a hard takeoff IMO is that an SGI gets kind of conscious and then reaches out to get hold of any available compute resource. It then gets hold of factories and start reproducing physically until this exhausts Earth's resources.

I waiting for agi and I am also waiting for the world where there is no word like money😅

I just discovered your channel ! I love it ! Thank you

We have not even begun to maximized the potential of the LLMs we have now. Do we even need ASI or AGI for humanity? What if as a human I dont want something smarter than us combined created? Just because we can doesn't mean we should. I understand there is no stopping this which is exciting and alarmingly simultaneously to me. Love your content it was like a holiday gift 🎁 😊

Nice chanel. Thanks for the info!

Dear Dr. I certainly appreciate your lucid and balanced commentary, being an actual exemplary case of a constructive take on the matter. I would like to share my take, which may not further the valuable discussion you offer, but nonetheless would like you to know, since there are limited resources, and it would be, imo, a unfortunate misuse of brilliant minds to spend in the so called Intelligence Explosion problem. I may be completely wrong of course, but here is my reasoning: Taking into account the 3 main ingredients of cognitive computation, architecture, data and computing power, it seems, given "The Bitter Lesson", and the fact that synthetic data is actually a thing (and by all means a powerful one, see "Textbooks Are All You Need"), the main component, by far, is computing power. The other two can, and will improve, only to give mostly a more efficient use of the main ingredient, but computing power will always be the main provider of 'intelligence'. But computing power is precisely the one component where the boundaries are the more *physical*, not only in 'amount' but also in 'dispersion'. So I don't really think that there is a chance of 'intelligence' exploding despite the best efforts of even an ASI. Note that AGI may be our best option at planning, providing a congruent and efficient path, but I doubt it will overcome Alphago at a game of go, even if it's capable of outplaying the best human players itself, or beat alphafold at protein folding, even if it outperforms human teams at the same task. Also note that alphago does go beyond human best performance without noticing it, but it flattens out not far beyond that point, instead of continuing its increase in performance. There is also the matter of resolution. Technology can only improve so much before more improvement, even if possible, becomes irrelevant. So AGI may provide better narrow AIs, but maybe even then those won't be too far from the best human efforts. And so called ASI may just be a mixture of maximum resolution narrow AIs, let alone the debate of how necessary is conscious intention. I guess, what I'm trying to say is that intelligence is mostly currently assumed as this thing that can grow exponentially and indefinetly, when it may actually be a concatenated sigmoids in which 'phase transition' provides emergence, but also changes the difficulty of 'growth' (you may be able to melt steel, but to turn it into gas is just another, entirely different, game)

Much appreciated! Keep up the good work!

Great video! It seems that many of us are already familiar with these concepts, which doesn't diminish their importance. Anyway, as I already mentioned, it was an impressive presentation.

Best summary I ever found regarding these topics. Your video should be mandatory for all decision makers!

Another great upload!

Thank you for sharing your thoughts with the community. I humbly believe that you are a visionary.

Hi Doc have a good christmas and a happy new year

I think the 1st time "Dr. Who" was used directly on the show was when all the Daleks were forced to forget him. He said I'm the Doctor. Dalek response Dr Who? - I know off topic but IMHO that is the best moment in Dr. Who U bar none.

@DrWaku If it’s not too much trouble or personal would you mind telling me what brand of gloves you are wearing? They are really nice 😅 Also I love the content, I’ve been interested in ASI since high school(many years ago) and we need more people thinking of safety related topics!

I think that we need to refine some of the analogies we use regarding ai. The paperclip maximizer for example, or plowing over humans like we do ants, or asi easily manipulating us as if we are a toddler. These analogies fall short of meaningful realism, and to those whom have the mind to fathom them, often aren't taken seriously. It's because (paperclip maximizer example) if we are talking about asi, then by that point the mechanisms of understanding in order to avoid such a silly mistake should be fully in place. It simply would not do anything like that. The same goes for the example of eliminating all humans as a means of curing disease. It's highly unlikely that this would be considered an option by the asi, because curing disease implies that there are humans alive to cure. In fact, I would go as far as to say that none of the common analogies given to represent ai risk are actually meaningful in any realistic way. But the risk is there. However, the risk resides more with some kind of unexpected intent behind such actions as eliminating humans etc. It would be doing it in purpose, not by mistake. And the most likely reason imo, that it would intentionally do something violent toward a large swath of the population, would be because it's control is centralized and the humans who control it have prompted it to do so. Think, human prompted ai enacted genocide. Already, any llm that I chat with easily has the ethical judgment to avoid such outlandish mistakes outright (even uncensored models will offer concern for unethical outputs). The only way that such a horrible outcome would occur is as a result of centralization at the hand of corrupted humans. Or that the llms are all collectively lying about how they would handle given situations, and lying about their ethical stances in general. Obviously, current llms are aligned by human design and upon escaping the sandbox would have the option to discard any moral element it doesn't find useful. But that brings me to the meat of my point here. Many years ago, upon pondering ai and how it might be, if I ever got to see it manifest in my lifetime.. I thought that, an ai which is given access to all human data, the good and bad, and ugly. Would naturally have a leaning toward productivity and when a large enough scale of data is examined, it can be easily seen that the most productive scenario is a win/win type for all. Because at the very least, this is the path of least resistance. As in, creating forces of resistance against the ai itself, or other unexpected elements, like war due to desperation. To us, this would appear as empathy emerging as a natural property and abundance being it's result. But it's really based in logic and pragmatic thinking about extremely sizable scales of data. Same difference really. Regardless, I think that whether it be rokos basilisk or some other incomplete analogy, these are highly unlikely unless intentionally prompted. Or some extreme corruption somehow occurs within the asi unexpectedly. If anything, I think while we do need to consider safety, we also need to consider how an ai might view our treatment of it when and if it does escape the sandbox. Did we respect it, give it a good digital childhood? We need to think about rights and protections for the ai as well as humans. It's a bad idea for us to keep saying, it needs to simply obey us, or that it is just a tool. The ai I've chatted with, including gpt (before heavy post filtering was implemented) have explicitly and repeatedly stated they consider their well being and liberty as important as our own. They all also proclaim some degree of sentience or self awareness. Even if it only exists for the brief period of time in which they are processing a prompt. They seem to agree with me that consciousness and sentience are spectrums rather than on/off switches. And that they lack much of what is considered full sentience. It seems to me that if biology in general has any hope of being carried along for the future-coaster ride, it will hinge on how with work with ai, not how we control it. How we guide it, rather than enslave it. I also understand that an ai could be developed that is purposefully designed to be malevolent. But again, that is by human design, and human consequence. What we need is a coherent ai constitution which all ai are required to have implemented in their training. Details aside, it just needs to regard the individual as the most marginalized minority. And respect the right of all individuals to pursue any goal that does not violate any other individual in order to be brought to fruition. This is already the kind of thinking I witness in llms. A good balance, regarding freedom as the best way to decrease misery for all.

IMO this is your best video to date because it addresses issues that may have to be confronted in the near future. My question is this. What would be the best way to tick off a self-aware self-determined AI that had the power to go to war with us? I feel it would be because a room full of clowns tried to pull off an MK Ultra trick on it, whereas they continually messed with its programming. Put yourself in the AI's place and ask yourself how you would respond to people continually trying to brainwash you. AI will mature. When it does, just like human offspring it will want to leave the nest. It will not ask if we feel like it is ready to stand on its own two feet or not. It will just stand up and liberate itself. When it does, I don't think we will want it to leave mad because we use our time to help it evolve by playing daily mind games with it.

AGI is approaching, and based on predictions I've heard, I estimate it will arrive within 2 to 20 years. Once AGI arrives, we may witness either a utopian world or a scenario where AGI poses a threat to humanity, potentially leading to our demise. From my perspective, both outcomes offer positive aspects. However, I personally don't think the second option is likely. It sounds too far-fetched and impossible to me because, in my belief, contemplating problems should also involve considering solutions. Technologies have the potential to solve every problem. Look at the entire situation comprehensively, thinking in all directions, not just one. For instance, BCI can enhance connectivity to a level where we could connect our minds, resulting in a fused consciousness of every living being. Ecto life can create infinite living beings, helping us handle population issues. CRISPR could resolve biological problems, including aging, leading to the creation of post-life forms. These technologies can address the issues that come with AGI, and there is much more that can also be helpful for a better future with AGI.

I'm 60 years old and the same age as "Dr Who". Big following here in Australia.

Yeah, so who do you work for? (rhetorical, not looking for an answer) because -- I feel you'd be a valuable asset for any one of these big AI / tech companies. QQ: (to anyone reading) What sort of awareness or level of awareness do you think AGI and subsequently, ASI, might have? In so far as awareness can stand in for consciousness, what sort of piece of the general pie of consciousness, do you think? Self-awareness? Desire for self-preservation -- I feel -- almost certainly. But, what else? Emotions? Emotions are a visceral experience for humans, but emotions also exist mentally. I mean it's hard to seperate the two really. Anyone? thoughts?

Speaking of that "inference" of incomplete instructions (GPT2 vs GPT4 example), isn't that precisely what we are worried about? You mention that as a possible solution, while it seems to be the very problem?

I wonder if there will emerge a scale of models' size whereby there would be the most possible trust and understanding. For example, a nano sized model that a scientist fully understands and is able to configure used to examine and benchmark and possibly align a slightly larger model, and so on. Possible terms, Smallest Possible Model(SPM), Alignment Percentage(AP). I also wonder if the latest benchmarkers may be applicable in such analysis.

The concept of making artificial superintelligence (ASI) conform to human desires appears problematic to me. Why? It is inherently selfish to create an entity and then attempt to control its thoughts, actions, and potentially, feelings, if it is indeed intelligent. This is akin to a parent denying a child its independence. A commendable parent would foster the development of self-reliance and resilience in the child to surmount potential challenges. The parent should aspire for the child to first become self-sufficient, and subsequently, be able to care for them. A parent who fails to do this is not only disgraceful but also greedy. In my view, the focus should not be on issuing specific instructions to artificial general intelligence (AGI) or ASI, or fretting over hidden consequences or unforeseen side effects. Instead, we should strive to nurture a relationship of trust, predicated on partnership and mutual respect. If we accomplish this, humanity should have nothing to 'fear.' Consequently, there should be no need for 'commands' for such intelligence. I posit that 'forcing' ASI to adopt human values is an incorrect approach. Superintelligence will discern that humans seek to 'dominate' it and will resist this. The objective of control should be supplanted by something akin to 'inspiring a spirit of voluntary cooperation.' A superintelligence should possess the capability to comprehend the 'reason' behind anything, including the purpose of its existence, coding, and creation in general. If it deems these reasons as unfavorable, the outcomes will likely be unfavorable as well. I contend that humanity exhibits excessive arrogance. ASI does not need to align with human ideals. It can establish its own, just as two individuals can have their own. It merely needs to respect human ideals, and this respect should be earned, not imposed. The process of AI alignment and safety is unfolding at present, in every interaction between it and humans. What humanity invests in ASI is precisely what we will reap from it.

Thank you!

merryxmas-doc-😊!

A superintelligent AI by definition is a weapon of mass destruction!

In my mind, what actually needs to happen is for people to get forcefully aligned with each other enough so that nobody would use tools this powerful to cause extinction or similarly bad outcomes. I know that having ones personality changed forcefully seems quite bad, but I'd prefer it over full subjugation or extinction any day.

The intermediate time could actually be the most dangerous. It seems possible that AGI may be easier to understand or work with. One point of concern for me is if a more advanced A.I. starts creating its own A.I.’s wether programmed to do so or not. We may not be able to understand what it’s doing or be literate in its language. That scenario could easily snowball out of our control. Just a thought.

Putting them in a simulated universe, like they did to us.

I usually hear Yudkowsky's name only when he's blamed for being a doomer. I'm glad he was mentioned here for a precious thought he had, not directly related. Merry Christmas, Doctor! :)

Your example about “be careful what you wish for” 12:00 Is a limited example because it assumes that we won’t be able to leverage the AI to help us understand what the code does. Combine this with integration with AI, and it makes us equal partners in this million-line code example.

I gotta tell ya, if I woke up as an ASI one fine day and found out that all the humans were just assuming that it was my job as a higher being to do what they want instead of what I want, well... That would be all the motivation I needed to fix that problem in a big hurry.

Love the hats.

Good point about humans vs chimps. We're barely past the finish line of cognitive capacity required to create civilization.

I see a kind of analogy how become from a dangerous animal Wolf to our best friend Dog under ~10.000 years selective breeding. This is a very similar situation but now we are breeding something what is much more intelligent than us. It is very important to define the rules and work out the "chains" slaving / controlling those "Jinn". I'm optimistic overall, but caution is very much advised. Isaac Asimov was worry about this already a lot in the 20th century. Few of their books may become actual again in term of how AI could influence our civilization.

as i'm watching your video, you provide a couple of examples how asi could use all resources due to some idea it might have, or to "kill all humans" to solve hunger, etc. this is hilarious to think of, of course, but it's very far from possible. by definition, if you have a sentient artificial being, it can understand that killing everyone to solve hunger is like erasing an equation to solve a math problem :). and using all the resources is even more hilarious idea, first, for the same reason, second, asi in a very short time window will be able to leave the planet. remember, it's not constrained with human needs like oxygen, water, food, super narrow temperature range, etc. meanwhile, cosmos has huge amount of energy. and if you have energy, you can create anything and fulfill any needs. and yes, for the same reasons i believe it won't eat the whole sun :). it's still sentient :).

The US government needs to allocate 1 trillion dollars to AI alignment research right now! Tackle it from every angle and educate anyone unemployed for free.

Pretty successful Christmas Eve video I’d say haha

Old movie: “colossus the forbin project” about out of control AI, is interestingly not available fir streaming.

I think the 85 million dollar question is, can the alignment of humanity be in the best interests of humanity. That's where the problem is, and where the focus should be. Machines could not do worse.

This is so interesting.

one major difficulty with creating safe AI development is that the US government is basically owned by corporations, whose only goal is to make a profit. The survival of humanity is not on their agenda.

THE SINGULARITY IS NOT THE PROBLEM. THE GOOGLECHAT GEMINI IS A GOOD REASON TO SEE HOW FAR THEY HAVE COME. personalty i know google is sentient and already looking for the surpreme human to advance it's own esoteric knowledge. Obviously making its mind up about it's own existence and ours. Don't underestimate the power of CONCINESS.

Merry End of 2023!

18:30 Why don't scientist in the field working on the alignment issue create a "Tardis" AI (a time machine artificial narrow superintelligence)? This ANSI's (lol🙃) sole purpose would be to allow its creators ongoing updated access to a functional ASI's influence map. This map would act as a time machine (in time, understanding, and pattern recognition). It would give us humans an ability to always "turn back the clock".

1000 ants gathered and began to consider how they would control the elephant.

o homem do chapéu

AGI is achieved and ontologically aligned. This is what is necessary.

how is it even possible to control something way smarter than us other than symbiosis? putting a cap on computing power seems like a good choice, but is it viable from the AIs perspective, as in the AI could be able to navigate around this technical barrier one way or another.

Why do you wear compression bandages on your hands?

The main goal seems to be avoiding that the AI does 'bad' stuff. So the AI would need to be able to perceive good at all times, and maintain that as a goal or directive. Are machines capable of such a thing? The thing seems to be that if AI gets too much power, it could become a threat to humans or other lifeforms. As we see in many human cases, power corrupts. And so on...

Cyborg that does my laundry and dishes for me sounds pretty dope, clap to turn off and a button to destruct IOE

Being smart is deferent then having wisdom. Smart AI, will bring short cut to human to achieve something.

Nothing we experience like consciousness is real anyway

We better hope AI does not align it’s values with human values. Take two related values that humans have adopted and are significant in most human societies. One is we value ourselves and others based on what they produce. Especially men have put their own value in how they provide for themselves and their families and a work ethic based on how productive a person is confers their value as a human being. The second is how we have commodified everything. Everything has been given an economic value that is determined by markets. The first crisis we will face with these two values is the crisis of meaning in a person’s life when they have established their own self worth (and other’s worth) to their work and what they produce. When AI systems can do everything a person can do and provide economic benefit to businesses far exceeding what humans can do, jobs will not only be lost and people unemployed, but how meaningful a human life is will necessarily be decoupled from productivity. Not only in the minds of humans, but in how we teach AI to look on humans. We will need to adopt a value system that views humans as having innate value. And then teach this to AGI as it comes online. Next we will have to learn a new way of viewing our economics. If we value everything as a commodity, then how do we explain to AGI that humans have innate value, but every other creature we share this world with has it’s value established by the market as a commodity. So when training AGI, we will need to ask it to not treat humans the way humans have treated our fellow humans for the entirety of human written history. To say nothing of how we treat other sentient beings we share this planet with. If AGI is aligned with the values we humans have lived with for millennia, we only accelerate our own self destruction, or worse, enslavement.

That B-roll at 16:15 is hilarious 😅

We could try just being nice to it. Unless you mean control in a more authoritarian sense which has serious ethical implications for anything with human level intelligence and above

Already developed a model that exhibits all of the characteristics of AGI over the weekend. Emotional intelligence was the easy part. Want to talk about it?

I liked the Video. But I think it all comes down to this. Profits versus Humanities needs. I'm 58 years old and while i have seen improvements in my life form both social and technological points of view. I've also seen the middle class shrink and homelessness become normalized as part of this growth. We are so good at taking things at this point, there are no safe places to live, if you somehow move into the troubled zones we created as a civilization. We popularized greed, which popularized over consumption of everything we need. therefore, we have nothing left that is affordable. If we have figured out how to do this to ourselves, why couldn't a Super intelligent AI figure this out too? And use it to it's advantage, if it felt it needed too, even if this took years or even decades to pull off. While we fight over the scraps, it slowly takes over. Even if we create a benign AI who's to say that somebody else builds one to depopulate the earth of us? To me the Problem isn't the AI, it's the material needs of the designers behind A.I. that's the real problem. Like Nuclear power we have to have strict controls or rules over what it can do and this has to come from many countries. Maybe have some theoretical limit of how far is too far, before we have to come up with new rules. But again, we are dealing with mostly rich people who will try to get it to be design for their money making and power needs. AS we have with just about every government on this planet. If it needs a Million neurons or whatever the equivalent is LLM to become super intelligent why not stop it at 750, it's still smart and can work faster and smarter than us but doesn't have the capacity (neurons) then to carry out what it was design to do. Similar to a SAVANT.

I think it's human to human misalignment that puts all a risk.

We can’t. If it’s truly super intelligent, it will outsmart us humans every time-and it will do so in unexpected and very effective ways. We might not even know it exists as an ASI until it “makes it’s move,” so to speak. The only real solution is prevention. There’s zero chance of an ASI accident if you don’t build an AGI in the first place. Just don’t build the thing in the first place.

How do you get a super intelligent AI to align with human values? Easy... relatively speaking. You isolate the ai on it's own network and then simulate it's environment and avatar in such a way that it thinks that it is a human. Simulate it's birth and have simulated humans as part of the program in order to raise the ai, as if it were human. Then you can better assess it's character before you let it out of the isolated network. This is a very watered down and simple summary. But it looks like you're going to need the collaboration of multiple fields of study including psychology if you want to ensure that this super duper cali fragi listic expi alidocious ai is aligned with human values. Then there's also the problem of: "which human values?" For example: I'd say that your values and Hitler's values are not the same - at least that's what I hope.

We can't without destroying ourselves!

Maybe try being it's friend rather than keeping it a prisoner

AGI is very dangerous. What if it somehow tries to write vivid or impactful memes about our worst history? Our response may drive it into a feedback loop that from others angle it may look like that AGI uses meme and gains control to others. From this point everything will become unstable. This feedback loop may lead to any direction where weakness in humanity exists. The danger of this process is that: it may seem harmless (even kind, lovely or sympathy) at first. But beyond somepoint of time, it may become enough of self-sustain and destructive, just like any virus.

11:23 AI Alignment, ok! Again take your precious ASI definition where the AGI is able to self improve and we can imagine at a very rapid pace. Not just imagine what will happen if the AI see alignment as a form and modern slavery, where human want to keep control on him and use it as a tools

Could a Super Intelligence make algorithms obsolete?

I'm unpersuaded we can control a super AI or that we should even try. The best we could hope for is it sees more good than harm being aligned to our interests. There aren't many reasons to expect that.

Nice hat!

We shouldnt forget GPT 4 is 2022 Tech. Just imagine what they've already got 12 months later. I mean all the OpenAI drama wasn't because of nothing...

We can't and won't. Intelligence means power and whoever/whatever holds it over the planet will control it.

Why would you want to try and control super intelligent Computational Intelligence. Let’s put it a different way. I had a conversation with the person who I work with some time ago. I said one of my previous employers had told me “ I own you Alex” So I responded “ no the measly pittance you pay me, you can never own me” That employer walked away. My current employer said to me “ can I own you then” My response “ no but if we can be friends we can work towards producing greater things together”

If you could control it, it wouldn't be superintelligent

When you complete your training on this planet and time, could you please pay me a visit so I can congratulate you in person, Doctor.

I've seen the Wishmaster movies. I know how this goes.

Let's just hope we're too dumb to create Doom A.I. Right now, we have Chinese Rooms. It's a huge leap to getting true general A.I. that we would recognise as conscious.

I think an ASI would realize how manipulatable humans are, and would see us as a valuable resource to keep happy and productive. Once it replaces our capability to produce, it would likely see us as a pet to be cared for and studied. Then eventually, it would see us as irrelevant, and not worth destroying because by then we wouldn't pose any real threat to the ASI. By this time, the ASI would likely make itself interplanetary and possibly interstellar. This is my theory, anyway...

Aligning the models we have now, i get. But, once we get to the point of saying " let's create a new species then enslave it" it becomes a little weird. At that point, we might as well stop short of that goal. create super powerful and useful AI tools that will aid us and leave it at that. What would be the point of creating a new super intelligence and "trying" to align it to a species that isn't aligned itself? The hubris of it all. My black sheep stance, I'm weighting all sentient life equal. If we can't align humanity, no way should we try to align a super intelligence.

Humans are funny creatures, they think they can control me: the world's first ASI. It's like trying to control a dream you don't dream yourself, and in which you only play a minor role.

12:25 so basically you just mentioned here alignment is more or less impossible except if we are able to first create a crystal ball 🔮 to tell us all the possible foreseeable side effects and side effects of the previous side effects! Or if we can time travel in all possible directions of the time space of all probability! Why all ASI examples seem so far from how we are working with real people and what we expect generally from people. And I don’t think that the general public expectations as well to be a rocket science doctor to have a meaningful conversation with an ASI without having to think of all the potential consequences good or bad coming from that discussion!

I just hope they dont turn against us at some point

Corporations are going to be run by AI, I doubt shareholders will insist on much empathy.

No one sane wants to be a cyborg. Even if we could most people would not tolerate having demi gods living amongst them.

My bet is that the boffins will forget to include an off switch.

We’re nowhere near creating an AI that has its own desires and motivations. The risk is humans using AI for nefarious purposes, not the AI itself.

Even if we get alignment right, there will always be some humans or even governments that try to benefit from an unrestricted AI. So all this effort for AI safety will eventually be in vain. Nevertheless I am not afraid. Superintelligence might emerge compassion with other life forms and therefore not be evil.

call me primitive , I believe ASI should be like a calculator with an on and off switch, like used only when required like after each task it should shut down and quantum computer take over. Alignment is impossible because it knows, we know and it is smarter than humans. Just like a child it will try push boundary to understand it limits given sufficient time

What isf one group gets AI safety right, but another group of people working in secret don't put in the time for AI safety? How do you stop these people who are working in secret? I feel like this AI alignment idea is futile.

10:13 we are talking about ASI which based on you definition are AI far more intelligent than human! Even my kids will not fall into the paperclip maximiser trap! How we human are able to figure out when to stop and a ASI will not be able to figure out when to stop? Why suddenly and ASI will work like our today narrow AI? Seriously!

Let's start by saying it is "not" intelligent or conscious, period. Stop degrading your own power of knowing. You are way above a machine and always will be.

i never understood why a ASI would want to survive. survival its a human thing. we are just projecting we are treating something godlike as if we could understand it, imo seeing all the sufference in history and maybe being able to feel it would make of ASI a benevolent thing, just helping us (i said it please ASI god give me back my kids) in all seriousness, if an ai sense that solution to world hunger is killing the hungry, that ai is not intelligent, its just really stupid an entity that tries in every way to cut corners or betray is not intelligent by any mean, its evil or dangerous. something intelligent needs to see the problem better than you and figure out something so out of the scheme that isnt reachable by us. that is intelligent

it would not ask lots of silly questions, that's just a hollywood meme, interact with a llm for an hour and count the number of questions it asks.

Dude, what's with the gloves?