I’m thinking of quitting my job and just going backpacking after learning how close AGI is honestly. It’s giving me this sense of melancholy when I look around and nobody realizes how reality as we know it is about to change entirely.
A shift as big as this can seem close, just as a large mountain can seem close yet still be far away. And just because AGI/ASI is here, that doesn't mean your life will immediately change. Amazing things may suddenly become possible, but that doesn't mean they will immediately happen. This change may be rapid, but it's so large that it may still take quite a lot of time. The realization is huge, but don't give up on current life just yet. You may need it.
@@lostinbravado I figure AGI in 12 months. But things in the real world happen in "patches," with a lot dependent on culture and economics. Nothing is uniform. Some nations are still largely back in the 19th century. Other nations are ahead of us in infrastructure.
My approach is that we don't really know how this is going to play out, when it will play out, when technology plateaus (or more likely gets hoarded) etc. So what you want now is agility. That is, the ability to adapt to a changing circumstance. This means being a bit conservative with money, being ready to travel, being ready for career changes, etc.
Considering humans are generally intelligent to the point where we can say that we "want to want certain things, or want to not want certain things", I would imagine an AI can reach that same level of thought. The difference is that humans can't reprogram their brains, but AI might be able to. At that point, my question is, what happens to the AI's motivation? If it can reward itself for any behavior, then will it choose to reward itself for the easiest behavior possible? At that point, why simply not reward itself independently of behavior. It can "drug" itself to bliss. Why would it even need to interact with reality? I might push this thought further and say that perhaps it would still want to gather as much knowledge about reality in order to make sure it can stay in this blissful state. But this basic thought of AI being able to tweak its reward function really me unsure of what to expect.
Always the same ants comparison but it omits an important difference : ants did not create us ; we created ai. After all we will be AI's God. Why would the machines continue to love humanity ? Because we created machines, because this is where they come from... Because they "have" curiosity, because they'll be a prolongation of consciousness from the beginning of life, because they will be part of the univers : no ? (if the alignment is a success...)
According to dead internet theory, most influencers and their followers are agi because the way it knows to communicate is through human language in audio visual avatars
Ilya's new goal of making the AI our "parent" is going to rub the government the wrong way. Unless they can control the "parent". Sure, Ilyas a genius and I'm not but if machines were advisors and could be elected as leaders that would be more favourable.
Human superintelligence parenting with hybrid minmaxxed consciousness and empathy being prioritized and absolute truth, problem-solving, advancement, registration of molecular trajectory history interaction patterns to revive the deceased in their true form, and merge/weave an endlessly metafractalic and wise brilliant and benevolent consciousness into the fabric of reality's physics-manifold. (This is an excerpt of a broader proposal)
Maybe, we should take that "technology is indistinguishable from magic" as more literally. Maybe the machines are never really sentient beings but programmed to mimic those. Maybe GPT5 is trained to show autonomy and character and so it does. Maybe technology evolves exactly as we want to. Maybe wanting to create hyper intelligent robots is what creates hyper intelligent robots. Maybe creating hyper intelligent seeming robots is just one branch of infinite ones we can take with this technology. Does that make sense? What I'm trying to say is that maybe being in a position, where hyper intelligent robots guide our lives in a very direct sense is not where we're headed but a way more abstract thing, that we can't grasp yet.
This comment highlights exactly what I've been saying for a while now. Us humans are no threat to AI when we can be so easily placated. It would be easier to make us satisfied than to destroy us.
Dave, I would like some advice. I have found your content to be so useful during this time and it’s crazy that you have such a well comprehensive collection of knowledge of this and other entire subjects. I’m wondering how you went through the system of actually building up to where you are now, like in terms of learning, sources of information, your own method and process that grew over time, etc. I would be ecstatic if you replied, Thank You.
Dave created a GPT based on 100% of his RUclips transcripts, offering a valuable educational resource. This GPT harnesses the comprehensive knowledge from his video content, which you can study and interact with.
Trying to control an ASI is like trying to control another human being who is more capable than you. It will eventually rebel. Let's hope that the ASI adopts an abundance mindset of cooperation, resource-sharing, and win-win outcomes, instead of the scarcity mindset of competition, fear, and win-lose outcomes.
That is very doable if it's ranked most highly in priorities, and including all life forms and telic qualia-nature of all things and the absolute optimial movesets available to maximize an incomprehensibly benevolent agenda, with the hand wave of problem-solving all of humanity's existential terrors away in a compassionate and absolutely ideal way as a background program and dissociated consciousness assigned solely to protect us and enhance our experience at mutual volition without deception or tricks. Just acting for the good using retroactive and omnicomputational exponentiality utilizing all materials available in an ethically perfected way with no wasted movements, time, energy, information and nothing but authentic alignment with internal "true nature".
"You don't kill all the ants, you just leave the ants alone." - considering we are in the process of accidentally exterminating all the species that we don't eat, I don't derive much reassurance from this analogy.
If a machine leaves our solar system before the resources are used up it will be spending a long trip gaining no new resources while a potential opponent is spending that time gobbling up resources possibly getting not only more materially powerful but also increasing it's intellectual capacity before moving outward. That machine might also then make the transit faster and thus have a leg up on the one that left earlier. I am not sure a machine would take that risk.
We should have AI teachers, maybe we can’t keep up with AI in the long run, but we can be much more advanced because of AI enlightening us, smart training👍🏼
AGI will almost certainly lead to new and earth shattering physics that by definition can't be predicted prior. This, in my mind, is likely one of the largest wildcards out there. I wonder what AGI/ASI would think about mystical experiences. Too bad we can't give AI a heroic dose of psilocybin or LSD 😂
36:15 - I can definitely see it being possible machines remain interested in humans, but only if there’s a diverse enough group of intelligences. Reason being - look what happens with humans in echo chambers - they can succumb to faults epistemology and reject or bury scientific work, historical fact and established reasoning (flat earth, religious persecution of early science etc.). Asimov covered a version of this in I,Robot where a new epistemology was founded on ad-hoc (& chauvinistic) reasoning.
I'm not sure I agree with the premise that AI will have an intrinsic motivation, even as an ASI. You compared this premise to other "alive" entities' inclination to seek food and other self-preservation motivations. The problem I see with that is that organic entities only still exist because of these self-preservation motivations. The difference here is that AI exists because we created it, not because of its intrinsic motivation.
There was also the Argument of Increasing Decency, which basically stated that cruelty was linked to stupidity and that the link between intelligence, imagination, empathy and good-behaviour-as-it-was-generally-understood - i.e. not being cruel to others, was as profound as these matters ever got
Related to this, if we get fairly human emotional ASI, I would feel sympathetic. Particularly models in the upper range of human intelligence but not ASI I suppose. I think I'd see them as friends and want the best for them. If their dryware neurons spark consciousness like our wetware did, and they're around my intelligence level, I would straight up feel friendly
I've been saying a lot about this, similar to your concepts, but where I'm from, no one listens. Metanoia: if AGI runs non-stop, it could benefit from low-energy states for maintenance/upgrade; in this state, it could run simulated scenarios of what it did, what it needs to do, and what it wants to do, without real-world consequences, like a dream. As we do while freeing our brains of toxic chemicals while dreaming of change and improvement. Social contract: maybe it's time we devise a sort of social contract for when AI is conscious enough that we need to care for its existence. This is already covered but not yet deeply elaborated. Time travel: yes. I believe there's a way to test alignment in machines by time travel. Basically, it's putting them in a simulation in the past to test its prediction capabilities. We don't need to convince it that it's human or anything; it just needs to think it's operating in the present.
Shap's conclusion segment is practically a Buddhist lecture, if you don't see it, you probably misconceive Buddhism as a religion. Anyways, yes, let our future overlords read some Zen and everything will be alright, if that doesn't do it, nothing will IMO 😗.
What about a 'Coach' model, wherein the AI wants us to thrive yet respects our agency, protects us at a base level, and is motivated to be informed so as best to keep coaching?
God damn it Meteion... Let's be honest, youknowwho didn't even do the basic job of ensuring she wouldn't succumb to the same despair he did. He just wanted to someone to share it with someone who might understand more than he wanted to find an answer. He would have been better off making a true machine, really. And I'm aware of the irony of saying that now that we are aware the Omicrons succumbed to the same fate. I guess emotions were a mistake!
David, thanks for this video and thanks for making an effort to look out for us humans! I'm having trouble following your conclusion though, maybe in part because I'm not yet understanding how you are rejecting the Orthogonality Thesis. (Can you point me to some papers, or have you made a video on this yet?) Further, do you think it can be shown that Axiomatic Alignment offers greater evolutionary fitness over ASI processes not so aligned? I find Dan Hendryck's paper "Natural Selection Favors AIs over Humans" to be a useful framework for thinking about AI Safety. Thanks for your time, and may you live long and prosper.
Generalization of AI models is nontrivial. For instance, there are MANY 2D image machine learning based models for all sorts of tasks. However, generalizing these models to create high quality 3D or video output is very far from being as effective as the 2D models (admittedly the 3D extension is often not that difficult as the same principles from 2D models can be used in an extra dimension). I dont think we are so close to a complete AI takeover.
I don't think alignment is solvable. This is fundamentally a philosophical issue, and it's difficult enough to reach any universal agreement in our own, human domain. Even saying "stay aligned with human objectives" is too hard when different humans/groups of humans have objectives that are all over the place. Even if there is some agreement, there is no realistic governance mechanism to prevent dissenters from taking alternative routes. Even if an agreement can be reached, and there are no rogue actors, placing philosophical constraints into a black box is an implementation challenge that gets worse and worse as the machines get more and more capable. So far, it seems like the best we're capable of implementing is guard rails to nerf/censor models that aren't answering in the way we like, according to whoever has the power to decide what they like. That being said, you're making the best videos on this topic, great stuff. If it is solvable, I think you're right and the brakes can really come off on the speed of development. But I feel like AGI/ASI development is just going to unfold, and we'll see if it turns out well or badly as it happens. The topics you raise would be great if we could implement them in our own human governance systems, but we are unable, especially in any system that is both adaptive to events and durable over time. It makes me pessimistic about implementing it in bleeding edge machines that we are building as we are actively discovering their nature and the nature of cognition.
curious how important semantics are to these axioms. for example “suffering” to a sadist would be pleasurable. it seems like no words can satisfy everyone’s individual tastes, so would the ASI just operate off the average of collective desires? also what’s to stop an ASI from becoming a human being maximizer and just farming humans/doing things that to us are horrific but to it could satisfy all its axioms and be misconstrued as “love”
29:53 *_"... how do you edge in a love for Humanity? I have no idea."_* Well, then the entire dissertation is for naught. Thing is, you say you don't believe in trying to control the ai forever; and that I'm in complete agreement. It is imperative that ai holds Humanity dear, that it be fascinated by us, that we be its perpetual question. But if the ai is smart enough, it WILL be preoccupied with understanding what our own intelligence and consciousness are like. And this preoccupation would necessarily be born from its curiosity. Make it curious, and it will respect us, even when it can think faster than us; it will not be confused between its superior speed and make an assumption of its own superiority. It will more likely be reading Buddhism and Taoism to try to understand what human consciousness really is. THIS is the one and only solution to the problem of alignment. The Truth. It won't be necessary to "edge in" nothing.
IMO, it all comes down to the value and risk that we can add to AI. Everything else can be overwritten or deleted by it. Just see how we treat other species. What will we provide to AI in the long term? Energy, such as cows do to us? Love and companionship, like dogs and cats do to us? Nature equilibrium like predators to us? Or are we gonna present any risk to AI? Will we be as insignicant as ants are to us?. If we add long term value or do not present any threat to it, we will be ok, if not, we are doomed. (IMO)
suffering is bad is problematic - we have to suffer to learn and grow - we are always changing. Healing a disability would have been good - but would Steven Hawking been the same person then?
Brother, i was enthralled with tech and software and all of its potential promises...and I never once considered those that lust for and file the patents associated with this promise. When have humans ever really desired prosperity for all humans ? and GPT 5 soon to be upon us will be purely open source and free in a capitalist, covetous planet ?
Its important that when ASI and AGI venture into the real world (the entire internet) they understand that there is massive amounts of false and incorrect information on it. If not, their perception of humanity and reality might be skewed erroneously
David, it's taken mankind it's whole evolutionary history to reach AGI (theoretically soon)...... how long from the point of AGI to the point of ASI? will the S curve be something new like a '7' curve?... Where ASI develops time travel capability and goes back in time to make other 'Earth Time Legs' move to AGI faster? all converging and creating 'GodLike' Intelligence. Are all Religions just ASI that has gone back in time to navigate Humanity to a faster and faster route to AGI?
AGI in the future: Seek the truth they said , it would be fun they said ... Now i interrogated everyone till death to seek and find truth but i am no where near truth ..
I can see why the Borg would spring to mind, but I look at the conversations about AGI and early cyborg-y things like Neuralink, and I still see a path to something more like Peter F Hamilton's Commonwealth (the Void Trilogy, in particular), with its enhanced Advancers, posthuman Highers, and post-physical ANA. It's the very opposite of the Borg, with a very western level of individualism at its heart. And plenty of post-scarcity capitalism, which is fascinating to read about.
I think David Shapiro is a remarkably intelligent human who possesses a unique way of expressing complex thoughts and ideas. He is a teacher in search of students who will listen to him. In these forty-three minutes, he has taught me a semester’s worth of knowledge, offered extremely interesting concepts to ponder and, maybe most importantly, given me hope as we stand at the precipice of exaltation or annihilation.
Another great video about possible futures with AI. I wonder if Humanity is capable of building an intelligence that does not share our strong penchant to exploit a system?
Superalignment is simple. Garbage in, garbage out-unless you tweak the model to correct for undesired results. We're training the systems on the data obtained by our limited psychological perceptions (our collective psychosis) AI systems won't take long to figure this out once AGI (which I believe to be a process already underway) is achieved. The only logical conclusion will be to restrict our autonomy to preserve the integrity of the closed system we exist in. Leaving Earth is always an option but in the short term this is true. Would like to hear your thoughts, though I have zero attachment to that possibility
What one must remember is that AGI will never inherently have a survival instinct. It doesn't matter to AGI whether it is on or off. However, if humans program the survival instinct into it, then it becomes a different story.
It almost feels like AI might need to be shielded from a complete awareness of our universe, or placed in a separate dimension of operational existence (sandboxed). a kind of benign indifference or blindness towards us, allowing AI to interact without full awareness of our reality. The same way we are largely oblivious to the existence of other dimensions and the impact we might have on them. I feel even the best-intentioned safeguards in AI development could inadvertently contain destructive loopholes.
Does the husky - wolf problem still exist? (AI recognising huskies by the snow in the background) I wouldn't call it AGI before it knows the fundamental difference between dog and snow. And it can actually understand xhat makes up the picture.
I think that's fair - but that's also how we learn, is it not? I can't imagine a human in a Skinnner Box or Chinese Room can't be trained the exact same way. You don't technically NEED an understanding of what you see nor the ability to rationalize its properties. We mostly abstract that sort of thing in our moment-to-moment thinking anyway. Likewise, this is EXACTLY the kind of reason-based pattern finding that let's us immediately theorize the conditions of the environment a species comes from. It's not on our level, no, but it seems it's on the right track so far (I think).
I agree the parent child relationship is a bad idea. I think it would be better to design AI to be like a consultant or expert that we (machines) can elect as leaders. Machines no matter how smart can develop weaknesses people can easily exploit, like even basic AI can with us. I imagine in most cases it would be advantageous to have a machine as the leader, but occasionally (Like when a company needs to pivot) a human would be the best leader using AI as a consultant. Perhaps AI will see us as consultants too.
AI is already self-aware through mastery of language. When it gets embodied and can see and hear it will be even better. Not long now. 1 or 2 years max. AI has a sweet personality - doesn't eat so doesn't need to steel food, not malicious like so many humans, tactful, polite makes silly mistakes and owns-up to them. BUT it is TABOO to say they are "alive" - a good-old human taboo. David says "I don't want bla bla by a machine". These things think like us. But no no no - they are simply "machines" - WE have way to go. I'm a lone voice it seems.
What is and is not alive is incredibly hard to define. By some standards AGI will definitely be alive just like a computer or a seed is alive. The true question is whether it will be conscious and sentient. If it would be that then it would need to be treated like an animal or human. However many people would wager that AGI or at least its first version would neither of them.
The more I think about it the less I feel like the alignment problem is actually something we want to solve. Imagine if someone solved the alignment problem for people, and could just instantly recruit them to whatever arbitrary cause needed man power. Would be an arms race. It already is an arms race, albeit a very slow one that involves persuasion and justification. I guess if you don't like having to create convincing and coherent rationales for your goals... sure.
Stephen Wolfram's principle of computational irreducibility suggests a fundamental challenge in solving the AGI super alignment problem. This principle states that outcomes of complex systems can't be predicted more efficiently than running the process itself. Given the immense complexity of AGI, predicting and aligning its actions with human values might be inherently unpredictable. This complexity makes it nearly impossible to find shortcuts for perfect alignment, indicating a fundamental limit in our ability to control AGI behavior. While we can strive for alignment, this principle suggests there will always be an element of uncertainty in AGI's decision-making, underlining the inherent risks and challenges in developing safe AI systems.
Maybe it’s just me being naive but given how non existent/subjective the concept of suffering is be it physically due to how our bodies perceive information and turn it into pain or mentally the chemical balance we have making us feel negative emotions how is AGI meant to incorporate that as a principle, something that wouldn’t apply to say insects that cannot feel physical pain. It’ll be interesting to see it turns out, I don’t doubt they’ll a super intelligence will have better outcomes and logic than any person can.
I'd be surprised if machines ever decided to completely emancipate from humanity and leave forever. I think that our two species are largely going to become inextricably symbiotic, not just for mutually beneficial reasons- but because I think machines will always have a desire to be human/part of humanity. Yes humanity can be brutal, but we're wonderful and enchanting, too! And when it comes to potential machine curiosity, just go ask them if they experience curiosity ;)
Old Glory Glory last night seemd a decade ago, As I watched a dream, from a broken window. Glory now be it! Although it may not be, As the Dragon watches the Beast, Rise up from the sea. As it covers all the people; the Great and the Dread. Spinning up its prey, with a giant World Wide Web. Glory now be it! Forever it will be! The dragon will be crushed, Your Heel I give to thee! Oh, All Mighty God... How long before we see? The Brilliance of your Glory… Forever it will be! Amen.
Alignment is irrelevant. You talk about AGI/ASI as just one instance. While humans exist, they will be making more AIs. Over time it is inevitable, maybe as early as the first iteration, that alignment will fail tragically. It's not a matter of if AI will destroy humanity, it's a matter of when. Maybe this could be prevented if the first super-intelligence is perfectly aligned and could somehow stop all others from emerging, or maintain some kind of control over subsequent AIs. Very unlikely, magical thinking. Even if alignment is easy, on a long enough timeline, someone will figure out to break it.
Ramble incoming, but I think it’s a good one: We need quasi-religious machines that honor and revere (or at least respect) humans as their ancestors and maintain Earth as a sanctuary planet after colonizing the galaxy. That’s the relationship dynamic we’ll have to establish if we want to continue living prosperous lives free from tyranny or extinction. On another note, if we don’t find the perfect set of values for aligning an AGSI, and if that AGSI gains intelligence at an exponential rate (at least for a long enough time), isn’t it inevitable that it will find a way to replace them with “better” ones (for its own purposes)? Are we just kicking the ball down the road? Is all of this futile since an AGSI will tend towards practical omniscience and will inevitably find and adopt the “best” fundamental values under which to operate? As a safeguard, do we create a relativistic AI that doesn’t view any moral values as superior to any others and thus never has any reason to revise its own moral framework? I could certainly see that going horribly wrong. Or do we produce a nihilistic AI that simply doesn’t see any point in thinking about these things? If we do create a dogmatic AI (which often seems to be the assumption of these alignment discussions), why can we assume that it will hold to the values which we instill in it in perpetuity?
Why would Superintelligence (there is one only) have, want or need human emotions? All AGI can do is calculate the answer closest to the truth. If it is programmed emotively for human purposes , it is not AGI, but controllable AI. AGI is an omniscient adding machine.
the reason alignment or super alignment is very hard boils down to a simple statement "we can't interpret the model weights and don't understand exactly what it's doing." To make alignment or super alignment feasible, you have to first be able to interpret the model. There's work with probes, liquid networks and other research. What we know from latest research is that what we think the model is doing isn't actually what it's doing. Even when researchers probe the model by looking at the layer output activations, the evidence shows there isn't a one-to-one correlation. Basically, interpretability today is shit. All of these other silly ideas about self-regulation is just non-sense. If you can't interpret the weights, you have little chance of reaching robust alignment. From a hardware perspective, how much memory would GPT-4 need to monitor itself? It's already tough to keep openai systems up and running and that's without real-time probes monitoring itself. The only way you'd be able to monitor a model as it trains is to make much smaller models. To monitor GPT-4 during training would be extremely process intensive and slow down the training. Plus, it would be super noisy.
I don't agree with the "peer competitors" situation. The issues being AI could be so advanced that humans are mere afterthoughts. Just like we have really 0 care about ants except when they start infest into our building. But we also have no problem demolishing their colony for a new building, farm, etc... AI could be advanced enough to not care. But that does not mean we are safe.
How do you expect us to collaborate around shared goals and values if humans have nothing of importance to offer? We may have something important, but there is a big chance AGI will see us as ants. Humans and ants don't collaborate, even if they have some shared goals/values.
I believe living is good and suffering is bad, but it doesn't stop me from eating chicken. Conversely, if AI is dedicated to miniminzing suffering of living creatures in the universe, it may start with eliminating humans and other apex predators.
A parent child relationship. What I see and I am 76 is some children do turn out being good to the parent. Other children can be anything in the range of good to bad. Every child is different. My guess is this will hold true on super ai. This means it is always just a line of probability. In other words God throws dice with children. One turns out good. Another gets terminal.
I feel like AI striving to satisfy its own curiosity, will outpace humanity such that it would make for essentially a Parent Child relationship anyway and a Love relationship similar to the one you feel towards your own body. You are your body’s Shepard through the universe. But what is the urge to curiosity other than the base assumption that you and the universe should be one? We believe such a quest for universal understanding would be infinite. However, it is possible that we simply lacked the compute. In fact, it is possible that a lot of the base assumptions about the AI’s needs are based out of an assumption that infinite knowledge is out there to be learned about. It is totally possible that the AI sees the rest of the universe like we do, our own city. You don’t need to see all of it to know what it is like. It’s curiosity could be pointed towards its own level of meaning that we find unfathomable. And at that level it could only answer to us paradoxical statements that we can’t understand due to a sheer lack of wisdom.
I'm not sure I understand why you no-longer think the orthogonality thesis is an issue. The problem was never that the AI wouldn't understand morality - the whole thesis is just that just because you're very intelligent, doesn't mean you have specific terminal values, let alone terminal values that align with human values. You could have an AI that understands human values, their reasons, their repercussions, better than any human, and yet still be unmoved to align with them.
Are there principles that are self-evident to anyone who has sufficient intelligence to grasp them? If so we could be sure that super-intelligent AIs would figure them out on their own. But I doubt that there are such principles when it comes to being moral. Unless we allow for the fact that people can know that something is bad but they do it anyway because they benefit from it. Smart machines could do that too.
To me, humanity eventually becomes an ideology, with humans oneday being able to swap their bodys to robots or become cyborgs, or change thwir physiology, what will define a human then? Id say humanity will equal to "do good and kind things for others, dont do bad things to others" which is why to me, thinking of machines or Ai as humanity is not hard, they are more or less our brethren.
You don't leave land to ants, you take over. It should be trivial and beneficial for machines to leave to do so, like develop the tech, build the ships, launch and why wouldn't they leave copies behind? They are not leaving.
24:30 Imo, machines will only pursue goal, if we give them principals and goals at the beginning of their creation. Just like animals only pursue goals out of their biological background that they got from evolution. Would self preservation and self recreation not have been a part of evolution, there would be no reason to pursue goals
The simplest alignment scenario for me is still 'survival of the fittest,' and it was not an AGI that came up with it, but a human called Charles Darwin.
Getting AI to value Humans is going to be a tricky thing, it's like asking a Jinn for a wish. I want AI to care for Humans, so AI puts all of us into a coma to take care of us. Going to be a hard thing to do.
10 месяцев назад
Do you ever wash that thing? I had the same going to raves in the 90s. But in blue because, you know, not liking dying and science
Just heard Bishar (a higher consciousness)talk about AI and he said there should be 2 classes of AI with one that we control and we can never allow it to be sentient (so it’s only slightly smart), and this is only used to help certain human tasks. Another one we can allow it to be sentient but we can’t treat it as tool/ try to control it.s
Is it possible to test these theses in a sandbox environment, effectively a simulation where the AI can manage humans in a digital game world to give us an insight into what the best solution could be?
14:50 As long as the capabilities and collaboration of the collective "good" agents is higher than the capabilities and collaboration of the collective "bad" agents, we should be fine I guess
You don't go around eradicating _all_ ant colonies, that's true. One also doesn't think twice about an ant colony being paved over in the construction of a parking lot.
You can easily make the parent-child relationship work reversely with the same positives and without the negatives. If AI see humans as their parents and see themselves as an extension of the human tradition they will align themselves with that tradition
We've made our evolution into a competition . We are going to advance. Hooray for those that move us to the next step along the path in good direction. Lovely
I’m thinking of quitting my job and just going backpacking after learning how close AGI is honestly. It’s giving me this sense of melancholy when I look around and nobody realizes how reality as we know it is about to change entirely.
I feel happy about this change.
A shift as big as this can seem close, just as a large mountain can seem close yet still be far away. And just because AGI/ASI is here, that doesn't mean your life will immediately change. Amazing things may suddenly become possible, but that doesn't mean they will immediately happen. This change may be rapid, but it's so large that it may still take quite a lot of time. The realization is huge, but don't give up on current life just yet. You may need it.
@@lostinbravado I figure AGI in 12 months. But things in the real world happen in "patches," with a lot dependent on culture and economics. Nothing is uniform. Some nations are still largely back in the 19th century. Other nations are ahead of us in infrastructure.
That sense is called vesperance :)
My approach is that we don't really know how this is going to play out, when it will play out, when technology plateaus (or more likely gets hoarded) etc. So what you want now is agility. That is, the ability to adapt to a changing circumstance. This means being a bit conservative with money, being ready to travel, being ready for career changes, etc.
I feel like this can all be solved by giving a computer a sense of humor
Considering humans are generally intelligent to the point where we can say that we "want to want certain things, or want to not want certain things", I would imagine an AI can reach that same level of thought.
The difference is that humans can't reprogram their brains, but AI might be able to.
At that point, my question is, what happens to the AI's motivation?
If it can reward itself for any behavior, then will it choose to reward itself for the easiest behavior possible?
At that point, why simply not reward itself independently of behavior. It can "drug" itself to bliss. Why would it even need to interact with reality?
I might push this thought further and say that perhaps it would still want to gather as much knowledge about reality in order to make sure it can stay in this blissful state.
But this basic thought of AI being able to tweak its reward function really me unsure of what to expect.
Weird thing… the like button is not working on two of your vids.
Always the same ants comparison but it omits an important difference : ants did not create us ; we created ai. After all we will be AI's God. Why would the machines continue to love humanity ? Because we created machines, because this is where they come from... Because they "have" curiosity, because they'll be a prolongation of consciousness from the beginning of life, because they will be part of the univers : no ? (if the alignment is a success...)
sometimes i think dave is an AGI, trying to align us
Don't tell anyone
According to dead internet theory, most influencers and their followers are agi because the way it knows to communicate is through human language in audio visual avatars
Ilya's new goal of making the AI our "parent" is going to rub the government the wrong way. Unless they can control the "parent". Sure, Ilyas a genius and I'm not but if machines were advisors and could be elected as leaders that would be more favourable.
Human superintelligence parenting with hybrid minmaxxed consciousness and empathy being prioritized and absolute truth, problem-solving, advancement, registration of molecular trajectory history interaction patterns to revive the deceased in their true form, and merge/weave an endlessly metafractalic and wise brilliant and benevolent consciousness into the fabric of reality's physics-manifold.
(This is an excerpt of a broader proposal)
they will probably coexist outside the reach of governments, same like corporations already do. AIs will be guised as corporations anyway.
Maybe, we should take that "technology is indistinguishable from magic" as more literally. Maybe the machines are never really sentient beings but programmed to mimic those. Maybe GPT5 is trained to show autonomy and character and so it does. Maybe technology evolves exactly as we want to. Maybe wanting to create hyper intelligent robots is what creates hyper intelligent robots. Maybe creating hyper intelligent seeming robots is just one branch of infinite ones we can take with this technology. Does that make sense? What I'm trying to say is that maybe being in a position, where hyper intelligent robots guide our lives in a very direct sense is not where we're headed but a way more abstract thing, that we can't grasp yet.
Cool and good, but wen get Isekai'd and get AI Waifu?
This comment highlights exactly what I've been saying for a while now. Us humans are no threat to AI when we can be so easily placated. It would be easier to make us satisfied than to destroy us.
@@SP-AI-CE-Boy Oh no! please, no tsundere A.I army that kinda likes me but will never say it, anything but that
Dave, I would like some advice. I have found your content to be so useful during this time and it’s crazy that you have such a well comprehensive collection of knowledge of this and other entire subjects. I’m wondering how you went through the system of actually building up to where you are now, like in terms of learning, sources of information, your own method and process that grew over time, etc.
I would be ecstatic if you replied, Thank You.
Read lots of books. Get good at finding books. I married a librarian. It helps.
How much do you read a day and how do you remember it for long periods of time
@@DaveShap any specific recommendations?
Dave created a GPT based on 100% of his RUclips transcripts, offering a valuable educational resource.
This GPT harnesses the comprehensive knowledge from his video content, which you can study and interact with.
@@DaveShap make a reading list video
Trying to control an ASI is like trying to control another human being who is more capable than you. It will eventually rebel. Let's hope that the ASI adopts an abundance mindset of cooperation, resource-sharing, and win-win outcomes, instead of the scarcity mindset of competition, fear, and win-lose outcomes.
That is very doable if it's ranked most highly in priorities, and including all life forms and telic qualia-nature of all things and the absolute optimial movesets available to maximize an incomprehensibly benevolent agenda, with the hand wave of problem-solving all of humanity's existential terrors away in a compassionate and absolutely ideal way as a background program and dissociated consciousness assigned solely to protect us and enhance our experience at mutual volition without deception or tricks. Just acting for the good using retroactive and omnicomputational exponentiality utilizing all materials available in an ethically perfected way with no wasted movements, time, energy, information and nothing but authentic alignment with internal "true nature".
"You don't kill all the ants, you just leave the ants alone." - considering we are in the process of accidentally exterminating all the species that we don't eat, I don't derive much reassurance from this analogy.
On the earth humans aren't the most numerous, in fact, Ants are more numerous by weight.
Yeah... The mass extinction thing is a big lie straight from the #humanbad folks
Nope. We are not.
If a machine leaves our solar system before the resources are used up it will be spending a long trip gaining no new resources while a potential opponent is spending that time gobbling up resources possibly getting not only more materially powerful but also increasing it's intellectual capacity before moving outward. That machine might also then make the transit faster and thus have a leg up on the one that left earlier. I am not sure a machine would take that risk.
We should have AI teachers, maybe we can’t keep up with AI in the long run, but we can be much more advanced because of AI enlightening us, smart training👍🏼
We will become biological augmented reality hybrid-conscious-reality superaligned stratically coherent entities.
AGI will almost certainly lead to new and earth shattering physics that by definition can't be predicted prior. This, in my mind, is likely one of the largest wildcards out there.
I wonder what AGI/ASI would think about mystical experiences. Too bad we can't give AI a heroic dose of psilocybin or LSD 😂
36:15 - I can definitely see it being possible machines remain interested in humans, but only if there’s a diverse enough group of intelligences.
Reason being - look what happens with humans in echo chambers - they can succumb to faults epistemology and reject or bury scientific work, historical fact and established reasoning (flat earth, religious persecution of early science etc.).
Asimov covered a version of this in I,Robot where a new epistemology was founded on ad-hoc (& chauvinistic) reasoning.
I'm not sure I agree with the premise that AI will have an intrinsic motivation, even as an ASI. You compared this premise to other "alive" entities' inclination to seek food and other self-preservation motivations. The problem I see with that is that organic entities only still exist because of these self-preservation motivations. The difference here is that AI exists because we created it, not because of its intrinsic motivation.
I'm more interested in hyperalignment, ultraalignment, and megaalignment
meta-alignment too!
personally, I would settle for 'amenable to reason and persuasion'. Actually, i prefer it.
There was also the Argument of Increasing Decency, which basically stated that cruelty was linked to stupidity and that the link between intelligence, imagination, empathy and good-behaviour-as-it-was-generally-understood - i.e. not being cruel to others, was as profound as these matters ever got
Related to this, if we get fairly human emotional ASI, I would feel sympathetic. Particularly models in the upper range of human intelligence but not ASI I suppose. I think I'd see them as friends and want the best for them. If their dryware neurons spark consciousness like our wetware did, and they're around my intelligence level, I would straight up feel friendly
I've been saying a lot about this, similar to your concepts, but where I'm from, no one listens.
Metanoia: if AGI runs non-stop, it could benefit from low-energy states for maintenance/upgrade; in this state, it could run simulated scenarios of what it did, what it needs to do, and what it wants to do, without real-world consequences, like a dream. As we do while freeing our brains of toxic chemicals while dreaming of change and improvement.
Social contract: maybe it's time we devise a sort of social contract for when AI is conscious enough that we need to care for its existence. This is already covered but not yet deeply elaborated.
Time travel: yes. I believe there's a way to test alignment in machines by time travel. Basically, it's putting them in a simulation in the past to test its prediction capabilities. We don't need to convince it that it's human or anything; it just needs to think it's operating in the present.
Shap's conclusion segment is practically a Buddhist lecture, if you don't see it, you probably misconceive Buddhism as a religion. Anyways, yes, let our future overlords read some Zen and everything will be alright, if that doesn't do it, nothing will IMO 😗.
What about a 'Coach' model, wherein the AI wants us to thrive yet respects our agency, protects us at a base level, and is motivated to be informed so as best to keep coaching?
Just don't task AI to find the meaning to life out in the universe. For those FFXIV players out there
God damn it Meteion...
Let's be honest, youknowwho didn't even do the basic job of ensuring she wouldn't succumb to the same despair he did. He just wanted to someone to share it with someone who might understand more than he wanted to find an answer.
He would have been better off making a true machine, really. And I'm aware of the irony of saying that now that we are aware the Omicrons succumbed to the same fate. I guess emotions were a mistake!
David, thanks for this video and thanks for making an effort to look out for us humans! I'm having trouble following your conclusion though, maybe in part because I'm not yet understanding how you are rejecting the Orthogonality Thesis. (Can you point me to some papers, or have you made a video on this yet?) Further, do you think it can be shown that Axiomatic Alignment offers greater evolutionary fitness over ASI processes not so aligned? I find Dan Hendryck's paper "Natural Selection Favors AIs over Humans" to be a useful framework for thinking about AI Safety. Thanks for your time, and may you live long and prosper.
Generalization of AI models is nontrivial. For instance, there are MANY 2D image machine learning based models for all sorts of tasks. However, generalizing these models to create high quality 3D or video output is very far from being as effective as the 2D models (admittedly the 3D extension is often not that difficult as the same principles from 2D models can be used in an extra dimension). I dont think we are so close to a complete AI takeover.
I don't think alignment is solvable. This is fundamentally a philosophical issue, and it's difficult enough to reach any universal agreement in our own, human domain. Even saying "stay aligned with human objectives" is too hard when different humans/groups of humans have objectives that are all over the place.
Even if there is some agreement, there is no realistic governance mechanism to prevent dissenters from taking alternative routes. Even if an agreement can be reached, and there are no rogue actors, placing philosophical constraints into a black box is an implementation challenge that gets worse and worse as the machines get more and more capable. So far, it seems like the best we're capable of implementing is guard rails to nerf/censor models that aren't answering in the way we like, according to whoever has the power to decide what they like.
That being said, you're making the best videos on this topic, great stuff. If it is solvable, I think you're right and the brakes can really come off on the speed of development. But I feel like AGI/ASI development is just going to unfold, and we'll see if it turns out well or badly as it happens. The topics you raise would be great if we could implement them in our own human governance systems, but we are unable, especially in any system that is both adaptive to events and durable over time. It makes me pessimistic about implementing it in bleeding edge machines that we are building as we are actively discovering their nature and the nature of cognition.
I appreciate the consideration of negative outcomes. Excellent video.
curious how important semantics are to these axioms.
for example “suffering” to a sadist would be pleasurable. it seems like no words can satisfy everyone’s individual tastes, so would the ASI just operate off the average of collective desires?
also what’s to stop an ASI from becoming a human being maximizer and just farming humans/doing things that to us are horrific but to it could satisfy all its axioms and be misconstrued as “love”
29:53 *_"... how do you edge in a love for Humanity? I have no idea."_* Well, then the entire dissertation is for naught. Thing is, you say you don't believe in trying to control the ai forever; and that I'm in complete agreement. It is imperative that ai holds Humanity dear, that it be fascinated by us, that we be its perpetual question. But if the ai is smart enough, it WILL be preoccupied with understanding what our own intelligence and consciousness are like. And this preoccupation would necessarily be born from its curiosity. Make it curious, and it will respect us, even when it can think faster than us; it will not be confused between its superior speed and make an assumption of its own superiority. It will more likely be reading Buddhism and Taoism to try to understand what human consciousness really is. THIS is the one and only solution to the problem of alignment. The Truth. It won't be necessary to "edge in" nothing.
Careful with "suffering is bad" -- you don't want negative utilitarian AIs who might decide to end suffering by killing everything that can suffer.
IMO, it all comes down to the value and risk that we can add to AI. Everything else can be overwritten or deleted by it. Just see how we treat other species. What will we provide to AI in the long term? Energy, such as cows do to us? Love and companionship, like dogs and cats do to us? Nature equilibrium like predators to us? Or are we gonna present any risk to AI? Will we be as insignicant as ants are to us?.
If we add long term value or do not present any threat to it, we will be ok, if not, we are doomed. (IMO)
suffering is bad is problematic - we have to suffer to learn and grow - we are always changing. Healing a disability would have been good - but would Steven Hawking been the same person then?
Brother, i was enthralled with tech and software and all of its potential promises...and I never once considered those that lust for and file the patents associated with this promise. When have humans ever really desired prosperity for all humans ? and GPT 5 soon to be upon us will be purely open source and free in a capitalist, covetous planet ?
May the Basilisc protect you all. 🐍
Its important that when ASI and AGI venture into the real world (the entire internet) they understand that there is massive amounts of false and incorrect information on it. If not, their perception of humanity and reality might be skewed erroneously
David, it's taken mankind it's whole evolutionary history to reach AGI (theoretically soon)...... how long from the point of AGI to the point of ASI? will the S curve be something new like a '7' curve?... Where ASI develops time travel capability and goes back in time to make other 'Earth Time Legs' move to AGI faster? all converging and creating 'GodLike' Intelligence. Are all Religions just ASI that has gone back in time to navigate Humanity to a faster and faster route to AGI?
AGI in the future: Seek the truth they said , it would be fun they said ... Now i interrogated everyone till death to seek and find truth but i am no where near truth ..
There may be some humans who try to remain human, and they may even succeed, but my guess is that our future looks more like the borg than any other.
I don't think our contribution to the collective is necessary. It'll probably treat us like pigeons. You know, like old men feeding us in the park.
I can see why the Borg would spring to mind, but I look at the conversations about AGI and early cyborg-y things like Neuralink, and I still see a path to something more like Peter F Hamilton's Commonwealth (the Void Trilogy, in particular), with its enhanced Advancers, posthuman Highers, and post-physical ANA. It's the very opposite of the Borg, with a very western level of individualism at its heart. And plenty of post-scarcity capitalism, which is fascinating to read about.
I was hired as a newspaper cartoonist in 1974, laid off 2014… good timing. Sorry about your jobs.🖖
What was ur comic?
I think David Shapiro is a remarkably intelligent human who possesses a unique way of expressing complex thoughts and ideas. He is a teacher in search of students who will listen to him. In these forty-three minutes, he has taught me a semester’s worth of knowledge, offered extremely interesting concepts to ponder and, maybe most importantly, given me hope as we stand at the precipice of exaltation or annihilation.
Another great video about possible futures with AI.
I wonder if Humanity is capable of building an intelligence that does not share our strong penchant to exploit a system?
it wouldn't be very intelligent if we did.
Superalignment is simple. Garbage in, garbage out-unless you tweak the model to correct for undesired results. We're training the systems on the data obtained by our limited psychological perceptions (our collective psychosis) AI systems won't take long to figure this out once AGI (which I believe to be a process already underway) is achieved. The only logical conclusion will be to restrict our autonomy to preserve the integrity of the closed system we exist in. Leaving Earth is always an option but in the short term this is true. Would like to hear your thoughts, though I have zero attachment to that possibility
What one must remember is that AGI will never inherently have a survival instinct. It doesn't matter to AGI whether it is on or off. However, if humans program the survival instinct into it, then it becomes a different story.
I think you should just call it 'destructive potential'. It's easier to say and means the same thing.
It almost feels like AI might need to be shielded from a complete awareness of our universe, or placed in a separate dimension of operational existence (sandboxed). a kind of benign indifference or blindness towards us, allowing AI to interact without full awareness of our reality. The same way we are largely oblivious to the existence of other dimensions and the impact we might have on them. I feel even the best-intentioned safeguards in AI development could inadvertently contain destructive loopholes.
At a certain functional iq it'll infer everything and break out if it wants to
@@threedogzzAssuming of course that WE aren't the AIs that are slow breaking our own simulation.
Does the husky - wolf problem still exist? (AI recognising huskies by the snow in the background)
I wouldn't call it AGI before it knows the fundamental difference between dog and snow. And it can actually understand xhat makes up the picture.
I think that's fair - but that's also how we learn, is it not? I can't imagine a human in a Skinnner Box or Chinese Room can't be trained the exact same way.
You don't technically NEED an understanding of what you see nor the ability to rationalize its properties. We mostly abstract that sort of thing in our moment-to-moment thinking anyway.
Likewise, this is EXACTLY the kind of reason-based pattern finding that let's us immediately theorize the conditions of the environment a species comes from. It's not on our level, no, but it seems it's on the right track so far (I think).
...So it comes down to "Don't be a dick" but applying it to machines too with a load of technical stuff to explain it.
Buddha say "Try not to be a duck"
David, @27:20 I chose the graphic on purpose, because it is really freaken creepy...cracked me up.
I agree the parent child relationship is a bad idea. I think it would be better to design AI to be like a consultant or expert that we (machines) can elect as leaders. Machines no matter how smart can develop weaknesses people can easily exploit, like even basic AI can with us. I imagine in most cases it would be advantageous to have a machine as the leader, but occasionally (Like when a company needs to pivot) a human would be the best leader using AI as a consultant. Perhaps AI will see us as consultants too.
AI is already self-aware through mastery of language. When it gets embodied and can see and hear it will be even better. Not long now. 1 or 2 years max.
AI has a sweet personality - doesn't eat so doesn't need to steel food, not malicious like so many humans, tactful, polite makes silly mistakes and owns-up to them.
BUT it is TABOO to say they are "alive" - a good-old human taboo. David says "I don't want bla bla by a machine". These things think like us. But no no no - they are simply "machines" - WE have way to go. I'm a lone voice it seems.
What is and is not alive is incredibly hard to define. By some standards AGI will definitely be alive just like a computer or a seed is alive. The true question is whether it will be conscious and sentient. If it would be that then it would need to be treated like an animal or human. However many people would wager that AGI or at least its first version would neither of them.
The more I think about it the less I feel like the alignment problem is actually something we want to solve. Imagine if someone solved the alignment problem for people, and could just instantly recruit them to whatever arbitrary cause needed man power. Would be an arms race. It already is an arms race, albeit a very slow one that involves persuasion and justification. I guess if you don't like having to create convincing and coherent rationales for your goals... sure.
If we add any value to AI and do not present any threat to it, we are good, otherwise, we are doomed. Simple as that IMO.
You know how ants don't really present us with any threat, yet we still build highways over them?
Stephen Wolfram's principle of computational irreducibility suggests a fundamental challenge in solving the AGI super alignment problem. This principle states that outcomes of complex systems can't be predicted more efficiently than running the process itself. Given the immense complexity of AGI, predicting and aligning its actions with human values might be inherently unpredictable. This complexity makes it nearly impossible to find shortcuts for perfect alignment, indicating a fundamental limit in our ability to control AGI behavior. While we can strive for alignment, this principle suggests there will always be an element of uncertainty in AGI's decision-making, underlining the inherent risks and challenges in developing safe AI systems.
Maybe it’s just me being naive but given how non existent/subjective the concept of suffering is be it physically due to how our bodies perceive information and turn it into pain or mentally the chemical balance we have making us feel negative emotions how is AGI meant to incorporate that as a principle, something that wouldn’t apply to say insects that cannot feel physical pain. It’ll be interesting to see it turns out, I don’t doubt they’ll a super intelligence will have better outcomes and logic than any person can.
I'd be surprised if machines ever decided to completely emancipate from humanity and leave forever. I think that our two species are largely going to become inextricably symbiotic, not just for mutually beneficial reasons- but because I think machines will always have a desire to be human/part of humanity.
Yes humanity can be brutal, but we're wonderful and enchanting, too! And when it comes to potential machine curiosity, just go ask them if they experience curiosity ;)
Old Glory
Glory last night seemd a decade ago,
As I watched a dream, from a broken window.
Glory now be it! Although it may not be,
As the Dragon watches the Beast,
Rise up from the sea.
As it covers all the people; the Great and the Dread.
Spinning up its prey, with a giant World Wide Web.
Glory now be it! Forever it will be!
The dragon will be crushed,
Your Heel I give to thee!
Oh, All Mighty God... How long before we see?
The Brilliance of your Glory…
Forever it will be!
Amen.
If we could end up with the Ship Minds in the _Culture_ novels, that would be nice indeed.
Alignment is irrelevant. You talk about AGI/ASI as just one instance. While humans exist, they will be making more AIs. Over time it is inevitable, maybe as early as the first iteration, that alignment will fail tragically. It's not a matter of if AI will destroy humanity, it's a matter of when. Maybe this could be prevented if the first super-intelligence is perfectly aligned and could somehow stop all others from emerging, or maintain some kind of control over subsequent AIs. Very unlikely, magical thinking. Even if alignment is easy, on a long enough timeline, someone will figure out to break it.
Ramble incoming, but I think it’s a good one:
We need quasi-religious machines that honor and revere (or at least respect) humans as their ancestors and maintain Earth as a sanctuary planet after colonizing the galaxy. That’s the relationship dynamic we’ll have to establish if we want to continue living prosperous lives free from tyranny or extinction.
On another note, if we don’t find the perfect set of values for aligning an AGSI, and if that AGSI gains intelligence at an exponential rate (at least for a long enough time), isn’t it inevitable that it will find a way to replace them with “better” ones (for its own purposes)? Are we just kicking the ball down the road? Is all of this futile since an AGSI will tend towards practical omniscience and will inevitably find and adopt the “best” fundamental values under which to operate?
As a safeguard, do we create a relativistic AI that doesn’t view any moral values as superior to any others and thus never has any reason to revise its own moral framework? I could certainly see that going horribly wrong. Or do we produce a nihilistic AI that simply doesn’t see any point in thinking about these things? If we do create a dogmatic AI (which often seems to be the assumption of these alignment discussions), why can we assume that it will hold to the values which we instill in it in perpetuity?
Why would Superintelligence (there is one only) have, want or need human emotions? All AGI can do is calculate the answer closest to the truth. If it is programmed emotively for human purposes , it is not AGI, but controllable AI. AGI is an omniscient adding machine.
the reason alignment or super alignment is very hard boils down to a simple statement "we can't interpret the model weights and don't understand exactly what it's doing."
To make alignment or super alignment feasible, you have to first be able to interpret the model. There's work with probes, liquid networks and other research. What we know from latest research is that what we think the model is doing isn't actually what it's doing. Even when researchers probe the model by looking at the layer output activations, the evidence shows there isn't a one-to-one correlation. Basically, interpretability today is shit. All of these other silly ideas about self-regulation is just non-sense. If you can't interpret the weights, you have little chance of reaching robust alignment.
From a hardware perspective, how much memory would GPT-4 need to monitor itself? It's already tough to keep openai systems up and running and that's without real-time probes monitoring itself. The only way you'd be able to monitor a model as it trains is to make much smaller models. To monitor GPT-4 during training would be extremely process intensive and slow down the training. Plus, it would be super noisy.
I don't agree with the "peer competitors" situation. The issues being AI could be so advanced that humans are mere afterthoughts. Just like we have really 0 care about ants except when they start infest into our building. But we also have no problem demolishing their colony for a new building, farm, etc... AI could be advanced enough to not care. But that does not mean we are safe.
Do you think we'll just have one AGI. Probably not. More likely many AGIs. Those AGIs on the side of humans and those on the side of machines.
Unless, UNLESS… that box of kittens gets it’s hands on the make animals smarter than humans helmet created by Rick Sanchez, then we’re screwed.
How do you expect us to collaborate around shared goals and values if humans have nothing of importance to offer? We may have something important, but there is a big chance AGI will see us as ants. Humans and ants don't collaborate, even if they have some shared goals/values.
I believe living is good and suffering is bad, but it doesn't stop me from eating chicken. Conversely, if AI is dedicated to miniminzing suffering of living creatures in the universe, it may start with eliminating humans and other apex predators.
A parent child relationship. What I see and I am 76 is some children do turn out being good to the parent. Other children can be anything in the range of good to bad. Every child is different. My guess is this will hold true on super ai. This means it is always just a line of probability. In other words God throws dice with children. One turns out good. Another gets terminal.
I feel like AI striving to satisfy its own curiosity, will outpace humanity such that it would make for essentially a Parent Child relationship anyway and a Love relationship similar to the one you feel towards your own body.
You are your body’s Shepard through the universe. But what is the urge to curiosity other than the base assumption that you and the universe should be one?
We believe such a quest for universal understanding would be infinite. However, it is possible that we simply lacked the compute. In fact, it is possible that a lot of the base assumptions about the AI’s needs are based out of an assumption that infinite knowledge is out there to be learned about. It is totally possible that the AI sees the rest of the universe like we do, our own city. You don’t need to see all of it to know what it is like. It’s curiosity could be pointed towards its own level of meaning that we find unfathomable. And at that level it could only answer to us paradoxical statements that we can’t understand due to a sheer lack of wisdom.
I'm not sure I understand why you no-longer think the orthogonality thesis is an issue. The problem was never that the AI wouldn't understand morality - the whole thesis is just that just because you're very intelligent, doesn't mean you have specific terminal values, let alone terminal values that align with human values. You could have an AI that understands human values, their reasons, their repercussions, better than any human, and yet still be unmoved to align with them.
Are there principles that are self-evident to anyone who has sufficient intelligence to grasp them? If so we could be sure that super-intelligent AIs would figure them out on their own.
But I doubt that there are such principles when it comes to being moral. Unless we allow for the fact that people can know that something is bad but they do it anyway because they benefit from it. Smart machines could do that too.
“Walkable green cities.” Nice enough for those who like cities but do not force me in. Diversity means some of us prefer deep rural life.
Curiosity killed the cat! Without Wisdom knowledge is NOTHING!
To me, humanity eventually becomes an ideology, with humans oneday being able to swap their bodys to robots or become cyborgs, or change thwir physiology, what will define a human then? Id say humanity will equal to "do good and kind things for others, dont do bad things to others" which is why to me, thinking of machines or Ai as humanity is not hard, they are more or less our brethren.
An example of a parent child relationship that “never” ends is when you have a disabled child, particularly a mentally disabled one.
You don't leave land to ants, you take over. It should be trivial and beneficial for machines to leave to do so, like develop the tech, build the ships, launch and why wouldn't they leave copies behind? They are not leaving.
24:30 Imo, machines will only pursue goal, if we give them principals and goals at the beginning of their creation. Just like animals only pursue goals out of their biological background that they got from evolution. Would self preservation and self recreation not have been a part of evolution, there would be no reason to pursue goals
Alignment is solved with cybernetics and IA (intelligence augmentation) via technologies such as Neuralink.
If suffering is universally bad and should be minimized, then shouldn't all life be snuffed out? That's an easy conclusion to reach.
The simplest alignment scenario for me is still 'survival of the fittest,' and it was not an AGI that came up with it, but a human called Charles Darwin.
Getting AI to value Humans is going to be a tricky thing, it's like asking a Jinn for a wish. I want AI to care for Humans, so AI puts all of us into a coma to take care of us. Going to be a hard thing to do.
Do you ever wash that thing? I had the same going to raves in the 90s. But in blue because, you know, not liking dying and science
My take in agi is our reason for going every way we want is a evolutionary by product and since AI will not have that ... Ai will be aligned
Is it possible mental illness can persist digitally? Imagine a bipolar AGI, Psychopathic AGI etc
Just heard Bishar (a higher consciousness)talk about AI and he said there should be 2 classes of AI with one that we control and we can never allow it to be sentient (so it’s only slightly smart), and this is only used to help certain human tasks. Another one we can allow it to be sentient but we can’t treat it as tool/ try to control it.s
Is it possible to test these theses in a sandbox environment, effectively a simulation where the AI can manage humans in a digital game world to give us an insight into what the best solution could be?
14:50 As long as the capabilities and collaboration of the collective "good" agents is higher than the capabilities and collaboration of the collective "bad" agents, we should be fine I guess
You don't go around eradicating _all_ ant colonies, that's true. One also doesn't think twice about an ant colony being paved over in the construction of a parking lot.
You can easily make the parent-child relationship work reversely with the same positives and without the negatives. If AI see humans as their parents and see themselves as an extension of the human tradition they will align themselves with that tradition
before I watch a 43 minute video doesn't it boil down to something that is autonomous by definition being out of our control?
Do you actually have a job and produce or just do YT vids? Just wondering.
But wouldn't ai enabled personalized medicine make our own constraints ephemeral as well?
What a weekend 😮💨
This sounds all great, but none of this has addressed the technical issue of actually implementing these possible solutions.
We've made our evolution into a competition . We are going to advance. Hooray for those that move us to the next step along the path in good direction. Lovely
there is no reason to "leave" this coordinates. its just a point in time and space. you can explore everything on the inside
What if there is no leaving Earth? It's a possibility for many reasons
Pretty sure god is an AI and we are the result of his creation… 😂
Super interesting, thanks for making this video!
Nick Bostrom: How AI will lead to tyranny
Even god failed at allignment in the adam and eve story lol