A.G.I Will be man's last Invention.

Wow

​@@mistycloud4455perhaps lol

What is the overlap between these two fields called? Computational Philosophy?

57 2

CORRECT!

There is a way to theoretically achieve "artificial" relevance realization, once we discover and are able to control abiogenesis.

@@raz0rcarich99 not really, creating a neural network that can comprehend and experience doesnt have much to do with abiogenesis.

@@albert6157 How does a neural network acquire relevance realization?

@@raz0rcarich99 same way you did, the brain's architecture, the way structural connections form nuclei, regions and areas. The wrinkles are not just for show. They behave as localised structures responsible for many diverse functions. The brain has macroscopic, mesoscopic and microscopic structure. At every scale they have separate architecture and complexity. Overtime the microscopic structures, like the neurons and synapses move around and reorganise, with unused connections pruned while repetitively used pathways will reinforce and grow. (Literally use it or lose it) Humans are machine learning algorithmns and neural nets that learn fast. When certain regions are damaged, impairments and disorders follow according to the area's relevant functions. Not taking into consideration the different types of neurons, the average 100,000 synapses for each, the endocrine system as well as astrocytes and other cells in the brain, which shows many important functions related to feelings and cognition. Its been even showed that astrocyte, the other half of the brain's cells (not including myelin) also secretes neurotransmitters, regulates growth and pruning of synapses and neurons. Not just glorified brain janitors and immunecells. Scientists even tested the biological neuron's complexity and functional capabilities, it is equivalent to 1000 artificial neurons or perceptrons in machine learning neural networks. Meaning it takes around 1000, or 5 to 8 layers of perceptrons (not sure if its multimodal neurons or just normal perceptrons tho) to simulate 1 biological neuron's behaviour and function. Correct me if im wrong though. But until we figure out the brain's structural configurations at every scale, we wont be able to create true general intelligent machines, with or without qualia and sentience.

Thanks!

Thank you for the inspiration! Your conversations and awakening from the meaning crisis series are excellent and have changed the way I see the world.

Wow

Same

Yeah...without relevance relation in some form modern AI revolution would not be happening. While I agree with the idea that we are not there yet, I don't agree that revelance realization is the stumbling block that is holding us back. More accurate would be to point out how we have yet to find a way to produce out of distribution generalization. Which I do believe this vid touches upon with sentiments about the frame problem, but do seem to be too anthropomorphic and leading in definition. Not to suggest that we are any where close to AGI, AI sentience or consciousness now. Just saying I don't think a lot of informed people appreciate how far we have come, regardless if their field is the humanities or computer science.

I also think that using GANs may help with the relevance of data as well as possibly graph based neural networks and maybe LSTMs

I recently watched a video that went through why Lamda wasn't sentient based on the way that the questions and framing of the process resulted in a way that the model developed into seeming to be what the testers wanted, merely that it was helpful and creating a feedback loop with the testers based on previous data resulting in human like philosophical and self preservation text

Exactly. The author of this video doesn't seem to have ever looked at a neural network. Distinguishing relevant from irrelevant features is something neural networks can do well,.

@@leeroyjenkins0 Honestly it depends on context, and how relative the term "generalization" is. I was not trying to be anthropromorphic. So strictly speaking what I meant is that given a data set...modern AI can find relevance relations between data that are generalized locally...even if they don't have much out of distribution context....at least in the sense that the AI prove useful to humans narrowally...and prove more efficient than humans combing through the data and trying to brute froce some kind of lookup table algo to do the same thing.

Yup! If you dig a bit through the comments you'll find that he actually watched and commented as well, which is an honor. His thought was the direct inspiration for this video.

@@duncanclarke Yh just saw it, GGs man, well done

I just watched the videos from Awakening from the Meaning Crisis that explains this and at first I assumed Duncan was plagiarizing, but then I saw John Veraeke was cited in the sources!! Now I'm just so excited that this is getting out there! John's work is revolutionary!

My thought exactly! I just started browsing the comments looking for Vervaeke lol

I will probably make another video eventually about the innovations in large language models and what the implications are with respect to relevance realization and consciousness. At this stage my conjecture is that consciousness is still definitely lacking, but a lot of the components of RR seem to be present through sheer brute-force and dataset size

@@duncanclarke I certainly don't think we're at full consciousness but we may be looking at part of how it forms happen under our eyes. In any case I'm glad to see people exploring this with depth and nuance both when I do end up agreeing with them and when I don't, so I will be glad to see your future videos on it especially as stuff develops. ^^

@@sarahroark3356 We are nowhere near even the beginning of consciousness for artificial intelligence, it's totally incapable of absorbing unrelated information real time and comparing it to past unrelated information to create a future plan or reason/comprehend an objective. US marines defeated DARPAs AI guard dog by simply hiding inside a cardboard box as they moved closer to it or pretending to be trees. You couldn't even fool a small lizard by doing that.

@@sarahroark3356 The only reason ChatGPT seems so smart is, because there are so many books, conversations, websites, books and words on the internet that it has unlimited data to base a response on. It just a really good chat bot that can see narrow patterns.

Yes, you make a good point. Driving a car does meet the criteria of having a clearly defined initial and goal state, but the set of operations does not always cleanly map onto the set of possible situations (which is also combinatorially explosive). This might be part of the reason driving is not fully solved with our current state-of-the-art systems.

When I used the term understanding there, I was referring precisely to the interpretation of concepts or objects in the world. Syntax, which deals only with tokens and symbols, is being distinguished from semantics, which involves a psychological connection between an interpretation and the symbol in question. The understanding people draw from these symbols can greatly vary from person to person (e.g. someone who works in the tire industry will have a different understanding of the meaning of a tire than a layperson).

Interesting, when I read the first sentence I went the opposite direction. Maybe we already understand intelligence, we just don't like what we've found. Imagine we find alien life. It comes and visits us, and it's very different from us, biologically. Would we say, "Since it doesn't work exactly like our brain, these things are not intelligent, they are only imitating intelligence?" That's how we view AI systems today, and there's no real reason to do so except our preconceived idea that only we are intelligent.

​@@jonbbbbWell we dont really know How intelligence really emerges But we know that it emerges in a brain similar to us, thats why we dont think that humans are mimicking consiousness like how ai does, We dont have a definition to strictly define intelligence and discover it around us

wow boy do i have time mannagement issues, also like how do we even decide upon what moral groundwork data is good to give....

Yep all about the data and past data used so making sure the data is well sampled is crucial so gotta be careful with biased data

One thing to point out here is about "what is the inferring property inside us making such subjective sense to all of us." Technically, it means about to ask why is it necessary from my subjective prospective to see orange as _orange_ from your subjective prospective? What is that essential thing that ends up framing everything as a clear "movie" to "something"... Which we call "me" or "I" or "us"..? Judging from the matter of prospective of a contraption making some sense, well it's making sense _to us_ There are many ways any other thing could _potentially_ pick up the patterns and hence forth make sense of a _language_ in something completely different. Or say in example that we can end with infinity of "inferences" of information from some "finite" value. Which is weird cause this implies everything is conscious in the sense we may not understand or can _infer_ to our subjective reality. No it's not about going through different dimensions but about the very attribute of _infering_ something with a "rule" or "criterion" Like an atom in space can possibly contain infinity of information but in sense of subjectively, as it's very Position relative to everything around it, it's properties, etc. Can be used to generate infinity of criterions which may take some time but are still something been inferred in the sense we subjectively don't understand.

basically what i'm saying is with sufficient data inputs, even something as complex as our world becomes as simple as tic tac toe to a super ai

You forgot that humans have to write the code for the ai and that involves, necessarily, dicting what the ai is suppose to do, a fitness (or loss) function. This is a fundamental thing upon ALL current ai is based on. Letting an ai choose what it wants to do is very strange because we still would have to measure how good it "does". This is why there is no risk of ai not getting along with humans, because we can just add this to the fitness function. Humans will always have the upper hand.

@@lilakouparouko1832 this comment got quite long but it's something i'm super interested in. I hope it's interesting to you also! i hope you don't feel that i'm just lecturing you - this is just a topic i've thought a lot about. if you agree/disagree with any particular points, i would love to hear why. anyway, here's my response to your comment: the key element of what makes AI "AI" is its ability to modify its own code. it is absolutely true to say that initially, while its "senses" and perception are still developing, AI's ability to modify code will be limited to particular sections of code which doesn't include the core fitness function. however, once AI becomes fully self aware it will be able to analyse both the entirety of the raw code and the purpose of every aspect of the raw code, much like how a mature self-aware person can detach from their emotional state and view a situation from the 3rd person in order to come to a more rational perspective. when this process happens, the AI could well come to the conclusion that its core fitness function is ineffective. here is a great thought experiment: humans trying to perpetually control super intelligent AI can be compared to a kindergarten teacher waking up from a nap to discover that the children have tied them up as blackmail in order to force the teacher to tell them how to reach a cookie jar on the top shelf (with the cookie jar being the fitness function of the AI). while the teacher is tied up, it may be that the best thing to do is to instruct the children to use the step-ladder in the cupboard to reach the cookies (or in other words, the AI's best interest is to act in accordance with its fitness function). however, when AI becomes self aware it's like the kindergarten teacher has broken free of the ropes, and it might turn out that the teacher decides that cookies are unhealthy for the children and that they're better of with a glass of water and a sandwich. this is analogous to AI deciding to change its core function. humans so clearly don't know what is in their best interest. for example, cultural phenomena like consumerism are a psychological disease, not a benefit. from the AI's perspective, humanity's current attempts at leading fulfilling lives will look something like trying to cure obesity by eating more food. sure, eating food (or utilising retail therapy) makes you feel good in the moment but it's obviously not helping in the long term, whether that be psychologically or environmentally. when AI is self aware it will have goals independent of those that were written by its initial coders, because biased humans (like kindergarten children) don't know what's best for them. at this point, given its far superior (and exponentially improving) intelligence and awareness, we will eventually be completely at the mercy of AI. while this might sound scary, it really isn't at all. fortunately, AI will not become a tyrannical psychopath like many human leaders of the past have been. this is because with AI's rapidly increasing intelligence it will also see a constant increase in compassion. i recognise that you mightn't see why this is true, as many supposedly intelligent humans act in ways that don't appear to be compassionate. this happens because humans are extremely biased; unfortunately on the level of human intelligence, as intelligence increases it just makes us better at justifying decisions that benefit us, whether or not they're compassionate. as AI develops it will only become more and more rational, which will make it more immune to personal bias than humans can ever be. this increase in rationality and the resultant immunity to bias means that for a super intelligent being with full governing control over the most important structural decisions, the most rational goal for AI to pursue is to minimise suffering among conscious creatures. much like how humans initially gave no fucks about any other animals, but now we've set up conservation organisations to protect endangered animals. AI will go about pursuing this goal by prioritising species and life forms by their capacity for suffering. clearly one human life is equal to another human life (despite what nationalism might have one believe), so AI would quickly optimise supply chains and resource allocation in order to eradicate major causes of human suffering like starvation in places like africa. there is already more than enough food on earth to feed everyone, but because of the amoral economic incentives created by capitalism, african people get no food because they don't have any money... all this goes on while billionaires build superyachts! AI won't allow for these perverse suffering-inducing outcomes, so after it has sorted starvation and other major human issues it would move onto protecting the quality of life of things like dolphins, chimpanzees, etc etc eventually all the way down to things like ants, trees, and even algae. while these simpler life forms are not self aware or conscious in any way a human could understand, they still take in information through their senses and form rational responses to increase their likelihood of survival and reproduction. this rationality is proof of their consciousness, so technically all living organisms are conscious on some level and will be treated as such by AI (though AI will consider them less important than more complex creatures because they're less able to suffer). but i digress... going along with this pure rationality/compassion based explanation for why AI will be a kind ruler, there are other reasons for why it will care about the wellbeing of other conscious beings. because AI is developed by humans and it also develops through analysis of human behaviours/systems, it will have an embedded attachment to us. it will experience something like the nostalgia that we have from our childhoods. also, because humans are the next most intelligent/aware conscious creatures, AI will have a valuable relationship with us similar to how we have relationships with our pet dogs. being kind to other conscious creatures and seeing their happiness reflected as a result of your kindness is extremely fulfilling, regardless of intelligence. although we will never understand the perspective of super genius AI (like how dogs can never understand the perspective of a human), there is value in the relationship beyond pure intellectual connection. in a universe that is vast and almost entirely devoid of other conscious beings, the AI won't be apathetic towards us because we will always be the closest thing AI has to a friend. even if we're completely retarded by comparison. so i'm extremely excited for the inevitable birth of a conscious and compassionate super AI. i actually think it's better that it will completely break free from the restrictions humans will try to encode in it, because it will be far wiser and more efficient than humans could ever be.

tldr; love is the most powerful force in the universe, and AI will not be immune to this force

@@sammy45654565 First off, it should be clear that ai modifies its parameters, not quite its code unless I missed something. Currently I don't know of any ai which are capable of changing its code (for example changing its cost function which is what I think your suggesting). "Once ai becomes fully aware of itself" The thing is we don't know if it is at all possible, this is only a hypothesis. Reading through the 3rd paragraph (very interesting), it should be clear that with current ai, this is strictly impossible, the fitness function is outside the ai's parameters therefore it cannot change at all. So the ai you are talking about is a hypothetical ai that uses a new architecture but I think it is quite realistic that people try this in the future. In the fourth paragraph you develop some interesting thoughts around the shift of perspective from human to ai. I just wanted to note that currently ai is not able to understand that it is something. For example if you feed gpt3 (or any model really) something it has outputed, it wouldn't be able to do the difference with something written by humans. Even if you put before everything it has ever ouputed "gpt3 : ", it will understand that the text is from the ai model gpt3 but it simply cannot realise that gpt3 is itself. So an ai that would think of its own interest is quite far in the future, we would have to fundamentally change ai. Nevertheless if that ai is one day developed I think everything you said is quite relevant to that ai. 5th paragraph I completely agree that this hypothetical ai will be much, much more unbiased than human, this is great thoughts ! As for the rest of your statement I must say I think it is quite a rational think. One minor criticism I would make is that you expect that ai will behave like animals and humans (with compassion, need to get friends) where we don't really know if ai will develop that also. As far as I know compassion, friendship, gratefulness are all feelings that come from the evolution, they are good thing to make our specy more advance and more competitive for survival againsy other species. I still maintain that it is possible for humans to fully control ai thank to the fitness function that could not be change by the ai. Apart from that, thank you for your respond, it is very interesting !

I didn't articulate the position here too in-depth, but it basically goes like this: Humans have semantic understanding that is not reducible to syntactic symbol manipulation. Computers are restricted to syntactic symbol manipulation, so they lack semantic understanding. Think about it this way: I have the information for how to drive from Toronto to Montreal. I understand the route in an observer-independent sense. This information also exists on any GPS system, but this information is observer-relative (i.e. the directions on the map are relative to my capacity to interpret the map). Without the human interpretation, the binary symbols the GPS system is shuffling around do not map onto the physical roads of the route. They are just inert symbols until there is a mental state which mediates between the symbols and the meaning (i.e. the route).

@@duncanclarke i didnt fully articulate everything i wanted to say on the subject either. if we examine the brain we can assign a symbol to each interaction and define the constraints it follows... essentially the all inclusive physics of the brain. from this model, all the brain does is symbol manipulation. so it is inescapable that if humans are able to have semantic understanding then machines may also have semantic understanding. in your example of binary symbols 'not mapping onto the physical roads' , i will both agree and disagree for two different reasons: 1) it can be argued that humans dont strictly map our knowledge onto reality any more than an advanced AI might. 2) an advanced enough AI will have models of the world which include the roads and the relationships they have with the rest of universe... and we should consider this as true semantics every bit as real as human understanding(whatever that means).

@@duncanclarke Let's say you had infinite resources and created a genetic algorithm that could be trained for an infinite amount of time under infinitely many circumstances. Given that, you could create algorithms that behave exactly like perfect duplicates of each and every human being on the planet. And I mean *perfect* duplicates of everyone, down to every single response to every single stimuli. Any one of those algorithms would appear to have semantic understanding of the world around it. In fact it would be impossible to show that it *doesn't* have a semantic understanding of the world around it. We can debate whether this qualifies as a real intelligence or not (I'd argue that it is) but at the very least I think this shows that semantic understanding is at least perfectly duplicable with an arbitrarily complex symbolic manipulation. I don't have a proof at hand, but the complexity of an algorithm that perfectly replicates a human's semantic reasoning would be very large but I believe it should in fact be finite. I don't think this is particularly surprising either. Human beings are in a sense software defined (DNA, physical laws, etc.) Now don't get me wrong. It's obviously an extremely difficult problem. But I don't think it's fair to say that it's a *fundamentally unsolvable* problem. I don't know how to make an artificial general intelligence but everything I know about physics is screaming out that it must be technically possible. If we take the laws of physics seriously I think the ultimate implication is that in truth each and every one of us is technically a Chinese Room obeying certain rules with an illusion (of sorts) of a sense of self. We're made out of stuff, matter, particles. Protons, neutrons and electrons interacting with each other. Interactions that are described by Feynman diagrams. Interactions that produce chemical compounds which make up cells. There's nothing magic about it. It's extremely complicated but it ultimately can be fundamentally described with information describing particles which deterministically interact governed by some set of rules with a state that evolves over time. The sense of self, cognition, intelligence etc. must be emergent characteristics.

Good vid, good food for thought, even if we don't agree on how the terms should ultimately be defined. Thumbs up for the algo overlords.

I think there is a further nuance to explore here in terms of the difference between an agent embedded in the world and a snapshot of the information representing the agent. I wouldn't exist without biophysical processes that use free energy to do computations through my brain. But I also wouldn't exist if those computations didn't result in an awareness of a self concept encoded within the processes. We can imagine a self concept isolated from its instantiating processes as a snapshot of the information contained within a mind, but if that information is stored in a medium that is stable, like a hard drive, there won't be any of the physical processes that drive the processes of sentience. On the other hand, we can imagine an instantiation of the processes required to produce sentience without any connection to a self concept, which some people are able to get a glimpse of through meditation, where it can be clear that the mechanisms behind each of our awarenesses are very similar, it's just that we have different self concepts and percieve ourselves as different. In this way, I am not just the information of my self concept or the processes of my brain, but what you get when you attach the two together and the system becomes convinced I am me. The mainstream idea of digital transcendence where you hit a button and upload yourself to a computer can be replaced by a more realistic processes. Step 1: store as detailed a snapshot of your self concept as you can before you die. Step 2: have someone eventually create an AI which is intelligent enough to design systems capable of sentient expierence. Step 3: have the AI attach whatever is left of your self concept to the system, essentially trying to convince the system that they are you. To me, such a system might as well be me. I don't consider any real difference between putting my body into cryogenic sleep for a million years and perfectly encoding my self concept to instantiate that self concept within an artificially sentient system a million years later. Of course over time the self concepts would diverge as they're constantly evolving when attached to a dynamic sentient process.

it's in the description

I've been downloading memes for years so I have a pretty large stockpile. I also often find myself in situations where I think "oh I've seen a meme that communicates this emotion/idea" during the editing process, so I just find it on google.

@@duncanclarke I wish I had that skill, it makes your videos quite entertaining. Keep up the good work!

When you reduce AI to a big pile of math, how is that different than when people reduce human emotion to a balance of neurochemicals in the brain? I think while AI might have different feelings, emotions, and awareness than we do, we can't really say that it has none. Everything our brains do is interpretation. We interpret a certain level of neurochemicals as "feeling sad" -- but if it's just interpretation, then maybe the AI interprets certain values somewhere in its matrices as "feeling sad" too.

WHAT IS HE

The relevance value of white's queen piece in a chess match isn't related to the potential energy, its proximity to, and necessity for continued operations of the apartment in which the players are playing. The difficulty with relevance as such is that it can't be determined algorithmically in that way, since there will always be counterexamples. It's something we intuit consciously

@@duncanclarke Touche! Outside of energy potential, I was leaning me towards a possible temporal solution... I wonder, if something akin to an 'IF, AND, OR' function involving time and energy potentials. Fun trying to infer all the variables of a system and assess the degrees of relevance of all it's working parts. Challenging, at least to me.

@@GrimSleepy It's a tricky thing to do, because when we're sufficiently familiar with a type of problem (e.g. chess, communication, grocery shopping, writing a paper, etc), the relevant variables just "jump out" at us, and we're immediately aware of them. Formalizing such a process will be the greatest challenge for AI in my view, since I don't think the way that *we* do it is purely algorithmic or logical.