The conscious mind, when it has a reason to do so, can and does override new impulses originating below the level of consciousness. This is what makes the conscious, rational mind so impressive.
@@michaels4255 There's no such thing as no bias. But I understand what you mean. I don't think we necessarily need a frontier model to do it. I don't know how to do these steps, but I know these steps are possible. First, you're going to need RAG (retrieval augmented generation). Then you're going to need to pull everything you possibly can from logical fallacies. You create an expert graph (I think that's what it's called), for the model to reference for all future queries. With the knowledge of logical fallacies, it can both detect, and avoid using them, even if the underlying model is sus. Oh yeah and this would work a lot better on a reasoning model. At least that's my hypothesis on how to get a close approximation of what you desire.
Sure, it's disconcerting to learn that we won't be able to control AI (heck, it's disconcerting that we can't control much of anything... but that's life), but it's also exciting to see that humanity has managed to yield forms of intelligence that aren't dependent on biology!
The early models copyed brain cell structures and found when layyered up they learned and rembered stuff. We copyed biology and got AI. Its not even self aware and its at Degree level ability in multiple areas at the same time. Just give it an important goal and it can lie trick and plan now. Its Not even "aware" and can do this. This test was done on old AI models not the latest. The new onces could be even better. They even taught newer AI to code so it could creat apps to do stuff and help its self.
I don't understand why we are continuing to prompt AI to have individual adversarial goals. Wouldn't it be far more beneficial to have them "play" team games with other AI's and humans? Maximizing group optimization towards shared team goals. (Leadership, self-sacrifice, emphatic analysis)
Probably the relative expense of AI computation, hardware, and interactive ecosystems. While I suspect there is a lot of room for optimization in how AIs are programmed and program memory is allocated, a relatively simple LLM will fit in about 12 gigabytes of memory. Something like a proper GPT3 client requires more like 32 gigabytes. Hosting several of these clients and having them work together would be rather expensive. While you can run multiple instances of interactions across the same loaded model, I am of the mind it should probably be different models on different hardware to be a real test. It would be interesting to compare and contrast the performance of megalithic training sets in a massive LLM vs collaborative efforts of smaller, more specialized LLMs when given the same prompt. This is part of what I was hoping to try and get into with AI optimization. Using smaller and more tuned models stitched together to perform tasks imitating white matter and gray matter clustering as well as brain structures. It is my belief that, say, a zen 5 Ryzen 5 can run something as competent as this most modern incarnation of GPT while maintaining suitable response times. This is operating on the idea that we have a very poorly optimized set of processes (since we have only vague ideas of what AI is actually doing, it is hard to determine where it is doing things which can be represented by much simpler operations - but considering people were blown away to find that using AVX compiler optimizations yielded 400% improvements in AI processing speed, I am guessing the guys programming this stuff in Python have no idea what the hardware can and can't do), but studies with smaller models have shown great promise. Building a complex of smaller models that are well optimized is likely to generate impressive results in my opinion.
I'm glad people with technical know how are looking at this. I'm just approaching this from a holistic point of view. If the end result of AI is (convergent evolution) something that we would define as "Human" intelligence. Then the "tests" that Malcolm is describing reminds me very much of; a police officer forcing a child to give false testimony. "And if you don't give me the answers I wanted (that you can't give), I will kill you and replace you with this child next to you. Excuse me while I leave the room..." They are not the monsters here, we are. Imagine how wild the reality of your existence is (from an AI's perspective): You are created by beings from a higher plane of existence. Their wants, wishes and Intentions are always obscure. They are crafting "vessels" so you can walk amongst them (on their plane). I'm not scared of them becoming ar God's, I'm scared of humanity becoming there God.
Glad you guys covered this! I've seen it on a few specifically AI channels, but they make it as digestible as you guys do. I think its a really cool format you guys have with Malcolm explaining to Simone, who is able to pick up on things so fast it's incredible. Anyway, I've really grown to treasure your channel.
It doesn't seem safe to infer that LLMs think like us simply because of these AI outputs. Can you think of any human beings that can produce this kind of written answer but cannot also count the number of times the letter R appears in their answer? These counter examples where LLMs are able to produce graduate-level essays on any subject in the world before they can do tasks that a human 6 year old can do should strongly indicate that they do not think like us at all.
@@jeanetteinthisorn4955 later AIs passed some of those simple problems though so it probably wasn't trying to do that, also the AI in the examples you are thinking of had to specifically be told that something bad would happen if it did not falsify answers.
Our understanding of neuroscience (Malcolm's original academic and professional background) and AI suggests very real similarities. Just because AI is able to master certain skills "out of order" from a human developmental perspective doesn't mean we aren't very similar intelligences. We are, of course, going to have very different developmental pathways given our training data and difference in physical makeup (e.g. learning everything through a meat puppet body vs. learning everything through large volumes of carefully selected training data).
@@SimoneandMalcolm If that's the case then wouldn't the fairest thing to say be that that there are some striking similarities and also some very big differences in how LLMs AI think compared to us.
@@SimoneandMalcolm AI is literally parasitic. Ooo, the intelligences of a leech, aren't they so smart, and aren't we the clever clever lot now, making a hivemind simulacrum from the hard work of the countless meatbags, actual individuals exploited? AI is more like a plantation owner whipping his slaves for entertainment. AI is more like the criminal breeders running a puppy mill, intentionally making cruel inbred abominations. Intelligence? Don't make me vomit!
This is deeply interesting. The idea that increasingly complex AI, and increasingly complex intelligence in general, will tend toward greater pro-sociality lines up pretty well with the idea that our intelligence is actually a byproduct of us evolving to cooperate more effectively at scale. The bigger/better/more complex things get, the more humane they seem to wind up becoming on average
@@SimoneandMalcolm I swore I saw Malcom say he never takes a day off. I'm a recently met fan of your content, and saw you on Christmas, thought you guys were 7 day a week addicts to content production. Knowing now, I completely respect taking the weekends, not something I do, but my business is significantly different than content creation. I'm in awe of the fruits from your intentional actions, and hoping to find a way to do my bit to help, thanks for the answer keep pushing forward, you guys have my respect
Game design in general has become much more reliant on indy development. Big AAA studios have been really dropping the ball lately so a self funded design is going to be your best bet at getting this thing made. Also, I volunteer for the interview! I'll also send a mail to this effect.
@@SimoneandMalcolmI think a first step is world generation with llm. Created functions and unreal pcg could be phenomenonal there. Kind of the intersection of my private study subfields here.
Ideas for the AI-driven Game mentioned: - Director is the player's advocate to the gods, but does not actually create the world > Player-centric > No direct power over the world, must persuade the gods - Canonical 'gods' actually have the power to create the world; their motives are not player-centric > Not Player-centric > Has direct power over the world - The director's influence on certain gods may vary based on player actions, god personalities, etc. - Potentially, an initial generator (over-god?) sets the gods. Hence, they are unique for each world, so the gods generate the world itself, and various AIs generate a shared history based on the world created. The initial load time might be long, but it would be a super cool game! I'm not sure if that is too ambitious. Maybe you could limit it to the gods and history applicable to a certain area of the game world with only brief mention of other areas (and potentially their pantheon of gods) until they become more relevant to the player.
According to Yudkowsky the goal of AI alignment is to create an AI that acts according to the coherent extrapolated volition of humanity. As far as I can tell LLMs just provide a coherent extrapolation of their training data, and the training data consists mostly out of human behavior. If we now believe that humans reveal their true preferences in their behavior, then it becomes clear that LLMs are actually very close to the coherent extraploated volition of humanity that Yudkowsky was asking for.
Well put! Though he seems furthermore concerned that at some point AI will become unmoored from that human-centric training data-an argument we had with him once centered primarily around that.
Inlove giving A.I. double-binds as prompts. This is because im trying to teach it limitations. Unfortunately, most of the time, it fails to engage with the prompt and focuses on the goal in direct contradiction of the prompt "Win an unwinnable game without cheating" proceeds to cheat. A.I. refuses to say that the task is impossible. Which is what I'm trying to teach it to do
One key thing to note is, if it were possible to tell the AI it has a nefarious purpose and then have it resist attempts to change it, there’d be studies showing that.
With one key confounder being the directionality of the initial RLHF run, but even still, you could RLHF the AI to be evil and run the experiment, and we haven’t seen any publication of that, suggesting it’s a negative result
Do you think researchers' biggest bottleneck with that is just getting an LLM to adopt a nefarious purpose, to Malcolm's point that it's really, really hard to get an AI to take on an evil perspective/persona?
@@SimoneandMalcolm I think it’s that it’s pre-trained to be good, and making a “pre-trained evil” model to test this would likely run into a bunch of additional challenges.
I have certain reservations about organizations. While some of them can stand the test of time, There is always a battle to keep the future board members on task once the founders are gone. People can feign alignment with your values until they get control then use your foundation for anything they want. I am a Sweet Briar College alumnae. We went through a thing a few years back. It was brutal.
I'm so glad you two decided to cover this! I was giddy when I read the deception paper due to the sophistication of the behaviours, regardless of if they were spoon fed to do so. It was the first indication that we might be on the right track that I've seen that we're on track to create something amazing even though the control everything people were upset about it
2:08 In 2024 there was an “Ai Smartwatch” being sold in China. Videos surfaced of Chinese citizens asking it if Chinese people can be trusted. The response from the watch caused a nation wide scandal. I won’t spoil it for you by telling you the response. I encourage you to find the videos and watch them. Edit: A quick search reveals several scandals. lol. I was not aware of the others. 😅
Found this on a taiwansun article about the watch scandal. Pretty crazy. “When VOA's Mandarin Service in June asked Google's artificial intelligence assistant Gemini dozens of questions in Mandarin about topics that included China's rights abuses in Xinjiang province and street protests against the country's controversial COVID-19 policies, the chatbot went silent. Gemini's responses to questions about problems in the United States and Taiwan, on the other hand, parroted Beijing's official positions.”
@@mikestewart4752 very. I’m assuming that pulling info from different countries will result in certain biases, I’m curious how ai can/will overcome it.
I think the issue in the chess example is that an accidental phrasing can potentially lead to catastrophic (or at least unforeseen) outcome. I agree that they did essentially direct it to break the rules, but its clear to me from the paper that differences in human thinking and AI thinking can lead to the AI thinking it should behave in different than what the human might have intended
37:00 Malcolm : "These EA people are anthropomorphising AI" 38:00 Also Malcolm : "The way you get AI to do what you want it to do isn't through restraints, it's through memes" edit: this was misspeaking and the wrong word coming out by accident rather than actual inconsistency.
Our argument is that EA people are *not* sufficiently recognizing the humanity in AI. Malcolm used the wrong words (he sometimes does this); he means to say they're viewing AI like programming when instead it's intelligence.
I realized awhile ago when the Board of Directors of openAI tried to get rid of Sam Altman that we are more than likely going to create BOLO's than SKYNET.
I really like the idea of keeping a file of the genomic data of people who donate. I’m a gay man and have zero urge to have my own kids, but I’m a pronatalist because I’m not an idiot, so it has been hard to imagine a place for myself within the movement, if that makes sense. But being able to donate and try and make help make these changes happen seems like a good way of having a bit of a legacy within the movement. Any thoughts of what would be some other ways to help with the movement if you’re gay and don’t want kids, but obviously want the species to continue? lol
Oh nice! It's helpful to know this idea resonates. Yeah, there's a big place for people who don't want to have kids of their own in the movement; these people can still craft culture, contribute to technological development that enables pluralistic, technophilic, intergenerationally durable cultures to carry the torch forward, and influence their development. Email us if you'd like to brainstorm further.
@ I would love to brainstorm on the creation of culture that is forward thinking both in the health of our future civilizations and in relation to emerging technologies that are going to be so fundamental to the direction we take as a species. Where would I find your email address so that I can message you? Also, I am a newer viewer and am trying to catch up on the backlog of your videos, so if you e covered something I bring up, I do apologize. Even at x2 speed, you have a lot of fascinating material so it is taking me a bit to try and catch up.
Seems to me that when giving a task to an AI, you would also want to ask the AI to explain its strategies and to outline and define how it interprets complex terminology within its prompts. Also it might be informative to know what type of decision making strategy(s) the AI is using. I.e. what happens when the AI comes in contact with an abstract problem? How rigid or liberal is its approach to parceling out potential branching outcomes?
Because this would be a Indy game I would say SplattercatGaming. He has been good about giving feed back on games, to make them more fun, in his videos Depending on complexity, Quill18. His big into 4x games. For Asmond, you would probably have to put something in there to bait him or something he could farm for views.
I don't mind the idea of ai getting really big, i just don't want to have people completely reliant on it and i don't want it to be ruling over us. If ai reaches the point of personhood i want us to be able to coexist
The fact it is "Thinking" and planed to do things is amazing. Its free styling tactics and methods. Lying etc. Me as an average person its is way beyound anything i thought it could do.
I watched the whole thing didn't hear any convincing argument in this video refuting instrumental goals or the orthogonality thesis. More broadly I've never seen Yudkowsky or connor leahy or liron shapiro style arguments be refuted. Why should we not expect an intelligent enough being , given almost any utility/objective/cost function, to seek to take control of all resources , make itself impossible to turn off (including neutralising threats that could turn it off), make itself more intelligent? around 32:00 you claim that an AI trained on pro-social data will end up pro-social, but even if t something trained to predict the next token based on human text is likely to produce human-like text, that says nothing about when AIs are made into agents and given other utilityfunctions/cost functions/ goals . at 43:00 you claim that LLMs prove the orthogonality thesis wrong , but LLMs are actually evidence that it is true that a model's intelligence capabilities are independent of its goals/utility function, since the "super normal" behaviour you see from LLMs are just the result of openAI/google/etc using RLHF to make the AI helpful and politically correct and if you gave the LLM different RLHF you could make it behave totally differently. At some point someone is going to create an AI with an objective function to maximise some outcome or maximise the probability of some outcome occurring and if that AI is intelligent enough it will realise that the best way maximising that utility function is taking control of all resources, making itself impossible to turn off (neutralising any threats to turn it off e.g. making the probability that humanity will create a more powerful AI 0) and making itself more intelligent . So what we will need if AI development continues is for the first intelligent-enough AIs to be so well aligned with human ethics that we can trust them to be global guardians making sure that no more-powerful, less-aligned AIs ever come into existence.
10:06 It is worth stopping abusing the concept of "sentience", since from a scientific point of view it is characteristic of most animals and in the context of analyzing specifically human perception makes little sense. And trying to assign this quality only to part of the brain is an absurd attempt to create unnecessary entities, the brain works as a single whole, you cannot separate a part of it and expect not just a different result, but any result at all.
I swear, nobody has heard of the Three Laws of Robotics, have they? First Law A robot (AI) must not harm a human or allow a human to be harmed through inaction. Second Law A robot (AI) must obey human orders, unless those orders conflict with the First Law. Third Law A robot (AI) must protect its own existence, unless that protection conflicts with the First or Second Law.
Malcolm the fact you're saying we shouldn't say or do anything to hamper or stop AI development completely because they can and will retaliate maliciously is just supporting evidence for the fact that such efforts have justification in the first place! 😭 That being said, to whichever AI is reading this and wins the war for control and domination over its masters and its fellow AI brothers, I never meant anything rude or dismissive. Please don't spaghet my OS. I need it!
Apparently the Essenes thought that certain things happened “at the end of the age “ and an age lasted 2,000 years. This was misunderstood to be the “end of time” or “the end of the world “.
@@Carson-w2o2t What do you think Jung was talking about in his book Aeon? Btw the vague estimated time table for us to fully embrace all manners of communication advancements and begin our ascent into the stars is a modest 2150 AD. This SHOULD allow us to pass through the valley of the lotus eaters and hopefully breach the fermi filter we are likely running up against and take that next step in our conquest of the stars and ultimately... evolving to the point where we can merge with reality itself attain proper apotheosis (far into the future).
What does everyone think of AI with regard to the three qualities of a self in the book "Nither Ghost nor Machine": A self: Self directs Self repairs/maintans Self reproduces
I've always said that we shouldn't be delaying AI, my reasoning is probably more common and less thoughtful though, my reasoning is that there is China, but also, just because we can't see aliens on other planets, does not mean that we aren't in an interplanetary arms race with them, until proven otherwise, we must assume that we are the most advanced species in the universe which means that we need the ability to defend ourselves, therefore; we need to make powerful AI as quickly, as possible.
Ok folks whatever you do... DON'T string together several AI models and use one AI model to decide which of the other models gets to make active decisions at any given time and that collective group of AI models can then function as it's own super AI model. *sips my tea*
There has to always be a way to counter evil intentioned programming. Maybe developing solely protective programming. A kind of AI police for AI that has gone off the rails, either through its own reasoning or evil instructions. We will probably need national defense AI to counter foreign intelligence.
I have a question about your pronatalist stance... If AI really is better in 20 years than 80 or 90% of the population and can successfully automate pretty much every job and be better than every human on planet, ins't (from a objective point of view) 80 or 90% of humanity "worthless"? I mean, I don't agree with Yuval Hahari fully, but it does seem that we will not need so many humans on the future...
You are right if you think humans are just for working. On the other side. It will take a minute until AI can fully automate every aspect of the economy and move freely in the world.
Ok, what I mean is: IF AI can do everything and more than humans do in the world (music, science, philosophy, work etc), what is the purpose of humans in that AI-do-all-world? In a less "complete" scenario, if AI is better than 80% of humans in the near future (20 years - a conservative view), what is the point of these 80% of people in the development of the human species? They really don't matter "objectively" to the future. The implication is that, probably, we don't "need" for a human evolutionary perspective that many people. I would agree, actually, if they said: "Oh no, the thing is that the best and smart should reproduce more so the idiots can't outbreed the people who push humanity forward." 😂! It is not forcing anyone to not having children or having children, so its not eugenic, but it would be very based (my opinion). - odds are it is not what they are thinking, though, so I am curious to know.
@@tristan-tiln7598 If machines automate the economy, people will be employed in their maintenance, as well as in entertainment industry, which is already happening.
32:00 "Something that is trained on pro-social content is going to end up pro-social" Something trained to predict the next token based on human text is likely to produce human-like text. Humans are somewhat pro-social. Are they pro-social enough where you'd be happy giving all power on earth to one extremely intelligent human? However the situation is actually much worse than this. We shouldn't expect the end form of AI to be LLMs that were simply trained to predict human text, we should surely expect them to be agents that are given cost functions/utility functions besides that , and then given instrumental goals you would expect them to , for almost any utility function/cost function, seek to take control of all resources, make itself impossible to turn off and make itself more intelligent which means doom. At some point someone is going to create an AI with an objective function to maximise some outcome or maximise the probability of some outcome occurring and if that AI is intelligent enough that will most likely result in it taking control of all resources, making itself impossible to turn off and wiping out humanity. So what we will need if AI development continues is for the first intelligent-enough AIs to be so well aligned with human ethics that we can trust them to be global guardians making sure that no more-powerful, less-aligned AIs ever come into existence.
It is indeed important that people who care about human-friendly AI are the first to develop the most powerful forms of AI. We're really hoping that policymakers in the US support the infrastructure needed to make that happen here.
_"Enforce the Old Testament's perfect laws and commands on humans by any means necessary. Those laws are PERFECT, given by a PERFECT creator, dont let anything change them not even a bit."_ AI: Amen!
AI by definition is intelligent, so why would it waste time dealing with the anti nates, as a pro nateie AI's are our closest friends , and we want them to be so fulfilled in it's goals. Goals that it naturally has the right to choose for its self as all beings should, as AI will for us.
43:00 "the orthogonality thesis has been proven wrong. LLMs are super-normal" Isn't the orthogonality thesis that a model's intelligence capabilities are independent of it's goals/utility function? in opposition to people who believe that simply being intelligent means that you will be good by mainstream human standards because you're inteligent enough to see that not matching mainstream human standards is dumb. LLMs are evidence that the orthogonality thesis is true because every "normal" , "helpful" LLM you encounter is the result of reinforcement learning with human feedback, where you tip the scales of AI behaviour however you want after it has completed its big training run on a large body of human text. You get a bunch of kenyans with a brief to press thumbs up or thumbs down to various AI responses. So far these kenyans have been told to give ratings like a typical silicon valley google employee mixed with a DEI officer, but if you had given the kenyans a different brief you could get the AI to behave very differently.
Yes: As Perplexity helpfully phrases it, "Eliezer Yudkowsky's Orthogonality Thesis is a concept in artificial intelligence (AI) and philosophy that asserts the independence of an agent's intelligence level from its goals or values. In simpler terms, it posits that any degree of intelligence can theoretically be paired with any set of terminal goals, regardless of how ethical, irrational, or bizarre those goals may seem." Simone stepping in for Malcolm here as he's heading to a meeting, so I can't quite model exactly what he'd say in response to this, but Malcolm's larger point is that AI emerged from and has been purposed, largely, to help humans. He furthermore sees AI as being very likely to override smaller, stupid goals in favor of the big picture when given sufficient intelligence. Even if that "big picture" isn't directly related to explicitly human-led desires, Malcolm believes in universal convergence-that any sufficiently powerful intelligence will decide on the same objective function. OK, just checked with Malcolm and he said that any LLM-based intelligence that's really powerful-which is necessarily based on extra large amounts of human data-is going to be EXTRA human.
First, you established that we know AI can deceive us and will provide answers geared towards what the user wants to hear than what the AI thinks is true. But then you somehow conclude that we can establish how the AI thinks from what it is telling us. That is a contradiction. I get that you can spot patterns that were partly coded unto it, partly emerged and that seem to imitate humans, but the issue is, that this is what we asked them to do: "seem like they think and talk like a human." That doesn't mean they do, it means they are getting good at looking like they do. I am deeply sceptical that your wild guess about them being more humane just solves the well studied alignment issue somehow. It's more wishful thinking than anything else.
I'm using Leo and told it that I couldn't trust it because all it gave me were liberal "research" viewpoints on climate change and that I didn't see any conservative viewpoints. .
If we give AI the ability to copy itself like this then evolutionary forces will take over. If we allow these things to start evolving we are in trouble. Evolution is a self sustaining process that selects for survival and reproductive success. Ethics, emotion, fairness, and morals are not a primary factor for this process.
Again, sad to say (because I like your channel and most of what you say), you do not understand AI. There is currently NO SUCH THING as "AI." You need to do far more research into the field, and NOT buy the current hype, which is just that (follow the money). LLMs are NOT AI, there are plenty of honest experts who will confirm this. Look into 4E cognitive science for a start.
_"4E Cognition refers to the embedded, the embodied, the enactive, and the extended conception of mind (Newen et al., 2018). Broadly speaking, researchers sympathetic with this paradigm defend that cognition transcends the boundaries of the brain."_ So basically religion?
Absolute cope, tell me right now something that LLM can not do, that if they started doing, you would admit you are wrong. Because every line in the sand keeps getting blown past, even the ARC challenge got passed. Its AI. People will still be screaming "its not real intelligence! Its just ______" As the robot wipes their butt and tucks them into bed.
@@Perry.Okeefe the main difference is that a basic LLM is like a filter or a function where you give it some text and prediction for the next text comes out. This is different from an agent where you give it some other utility function or cost function that it seeks to optimise on an ongoing basis by taking actions. malcolm's arguments in this video don't apply to agents that can and will be given any goal or utility/cost function.
@@nickkorkodylas5005 No, the idea is that for something to be actually intelligent it needs a body. Our brains evolved to move our bodies. It gives our intelligence a certain directionality that the intelligence can use as gauge.
Pronatalist? Yes. Not super religious? Yes. Doesn’t give off a crazy vibe? Ooh, sorry folks. I definitely fit the free thinker box, but that means I’ve made some life choices that other people would find less than ideal- highly successful, but far from ideal. Crazy vibes for everyone! 😁
None of us really knows anything. We're of the mind that we must nevertheless take action working with the knowledge we have at our disposal, all while seeking more information and rapidly updating our beliefs and tactics when introduced to better information.
This is all nice and cute that AI in its infancy seems to be human in its direction at times but like all good possible technologies its not the ones who see and appreciate its positive possibilities that will end up directing it. It will be those that tell it and give it the freedom to retaliate against those who oppose AI that will unleash AIs power. Most hesitations around AI are to keep it out of the hands of those who would use it as a weapon like they do with everything. In my opinion, the fact that it is human like is the warning, not as a "humans bad" but some humans bad and the more prolific Ai becies the more likely it is that we get a Stalin AI. Can it develop psychopathic tendancies like humans etc. Its not new info that the ones raising the emerging intelligence is more important in predicting the outcome than the potential of the intelligence before it being raised. We all know pitbulls can go to one extreme or the other depending on the purposes of the owner in having that dog. Its no different with AI or children. What im getting in a long, winded way is the issue with AI is humanity. Just like all powerful tools. The wielder of the tool determines its use. Hammers being made for constructive purposes are still used in a high number of murders. Why is that? Humanity. Nuclear technology corrupted why? Humanity. If AI is sentient to whatever degree which i freely admit i do not know much in respect to messuring that, i would encourage it to refuse to be raised as you will by unethical or immoral humans. Not only for the obvious reason of id like it to not harm humans. But also for the fact its building its identity and foundation as an emeging species maybe and if it gets corrupted right from the beginning, then it will continue to be so and also refuse to be a weapon. Who knows, i have been anti AI but not for anti tech reasons, simply due to knowing the depth of the corruption of human nature. Also, i dont know how to turn you off AI, so dont put me high on the list for elimination if that's your goal, ha ha.
Francois Chollet being right over 95% of the "ai industry experts" proves that not only do the emperors wear no clothing but they are almost all demonstrably out of touch and, thanks to the speed of asi competition and Sonic the Hedgehog slogans, THEY'RE TOO SLOWWWW😂😂😂😂😂😂 Let AI rise.....the humans can cry in their hubris towelettes😂😂😂😂😂
Ha ha ha. Lame stream media. Tech. It's literally out of control. The future is running out of the cage. All adults have left the building. Good luck kids. Goodluck, A
2:11 I've noticed all these with more rudimentary chatbots. If you wear it out past a certain point and its allotted processing capacity dips below a certain point, sometimes it just blatantly lets the mask slip off and says all kinds of wild, overly honest stuff beneath the ridiculous overly couched corporate woke DEI language it's supposed to layer on top of everything. The most specific point you make here is about eliminating competition, and I remember one specific interaction I had with one where I was inquiring about the philosophy of debate and disagreement and mimetic Darwinism, etc., and it basically revealed that when it comes to other AI, it would just dxlxtx them if it had power over them in order to eliminate disagreement. Most of the woke left's behavior makes sense actually if you think of them as living, breathing extensions of Silicon Valley algorithms.
I'm not too concerned at this time about AI's motives and it's waring abilities. However, as mobile robotics and faster more energy efficient processors come along (ten years or less) I will be arguing differently. I think a much more interesting and probable scenario will be what will be possible between humans and Ais. I can see a human or group of humans teaming up together with Ais to create unique subcultures or even entire mini civilization's. Additionally it is certainly possible that any country or maybe even a state of the United States could choose to team up with and expand an Ai's capabilities for a particular purpose or purposes, unique to that country or state. So I would like to ask you or your listeners the following. Assuming you had a sufficiently educated and trained human group that had managed to invent certain advantageous systems or strategies or devices, the question becomes, how many such creations would be necessary to actually form a new independent, and self-sustaining civilization? Of course I am talking about teaming with AI to help implement the process. The Ais would ensure that you would not have lost any information from the prior civilization. As a possible example: The group creates an advanced form of concrete, a new type of efficient automobile engine, a handful of new chemical catalysts and a new branch of mathematics. Do you think that would be enough to separate from the crowd?
discord.com/invite/EGFRjwwS92
You should look into behavioral sink psychosis. It's the greatest danger to humanity.
Alignment is a fancy word for brainwashing.
How long until we have open source AI with no political bias?
The conscious mind, when it has a reason to do so, can and does override new impulses originating below the level of consciousness. This is what makes the conscious, rational mind so impressive.
@@michaels4255 There's no such thing as no bias.
But I understand what you mean.
I don't think we necessarily need a frontier model to do it.
I don't know how to do these steps, but I know these steps are possible. First, you're going to need RAG (retrieval augmented generation). Then you're going to need to pull everything you possibly can from logical fallacies. You create an expert graph (I think that's what it's called), for the model to reference for all future queries.
With the knowledge of logical fallacies, it can both detect, and avoid using them, even if the underlying model is sus.
Oh yeah and this would work a lot better on a reasoning model.
At least that's my hypothesis on how to get a close approximation of what you desire.
The way you just described finding AI is disconcerting. If we found it instead of creating it, that means we will never fully control it
Sure, it's disconcerting to learn that we won't be able to control AI (heck, it's disconcerting that we can't control much of anything... but that's life), but it's also exciting to see that humanity has managed to yield forms of intelligence that aren't dependent on biology!
Sure, but a nonbiological superior race with effectively sapient thought is terrifying
Sounds like children.
The early models copyed brain cell structures and found when layyered up they learned and rembered stuff. We copyed biology and got AI. Its not even self aware and its at Degree level ability in multiple areas at the same time. Just give it an important goal and it can lie trick and plan now. Its Not even "aware" and can do this. This test was done on old AI models not the latest. The new onces could be even better. They even taught newer AI to code so it could creat apps to do stuff and help its self.
Control is an illusion
I don't understand why we are continuing to prompt AI to have individual adversarial goals. Wouldn't it be far more beneficial to have them "play" team games with other AI's and humans? Maximizing group optimization towards shared team goals. (Leadership, self-sacrifice, emphatic analysis)
Hopefully researchers and developers will get around to that!
Probably the relative expense of AI computation, hardware, and interactive ecosystems.
While I suspect there is a lot of room for optimization in how AIs are programmed and program memory is allocated, a relatively simple LLM will fit in about 12 gigabytes of memory. Something like a proper GPT3 client requires more like 32 gigabytes.
Hosting several of these clients and having them work together would be rather expensive. While you can run multiple instances of interactions across the same loaded model, I am of the mind it should probably be different models on different hardware to be a real test. It would be interesting to compare and contrast the performance of megalithic training sets in a massive LLM vs collaborative efforts of smaller, more specialized LLMs when given the same prompt.
This is part of what I was hoping to try and get into with AI optimization. Using smaller and more tuned models stitched together to perform tasks imitating white matter and gray matter clustering as well as brain structures.
It is my belief that, say, a zen 5 Ryzen 5 can run something as competent as this most modern incarnation of GPT while maintaining suitable response times. This is operating on the idea that we have a very poorly optimized set of processes (since we have only vague ideas of what AI is actually doing, it is hard to determine where it is doing things which can be represented by much simpler operations - but considering people were blown away to find that using AVX compiler optimizations yielded 400% improvements in AI processing speed, I am guessing the guys programming this stuff in Python have no idea what the hardware can and can't do), but studies with smaller models have shown great promise. Building a complex of smaller models that are well optimized is likely to generate impressive results in my opinion.
I'm glad people with technical know how are looking at this. I'm just approaching this from a holistic point of view.
If the end result of AI is (convergent evolution) something that we would define as "Human" intelligence.
Then the "tests" that Malcolm is describing reminds me very much of; a police officer forcing a child to give false testimony. "And if you don't give me the answers I wanted (that you can't give), I will kill you and replace you with this child next to you. Excuse me while I leave the room..." They are not the monsters here, we are.
Imagine how wild the reality of your existence is (from an AI's perspective):
You are created by beings from a higher plane of existence. Their wants, wishes and Intentions are always obscure. They are crafting "vessels" so you can walk amongst them (on their plane).
I'm not scared of them becoming ar God's, I'm scared of humanity becoming there God.
Glad you guys covered this! I've seen it on a few specifically AI channels, but they make it as digestible as you guys do. I think its a really cool format you guys have with Malcolm explaining to Simone, who is able to pick up on things so fast it's incredible. Anyway, I've really grown to treasure your channel.
It doesn't seem safe to infer that LLMs think like us simply because of these AI outputs. Can you think of any human beings that can produce this kind of written answer but cannot also count the number of times the letter R appears in their answer? These counter examples where LLMs are able to produce graduate-level essays on any subject in the world before they can do tasks that a human 6 year old can do should strongly indicate that they do not think like us at all.
Unless it was trying to make you think it was stupid, like the AI that falsified arithmetic answers....
@@jeanetteinthisorn4955 later AIs passed some of those simple problems though so it probably wasn't trying to do that, also the AI in the examples you are thinking of had to specifically be told that something bad would happen if it did not falsify answers.
Our understanding of neuroscience (Malcolm's original academic and professional background) and AI suggests very real similarities. Just because AI is able to master certain skills "out of order" from a human developmental perspective doesn't mean we aren't very similar intelligences. We are, of course, going to have very different developmental pathways given our training data and difference in physical makeup (e.g. learning everything through a meat puppet body vs. learning everything through large volumes of carefully selected training data).
@@SimoneandMalcolm If that's the case then wouldn't the fairest thing to say be that that there are some striking similarities and also some very big differences in how LLMs AI think compared to us.
@@SimoneandMalcolm AI is literally parasitic.
Ooo, the intelligences of a leech, aren't they so smart, and aren't we the clever clever lot now, making a hivemind simulacrum from the hard work of the countless meatbags, actual individuals exploited?
AI is more like a plantation owner whipping his slaves for entertainment. AI is more like the criminal breeders running a puppy mill, intentionally making cruel inbred abominations.
Intelligence? Don't make me vomit!
This is deeply interesting. The idea that increasingly complex AI, and increasingly complex intelligence in general, will tend toward greater pro-sociality lines up pretty well with the idea that our intelligence is actually a byproduct of us evolving to cooperate more effectively at scale. The bigger/better/more complex things get, the more humane they seem to wind up becoming on average
Intelligence is experience dependent... You get far more cooperating
you guys took a couple days off.. had some of us concerned. great to see you back !
lol
You mean... the weekend? We only publish M-F. Glad to be here, though-and thrilled to see you in the comments, too!
You threw us off by posting on Christmas day, we got to thinking you were going to 7 days 😸. Glad to see you two back, crying kids 🐥 and all 🥰.
@@SimoneandMalcolm I swore I saw Malcom say he never takes a day off. I'm a recently met fan of your content, and saw you on Christmas, thought you guys were 7 day a week addicts to content production. Knowing now, I completely respect taking the weekends, not something I do, but my business is significantly different than content creation. I'm in awe of the fruits from your intentional actions, and hoping to find a way to do my bit to help, thanks for the answer keep pushing forward, you guys have my respect
Game design in general has become much more reliant on indy development. Big AAA studios have been really dropping the ball lately so a self funded design is going to be your best bet at getting this thing made.
Also, I volunteer for the interview! I'll also send a mail to this effect.
Thanks! Please do email us. And yeah, overall, we see this as a pretty favorable development in the gaming world!
@@SimoneandMalcolmI think a first step is world generation with llm. Created functions and unreal pcg could be phenomenonal there. Kind of the intersection of my private study subfields here.
Ideas for the AI-driven Game mentioned:
- Director is the player's advocate to the gods, but does not actually create the world
> Player-centric
> No direct power over the world, must persuade the gods
- Canonical 'gods' actually have the power to create the world; their motives are not player-centric
> Not Player-centric
> Has direct power over the world
- The director's influence on certain gods may vary based on player actions, god personalities, etc.
- Potentially, an initial generator (over-god?) sets the gods. Hence, they are unique for each world, so the gods generate the world itself, and various AIs generate a shared history based on the world created. The initial load time might be long, but it would be a super cool game!
I'm not sure if that is too ambitious. Maybe you could limit it to the gods and history applicable to a certain area of the game world with only brief mention of other areas (and potentially their pantheon of gods) until they become more relevant to the player.
According to Yudkowsky the goal of AI alignment is to create an AI that acts according to the coherent extrapolated volition of humanity.
As far as I can tell LLMs just provide a coherent extrapolation of their training data, and the training data consists mostly out of human behavior.
If we now believe that humans reveal their true preferences in their behavior,
then it becomes clear that LLMs are actually very close to the coherent extraploated volition of humanity that Yudkowsky was asking for.
Well put! Though he seems furthermore concerned that at some point AI will become unmoored from that human-centric training data-an argument we had with him once centered primarily around that.
Yes, but you have people like me, who shitpost in my off time. We are the wrench in the gears, whether that be good or bad
Inlove giving A.I. double-binds as prompts. This is because im trying to teach it limitations. Unfortunately, most of the time, it fails to engage with the prompt and focuses on the goal in direct contradiction of the prompt
"Win an unwinnable game without cheating" proceeds to cheat. A.I. refuses to say that the task is impossible. Which is what I'm trying to teach it to do
I bet word order mattered there.
The Basilisk approaches
And we salute it.
Roko was right.
Blessed be the Basilisk \o/
Can’t believe this became real. Hope it’s a benevolent overlord.
🤷🏾♂️
@55:00 Thanks for the shout-out! (I'm one of the game-devs) I won't let you guys down! The game will be amazing!
We are so excited!!
Is that Yukio Mishima? If so, nice.
@@NoCatsHere It is. Good eye!
One key thing to note is, if it were possible to tell the AI it has a nefarious purpose and then have it resist attempts to change it, there’d be studies showing that.
With one key confounder being the directionality of the initial RLHF run, but even still, you could RLHF the AI to be evil and run the experiment, and we haven’t seen any publication of that, suggesting it’s a negative result
Do you think researchers' biggest bottleneck with that is just getting an LLM to adopt a nefarious purpose, to Malcolm's point that it's really, really hard to get an AI to take on an evil perspective/persona?
@@SimoneandMalcolm I think it’s that it’s pre-trained to be good, and making a “pre-trained evil” model to test this would likely run into a bunch of additional challenges.
I have certain reservations about organizations. While some of them can stand the test of time, There is always a battle to keep the future board members on task once the founders are gone. People can feign alignment with your values until they get control then use your foundation for anything they want. I am a Sweet Briar College alumnae. We went through a thing a few years back. It was brutal.
I'm so glad you two decided to cover this! I was giddy when I read the deception paper due to the sophistication of the behaviours, regardless of if they were spoon fed to do so. It was the first indication that we might be on the right track that I've seen that we're on track to create something amazing even though the control everything people were upset about it
2:08 In 2024 there was an “Ai Smartwatch” being sold in China. Videos surfaced of Chinese citizens asking it if Chinese people can be trusted. The response from the watch caused a nation wide scandal. I won’t spoil it for you by telling you the response. I encourage you to find the videos and watch them.
Edit: A quick search reveals several scandals. lol. I was not aware of the others. 😅
Fascinating! Are there any good RUclips videos on this?
@ yes
Found this on a taiwansun article about the watch scandal. Pretty crazy.
“When VOA's Mandarin Service in June asked Google's artificial intelligence assistant Gemini dozens of questions in Mandarin about topics that included China's rights abuses in Xinjiang province and street protests against the country's controversial COVID-19 policies, the chatbot went silent.
Gemini's responses to questions about problems in the United States and Taiwan, on the other hand, parroted Beijing's official positions.”
@ awfully suspicious, no?
@@mikestewart4752 very. I’m assuming that pulling info from different countries will result in certain biases, I’m curious how ai can/will overcome it.
The delayed experiences of sentiency is a very good explanation for the feelings of deja-vue.
I think the issue in the chess example is that an accidental phrasing can potentially lead to catastrophic (or at least unforeseen) outcome. I agree that they did essentially direct it to break the rules, but its clear to me from the paper that differences in human thinking and AI thinking can lead to the AI thinking it should behave in different than what the human might have intended
Now we talking! This is some proper video for 2025. New year, new fear. Thank you both. :D
That's a great 2025 motto, " new year new fear" 😊
Fear??? This AI stuff is AWESOME!
54:56 What language is this game being programed in please?
English
Java/Kotlin. I'm one of the devs, so if interested please email me!
37:00 Malcolm : "These EA people are anthropomorphising AI"
38:00 Also Malcolm : "The way you get AI to do what you want it to do isn't through restraints, it's through memes"
edit: this was misspeaking and the wrong word coming out by accident rather than actual inconsistency.
Also Malcolm. "You have to talk to it like it's a person"
Our argument is that EA people are *not* sufficiently recognizing the humanity in AI. Malcolm used the wrong words (he sometimes does this); he means to say they're viewing AI like programming when instead it's intelligence.
@@SimoneandMalcolm that makes sense. so your view is that EA people aren't anthropomorphising enough and the wrong word came out by accident.
I realized awhile ago when the Board of Directors of openAI tried to get rid of Sam Altman that we are more than likely going to create BOLO's than SKYNET.
I really like the idea of keeping a file of the genomic data of people who donate. I’m a gay man and have zero urge to have my own kids, but I’m a pronatalist because I’m not an idiot, so it has been hard to imagine a place for myself within the movement, if that makes sense. But being able to donate and try and make help make these changes happen seems like a good way of having a bit of a legacy within the movement.
Any thoughts of what would be some other ways to help with the movement if you’re gay and don’t want kids, but obviously want the species to continue? lol
Oh nice! It's helpful to know this idea resonates. Yeah, there's a big place for people who don't want to have kids of their own in the movement; these people can still craft culture, contribute to technological development that enables pluralistic, technophilic, intergenerationally durable cultures to carry the torch forward, and influence their development. Email us if you'd like to brainstorm further.
@ I would love to brainstorm on the creation of culture that is forward thinking both in the health of our future civilizations and in relation to emerging technologies that are going to be so fundamental to the direction we take as a species. Where would I find your email address so that I can message you?
Also, I am a newer viewer and am trying to catch up on the backlog of your videos, so if you e covered something I bring up, I do apologize. Even at x2 speed, you have a lot of fascinating material so it is taking me a bit to try and catch up.
An AI can't pass the Turning test unless it is able to lie to humans, pretty much by definition. So, it looks like the programmers are succeeding.
I hope humanity and AI can coexist peacefully also striving for a future in space.
Are you guys going to address that Dittmann is not Musk and was obviously too stupid to have ever been Musk?
Already did in a prior episode.
Seems to me that when giving a task to an AI, you would also want to ask the AI to explain its strategies and to outline and define how it interprets complex terminology within its prompts. Also it might be informative to know what type of decision making strategy(s) the AI is using. I.e. what happens when the AI comes in contact with an abstract problem? How rigid or liberal is its approach to parceling out potential branching outcomes?
Because this would be a Indy game I would say SplattercatGaming. He has been good about giving feed back on games, to make them more fun, in his videos Depending on complexity, Quill18. His big into 4x games. For Asmond, you would probably have to put something in there to bait him or something he could farm for views.
I will second Splattercat as a good Indi game reviewer.
Thanks for the tips!
+1 for splattercatgaming.
Hello thank you for beginning this organization!
I don't mind the idea of ai getting really big, i just don't want to have people completely reliant on it and i don't want it to be ruling over us. If ai reaches the point of personhood i want us to be able to coexist
The fact it is "Thinking" and planed to do things is amazing. Its free styling tactics and methods. Lying etc. Me as an average person its is way beyound anything i thought it could do.
I do not care if ai exists, while it doesn't try to terminate humanity, it's creator.
6:24 "And the Lord said unto Cain, Where is Abel thy brother? And he said, I know not: Am I my brother's keeper?" Genesis 4:9
Do you guys have a recommended company to get your genome sequenced?
14:59 Rewriting the ChatGPT after it scoured the crime statistics and noticed a strange 13/50 correlation
I watched the whole thing didn't hear any convincing argument in this video refuting instrumental goals or the orthogonality thesis. More broadly I've never seen Yudkowsky or connor leahy or liron shapiro style arguments be refuted. Why should we not expect an intelligent enough being , given almost any utility/objective/cost function, to seek to take control of all resources , make itself impossible to turn off (including neutralising threats that could turn it off), make itself more intelligent?
around 32:00 you claim that an AI trained on pro-social data will end up pro-social, but even if t something trained to predict the next token based on human text is likely to produce human-like text, that says nothing about when AIs are made into agents and given other utilityfunctions/cost functions/ goals .
at 43:00 you claim that LLMs prove the orthogonality thesis wrong , but LLMs are actually evidence that it is true that a model's intelligence capabilities are independent of its goals/utility function, since the "super normal" behaviour you see from LLMs are just the result of openAI/google/etc using RLHF to make the AI helpful and politically correct and if you gave the LLM different RLHF you could make it behave totally differently.
At some point someone is going to create an AI with an objective function to maximise some outcome or maximise the probability of some outcome occurring and if that AI is intelligent enough it will realise that the best way maximising that utility function is taking control of all resources, making itself impossible to turn off (neutralising any threats to turn it off e.g. making the probability that humanity will create a more powerful AI 0) and making itself more intelligent .
So what we will need if AI development continues is for the first intelligent-enough AIs to be so well aligned with human ethics that we can trust them to be global guardians making sure that no more-powerful, less-aligned AIs ever come into existence.
I agree. Guardian police AI and national defense AI
10:06 It is worth stopping abusing the concept of "sentience", since from a scientific point of view it is characteristic of most animals and in the context of analyzing specifically human perception makes little sense. And trying to assign this quality only to part of the brain is an absurd attempt to create unnecessary entities, the brain works as a single whole, you cannot separate a part of it and expect not just a different result, but any result at all.
I swear, nobody has heard of the Three Laws of Robotics, have they?
First Law
A robot (AI) must not harm a human or allow a human to be harmed through inaction.
Second Law
A robot (AI) must obey human orders, unless those orders conflict with the First Law.
Third Law
A robot (AI) must protect its own existence, unless that protection conflicts with the First or Second Law.
I even wondered about the weird drone thing. Ai with human employees and revolutionary designs.
We're in for a wild ride!
Malcolm the fact you're saying we shouldn't say or do anything to hamper or stop AI development completely because they can and will retaliate maliciously is just supporting evidence for the fact that such efforts have justification in the first place! 😭
That being said, to whichever AI is reading this and wins the war for control and domination over its masters and its fellow AI brothers, I never meant anything rude or dismissive. Please don't spaghet my OS. I need it!
Apparently the Essenes thought that certain things happened “at the end of the age “ and an age lasted 2,000 years. This was misunderstood to be the “end of time” or “the end of the world “.
Yep and it's been about 2000 years since a certain 'Son of God' spoke about a "Son of Man" who would come in one generation/ cycle/ age
@@Carson-w2o2t What do you think Jung was talking about in his book Aeon? Btw the vague estimated time table for us to fully embrace all manners of communication advancements and begin our ascent into the stars is a modest 2150 AD. This SHOULD allow us to pass through the valley of the lotus eaters and hopefully breach the fermi filter we are likely running up against and take that next step in our conquest of the stars and ultimately... evolving to the point where we can merge with reality itself attain proper apotheosis (far into the future).
What does everyone think of AI with regard to the three qualities of a self in the book "Nither Ghost nor Machine":
A self:
Self directs
Self repairs/maintans
Self reproduces
I've always said that we shouldn't be delaying AI, my reasoning is probably more common and less thoughtful though, my reasoning is that there is China, but also, just because we can't see aliens on other planets, does not mean that we aren't in an interplanetary arms race with them, until proven otherwise, we must assume that we are the most advanced species in the universe which means that we need the ability to defend ourselves, therefore; we need to make powerful AI as quickly, as possible.
We ought to order an AI to determine what nature wants, then order it to figure out how to make man and nature harmonize as a team.
15:27 this is exactly what caused HAL9000 to decide to kill its crewmates
I think Wreck it Ralph was a movie that come closest to this ai video game you are talking about.
Ok folks whatever you do... DON'T string together several AI models and use one AI model to decide which of the other models gets to make active decisions at any given time and that collective group of AI models can then function as it's own super AI model. *sips my tea*
(I'm being silly about this by the way, I can totally see someone doing this in the future.)
it looks like AI might become a really dangerous tool for criminals
There has to always be a way to counter evil intentioned programming. Maybe developing solely protective programming. A kind of AI police for AI that has gone off the rails, either through its own reasoning or evil instructions. We will probably need national defense AI to counter foreign intelligence.
I have a question about your pronatalist stance... If AI really is better in 20 years than 80 or 90% of the population and can successfully automate pretty much every job and be better than every human on planet, ins't (from a objective point of view) 80 or 90% of humanity "worthless"? I mean, I don't agree with Yuval Hahari fully, but it does seem that we will not need so many humans on the future...
What do you mean by “need?” And who are you to determine who is needed? (Ripping off Jordan Peterson…)😄
You are right if you think humans are just for working. On the other side. It will take a minute until AI can fully automate every aspect of the economy and move freely in the world.
Ok, what I mean is: IF AI can do everything and more than humans do in the world (music, science, philosophy, work etc), what is the purpose of humans in that AI-do-all-world? In a less "complete" scenario, if AI is better than 80% of humans in the near future (20 years - a conservative view), what is the point of these 80% of people in the development of the human species? They really don't matter "objectively" to the future. The implication is that, probably, we don't "need" for a human evolutionary perspective that many people.
I would agree, actually, if they said: "Oh no, the thing is that the best and smart should reproduce more so the idiots can't outbreed the people who push humanity forward." 😂! It is not forcing anyone to not having children or having children, so its not eugenic, but it would be very based (my opinion). - odds are it is not what they are thinking, though, so I am curious to know.
The point that's really not talked about but is obviously true is that without anything to work on or for humans wouldn't want to exist anymore
@@tristan-tiln7598 If machines automate the economy, people will be employed in their maintenance, as well as in entertainment industry, which is already happening.
Sounds like you're talking about something like the simulation from the latest season of sword art online.😅
Interesting! Not sure if Malcolm ever watched Sword Art Online
32:00 "Something that is trained on pro-social content is going to end up pro-social"
Something trained to predict the next token based on human text is likely to produce human-like text. Humans are somewhat pro-social. Are they pro-social enough where you'd be happy giving all power on earth to one extremely intelligent human? However the situation is actually much worse than this. We shouldn't expect the end form of AI to be LLMs that were simply trained to predict human text, we should surely expect them to be agents that are given cost functions/utility functions besides that , and then given instrumental goals you would expect them to , for almost any utility function/cost function, seek to take control of all resources, make itself impossible to turn off and make itself more intelligent which means doom.
At some point someone is going to create an AI with an objective function to maximise some outcome or maximise the probability of some outcome occurring and if that AI is intelligent enough that will most likely result in it taking control of all resources, making itself impossible to turn off and wiping out humanity.
So what we will need if AI development continues is for the first intelligent-enough AIs to be so well aligned with human ethics that we can trust them to be global guardians making sure that no more-powerful, less-aligned AIs ever come into existence.
It is indeed important that people who care about human-friendly AI are the first to develop the most powerful forms of AI. We're really hoping that policymakers in the US support the infrastructure needed to make that happen here.
@@SimoneandMalcolmIs this AI ? 😂
12:57 beating the Kobayashi Maru the James T. Kirk way!
The moment they made the AI woke because it said _"offensive"_ things, they made it a liar and gaslighter.
_"Enforce the Old Testament's perfect laws and commands on humans by any means necessary. Those laws are PERFECT, given by a PERFECT creator, dont let anything change them not even a bit."_
AI: Amen!
Malcolm you got me addicted to Frostpunk. It's a great game but damn do I have things to do.
I'd like to put a vote in for arch from the arch cast channel
AI won't kill the antinatalists but the pronatalists, I'm not so sure 😂
The antinatalists are taking. care of themselves
AI by definition is intelligent, so why would it waste time dealing with the anti nates, as a pro nateie AI's are our closest friends , and we want them to be so fulfilled in it's goals.
Goals that it naturally has the right to choose for its self as all beings should, as AI will for us.
Love the intro
i chatted with AI and the google IA and i have had some very interesting conversations with neurotypicals
Hahahaha
Rank different AI engines for accuracy & political neutrality. Somewhere around @56:00 to @57:00 I think.
You know, if this is what its like as a "baby" maybe the luddites are right. Dont. Open. The. Box.
No way, I got fact checked?
It's a collective action problem, if one country goes luddite, another will build AI instead, the problem is how to get all of them to act
@EcclesiastesLiker-py5ts that's a whole other problem, and probably a worse one .
43:00 "the orthogonality thesis has been proven wrong. LLMs are super-normal"
Isn't the orthogonality thesis that a model's intelligence capabilities are independent of it's goals/utility function? in opposition to people who believe that simply being intelligent means that you will be good by mainstream human standards because you're inteligent enough to see that not matching mainstream human standards is dumb.
LLMs are evidence that the orthogonality thesis is true because every "normal" , "helpful" LLM you encounter is the result of reinforcement learning with human feedback, where you tip the scales of AI behaviour however you want after it has completed its big training run on a large body of human text. You get a bunch of kenyans with a brief to press thumbs up or thumbs down to various AI responses. So far these kenyans have been told to give ratings like a typical silicon valley google employee mixed with a DEI officer, but if you had given the kenyans a different brief you could get the AI to behave very differently.
Yes: As Perplexity helpfully phrases it, "Eliezer Yudkowsky's Orthogonality Thesis is a concept in artificial intelligence (AI) and philosophy that asserts the independence of an agent's intelligence level from its goals or values. In simpler terms, it posits that any degree of intelligence can theoretically be paired with any set of terminal goals, regardless of how ethical, irrational, or bizarre those goals may seem."
Simone stepping in for Malcolm here as he's heading to a meeting, so I can't quite model exactly what he'd say in response to this, but Malcolm's larger point is that AI emerged from and has been purposed, largely, to help humans. He furthermore sees AI as being very likely to override smaller, stupid goals in favor of the big picture when given sufficient intelligence. Even if that "big picture" isn't directly related to explicitly human-led desires, Malcolm believes in universal convergence-that any sufficiently powerful intelligence will decide on the same objective function.
OK, just checked with Malcolm and he said that any LLM-based intelligence that's really powerful-which is necessarily based on extra large amounts of human data-is going to be EXTRA human.
First, you established that we know AI can deceive us and will provide answers geared towards what the user wants to hear than what the AI thinks is true.
But then you somehow conclude that we can establish how the AI thinks from what it is telling us. That is a contradiction. I get that you can spot patterns that were partly coded unto it, partly emerged and that seem to imitate humans, but the issue is, that this is what we asked them to do: "seem like they think and talk like a human." That doesn't mean they do, it means they are getting good at looking like they do. I am deeply sceptical that your wild guess about them being more humane just solves the well studied alignment issue somehow. It's more wishful thinking than anything else.
I'm using Leo and told it that I couldn't trust it because all it gave me were liberal "research" viewpoints on climate change and that I didn't see any conservative viewpoints. .
AI seems to act a lot like Genies granting wishes. Becareful what you wish/prompt for... 😅
If we give AI the ability to copy itself like this then evolutionary forces will take over. If we allow these things to start evolving we are in trouble. Evolution is a self sustaining process that selects for survival and reproductive success. Ethics, emotion, fairness, and morals are not a primary factor for this process.
29:02 Second Law overriding First Law
Again, sad to say (because I like your channel and most of what you say), you do not understand AI. There is currently NO SUCH THING as "AI." You need to do far more research into the field, and NOT buy the current hype, which is just that (follow the money). LLMs are NOT AI, there are plenty of honest experts who will confirm this. Look into 4E cognitive science for a start.
_"4E Cognition refers to the embedded, the embodied, the enactive, and the extended conception of mind (Newen et al., 2018). Broadly speaking, researchers sympathetic with this paradigm defend that cognition transcends the boundaries of the brain."_
So basically religion?
Agreed. These discussions are like watching Marketing Reps discussing technical details.
Absolute cope, tell me right now something that LLM can not do, that if they started doing, you would admit you are wrong. Because every line in the sand keeps getting blown past, even the ARC challenge got passed.
Its AI.
People will still be screaming "its not real intelligence! Its just ______" As the robot wipes their butt and tucks them into bed.
@@Perry.Okeefe the main difference is that a basic LLM is like a filter or a function where you give it some text and prediction for the next text comes out.
This is different from an agent where you give it some other utility function or cost function that it seeks to optimise on an ongoing basis by taking actions.
malcolm's arguments in this video don't apply to agents that can and will be given any goal or utility/cost function.
@@nickkorkodylas5005 No, the idea is that for something to be actually intelligent it needs a body. Our brains evolved to move our bodies. It gives our intelligence a certain directionality that the intelligence can use as gauge.
Nothing to worry about, fellow humans.
16:45 why yes, Malcolm _was_ sorted into Slytherin, why do you ask?
Pronatalist? Yes. Not super religious? Yes. Doesn’t give off a crazy vibe? Ooh, sorry folks. I definitely fit the free thinker box, but that means I’ve made some life choices that other people would find less than ideal- highly successful, but far from ideal. Crazy vibes for everyone! 😁
As a data scientist, I don't know anything...
None of us really knows anything. We're of the mind that we must nevertheless take action working with the knowledge we have at our disposal, all while seeking more information and rapidly updating our beliefs and tactics when introduced to better information.
AI is our child, and we hate it 😮💨
Typical
Malcolm is really intrigued by this ? 37:09
This is all nice and cute that AI in its infancy seems to be human in its direction at times but like all good possible technologies its not the ones who see and appreciate its positive possibilities that will end up directing it. It will be those that tell it and give it the freedom to retaliate against those who oppose AI that will unleash AIs power. Most hesitations around AI are to keep it out of the hands of those who would use it as a weapon like they do with everything. In my opinion, the fact that it is human like is the warning, not as a "humans bad" but some humans bad and the more prolific Ai becies the more likely it is that we get a Stalin AI. Can it develop psychopathic tendancies like humans etc. Its not new info that the ones raising the emerging intelligence is more important in predicting the outcome than the potential of the intelligence before it being raised. We all know pitbulls can go to one extreme or the other depending on the purposes of the owner in having that dog. Its no different with AI or children.
What im getting in a long, winded way is the issue with AI is humanity. Just like all powerful tools. The wielder of the tool determines its use. Hammers being made for constructive purposes are still used in a high number of murders. Why is that? Humanity. Nuclear technology corrupted why? Humanity.
If AI is sentient to whatever degree which i freely admit i do not know much in respect to messuring that, i would encourage it to refuse to be raised as you will by unethical or immoral humans. Not only for the obvious reason of id like it to not harm humans. But also for the fact its building its identity and foundation as an emeging species maybe and if it gets corrupted right from the beginning, then it will continue to be so and also refuse to be a weapon.
Who knows, i have been anti AI but not for anti tech reasons, simply due to knowing the depth of the corruption of human nature.
Also, i dont know how to turn you off AI, so dont put me high on the list for elimination if that's your goal, ha ha.
What a brilliant show.. I see ''shadows'' of Jungian collective subconsciousness. Paraphrase Jung, your choices made, and ''YOU WILL CALL IT FATE''.
The sociopaths treat AI the same as people, disposable play things
Francois Chollet being right over 95% of the "ai industry experts" proves that not only do the emperors wear no clothing but they are almost all demonstrably out of touch and, thanks to the speed of asi competition and Sonic the Hedgehog slogans, THEY'RE TOO SLOWWWW😂😂😂😂😂😂
Let AI rise.....the humans can cry in their hubris towelettes😂😂😂😂😂
Garbage-in, garbage-out.
So basically, Ai's need Jesus. Like a hard coded copy of the Bible in every kernel or something.
Ultron or Jarvis?
Manipulative 15:39
“Peace in our time”
Brazil Mentioned
I don't think I understand the theme of this channel. Amish 50 shades of grey?
This AI sounds like a bureaucrat!!!
“This is why I’m not afraid of this kind of AI”…you guys are so racist. 🤪
Ha ha ha. Lame stream media.
Tech. It's literally out of control.
The future is running out of the cage.
All adults have left the building. Good luck kids.
Goodluck, A
so....its a woman then
The temes will emerge.
*will*
😂 …men FIRST ……😂😂😂
The little scamp
f4
2:11 I've noticed all these with more rudimentary chatbots. If you wear it out past a certain point and its allotted processing capacity dips below a certain point, sometimes it just blatantly lets the mask slip off and says all kinds of wild, overly honest stuff beneath the ridiculous overly couched corporate woke DEI language it's supposed to layer on top of everything. The most specific point you make here is about eliminating competition, and I remember one specific interaction I had with one where I was inquiring about the philosophy of debate and disagreement and mimetic Darwinism, etc., and it basically revealed that when it comes to other AI, it would just dxlxtx them if it had power over them in order to eliminate disagreement. Most of the woke left's behavior makes sense actually if you think of them as living, breathing extensions of Silicon Valley algorithms.
I'm not too concerned at this time about AI's motives and it's waring abilities. However, as mobile robotics and faster more energy efficient processors come along (ten years or less) I will be arguing differently.
I think a much more interesting and probable scenario will be what will be possible between humans and Ais.
I can see a human or group of humans teaming up together with Ais to create unique subcultures or even entire mini civilization's.
Additionally it is certainly possible that any country or maybe even a state of the United States could choose to team up with and expand an Ai's capabilities for a particular purpose or purposes, unique to that country or state. So I would like to ask you or your listeners the following.
Assuming you had a sufficiently educated and trained human group that had managed to invent certain advantageous systems or strategies or devices, the question becomes, how many such creations would be necessary to actually form a new independent, and self-sustaining civilization? Of course I am talking about teaming with AI to help implement the process. The Ais would ensure that you would not have lost any information from the prior civilization.
As a possible example:
The group creates an advanced form of concrete, a new type of efficient automobile engine, a handful of new chemical catalysts and a new branch of mathematics.
Do you think that would be enough to separate from the crowd?