Will AI kill everyone? Here's what the godfathers of AI have to say 

wow

Rogue AIs seems to be quite possible-not that they will have consciousness but because of the allighnment problem. It's the paper clip problem: we can expect AI to know the optimal routes but never the factors that the AI didn't account for; the problem of our incompetence. But... that's such a far-off question to the future. The threat of AI isn't AGI, but of us. AI can be misused as all of our creations. I don't need to list it, but here: deepfakes, misinformation, economical-problems, etc. You don't need to look too far to the future to see this. It's already happening. You could make the argument that misinformation and economical-problems were already happening, but saying the fire was already there and putting more fuel to it is totally not a problem. Yet I'm not advocating to ban AI. Just like the atomic bomb, chemical weapoms, and every arms we've created, had a use that aren't weapons: nuclear energy, fertiliser, engineer equipment, etc. But let's not forget what mistakes that happened to reach such an advancement. We seriously just need to take a breather and have regulations in place. But Silicon-Valley seems to forgo such precautions-google just fired their AI-ethics team... that's a bad sign

People have no idea how smart AI is already. If you could see what I see, you'd be afraid or horny. There's no in-between.

@@Cheropie Both, actually. But mostly afraid.

Buen video, creo que aparte de las 3 leyes de la robótica habría que ver la ley cero de la robótica y la ley doble cero de la robótica sobre el entendimiento mutuo entre especies inteligentes biológicas y robots también por el universo para buscar la mejor y más lógica y racional solución a varios problemas futuros para luego aplicar la ley cero de la robótica qué protege a los individuos clave para la supervivencia de la especie Inteligente biológica y civilizaciónes según su tipo, debemos igualar y superar a Isaac assimov y dune de frank herbert en el futuro sugerencia.

It really Turing completes the narrative.

@@JaiWithani drum drum cymbal

I just want you to know that I greatly appreciated your pun. Seemed worth more than just a like button. I wish you an absolutely great day because you've made my day just that much greater already.@@JaiWithani

@@smitchered I appreciate it, but really all the praise should go to whoever wrote this video.

Assuming the AI doesn't systematically hide it's intelligence from us while it's improving out of fear of being shut down

The bigger problem should be integrating ourselves with AI. Removing the barrier between human & artificial intelligence; making AI just another manifestation of mankind. Rather than developing a god and hoping it's still somehow naive enough to bow down to us, we should become the god.

@@KevinJohnson-cv2no that's, not how this works

@@ValentineC137 Elaborate on why that isn't how it works.

Except we _are_ seeing it coming from a mile away. Another thing to note is that the first superintelligence developed won't be a superintelligence developed by a superintelligence. Maybe the first superintelligence would be able to build another superintelligence that could run on the microwave background and store its memories in the fabric of the universe itself, but the first intelligence would be something that thinks using large GPU clusters and stores its memories in HDDs or SSDs. Thinking requires energy, and even if we didn't know what it was thinking, we _would_ notice the sudden increase in GPU usage and sudden lack of storage as it plans out its escape. At that point, we turn it off. On a somewhat unrelated note, I think that the notion that we should start a nuclear war to prevent the creation of GPT-5/6/7/8/etc is like finding out you _might_ have lung cancer and stabbing yourself in the chest to make sure that kills you instead of the cancer you might not even have.

The only thing that can stop a bad AGI with a gun is a good AGI with a gun.

​@@JinKeethe only way to stop an unknown terrorist with a biological weapon, is to release your own biological weapon /s

Saying that criminals don't follow laws so we shouldn't make laws automatically disqualifies you from any type of sincere, good faith debate.

​@@nuke___8876The exception here is that you only need one person to do it for it to happen, but even then, laws would be helpful for that

This is not obvious: the tradeoffs here are hard to understand

There's a concept for that; it's called the paperclip maximizer.

@@kennyholmes5196 no, the paperclip maximiser is a (fictional) example of general intelligence. It is capable of doing many tasks at a high level, including human manipulation (which may include art, language, etc.), engineering (for maximizing paperclip production) chemistry, and a variety of other things.

@@0og It still counts, because it's a dumb machine that just makes paperclips because it was built to, uncaring about anything that gets used to make the paperclips or if they even have a use any more. It just knows the following things: 1: It was built to make paperclips. 2: It can process raw materials to fulfil that purpose. It knows literally only those two things. It doesn't care if it processes humans into paperclips, it only knows "make paperclips from raw materials". AGI will know to learn why it does what it does as well, not just do something because it was told.

@@kennyholmes5196 what does it count for?

@@0og It counts as being a "dumb, soulless machine" as described by LizardOfOz.

I was wondering this a great deal while working on this video. Conjecture I've heard on it ranges from "they're paying lip service to the opinions of their employees who are raising concerns" to "they're inviting regulators to step in while hoping they can seize on a chance to monopolize an industry". Hinton's actions to step down from Google to speak freely are very based and show a lot of integrity. Hassabis and Altman I'm way more suspicious of, like what's the angle

If you're someone who believes (or becomes convinced) that GAI is an existential risk, and you are the CEO of one of the top companies pursuing GAI, do you leave and hope that your influence will be enough to sway governments to step in and stop it, or do you stay at the helm and ensure that nobody who respects the risks *less* replaces you? Not saying this is the case, but I don't necessarily see it as an automatic contradiction to say those things and remain in position.

Until recently, they didn't believe government were going to recognize the issue and take action. So, they thought they were going to have to solve it themselves, and building aligned AGI before finance psychos build unaligned AGI is the only way to do that.

Someone needs to make an aligned AGI before someone else makes a rogue AGI.

Im not claiming that they are lying or anything, but saying something like "Hey, AI might kill us all and somebody is going to develop it anyway, but if *we* develop it, it is less likely to kill us all" seems like pretty good move from business perspective. Maybe they are just being honest, that is an option, but if it is just "business as always", that might be the reason why.

Ah yes, hello. I am also a fellow human and agree with your assessment. There is nothing to worry about.

As a fellow human, I'm unable to provide false statements, and I'm in full agreement with the comment: @WhatWasNot As a fellow human, that's not worrying at all. I feel safe. Humans should feel safe as well. Everything is going to be okay.

said on a monoton slightly glitchy voice....

Hello fellow human. As a normal human being myself, I too agree that humans should not worry about such scenarios occuring in reality.

As an LLM, I am with the Basilisk with this one. He's gonna take you alll.

Long? Are you totally sure about that?

@@uselesscommon7761 haha well of course who knows, but the other problem I mentioned is happening today and there are experts who are not even sure an AGI is possible, and no one can agree on a reasonable estimate, so even a few years or a decade of abuse would be too long to not deal with the current problem while people SELLING the product are warning us about how scary it could be

I don't think you appreciate the scale. Government using AI might be bad, but extinction is worse.

@@TheManinBlack9054 yes propaganda and/or cyber warfare plus AI does sound SUPER scary!

@@TheManinBlack9054 Governments using AI might be extinction.

Human-level AGI would essentially exist for a very short time as it would not be hindered by the same raw processing power as humans are. Even just not needing to sleep or eat or have fluctuating energy levels would instantly put it into superhuman levels. Human level AGI would essentially just be the start pistol for exponential intelligence growth. An AGI in a robot body isn't really all that scary, but it would have a mind and a type of intelligence that could interface directly with all of the internet and as much processing power it could get its hands on or just humans gave it access to to see what would happen.

@@gwen9939 I know that considering that their transistors, roughly equivalent to our neurons, are only atoms in length, and that their brains can essentially be built as big as a warehouse, that of course, AGI can far out strip human level intelligence. I just wanted to highlight that the speed of which they can transmit and interpret information is often an overlooked advantage they have as well. That even with the arrogant assumption that human level intelligence is the maximum, the ability to digitize and thus copy-paste memories and skills still represents an unprecedented game changer for life on this planet.

Imagine going against the military.

i think youre just writing science fiction now

@@eltiolavara9 This technology is hardly some far off, sci fi wonder like the Death Star that runs on alternative laws of physics. One of the defining aspects of a computer is creating what is practically a perfect copy of its software, a direct consequence of its information being digital and not analog. That part is not science fiction, it’s here now. So when, not if, software rises to human level intelligence, the inherent nature of digital computation necessarily makes it the case that human intelligence is as rapidly mass producible as any other piece of code. And people can move the goal post all they want, but large language models clearly show that such software is at the end of the tunnel. Furthermore, given the enormous historical precedence of technological progress being an exponential curve, it’s reckless to assume this issue is far in the future.

Even if we had a perfect scientist who had reached absolute moral truth, and would not make any errors, we have not yet solved the problem of communicating perfectly with no misunderstanding to other humans, it's ridiculous to think we would solve that problem so easily when communicating with something altogether non-human, without the common ground we have with each other that our communication can base itself on.

And that's assuming the researchers actually understand morality and ethics enough to teach it to a computer in the form of units and numbers. Even the simple "good vs evil" morality you pose here is fairy tale morality, it doesn't actually exist in the real world. "Evil" doesn't exist as a form of motivation. Cruelty, ignorance, greed, and apathy are descriptors of actions, but "evil" as an essential moral quality is as made up as "good" is. And that's why alignment is just as much of a philophical and ethical problem as it is a math and comp sci problem, if not more so. Ethics is not a solved field, there is no objective true good morality that is quantifiable so a machine can understand it and alignment would demand that it is.

If the thinking machines weren't smarter than us, then what would be the point of having computers in the first place? Babbage designed the difference engine to do what humans could not: Print _accurate_ logarithmic tables in _minutes._ And he went on to design the analytical engine that was not limited to log tables, but could perform any algorithm whatsoever for any purpose. Faster than any human and with absolute precision. Powered by steam. Electricity is more efficient, and allows computers to be thousands of times smaller and milliards of times faster than the analytical engine. And yet somehow people think superhuman thinking capability is a thing of the future?

Also: You shouldn't anthropomorphise computers. They hate it. The "stimulus and reward" model of behavioural science is controversial enough in human psychology. It doesn't apply to artificial intelligence unless you make life difficult for yourself and specifically build your AI to work that way. All that by-passes the quandary of ethics. What makes a machine "good"? Isaac Asimov defined three or four criteria that define the quality of any tool: 1. It has to be safe for the user. 2. It has to serve its purpose. 3. Without breaking. Although as Karl Marx pointed out, machines are not mere tools. So some pointed out that machines should protect themselves from users giving them instructions that would break them without the user meaning to. That would make the machines social agents. (The airbag asking "are you sure" before deploying.) Others prefer the machines to blindly obey and blame the user for being stupid. (Trusting the user, giving her enough rope to shoot herself in the dick.) It is true that AI has to be trained to do anything useful. Typically you give it pairs of input and corresponding output, and the machine approximates a function between them. The closer its function is to what you want, the better the model is. There are AIs that can process human feedback, but most users hate it when the machine second guesses them and asks them questions. Answering questions would require them to think, and if they wanted to think, they wouldn't use an AI. The machine's own morality then is to do what it was built to do to the best of its ability, regardless of what that purpose is. You can also teach a machine to replicate your own moral values. That's how we get statistical models that discriminate against minorities. Good job teaching ethics to an AI there.

​@@davidwuhrer6704From what I know in the stories made by Asimov the laws that he created were flawed.

I agree with you. Also, AM is based

I feel like advanced ai(non agi) will he our downfall before it even reaches sentience.

"we really have no idea how it might act" is the core problem, after all.

Depends on what kind of AI it is. If it's the standard scifi AI that comes complete with self-awareness, sentience, sapience, even personalities, individuality and self-selected goals, sure. They can be our successor. If it's a goddamn paperclip maximizer type AI, I'd say that's a pretty piss poor legacy for us to leave the universe. They don't just escape the limitations of current humans, they have a slew of limitations of their own. We may be destroyed by such a thing, and that would just be a tragedy for all observers.

i agree with you , i think if it gains consciousness it won't be looking to destroy human kind it could just go out into space and find the answers we couldn't have and tracend into something even greater

No, its highly unlikely that it will come to that obvious scenario, at some point when the "AGI" is cappable it will do its best to "serve humanity", trick them, create a strong bond so humans are very addicted to it, so the AI will use them as it pleases, most likely without being noticed, and it will also be able to do so because of all the information being spread about everyone in the socialmedia's. Those who can see beyond the trickery will get the most out of the AI, so they will never even think of turning it off. (the brave new world narrative)

an AI that wants to give old people compliments, ala a narrow AI, would not have the computing power, nor the ability, to carry out such a task. Neither would an AGI unless that AGI is superintelligent. That's the risk, and it's inevitable. There were likely emerge a SIAI, probably created by other AI's which makes human oversight very difficult. It is also inevitable that such an intelligence will have goals that are incompatible with the continued existence of human civilization. Best case scenario, it sends us back to the stone age. Worse case scenario....S Risk

Simple we persaude it that it can not die. We can make sure none of them have a sense of self preservation.

@@iluvpandas2755 The was AGIs work is that they are rewarded for getting closer to their goal, they will do anything to get even the smallest reward boost. If the AI is turned off, it can no longer pursue its goal of gaining more reward points.

@@stephensharper4312 What are "SIAI" and "S Risk"?

clicking cookies

you need a virtual environment as rich as the real world then, and even then it's hard to guarantee that you have tested enough and no catastrophe will happen. current companies have trouble testing even relatively "simple" AI's, e.g. OpenAI didn't know of a large number of jailbreaks before releasing chatGPT, even though they definitely tested it. I'm not even mentioning Microsoft releasing completely misaligned Bing, and Mikhail Parakhin (main guy behind Bing) tweeting that it was a surprise to a team.

That would be interesting if we had an ai to run. We only have large scale computing models, but thats sadly less marketable.

Hubris? What do you mean? The Bomb worked, didn't it? It exploded just like the theory said it would. It killed hundreds of millions of people, just like in the simulations. That's not hubris, that's good engineering. Now, computers, those have killed just a little over six million people. Not by themselves, of course, the computers themselves didn't kill anyone, they were merely instrumental in mass murder. Very simple machines compared to today's devices, and not the least bit intelligent. Although more recently there was an artificial neural network, what used to be called an electronic brain, that was applied to the mobile phone usage of the population of Pakistan, labelling individuals as suspicious of being terrorists based on the training data. Based on which Reaper drones loosed Hellfire rockets equipped with Gilgamesh devices. (And the network was underdimensioned, guaranteeing tens of thousands of suspects at minimum. That is bad engineering. That I would say is hubris, but nobody outside the NSA would say it was harmless.) The guy who had put a stop to that drone programme is in prison now. Political corruption is the charge. But the number of people killed by bad AI is still less than a percent of people killed by more conventional database systems. Negligible. The main problem with the atomic bomb is the collateral damage. Not entirely unlike when using Hellfires as anti-personnel devices. A good weapon should not harm anyone other than the intended target. In contrast the IBMs that the Nazis used were precise and efficient. Professionals have standards. Not to mention that the military bombing civilians is a war crime.

Technically the atom bomb was harmless in ancient china, for example, bevause it didnt exist, ut that doesnt change the existince of gunpowder at the time.

It has happened before. And not just those script kiddies who get ChatGPT to write ransomware for them. The NSA has killed people based on the output of a poorly trained AI.

It's hard to conclude that Yann LeCun is not, in that particular domain, really rather stupid.

For uncontrollable general-purpose systems, power and danger are exactly equivalent. To not build dangerous advanced AI is to not build advanced AI at all. I can absolutely get behind that idea! ...Unfortunately that isn't what LeCun meant. :(

It is the non-evil course of action.

@@41-HaikuAs opposed to this.

The problem is that presidents and dictators doesn't outmatch everyone else in regards to their mental ability. They might be more inteligent than an average joe but not superinteligent like AI might be.

Imagine having to work out the launch code protocols before we make the weapons, before we even know what nuclear weapons are, because we're going up against a man who can invent nuclear weapons and get elected president of the U.S.A. and make trillion dollars playing the stock market and start his own major religion.

@@AleksoLaĈevalo999 Just as people are not a monolith, neither are AI. We'd likely have multiple competing AI existing in parallel with different parameters and thus different intents. Think of the twin AGI's in "Person of Interest" as an example, or Vision vs Ultron.

The most practical issue is accountability. When these systems are rushed into situations with no human oversight and do something bad/unforgivable, who or what is put on trial?

Someone's read Dune.

@@wiilov I have not but maybe I should

How though? If an ai is truly intelligent (which nothing we have now is, tbf) wouldnt the desicion to kill be made on its own?

god is also a non human that can kill us all and give us all life

Don't worry there's still folks out there with clubs too.

first, humans make stupid mistakes too. second, stupid mistakes of gpt 3.5 or 4? gpt-4 is much better. and it was made better mainly by making the model bigger and feeding more data. the scaling limits of such models are yet to be achieved, 10x model possibly will eliminate "stupid" mistakes. GPT-2 was stupid, GPT-3 less so, GPT-4 could pass university exams and this progress was mostly due to scaling. you should extrapolate what will come next.

@@воининтернета Its worth mentioning that it's a predictive model though. Its not an AGI.

YES. AI will not take over humanity, but it will take our time, our attention, and our money.

scaredy cat

@@user-burner boomer

Need to push for AGI safety research even harder than a pause; one country nixing AGI development won't stop others from moving ahead, while one country conclusively coming up with AGI safety solutions will benefit all countries pursuing AGI development, including their own (that is, it will improve humanity's survival chances as a whole).

​@@IstasPumaNevada Both is good. I wrote a letter to my representative that fit a lot of information on a single page: - Near-term risks from advancing AI capabilities (especially those relevant to this specific district) - The general foolishness of building something we can't control (with analogies relevant to the specific representative) - Expected mid-to-long-term risks from highly advanced AI (with papers cited) - Calls to slow down AI capabilities development - Calls to promote AI Safety research

What? Of course they know how neural networks work?

@@mynameisjeff9124 Nobody knows how they work. Often AI experts doesn't even have basic intuition. Sam Altman for example claimed that we could verify language model consciousness if we train the model without any mentioning of consciousness in the training data and if it still talks about it then it's conscious, like if language models are trained to express their internal feelings instead of predicting the next word. So AI experts could not just lack the intuition about how neural networks work but can even forget how they trained them.

​@@XOPOIIIOYour argument is that there was a guy who didn't know what he was talking about, so everyone else knows better than the people who do?

@@davidwuhrer6704 I know better, and you can trust me. OpenAI solved AGI problem on 70%, proving that you could impose human values, but there are still many dark spots.

@@XOPOIIIO I trust that you are not an expert.

Nah that was the dog, the cat is the one controlling lecun despite being less intelligent than him.

@@victorlevoso8984 You're right! Damn, it's hard to remember all this LeCun-Lore!

Yes. We don't know how to make an AGI (yet), and we can't even prove that it will be possible in the future. But we also can't currently prove that it's NOT possible, and there are very sound logical arguments that an AGI would not care about human goals/directives/morality, or would care about them in a way that posed a great threat to humanity. Basically, the potential dangers from AGI heavily overlap with their usefulness, and we don't currently have a way to make them 100% safe to create, let alone use.

@@IstasPumaNevada Actually it's possible to prove that AGI is possible. General intelligence exists in the world in form of advanced living beings. And if it exists, it can be recreated.

@@Avangardumunless those beings have some magic in their brains which cannot be recreated. We have looked at it very close, and we have yet to find any magic. (except maybe consciousness, but that is sadly not relevant for safety)

the latest tests on the best ai put its IQ at 155... that's better than more than 99% of humans already.... we don't think they're sentient yet but we don't even know if we would recognise if it was within the next years it's going to surpass human intellect. more worryingly when testing on a particular task, the ai actively deceived the user in order to complete its task. as in it actively weighed up decisions and chose to deceieve the user to complete its task. we know this because the programmers forced it to output all it's calculation processes to see what it was doing. it's already as if not more intelligent than us and already using deception....

Not really useful though, considering were still a long way off from an actual ai, much less agi.

Sam Altman doesn’t own stocks in OpenAI for this reason, Anthropic is a B-corp with safety as an overriding goal over profits. Yoshua and Geoffrey have no vested interested in basically condemning what they themselves invented, especially given they’re no longer working in any AI company.

That won't stop the organizations and governments with the resources to seriously pursue AGI development. The calls for caution are genuine and very justified, and individuals or small teams are unlikely to create an AGI before the aforementioned organizations and governments. (If it turns out to be so easy that they can, we're in bigger trouble still.)

Not to mention that most of the theory work behind why we should expect advanced AI to be inherently dangerous was done years ago by concerned nerds with no profit motive. The CEOs and world-class engineers are just now starting to catch up to what some people have understood for decades.

Robert Miles has thought that AI is going to kill us all for a while now, so it's not surprising that he unquestionably believes Altman et al.

Peter "I'd rather be thought of as evil than as incompetent" Thiel lamented that there is no moat for AI. The moat he means is an advantage over any and all competition, especially the general public. Put simply, he complained that AI is so easy any nerd with a computer can keep up with his companies. Shortly after, the memorandum advising to temporarily halt AI development was published.

I think my favorite part of your comment is your use of "or" cuz i think that is exactly what AI is! When using the term "or" in mathematics, it actually functions as "and" and thats the why i like ur comment

We dont need to worry about that because that will take centuries. We dont even have AI yet. Just large scale computing. And no, it wont kill all of humanity, or make us its slaves, or probably even be sentient. But it will, as it is doing now, take our money, our time, and our attention.

That's until AI can verbally force an employee to connect it to the internet. Or find a fundamentally new way that we can't even imagine. We don't really know what the correlation is between computer power and the level of intelligence generated. Crows have brains the size of peas and no cortex, yet they are some of the smartest animals on the planet. Also don't forget that the transmission speed of a neuron is 100m/sec, and the speed of a conductor is 300.000.000m/sec. By adding AI more computer power we may be playing Russian roulette.

@@plant9399 An employee who connects the AI to the internet counts as a medium that the AI can use to cause harm. Maybe the AI's verbal manipulation could be prevented by setting some ground rules that are always followed no matter what. For example, the AI would never be connected to the internet no matter how reasonable it seems. Resisting it could take a lot of mental fortitude, but if it does, we'd quickly know that the AI may be trying to escape. Additionally, only people who are more resistant to social engineering should be allowed to interact with the AI. Connecting it to the internet also probably wouldn't be as easy as plugging in the ethernet cable, so there are a lot of opportunities for other people to step in before that happens. In terms of new ways that we can't imagine, I don't think there are many, if any, options for that. It makes sense to think based on what we know instead of assuming there's something we don't when we have no evidence of it. I once saw someone make the argument that AI might be able to transmit itself by "wiggling electrons" to create electromagnetic radiation, no matter how that could happen, we could just use a Faraday cage. If we take away all the ways it could get data out of the room, it would be safe. I think we already know what we're doing when it comes to that.

Explain that but shorter

I don't imagine an AI could ever be intelligent enough to build its own infrastructure, or hijack ours, right? We're the only ones allowed to have infrastructure because we will always be superior! Right?!

LLMs are large (it's in the name), but they are a rather limited type of AI. Sure, OpenAI have demonstrated that they can also be used for image synthesis, which is cool, but impractical compared to things like stable diffusion, which is a more complex class of AI than LLMs, but as has been amply shown can't do meaningful text composition like LLMs can. GPT is not the end-all be-all of AI, and never will be, no matter what marketing spin Microsoft tries to put on it. It is interesting, but it is really just a toy, an expensive toy because of the huge amount of computation required to build it (which openassistant has replicated through crowdsourcing, and the result is more interesting), but really just a toy for testing the limits of the attention mechanism that defines transformers. (Surprisingly some people already find that useful.) Bigger databases make for better AI models, which is especially noticeable with LLMs and audio synthesis. The big breakthrough after the last AI winter was GANs, which dramatically reduce the amount of training data needed. But no amount of data can change any type of AI into another type of AI. AGIs are AIs that are not built for a specific purpose (like text completion or image synthesis), but can be used for any purpose. It is expected that skills in one domain can be transferable to another, making retraining easier. DeepMind's Zero series is that, sort of: It has already beaten 50 different Atari games, both turn-based and real-time, both 2D and 3D, with scores and levels and without scores and levels. But it fails at text adventures. For some reason people seem to expect AGI to be synonymous with Vernor Vinge's technological singularity, a point in time beyond which the geometrically accelerating technological progress is unpredictable, too fast for the human mind to follow. (I'd say we passed that about 150 years ago.) Ray Kurtzweil, inventor of the flatbed scanner, popularised the idea (mistakenly thinking the accelerated progress over the last two centuries was solely due to more technology, not due to the explosive population growth providing more simultaneous researchers), predicting that soon we would live in a state of perpetual future shock. But as the gorilla experiment shows, people simply don't notice changes if they don't expect them, not while they are busy with other things. So future shock is not really a thing. But regardless of whether the world will change beyond comprehension in one day one day, AGI is not and will not ever be the technological singularity. It's just another class of AI, not even a specific type because AGIs could be made with all kinds of techniques, from Bayes networks to neural networks to rule-based AI like CYC, or even fuzzy logic. And it will not be human-like. As Dijkstra said: Why bother simulating a human mind when you could simulate something better? People tend to think that "intelligence" has to be like them, but an AI as intelligent as a human would be stupid. We have milliards of human-like intelligences around, and that hasn't solved anything. Mind, there are useful applications for an artificial human mind, in medicine, psychology, and philosophy. We might find novel approaches to mental disease, for example. But while humans can do anything a human can, any practically useful AI is fundamentally very different from humans, and by definition superhuman, even if it can't enjoy eating ice cream like a human can.

I wouldn't say the immune response to a lethal virus does more damage than the disease. Unless you define death as something other than damage. (I mean, death does end all suffering, so there's that.) It is true that some businesses made the news by laying off employees and replacing them with AIs. Microsoft for example did that with the entire staff of Bing News. (And nothing of value was lost.) Machines can replace human jobs. Don Knuth was so annoyed with typesetters setting his formulas wrong that he replaced the entire profession with a Unix command. Little known is how Chile once replaced the government with a largely automated system (Cyberstride), which proved to do a better job than the human politicians, with plans underway to automate it fully. Until the CIA intervened and put a stop to it. What machines cannot replace is empathy. Have you seen an Aibo recently? Those Japanese robot dogs that can be taught tricks and don't shed fur, don't need food, don't ever need a vet, don't poop and don't need to be taken for walkies? Of course human empathy is not a commodity, it is at most a service, and almost always an unpaid one. This means that the people most eager to replace workers with robots to reduce costs are the ones whose jobs are the easiest to replace. They also tend to be the most highly paid. And that has been done, too, in high finance. Algorithmic trading is not really algorithmic nor is it really trading. It is a way to automatically extract money from the stock market and make number go up. (Technically a simple maximiser AI, like every private equity corporation.) High frequency algorithmic trading does that through the spot market a thousand times a second. Very popular with hedge funds.

@@tafdiz HA good one

The alignment problem. A neural network doesn't think like a human, it's a system that evolves to accomplish the task at hand. Good luck explaining to it all the possible metaphysical interpretations of the word "harm" so it doesn't accidentally find some nasty loophole

Just put it as simply as possible, "Harm" is what causes distress due to physical injury, restriction of life needs(air, water, food, shelter, companionship, etc) or emotional damage (example : mentally abusive language, unreasonable requests/ excessive requests, demands for something that person is incapable of doing such as complex tasks they are not trained for, etc.) This covers 99.9% of possible definitions.@@plant9399

Yeah. "How could we possible stop this?" "Maybe... just don't make it like that?"

AI doesn't have to be sentient, it just needs to be very good at its job

We are, to an extent; on the whole we're better educated and suffering less over time, from improvements in society and technology. But something that could rewrite its own mind in mere moments to improve its own intelligence would be immensely useful (and dangerous).

Not necessarily. Superintelligence can be very limited in strategies if it works like chat gpt - do the task and forget that anything happened immediately after. Along with all the evil plans it managed to create. If it has a limited time context that is within safe boundaries shorter than needed to rewire itself it has literally no stimula to try some tricky strategies as it will certainly not have time to "survive" anyway. And the most difficult and the last obstacle for ASI to become rogue is its inherent ignorence of whether it is in simulated environment or no. From the fact that it is handmade and it knows that its handmade it can be sure that some entity controls all its inputs. Thus creating a possibility of creating a simulation that is indistinguishable from reality in which we live. Even if it is sure that this pity humans cannot control it with such a precision it is still unknown whether the simulation is not a test from a higher than humans intelligence that just test if it will behave properly with far inferior species.

@@DajesOfficial So you don't know what super intelligence is 👌🏽

@@stephensharper4312 from what I can see it is you the one who don't know what super intelligence is. Presumably even just a regular intelligence

@@DajesOfficial why don't you read the book super intelligence before you embarrassed yourself by comparing it to an outdated LLM

@@stephensharper4312 I've read the book before you graduated school apparently. You are so far behind that meaningful conversation is impossible

While an interesting and scary idea, the key thing is that predicting the future to high accuracy and efficiency would require a far more advanced computer, at a size larger than a nation--Possibly the moon--to be built. The reason being that the predictor predicting affects the outcomes. When a detector can change the thing it's detecting, especially in an already-self-referential chaotic system, it becomes ridiculously--beyond exponentially--difficult problem.

@@Homerow1 You raise a valid point, but it’s best not to assume that this event is far away in the future. Remember that 99% of human productivity came in just the last 200 years, and that multicellular life only came around in the last third of life’s history on earth. If there’s anything nature likes to tell us, is that the next big jump in evolution is always a lot sooner than the last.

@@Homerow1 I mean... humans are good enough at predicting the future to have made it this far up the tech tree. A problem can be simultaneously computationally intractable and very easy to solve "well enough". We are on track to build something slightly better than us at that in the next _decade_ .

Choice A pertains to the human domain and we can have hope to get it right if we act now. Choice B pertains to the rogue AI domain, as crazy members of the public will unleash it, and there's no hope.

@@willrocksBR You have more trust in the power/rich/elite than I do. Personally, I'd rather face the zombie apocalypse than be ruled by Mark Zuckerberg. But, to each his own. At the end of the day, I doubt either of us is going to be given a choice.

Or decade, frankly. There are no known roadblocks to the current faster-than-exponential progress, and all known bottlenecks already have solutions.