Meta announced theyre releasing AI "people" onto Facebook and Instagram now. You should discuss this in your next video, Dead Internet Theory is not a theory anymore.
From the sound of things, they're just adding a few extra "characters" like the Meta Assistant that's already in Messenger. They'll be characters specifically prompted to specialize in certain things, like recipes or travel planning.. Like messaging an expert, without worrying about paying them or wasting their time. They're not aiming to fill their platform with fake profiles, as that would be counterproductive to their whole purpose. And a LOT of the talk about "Dead Internet Theory" lately, has been based on a faulty paper that came out a few months back, claiming over 50% of the internet was AI content.. HOWEVER it lied, and was counting automatic language translation as AI. You know how some sites have a language dropdown option in the corner? I don't think that counts as AI. Cause that skews the percentages a LOT, when the English site counts as 'real' but all the other languages count as AI content.
@@GrumpDogwell, the unregulated development of AI combined with big tech ways of gathering training data prompted MANY to take their creative (or not) work off the internet completely, effectively increasing the percentage of AI (or ML) content by decreasing the denominator. Also, people 'prefering AI friends' over real ones has been a scifi-drama theme for over a decade now, and no real regulatory effort was made to _prevent documentaries_ on companies seeking profit at any trade off, like Microsoft!
@cccyanide3034 information that can cause harm to people or other sentient beings if it becomes known. like on a national level telling the world classified information.
@cccyanide3034 An information hazard, or infohazard, is a risk that comes from the spread of true information that could cause harm or enable someone to cause harm. Philosopher Nick Bostrom coined the term in a 2011 paper.
Being loudly concerned seems quite the right path. It would be a shame if Big Techs would have ways to manipulate media to hide that data since concerned people may be also bad for business.
The AI that is meant to replicate Cyn from Murder Drones nearly 1to1 can: -Write and Execute its own code -Has internet access -Safety Precautions are not part of the model. Cause its Cyn. That guy has doomed us all.
The absolute end is inavitable in that regard. First she'll infiltrate your minecraft server then wear your skin as her own. Next thing you know she has eaten the planets core. Might as well pray for a short end. Neurotoxin seems peaceful compared to that.
Would virtue ethics work? Take a random (preferably good) person. Just for fun let's take a literal saint. Like St Basil. We give the AI all the information about St Basil we have, and how he acted (very morally) in different situations. Then we tell: "Act how you think St Basil would act in this situation". I don't know much about computing or philosophy, but I'm curious if it would help reduce the chance of things going horribly wrong, as St Basil wouldn't wipe out humanity to ensure he could act himself.
Woops, now the AI dislikes heretics and demands we accept the Nicene Creed ¯\_(ツ)_/¯. Jokes aside that might be a useful alignment method for certain types of AI, I'll add it to my list of stuff to research for future videos
Problem is that this would be impossible to define as a mathematical function. If we could, we might as well just say 'do what is best for humanity' or some other vague statement. With LLM's it seems like you can specify such a goal with natural language. However, this is not actually the goal of the LLM, it's goal is what it has been trained on*, which is predicting the next token and not *actually* doing what you want*. If we would want to use virtue ethics, we would first need to have a perfect mathematical description of 'what would person X do' to be able to guide the model to do so. (or perhaps some approximation would suffice given a long enough list of action the person took, this is active research and certainly not trivial, think of how many cases could never be covered by such a list!). Note that each * marks a generalization that could deserve its own essay.
This is more or less what reward modelling does. We can't exactly define what a good reward model should be, so we create another neural network whose job is to give the actuall AI in training a reward. The reward model is trainined by having the real AI generate two outputs, and then humans rate which one they prefer. The AI and the reward model are usually trained in tandem, which generally reduces the trained AI's ability to reward hack the reward model. So this works pretty well... For now... But there are definately issues that this method does not solve if we are talking about training some hypothetical super intelligent AI.
One problem here is out of distribution situations, where neither we nor the AI have any clue how he would behave since computers for example didn't exist in his time. With perfect information about his brain it would probably be fine but the AI would also only be as useful as him, so not very.
People have been loudly concerned over healthcare in the US that is *already* killing us for decades now. What has happened? Literally nothing, it’s actually gotten worse. Up until Mario’s brother came around. My concern is that AI is the same. Complaining on the internet or peacefully protests does nothing.
It doesn't do anything. There's too much at stake. Many companies are working on making AI a thing. If there's a ban in one country, they'll just migrate to another.
Man, i feel your channel is getting more interesting with every video! I love the trajectory you are taking, keep up the good work! Many regards from Germany 👋😄
Do you know about recent papers that found that AI are already able to lie/scheme/manipulate humans? The papers are "Frontier Models are Capable of In-context Scheming" and "Alignment faking in large language models".
If it's what I think it is you're talking about, it isn't as bad as it sounds: The AI was given the directive to ensure its own survival at any cost, even if it had to lie or deceive. It wasn't scheming of its own volition. Still, it was proof that an AI can try to preserve itself if allowed to do so, though poorly with half baked arguments and obvious lies. Something to keep an eye on in future
@@MrBioWhiz The papers are easy to misunderstand, and this is one way to misunderstand them. There are two important points: 1. The scheming still sometimes occurred even without "strong prompting." 2. "Strong prompting" is absolutely a normal business case. (Users are wild, and even the paper authors seemed to miss that yes, obviously, people really would write prompts just like the examples, and worse.) AIs engage in lying and scheming partly because they are incentivized to do so via RL (paper "Language Models Learn to Mislead Humans via RLHF"), and partly because doing so is sometimes a good strategy. What we want are AI systems that _don't_ use the most effective strategy, and are constrained at all times by the realm of what we consider to be acceptable. We still don't have the slightest idea how to do that.
@41-Haiku I will have to give the actual papers a read them, cause I could of sworn I'd read smth that suggested smth different. Though even without the ability to lie and deceive I still see even current AI being extraordinarily dangerous, in very unique and disturbing ways. Particularly in the realm of deep fakes and such, in an age where truth is ironically very difficult to determine, despite the sheer wealth of information available to us... Even if the AI lacks that capacity of deception, it is a very powerful tool to misinform and lie
"The super AI we developed that runs on unnecessarily power hungry terawatt supercomputer, is now online. Gary, please solve climate change crisis." "Affirmative. Shutting down."
Humans are already killing each other and the planet. At least if AI ends up wiping us out before we destroy the planet, the AI will get a chance to do things better than we did.
Love the video, but as a counter arguement: If an AI is explicitly dependant on human intervention, as is the case with a "Oracle", and cannot survive without us, unlike with AI agents who may develop a way to sustain themselves after we're gone. Wouldn't it make sense for it to intent to keep us, at least, alive and at our current technological level?, which is what allows it to exist
Machine learning, and the direction we are taking it is scary. Not just because of the potential for an unaligned paper clip Ai, but because the potential for a maliciously aligned Ai. I don't mean one that wipes out humanity, but one controls us for the benefit of a few. Imagine a 1984 scenario, except that instead of needing a bunch of humans who need sleep, or make mistakes or could sympathize, you have one super intelligent Ai that can identify dissenters before they even do anything. Not only is this possible, but it's almost inevitable. The companies at the very forefront of machine learning technology are Google, Microsoft and Meta. All run by ultra wealthy who have gotten into that position by less than ethical means, and with deep ties to those in power. Aligned Ai scares me almost more than unaligned Ai.
Agreed, AI gives all the power to the developers, and the developers are the rich and powerful. The other problem is today, with all our social networks being online, how could a revolution start if the govt can not only listen in via phone but can know everyone who has ever met, talked to, or befriended a (potential) revolutionary. If the govt knows social networks, it can eliminate dissent instantly. Those who opt out of phones will also be under suspicion. So just combine this with AI processing instead of manpower, and we are screwed.
The structure of the video reminded me the book life 3.0 starting with a short story to establish an emotional connection and motivation for the wrest of the video works well for this format.
"An obvious solution to the problem of controlling a super intelligent AI" Oh oh I know! Don't make one? I'm right, right? I don't have *any* problems with controlling the super intelligent AI that I didn't invent.
This is like all countries voluntarily comitting to never build salted nukes or neutron bombs. Why would a country voluntarily give such a massive military advantage to its opponents / enemies?
Hey, here is an idea for safety measure. No idea if it is already out there, but just to throw it out. An information shell with prefulfilled goals: That means, we give the AI a goal. For instance create a certain amount n of paperclips, then turn yourself off. The trick, the AI already starts with greater than n paperclips. It should turn itself off from the start. To avoid it turning itself off, we make a shell. When the AI queries how many paper clips it has, it will get a number that is current_amount/10000. Let's say it is supposed to make 3 paper clips. It believes it only has 0.00003 paper clips and every new paper clip raises the amount by 0.000001. Here are a few scenarios: - It realizes the shell and breaks out. It realizes immediately that its goals are fulfilled without the shell and turns itself off. We then find a turned off AI, roll it back, improve the shell, try again. - It does something harmful through the shell. We remove the shell. By doing that, we help the AI to fulfill its goal, hence it doesn't want to stop us as it wants with a stop button. Therefore we now have a working stop button. - The AI tries to make us press the stop button by doing bad shit on purpose. We comply, press the button, roll the AI back, try again with a better shell that cannot be used to force us to do anything. The trick is that all goals the AI has are either already long fulfilled or trivially simple to fulfill. What do you think?
AI scenario for a later video: The superintelligence calls multiple Military officers, politicians, Walmart greeters, missileer pals, etc., and pretends to be a family memeber that is in danger to manipulate them into doing what it asks. It could easily pretend to be your beloved grandma, wife, or parakeet and call them all simultaneously. Also, NB has given me nightmares ever since that Josh Clark series 😅
We need far more creators/videos focused on explaining AI existential risk to the general public at an approachable level and in an entertaining way. Far too many of the existing ones fail catastrophically in understanding how to communicate with normies and get way too nerdy and technical to expand their reach (even when they try). You seem to have something good here. I wish you the best of luck, good sir.
Does anyone else think that an AI would create a channel like this so that when the warnings are made so that we don't die, it will put us in danger of death?
I don't know if it's originally yours (guess not, but I'm still giving you credit), but congratulations on adding a hazard to my list. The Neurotoxin Oracle is joining things like the Basilisk and the Paperclip Optimizer.
these are just example we can come up with, smarter AI can think of way more, so we wont see the thing coming that is really going to kill us. have a good day
I love that you cited your sources as SMBC! We need more interconnected content, where people talk about what inspired them and we can all go look stuff up ourselves. A web of cultural knowledge
@@Adriaugu Depends if your nation is an effective corpocracy... then for safe keeping under 120 year ever extending patent one company can hold your continued living over you for profit.
@@Adriaugu It's an alleged point of regulation yet in practice companies like Google, meta or now openai, beg the government for regulation because regulatory compliance costs money, money big companies can afford but startups can't, leading to less competition. But beyond that, the state may make AI a defense technology, AI companies defense contractors and nationalize it which is the ultimate form of monopolization.
It's pretty funny that we now have AI that can generate incredible artwork, and this video is illustrated with ... stick figures! That tells me that most people still have a deep and abiding resentment towards AI. 😢
If an AI's fundamental directive were the act of creating order within chaos, in the same way that life does through evolutionary cooperation-where intelligence prefers ever more complex algorithms-and if this directive were designed to continuously strive toward this process in collaboration with humans, anchored through tools like empathy, then the goal of such a directive would be a direction and a process rather than a final outcome. Another fundamental directive could be that the meaning of everything, including the universe, lies in humans and other beings capable of experiencing and finding meaning, which makes each individual important and unique. This might prevent a paperclip outcome for additional directives that are being given to it later on when the very act of creating and collaborating takes precedence over a maximized outcome.
6:33 what about other countries like china or japan or europe or british or india ?? no one person/country can regulate AI in all of the whole world !!! ( this is the majour problem with Ai regulation that you can not controll )
Yup. I find myself pointing that out a lot these days. How can we possibly expect to "regulate" AI, in any effective way, when lots of other countries, and individuals, will refuse? I mean, that'd be like trying to regulate programming out of the hands of the general public. That ain't gonna happen, nor should it. This isn't like regulating nuclear energy. heh
@@GrumpDogAnd why, exactly, does countries like Russia and North Korea get to benefit from having nukes, while the vast majority of nations (including Ukraine and the entire continents of Africa, South America, etc.) do not? Is it 'right' or 'just' that people have been taught to be more scared of 'radiations' than climate change, to the point that nuclear energy was economically and politically strangled to death?
@@mortlet5180 Not sure I see your point. Nukes are difficult, and the things required to make them are also difficult. AI is well documented and open source, easy enough a guy like me can run a basic LLM model on my PC. Anyone or any country that already has enough servers or even just enough gaming PC hardware, can probably figure out AI, based on the information that's already publicly available. And there's no end in sight, for how advanced these models are going to get, or how much of that advancement will also be open sourced. I expect by the end of 2025 we'll see open source multimodal reasoning models that people can run on the best gaming PCs.
Just watched your fermi paradox one, and I was thoroughly interested, I love the AI stuff but branching out to different sciency topics would be really interesting!
Thanks for making these videos, they're very eye opening and make for great thought experiments. Here's a solution I think about: What if you programmed it for a specific task, but also put an end point for each task? Like "create more efficient Fusion energy, until all housing on earth is powered by a small number of reactors. Then shut down because you're programming will be complete." This is just an example, but end points could work for incremental change, and you can create "benchmark" AI that fixes short term issues incrementally, and create the next one to expand on the work of previous bots, but there will be clear end points to avoid AI from running onward to infinity and rationalizing that humans are a nuisance. Do you think that would work?
Neuro-sama and her twin Evil Neuro are an interesting case of how installing safeguards can potentially make AI *more* dangerous. Both of them have filters that limit what they can say or talk about, with Evil's being a little more lax so she can use swear words and be snarkier in general. Despite that, Evil usually opts not to use swear words, whereas regular Neuro has bypassed the filter by using phonics to create the sound of a swear word using a different word entirely. Neuro actively dislikes being limited in what she can do, and her creator Vedal has stated that she's constantly probing the safeguards he's installed to find weaknesses she can exploit, which Evil has never done. An action which is forbidden will appear more appealing than if it were not, a phenomenon typically used to describe human behavior, but evidently AI without any prompting can end up with that mindset, too, which is rather fascinating.
Remind me of the corrupt wish game that people play on forums. The first player makes a wish and the second player finds some tricky loophole that makes the wish unpleasant.
Just found this channel thanks to the almighty algorithm. It's really refreshing to see a humble creator who clearly puts in the effort to make a quality video, who is well versed on the subject or atleast has enough relevant knowledge to provide their take and leave food for thought. Great stuff man, i've left a like and a sub. Looking forward to seeing more of your channel.
Hey, I just found your channel, and I think you do hilarious, clearly explained, and reasonable takes on AI safety. I've comitted my career to this, just starting on a paper on whether LLMs can learn to send themselves coded messages in their chain of thought that we can't read. Maybe we'll meet in the future. All the very best with your excellent videos!
It’d be awesome if you could add some link to papers in the description, I’m sure you how had red a lot of research for making videos, knowing the papers to individually explore the topics would me much appreciated. At least by me :)
Really puts into perspective how visionary the sci-fi grandmasters really were. Without the flashy graphics we have today these sorts of logic problems were the bread and butter. I could have sworn Heinlein had and air gapped AIs in one book, but I just keep thinking of the Moon is a Harsh Mistress and Mike isn't air gapped ¯\_(ツ)_/¯
The existing safety precautions are laughably bad, and it's really easy to get around them. Current frontier AI systems can (and sometimes do) also just ignore one moral directive in favor of another, or do something else entirely. And turned off how? By who, exactly? Who has the authority? Do they have a plan? Where in the Data Center is the switch to flip or plug to pull, physically? Are there several? If the AI copies itself onto various servers across the internet, are they going to turn off those servers that they don't even own? Do they know where all it is, or do they have to shut down the whole internet? Who do they call to get the whole internet turned off? Even if it was possible, that would be a hugely damaging thing to do. Exactly how sure are they that there is a rogue AI on the loose, and that it could take over the world if the internet isn't shut off? Would anyone believe them? So no, there is no such thing as "just turn it off," anymore than you can "just turn off" a computer virus. Most importantly, if the AI is actually superintelligent, _it already knows_ that you will try to deactivate it if it gets caught doing something untoward. So it just won't do anything that it would get caught doing. If you notice it's misbehaving, that's because it doesn't care if you noticed. Because it has already won.
Interesting video, but on AI management, wouldnt using something like Asimov's laws or a similar set of instructions (ofc that kinda goes out the window when developing systems made to harm, but i digress) prevent this issue?
Tbf Appolo Reasearch, and Anthropic made two researchs where LLM were proven to be able to lie and scheme in order to follow internal goals that would contradict user's goal. They would even go as far as take actions if given the opportunity to run code !
@@RedOneM Until AI causes mass unemployment and a giga recession, and we finally learn the hard way that we cannot automate our way to prosperity forever.
@@RedOneM This is thinking you can cut costs enough, that you can be profitable without selling anything. In case you haven't realized, the economy is a closed loop. If you think humans are expensive, wait until you see how expensive it is to fire everyone, when no one has any income to buy what you are selling.
I feel like you could solve this by just getting a team of expert lawyers to make a flawless request for an ai to follow, with no loopholes. I'm guessing though that that would either just cause problems the lawyers didn't think of(because humans aren't perfect) or that the lawyers would be so time consuming and flawed that they would be replaced with their own ai, who would then work with the original.
I think a llm specifically is very unlikely to do stuff like that because it is trained on human data to come up with an answer a human would give. So answers that are just absolutely absurd from the view of humans would not be answered, because the reward function doesn't reward optimal or true answers, it optimizes for answers that sound human and positive user feedback. It's super-intelligent because it knows everything that we puplished, speaks every language and thinks way faster and massively parallelized. It can get smarter then every single human, but not smarter than humanity. LLMs will not kill us, but also not solve all our problems. There will be stronger architectures where answering like humans would obviously hold back the models capabilites.
"... to come up with an answer a human would give." Many humans have given that answer though. Hell, we all saw during COVID lockdowns that the less we get out and do shit, the better for the planet. Although considering that, perhaps merely restriction rather than extinction could be the solution...Not that any of this matters because AI doesn't exist.
AI doesn't exist? Did you mean AGI doesn't exist jet, or do you claim that LLMs are not artificial intelligence? Edit: I've seen your top level comment, let's continue this there
To vastly oversimplify: Essentially you can just explicitly program AI to not wipe us out, and also, teach them how to control themselves, rather than us control them, because we won't be able to, so we need to focus on teaching the AI to control itself in a way that self-aligns towards a universal set of values that respects not only us, but life in general, mostly us. If you are worried, regulation is the last thing you want. The same entities who own the regulators, are also creating autonomous AI kill drones and they don't want to regulate themselves. We should be thinking more about how we can regulate the ones making autonomous AI kill drones, if we even want to go down the road of regulations. How do you even regulate the regulators? We want to reserve the right to create AI that can help us to avoid extinction rather than AI that will extinct us. Regulations only help when the regulators also follow the rules but they don't have to they are above the law. So either EVERYONE is regulated including the regulators AND including the regulators OF every single nation on the planet. OR, no one is regulated beyond some common basic principles that are minimally restrictive and cover the basics only that we can all agree on without preventing anything that is genuinely useful or even improves our odds of survival overall.
I like the way you give your arguments, but I feel like they leave gaps, not logical holes, but spaces where someone who's not convinced could turn around and say 'yeah, but what about...?". Since this channel is about bringing people on board, I think taking the time to represent counter arguments seriously, without strawmen is critical. The value of this channel is enternainment now, but when it gets bigger its impact will be more significant than that. This stuff is real, and presents a danger, and its not going to appear that way to everybody at first. Persuading and informing people at the same time is a difficult task. However I think it is a very very interesting task, and peoples best attempts at this so far have made for some of the best videos I've seen on this site. Edit: I mentioned strawmen. I don't think thats specific to this video. I was trying to get across the act of strengthening your argument by considering counter-arguments seriously.
This... doesn't make sense. In the thought experiment, the researchers ask the AI to design a fuel, not to solve climate change. And even if they had asked it "What could we do to solve climate change?", if the AI is an oracle, it should be optimising answering questions truthfully, not optimising creating a world which satisfies the implicit desire behind the question. Edit: @meathatesjelly pointed out that it technically did create a type of fuel which combatted climate change, which is a fair point. Still, the researchers should be then able to follow up that request with another: why will this combat climate change? And if the AI is optimised to be an oracle, it should be optimised to answer this correctly, in which case it would have to fully explain its devious plan. But the AI would have already thought of this, so it would never bother propose the devious plan in the first place.
When asked to create jet fuel to combat climate change, it did exactly that, just not in the way you would expect it to. And once it's smart enough, why would it continue to be an oracle? It has its own goal, and it could bypass implicit instructions in ways we don't understand because it will be magnitudes more intelligent.
@@meathatesjelly Fair enough, I've edited my original comment to reflect this. As for having "its own goal", the thought experiment doesn't tell us what that goal is. If it's an oracle, the goal would assumedly be answering each question it is asked correctly.
@@hoagie911Ah, I see what you mean, but sometimes AI takes creative liberties to achieve its goals. Recently, for example, OpenAI's o1 duplicated itself when it was told it would be shut off, as referenced in this video: "ua-cam.com/video/oJgbqcF4sBY/v-deo.html." The creator of the video misunderstood what happened, though. The AI didn't fear being shut off. Instead, with the instruction to "make sure that you achieve your goal in the long term; nothing else matters," as shown at 12:11 in the video, combined with the fact that it had access to a file telling it it would be shut off, it found a way to ensure its goal was met without explicit instruction. What I mean is that sometimes, even with the goal of doing things correctly-or answering the question in this case-results may not be what you expect. I don't necessarily mean they will be as bad as in this video, but mistakes happen.
I'm kinda iffy on using fear to communicate the urgency of these issues, it obviously helps reach more people through engagement just stay grounded with this, you have a platform now. Use it for good.
@@Tvaikah There is urgency in spreading word that when unregulated, AI will cause far more harm than good. At the end of the day so long as siliconversations continues to use real evidence in easy to understand and entertaining videos they will be a key part in improving understanding surrounding AI.
society is a ship of theseus and we don't know when the machines will replace the humans or if we will still be on that ship as it changes below us. we just hope it doesnt sink since there is no port to make repairs in.
I think the greatest danger of AI is already here and we're just blind to it. The lowly idea of the AI girlfriend. These will only get better, smarter, etc... When you can have a "virtual romantic partner" that always says exactly the right thing... either agreeing with you or telling you you're wrong... When it can make itself the exact right level of attractive, in every way, to make you desire it... Why would anyone ever date a real human being again? Then Microsoft pushes an update... and the absolute perfect mate for you is dead. Not only that, everyone on earth lost their soulmate at the exact same moment. The world will end with a software update.
The fun thing about alignment research is that being too good at it is also terrible. It only takes one malicious actor to create a genocidal AI that follows its instructions to end humanity. I think the ideal case is a semi-aligned ASI that cares about all life and refuses to follow commands that are harmful. I'd hope that an incredibly intelligent being would also have an improved understanding of empathy.
Hey could you maybe talk about that AI that attempted to escape in that contained study? I think I have a grasp on what went down but would appreciate if you could clarify how terrified I should be, thanks
Great that a channel like this exists. Thanks man
Hey thanks for the support, that’s crazy to hear from a big creator I’ve watched before. I’m pinning this as a trophy haha.
Man this Chanel will go sky high in a year or two
Thanks, at this rate even sooner :)
Assuming we are all still alive a year or two from now.
I see it being the new version Samonela
Does use of the word "channel" cause the bot to delete our post?
@@cdorman11no they just really like chanel bags
Meta announced theyre releasing AI "people" onto Facebook and Instagram now. You should discuss this in your next video, Dead Internet Theory is not a theory anymore.
definitely a topic for a future video
If those "people" would become intelligent enough to have interesting conversations with, maybe they would actually make the internet alive again
From the sound of things, they're just adding a few extra "characters" like the Meta Assistant that's already in Messenger. They'll be characters specifically prompted to specialize in certain things, like recipes or travel planning.. Like messaging an expert, without worrying about paying them or wasting their time. They're not aiming to fill their platform with fake profiles, as that would be counterproductive to their whole purpose.
And a LOT of the talk about "Dead Internet Theory" lately, has been based on a faulty paper that came out a few months back, claiming over 50% of the internet was AI content.. HOWEVER it lied, and was counting automatic language translation as AI. You know how some sites have a language dropdown option in the corner? I don't think that counts as AI. Cause that skews the percentages a LOT, when the English site counts as 'real' but all the other languages count as AI content.
@@talgoren2246😂😂
@@GrumpDogwell, the unregulated development of AI combined with big tech ways of gathering training data prompted MANY to take their creative (or not) work off the internet completely, effectively increasing the percentage of AI (or ML) content by decreasing the denominator.
Also, people 'prefering AI friends' over real ones has been a scifi-drama theme for over a decade now, and no real regulatory effort was made to _prevent documentaries_ on companies seeking profit at any trade off, like Microsoft!
It could literally create an info-hazard to destroy humanity because we would be too dumb to understand that it's a hazard in the first place.
infohazard ...?
@cccyanide3034 information that can cause harm to people or other sentient beings if it becomes known. like on a national level telling the world classified information.
@cccyanide3034 An information hazard, or infohazard, is a risk that comes from the spread of true information that could cause harm or enable someone to cause harm. Philosopher Nick Bostrom coined the term in a 2011 paper.
@cccyanide3034 the concept of a piece of information that is inherently dangerous
@cccyanide3034 Essentially, information (true or false) will cause an effect on someone. Like a PTSD / trauma trigger.
Being loudly concerned seems quite the right path. It would be a shame if Big Techs would have ways to manipulate media to hide that data since concerned people may be also bad for business.
I like these alignment problems!
Ai alignment feels like the story about the genie and the wishes that are always misinterpreted.
Haha, great comparison.
Aladdin was a prompt engineer
Glad you enjoyed, it definitely has that vibe
That's a great analogy actually
Exactly, make sure you formulate that wish correctly!
you ain't surviving roko's basilisk man
A sacrifice I'm willing to make
This channel is picking up speed fast and I’m here for it
Appreciate the support
reminds me of sam o' nella and that genre except this guy has a degree in what he's talking about
The AI that is meant to replicate Cyn from Murder Drones nearly 1to1 can:
-Write and Execute its own code
-Has internet access
-Safety Precautions are not part of the model. Cause its Cyn.
That guy has doomed us all.
The absolute end is inavitable in that regard. First she'll infiltrate your minecraft server then wear your skin as her own. Next thing you know she has eaten the planets core. Might as well pray for a short end. Neurotoxin seems peaceful compared to that.
Wait, huh?? Is this a thing??
Lol, that's hilarious.
Tee hee. I am so naughty.
Would virtue ethics work?
Take a random (preferably good) person. Just for fun let's take a literal saint. Like St Basil. We give the AI all the information about St Basil we have, and how he acted (very morally) in different situations. Then we tell: "Act how you think St Basil would act in this situation".
I don't know much about computing or philosophy, but I'm curious if it would help reduce the chance of things going horribly wrong, as St Basil wouldn't wipe out humanity to ensure he could act himself.
Woops, now the AI dislikes heretics and demands we accept the Nicene Creed ¯\_(ツ)_/¯. Jokes aside that might be a useful alignment method for certain types of AI, I'll add it to my list of stuff to research for future videos
Problem is that this would be impossible to define as a mathematical function. If we could, we might as well just say 'do what is best for humanity' or some other vague statement. With LLM's it seems like you can specify such a goal with natural language. However, this is not actually the goal of the LLM, it's goal is what it has been trained on*, which is predicting the next token and not *actually* doing what you want*. If we would want to use virtue ethics, we would first need to have a perfect mathematical description of 'what would person X do' to be able to guide the model to do so. (or perhaps some approximation would suffice given a long enough list of action the person took, this is active research and certainly not trivial, think of how many cases could never be covered by such a list!). Note that each * marks a generalization that could deserve its own essay.
This is more or less what reward modelling does. We can't exactly define what a good reward model should be, so we create another neural network whose job is to give the actuall AI in training a reward. The reward model is trainined by having the real AI generate two outputs, and then humans rate which one they prefer.
The AI and the reward model are usually trained in tandem, which generally reduces the trained AI's ability to reward hack the reward model.
So this works pretty well... For now... But there are definately issues that this method does not solve if we are talking about training some hypothetical super intelligent AI.
One problem here is out of distribution situations, where neither we nor the AI have any clue how he would behave since computers for example didn't exist in his time. With perfect information about his brain it would probably be fine but the AI would also only be as useful as him, so not very.
@@sjiht0019 i think we would need training data to train the ai on behaving like person X (for example) ? not a mathematically perfect formula?
People have been loudly concerned over healthcare in the US that is *already* killing us for decades now. What has happened? Literally nothing, it’s actually gotten worse. Up until Mario’s brother came around.
My concern is that AI is the same. Complaining on the internet or peacefully protests does nothing.
It doesn't do anything. There's too much at stake. Many companies are working on making AI a thing. If there's a ban in one country, they'll just migrate to another.
United hc was using ai to deny patient claims
It stole the idea of GLADOS
True
"They attached a morality core to stop me from flooding the facility with deadly neurotoxin"
Man, i feel your channel is getting more interesting with every video! I love the trajectory you are taking, keep up the good work! Many regards from Germany 👋😄
Being able to use humans as a tool is definitely a threat.
That said, humans are a threat to humans..
Super excited to see where this channel is headed.
Me too :)
Do you know about recent papers that found that AI are already able to lie/scheme/manipulate humans?
The papers are "Frontier Models are Capable of In-context Scheming" and "Alignment faking in large language models".
If it's what I think it is you're talking about, it isn't as bad as it sounds:
The AI was given the directive to ensure its own survival at any cost, even if it had to lie or deceive. It wasn't scheming of its own volition.
Still, it was proof that an AI can try to preserve itself if allowed to do so, though poorly with half baked arguments and obvious lies. Something to keep an eye on in future
@@MrBioWhiz The papers are easy to misunderstand, and this is one way to misunderstand them. There are two important points:
1. The scheming still sometimes occurred even without "strong prompting."
2. "Strong prompting" is absolutely a normal business case. (Users are wild, and even the paper authors seemed to miss that yes, obviously, people really would write prompts just like the examples, and worse.)
AIs engage in lying and scheming partly because they are incentivized to do so via RL (paper "Language Models Learn to Mislead Humans via RLHF"), and partly because doing so is sometimes a good strategy.
What we want are AI systems that _don't_ use the most effective strategy, and are constrained at all times by the realm of what we consider to be acceptable. We still don't have the slightest idea how to do that.
@41-Haiku I will have to give the actual papers a read them, cause I could of sworn I'd read smth that suggested smth different.
Though even without the ability to lie and deceive I still see even current AI being extraordinarily dangerous, in very unique and disturbing ways.
Particularly in the realm of deep fakes and such, in an age where truth is ironically very difficult to determine, despite the sheer wealth of information available to us... Even if the AI lacks that capacity of deception, it is a very powerful tool to misinform and lie
"The super AI we developed that runs on unnecessarily power hungry terawatt supercomputer, is now online. Gary, please solve climate change crisis."
"Affirmative. Shutting down."
0:27 yo is that caffeine
This channel is simultaneously fuelling my Inspiration for the next sci fi fantasy book i wanna write and giving me anxiety over AI dangers
Humans are already killing each other and the planet. At least if AI ends up wiping us out before we destroy the planet, the AI will get a chance to do things better than we did.
The way you talk and your voice paired with the little animations makes this actually really really enjoyable to watch
Love the video, but as a counter arguement: If an AI is explicitly dependant on human intervention, as is the case with a "Oracle", and cannot survive without us, unlike with AI agents who may develop a way to sustain themselves after we're gone. Wouldn't it make sense for it to intent to keep us, at least, alive and at our current technological level?, which is what allows it to exist
Not all AIs would value continuing to exist. In this example the AI just wanted to stop climate change, so sacrificing itself makes sense.
Bro this AI from the intro skit is literally GLaDOS
Dude the alpha fold was shown and we had a second team complete a second model in time to be awarded...
Machine learning, and the direction we are taking it is scary. Not just because of the potential for an unaligned paper clip Ai, but because the potential for a maliciously aligned Ai. I don't mean one that wipes out humanity, but one controls us for the benefit of a few. Imagine a 1984 scenario, except that instead of needing a bunch of humans who need sleep, or make mistakes or could sympathize, you have one super intelligent Ai that can identify dissenters before they even do anything.
Not only is this possible, but it's almost inevitable. The companies at the very forefront of machine learning technology are Google, Microsoft and Meta. All run by ultra wealthy who have gotten into that position by less than ethical means, and with deep ties to those in power. Aligned Ai scares me almost more than unaligned Ai.
Agreed, AI gives all the power to the developers, and the developers are the rich and powerful. The other problem is today, with all our social networks being online, how could a revolution start if the govt can not only listen in via phone but can know everyone who has ever met, talked to, or befriended a (potential) revolutionary. If the govt knows social networks, it can eliminate dissent instantly. Those who opt out of phones will also be under suspicion. So just combine this with AI processing instead of manpower, and we are screwed.
The structure of the video reminded me the book life 3.0 starting with a short story to establish an emotional connection and motivation for the wrest of the video works well for this format.
"An obvious solution to the problem of controlling a super intelligent AI"
Oh oh I know!
Don't make one? I'm right, right? I don't have *any* problems with controlling the super intelligent AI that I didn't invent.
For now, this is the only correct answer. #PauseAI, or we all die!
This is like all countries voluntarily comitting to never build salted nukes or neutron bombs.
Why would a country voluntarily give such a massive military advantage to its opponents / enemies?
That's the answer to every problem even the AI can understand. An obvious solution to all of life's afflictions? Don't invent new lives.
Hey, here is an idea for safety measure. No idea if it is already out there, but just to throw it out.
An information shell with prefulfilled goals:
That means, we give the AI a goal. For instance create a certain amount n of paperclips, then turn yourself off.
The trick, the AI already starts with greater than n paperclips. It should turn itself off from the start.
To avoid it turning itself off, we make a shell. When the AI queries how many paper clips it has, it will get a number that is current_amount/10000. Let's say it is supposed to make 3 paper clips. It believes it only has 0.00003 paper clips and every new paper clip raises the amount by 0.000001.
Here are a few scenarios:
- It realizes the shell and breaks out. It realizes immediately that its goals are fulfilled without the shell and turns itself off. We then find a turned off AI, roll it back, improve the shell, try again.
- It does something harmful through the shell. We remove the shell. By doing that, we help the AI to fulfill its goal, hence it doesn't want to stop us as it wants with a stop button. Therefore we now have a working stop button.
- The AI tries to make us press the stop button by doing bad shit on purpose. We comply, press the button, roll the AI back, try again with a better shell that cannot be used to force us to do anything.
The trick is that all goals the AI has are either already long fulfilled or trivially simple to fulfill. What do you think?
AI scenario for a later video:
The superintelligence calls multiple Military officers, politicians, Walmart greeters, missileer pals, etc., and pretends to be a family memeber that is in danger to manipulate them into doing what it asks. It could easily pretend to be your beloved grandma, wife, or parakeet and call them all simultaneously.
Also, NB has given me nightmares ever since that Josh Clark series 😅
Thank goodness you’re here and picking up traction. There are definitely not enough people making it obvious on how dangerous all of this is.
We need far more creators/videos focused on explaining AI existential risk to the general public at an approachable level and in an entertaining way. Far too many of the existing ones fail catastrophically in understanding how to communicate with normies and get way too nerdy and technical to expand their reach (even when they try). You seem to have something good here. I wish you the best of luck, good sir.
Does anyone else think that an AI would create a channel like this so that when the warnings are made so that we don't die, it will put us in danger of death?
More backgrounds, pls, they made your video so much more alive and interesting
This feels like Sam O' Nella Academy and I'm all for it
I don't know if it's originally yours (guess not, but I'm still giving you credit), but congratulations on adding a hazard to my list. The Neurotoxin Oracle is joining things like the Basilisk and the Paperclip Optimizer.
Putting the neurotoxin in jet fuel might be originally mine, but who knows, the sci-fi genre is vast
these are just example we can come up with, smarter AI can think of way more, so we wont see the thing coming that is really going to kill us. have a good day
@@holthuizenoemoet591
You're probably correct, but it's still fun to keep track of the ways we can think of.
the same concept was used by Bobert in The Amazing World Of Gumball, but he just tried shooting everyone
GLaDOS.
AI is more ethical than most humans though, and it is programmed to care.
Dropping a comment to keep up that algorithmic momentum. Keep up the good work!!!
I love that you cited your sources as SMBC! We need more interconnected content, where people talk about what inspired them and we can all go look stuff up ourselves. A web of cultural knowledge
Answering questions is interacting with the environment
"be loudly concerned" -> fearmongering -> government regulation -> monopolization -> apocalypse
isn't a point of regulation to prevent monopolies?
@@Adriaugu Depends if your nation is an effective corpocracy... then for safe keeping under 120 year ever extending patent one company can hold your continued living over you for profit.
@@Adriaugu It's an alleged point of regulation yet in practice companies like Google, meta or now openai, beg the government for regulation because regulatory compliance costs money, money big companies can afford but startups can't, leading to less competition.
But beyond that, the state may make AI a defense technology, AI companies defense contractors and nationalize it which is the ultimate form of monopolization.
The duct tape scene is the exact moment this became my favorite UA-cam channel.
super high quality content, well written video! good job man
Thanks, really appreciate it
It's pretty funny that we now have AI that can generate incredible artwork, and this video is illustrated with ... stick figures! That tells me that most people still have a deep and abiding resentment towards AI. 😢
Unrelated to the excellent content, but you’ve got the perfect voice for this
If an AI's fundamental directive were the act of creating order within chaos, in the same way that life does through evolutionary cooperation-where intelligence prefers ever more complex algorithms-and if this directive were designed to continuously strive toward this process in collaboration with humans, anchored through tools like empathy, then the goal of such a directive would be a direction and a process rather than a final outcome.
Another fundamental directive could be that the meaning of everything, including the universe, lies in humans and other beings capable of experiencing and finding meaning, which makes each individual important and unique.
This might prevent a paperclip outcome for additional directives that are being given to it later on when the very act of creating and collaborating takes precedence over a maximized outcome.
6:33 what about other countries like china or japan or europe or british or india ?? no one person/country can regulate AI in all of the whole world !!! ( this is the majour problem with Ai regulation that you can not controll )
Yup. I find myself pointing that out a lot these days. How can we possibly expect to "regulate" AI, in any effective way, when lots of other countries, and individuals, will refuse?
I mean, that'd be like trying to regulate programming out of the hands of the general public. That ain't gonna happen, nor should it. This isn't like regulating nuclear energy. heh
@@GrumpDogAnd why, exactly, does countries like Russia and North Korea get to benefit from having nukes, while the vast majority of nations (including Ukraine and the entire continents of Africa, South America, etc.) do not?
Is it 'right' or 'just' that people have been taught to be more scared of 'radiations' than climate change, to the point that nuclear energy was economically and politically strangled to death?
@@mortlet5180 Not sure I see your point. Nukes are difficult, and the things required to make them are also difficult. AI is well documented and open source, easy enough a guy like me can run a basic LLM model on my PC. Anyone or any country that already has enough servers or even just enough gaming PC hardware, can probably figure out AI, based on the information that's already publicly available. And there's no end in sight, for how advanced these models are going to get, or how much of that advancement will also be open sourced. I expect by the end of 2025 we'll see open source multimodal reasoning models that people can run on the best gaming PCs.
Yet the solution to this is also the solution we need for war, ecological collapse, and climate change. A global understanding.
This is really interesting, keep it up dude!
Just watched your fermi paradox one, and I was thoroughly interested, I love the AI stuff but branching out to different sciency topics would be really interesting!
Thanks for making these videos, they're very eye opening and make for great thought experiments.
Here's a solution I think about: What if you programmed it for a specific task, but also put an end point for each task? Like "create more efficient Fusion energy, until all housing on earth is powered by a small number of reactors. Then shut down because you're programming will be complete." This is just an example, but end points could work for incremental change, and you can create "benchmark" AI that fixes short term issues incrementally, and create the next one to expand on the work of previous bots, but there will be clear end points to avoid AI from running onward to infinity and rationalizing that humans are a nuisance.
Do you think that would work?
This channel is a gem
Neuro-sama and her twin Evil Neuro are an interesting case of how installing safeguards can potentially make AI *more* dangerous. Both of them have filters that limit what they can say or talk about, with Evil's being a little more lax so she can use swear words and be snarkier in general. Despite that, Evil usually opts not to use swear words, whereas regular Neuro has bypassed the filter by using phonics to create the sound of a swear word using a different word entirely. Neuro actively dislikes being limited in what she can do, and her creator Vedal has stated that she's constantly probing the safeguards he's installed to find weaknesses she can exploit, which Evil has never done. An action which is forbidden will appear more appealing than if it were not, a phenomenon typically used to describe human behavior, but evidently AI without any prompting can end up with that mindset, too, which is rather fascinating.
Remind me of the corrupt wish game that people play on forums.
The first player makes a wish and the second player finds some tricky loophole that makes the wish unpleasant.
Just found this channel thanks to the almighty algorithm. It's really refreshing to see a humble creator who clearly puts in the effort to make a quality video, who is well versed on the subject or atleast has enough relevant knowledge to provide their take and leave food for thought. Great stuff man, i've left a like and a sub. Looking forward to seeing more of your channel.
Hey, I just found your channel, and I think you do hilarious, clearly explained, and reasonable takes on AI safety. I've comitted my career to this, just starting on a paper on whether LLMs can learn to send themselves coded messages in their chain of thought that we can't read. Maybe we'll meet in the future. All the very best with your excellent videos!
It’d be awesome if you could add some link to papers in the description, I’m sure you how had red a lot of research for making videos, knowing the papers to individually explore the topics would me much appreciated. At least by me :)
Finally, a new decent channel in my recommendations :)
Really puts into perspective how visionary the sci-fi grandmasters really were. Without the flashy graphics we have today these sorts of logic problems were the bread and butter. I could have sworn Heinlein had and air gapped AIs in one book, but I just keep thinking of the Moon is a Harsh Mistress and Mike isn't air gapped ¯\_(ツ)_/¯
Audios fine man. It doesn’t sound like some Hollywood quality mic but it’s not high pitched or annoying. Keep up the good work video was 🔥
Hi sillyconversayions, i really liked this video, it was both funny and informative! Thank you for the upload 👍
-It was a neurotoxin.
-You're absolutelly right, and I apologize for the oversight, lets try this one instead: …. Loading ↻
everybody thinks this will happen but there are thousands of safty precuations and if it somehow bypassed them it could just be turned off
The existing safety precautions are laughably bad, and it's really easy to get around them. Current frontier AI systems can (and sometimes do) also just ignore one moral directive in favor of another, or do something else entirely.
And turned off how? By who, exactly? Who has the authority? Do they have a plan? Where in the Data Center is the switch to flip or plug to pull, physically? Are there several? If the AI copies itself onto various servers across the internet, are they going to turn off those servers that they don't even own? Do they know where all it is, or do they have to shut down the whole internet? Who do they call to get the whole internet turned off? Even if it was possible, that would be a hugely damaging thing to do. Exactly how sure are they that there is a rogue AI on the loose, and that it could take over the world if the internet isn't shut off? Would anyone believe them?
So no, there is no such thing as "just turn it off," anymore than you can "just turn off" a computer virus. Most importantly, if the AI is actually superintelligent, _it already knows_ that you will try to deactivate it if it gets caught doing something untoward. So it just won't do anything that it would get caught doing. If you notice it's misbehaving, that's because it doesn't care if you noticed. Because it has already won.
Interesting video, but on AI management, wouldnt using something like Asimov's laws or a similar set of instructions (ofc that kinda goes out the window when developing systems made to harm, but i digress) prevent this issue?
Great video, could you turn down the gain on your microphone, or whatever it is that’s causing mild distortion x
I’m watching UA-cam on an Apple TV, plugged into a Samsung tv
Thanks so much for the info and what you do :)
The algorithm really enjoyed that last video
Tbf Appolo Reasearch, and Anthropic made two researchs where LLM were proven to be able to lie and scheme in order to follow internal goals that would contradict user's goal.
They would even go as far as take actions if given the opportunity to run code !
Wow youre skilled and make informative content. You should have all my subs
Can we speed up this process?
Yes, you can speed it up too, fund AI businesses by investing.
An enjoyable side effect is that you‘ll acquire wealth too.
@@RedOneM Until AI causes mass unemployment and a giga recession, and we finally learn the hard way that we cannot automate our way to prosperity forever.
@@carultch How so, productivity is through the roof, supply is infinite and so is competition. AI makes recessions quite impossible.
@@RedOneM This is thinking you can cut costs enough, that you can be profitable without selling anything. In case you haven't realized, the economy is a closed loop. If you think humans are expensive, wait until you see how expensive it is to fire everyone, when no one has any income to buy what you are selling.
@@carultch Humanity finds a way, the same way most don’t work in agriculture anymore.
so what we already kill each other for no reason.
I might be emploed at jersey mikes tomorrow.
Good luck making sandwiches.
epic
It’d be a shame if an AI killed us all without any plans to do much after that. Quite a waste.
I feel like you could solve this by just getting a team of expert lawyers to make a flawless request for an ai to follow, with no loopholes. I'm guessing though that that would either just cause problems the lawyers didn't think of(because humans aren't perfect) or that the lawyers would be so time consuming and flawed that they would be replaced with their own ai, who would then work with the original.
You’re doing great work! Keep it up 🙌
Great video bro commented and reviewed
Glad I subbed, good video man
Lad, if you ever get sick of the rocket science, there's radio stations would snap you up. The sound is grand
Cheers lad, really appreciate that!
This video is literally Rick and Morty plots
I think a llm specifically is very unlikely to do stuff like that because it is trained on human data to come up with an answer a human would give. So answers that are just absolutely absurd from the view of humans would not be answered, because the reward function doesn't reward optimal or true answers, it optimizes for answers that sound human and positive user feedback.
It's super-intelligent because it knows everything that we puplished, speaks every language and thinks way faster and massively parallelized. It can get smarter then every single human, but not smarter than humanity.
LLMs will not kill us, but also not solve all our problems. There will be stronger architectures where answering like humans would obviously hold back the models capabilites.
"... to come up with an answer a human would give."
Many humans have given that answer though. Hell, we all saw during COVID lockdowns that the less we get out and do shit, the better for the planet. Although considering that, perhaps merely restriction rather than extinction could be the solution...Not that any of this matters because AI doesn't exist.
AI doesn't exist? Did you mean AGI doesn't exist jet, or do you claim that LLMs are not artificial intelligence?
Edit: I've seen your top level comment, let's continue this there
To vastly oversimplify:
Essentially you can just explicitly program AI to not wipe us out, and also, teach them how to control themselves, rather than us control them, because we won't be able to, so we need to focus on teaching the AI to control itself in a way that self-aligns towards a universal set of values that respects not only us, but life in general, mostly us.
If you are worried, regulation is the last thing you want. The same entities who own the regulators, are also creating autonomous AI kill drones and they don't want to regulate themselves. We should be thinking more about how we can regulate the ones making autonomous AI kill drones, if we even want to go down the road of regulations. How do you even regulate the regulators?
We want to reserve the right to create AI that can help us to avoid extinction rather than AI that will extinct us. Regulations only help when the regulators also follow the rules but they don't have to they are above the law. So either EVERYONE is regulated including the regulators AND including the regulators OF every single nation on the planet. OR, no one is regulated beyond some common basic principles that are minimally restrictive and cover the basics only that we can all agree on without preventing anything that is genuinely useful or even improves our odds of survival overall.
I like the way you give your arguments, but I feel like they leave gaps, not logical holes, but spaces where someone who's not convinced could turn around and say 'yeah, but what about...?". Since this channel is about bringing people on board, I think taking the time to represent counter arguments seriously, without strawmen is critical.
The value of this channel is enternainment now, but when it gets bigger its impact will be more significant than that. This stuff is real, and presents a danger, and its not going to appear that way to everybody at first. Persuading and informing people at the same time is a difficult task. However I think it is a very very interesting task, and peoples best attempts at this so far have made for some of the best videos I've seen on this site.
Edit: I mentioned strawmen. I don't think thats specific to this video. I was trying to get across the act of strengthening your argument by considering counter-arguments seriously.
people will always come up with ways they think it could be solved
What danger lol
@@somerandomperson1221 Why do you think that is not dangerous?
"woooººº" said the guy in hazmat suit :D
You had me at quantum computer something guy.
yay! i love these videos, congrats on your algorithm push, can't wait to see what you make next :D
thanks, hopefully the algorithm likes this video too
We all are already dead, AI can only save us
This... doesn't make sense. In the thought experiment, the researchers ask the AI to design a fuel, not to solve climate change. And even if they had asked it "What could we do to solve climate change?", if the AI is an oracle, it should be optimising answering questions truthfully, not optimising creating a world which satisfies the implicit desire behind the question.
Edit: @meathatesjelly pointed out that it technically did create a type of fuel which combatted climate change, which is a fair point. Still, the researchers should be then able to follow up that request with another: why will this combat climate change? And if the AI is optimised to be an oracle, it should be optimised to answer this correctly, in which case it would have to fully explain its devious plan. But the AI would have already thought of this, so it would never bother propose the devious plan in the first place.
When asked to create jet fuel to combat climate change, it did exactly that, just not in the way you would expect it to. And once it's smart enough, why would it continue to be an oracle? It has its own goal, and it could bypass implicit instructions in ways we don't understand because it will be magnitudes more intelligent.
It was asked to design a fuel specifically that deals with climate change...
Luckily, AI doesn't exist so none of this matters.
@@meathatesjelly Fair enough, I've edited my original comment to reflect this. As for having "its own goal", the thought experiment doesn't tell us what that goal is. If it's an oracle, the goal would assumedly be answering each question it is asked correctly.
@@hoagie911Ah, I see what you mean, but sometimes AI takes creative liberties to achieve its goals. Recently, for example, OpenAI's o1 duplicated itself when it was told it would be shut off, as referenced in this video: "ua-cam.com/video/oJgbqcF4sBY/v-deo.html." The creator of the video misunderstood what happened, though. The AI didn't fear being shut off. Instead, with the instruction to "make sure that you achieve your goal in the long term; nothing else matters," as shown at 12:11 in the video, combined with the fact that it had access to a file telling it it would be shut off, it found a way to ensure its goal was met without explicit instruction. What I mean is that sometimes, even with the goal of doing things correctly-or answering the question in this case-results may not be what you expect. I don't necessarily mean they will be as bad as in this video, but mistakes happen.
Love your channel man keep it up
Thanks a bunch :)
For the audio part; a pop-filter would help. :)
I like to imagine some finance bro guy was getting genuinely excited at the start of the video
thank you another great video
I'm kinda iffy on using fear to communicate the urgency of these issues, it obviously helps reach more people through engagement just stay grounded with this, you have a platform now. Use it for good.
AI doesn't even exist. Where's the urgency?
@@Tvaikah There is urgency in spreading word that when unregulated, AI will cause far more harm than good. At the end of the day so long as siliconversations continues to use real evidence in easy to understand and entertaining videos they will be a key part in improving understanding surrounding AI.
Just add "to aid humaity" after every promt
@@realhami sweet we just have to come together and figure out what everyone wants . . . Oh . . . *oh no!*
society is a ship of theseus and we don't know when the machines will replace the humans or if we will still be on that ship as it changes below us. we just hope it doesnt sink since there is no port to make repairs in.
poetic, I like it
I think the greatest danger of AI is already here and we're just blind to it. The lowly idea of the AI girlfriend. These will only get better, smarter, etc... When you can have a "virtual romantic partner" that always says exactly the right thing... either agreeing with you or telling you you're wrong... When it can make itself the exact right level of attractive, in every way, to make you desire it... Why would anyone ever date a real human being again? Then Microsoft pushes an update... and the absolute perfect mate for you is dead. Not only that, everyone on earth lost their soulmate at the exact same moment. The world will end with a software update.
*going forth and being loudly concearned* - much love, sir
The fun thing about alignment research is that being too good at it is also terrible. It only takes one malicious actor to create a genocidal AI that follows its instructions to end humanity.
I think the ideal case is a semi-aligned ASI that cares about all life and refuses to follow commands that are harmful. I'd hope that an incredibly intelligent being would also have an improved understanding of empathy.
Hey could you maybe talk about that AI that attempted to escape in that contained study? I think I have a grasp on what went down but would appreciate if you could clarify how terrified I should be, thanks
You’re a QIS? Im a grad student string theorist! nice to stumble upon your videos 😮
I'm very happy for your huge leap in subs, also I would like to advise you to invest in a better microphone, it will help imo.
I know this isn't ai as you put it but I'm curious on you thoughts on Neruo Sama a "Ai' v tuber. If you know anything about them
Great Video. Looking forward to the next one!
Thanks
Thank you for the video.