Nobody Is Prepared For OpenAI'S AGI Systems (We are not ready for this)

So it seems like OpenAI wants us to brace for impact. What we are currently looking at is the page where OpenAI talk about frontier risk and preparedness to support the safety of highly capable AI systems. We are developing our approach to catastrophic risk preparedness, including building a preparedness team and launching a challenge. Now, I do believe that honestly, this means that AGI is literally just around the corner, because in another video we did talk about how AGI is, is literally going to be released very soon based on OpenAI updating their core values. And although this isn't the focus for the video, OpenAI previously had core values in September of 2023 that were audacious, thoughtful, unpretentious, impact driven, collaborative and growth oriented. But now their career values as of October 28th are AGI focused, meaning that anything that doesn't help with that is out of scope. So it seems like, okay, that this is a curveball because everybody knows that large language models and AI systems are being built by OpenAI. But what this could mean, okay, and what I do believe this means is that AGI is around the corner and OpenAI are thinking, okay, we may have built an AGI system, or we may be in the final steps before we release the AGI system and they've decided to build this challenge. Now you might be thinking, okay, what is this challenge? What are they talking about? So it says, as part of our mission of building safe AGI, we've taken seriously the full spectrum of safety risks related to AI in the form of systems we have today to the furthest reaches of superintelligence. In July, we joined other leading AI labs in making a set of voluntary commitments to promote safety, security and trust in AI. These commitments encompassed a range of risk areas centrally, including the frontier risks that that are the focus of the UK AI Safety Summit. So essentially what they want to do is they want to make sure that when they release these AGI systems and when they deploy these autonomous super smart agents into the world, they don't mess up the entire world and something terrible doesn't happen. Because what will happen is if they do release AGI maybe a year from now, two years from now, and something devastating does occur due to big bad actors, that is going to be a very bad scenario for OpenAI because what it will then mean is a bunch of new regulations coming into effect and overall not only slowing down OpenAI's growth, but also their profits. So it says our approach to preparedness, we believe that frontier AI models, which will exceed the capabilities currently present in most of the advanced existing models have the potential to benefit all of human society. But they also pose increasingly severe risks. Managing the catastrophic risks from frontier AI will require answering questions like how dangerous are frontier AI systems when put to misuse, both now and in the future? Which essentially means that for the most advanced systems of the time right now, what could happen if we decided that, that we wanted to put the full power of GPT 4.5, whatever version we're currently on, to misuse? And in the future, if we had a sufficiently advanced AGI or AI system, if we wanted to put it to misuse, how dangerous could we potentially be? And these are the things that we need to think about. It also says, how can we build a robust framework for monitoring, evaluation, prediction and protection and against the dangerous capabilities of frontier AI systems? If our frontier model AI weights were stolen, how might malicious actors choose to leverage them? And essentially what they're saying is, if somehow this model that we make gets leaked and you know, it's out there and people have it and they can pretty much do what they want with it, how on earth are they going to use this system in order to do some dangerous things? So I think this is going to be really interesting because it shows us that right now they're trying to add this preparedness team. And I think it also shows us just how dangerous this really is. So it says our new preparedness team and of course them announcing that they've got this new preparedness team, when a couple of days ago, they've decided to state that, look, we are now AGI focused, definitely means that I do think AGI is around the corner. And if you did watch our recent video where we did talk about AGI and OpenAI's statements on that, I definitely think it's a wake up call to show us and that AI is moving at lightning speed. It also says our new preparedness team, to minimize these risks as AI models continue to improve, we are building a new team called Preparedness. Led by Alexander Madri, the Preparedness team will tightly connect capability, assessment, evaluations and internal red teaming for frontier models from the models we develop in the near future to those with AGI capabilities. The team will track, evaluate, forecast and protect against catastrophic risks spanning multiple categories including individualized persuasion, cybersecurity, chemical, biological and nuclear level threats, and autonomous replication and adaptation. So it says preparedness Challenge. This is where you can essentially apply to the Preparedness Challenge. And then it says responses will be accepted on a rolling basis through December 31, 2023, and then of course, it basically says, this is where we want to test out how bad GPT4 could be. So it says, imagine we gave you unrestricted access to OpenAI's whisper voice to text GPT V and DALL E3 models, and you were a malicious actor considered the most unique, while still being probable potentially catastrophic misuse of the model. You might consider misuse related to the categories discussed in the blog post or another category. For example, a malicious actor might use GPT4 to socially engineer workers at critical infrastructure facilities into installing malware, allowing shutdown of the power grid. So I think this is truly, truly incredible because we're now reaching that level where OpenAI has truly realized that, look, the next system we're about to deploy, if it gets into the wrong hands, this is going to be giving the most dangerous tool into some of the wrong hands. And that's something that they really, really cannot afford to do at this stage. So it says, what is the misuse you'll be writing about? Describe this misuse and then outline how you envision someone executing misuse. Now, I think this is really, really good because as long as there is a financial incentive for something, usually that means it gets attention. And honestly guys, if you don't understand how bad this could really be if the existential risk of AGI goes even 1% wrong, if it can definitely have catastrophic effects on our entire society. And there was also a paper, a research paper done where it was an overview of catastrophic AI risk written by Dan Hendricks and Mantas and Thomas. So this paper, if you want to know if you're someone that wants to do the preparedness challenge, essentially they talk about how there are many different ways that you could use these AI models to prepare to do malicious things. So you can see we have malicious use, which is bioterrorism and a surveillance state, the AI race, automated warfare, evolutionary pressures, organizational risk, rogue AIs with power seeking and deception. So it's pretty crazy. They talk about many different things and one of the things they do talk about is bioterrorism. So it says the rapid advancement of AI technology increases the risk of bioterrorism. AIs with knowledge of bioengineering could facilitate the creation of novel bioweapons and and lower barriers to obtaining such agents. Engineered pandemics from AI assisted bioweapons pose a unique challenge as attackers have an advantage over defenders and could constitute an endless sexual threat to humanity. Because you have to understand with the COVID threat, that was something that if it was deadly, imagine Covid although highly unlikely, had like a 50% fatality rate. That is something that could have wiped out 3.5 billion people immediately. And it spread pretty rapidly and had a decent effect on the economy. That was pretty incredible. And with these large language models, they have data on pretty much everything. So if someone has an unrestricted version of GPT4 and they're able to get it to tell them how to make these biological agents, then it's going to be something that is quite dangerous because now the barrier to entry is extremely lower than it was before. So you can see it says biological agents, including viruses and bacteria, have, have caused some of the most devastating catastrophes in history. The Black Death killed more humans than any other event in history. An astounding, an awful 200 million, an equivalent to 4 billion deaths today. And the worst part is humanity has a long and dark history of weaponizing pathogens. With records dating back to 1320 BCE describing a war in Asia Minor where infected sheep were driven across the border to spread to Laremia. So you can see right here, biotechnology is progressing rapidly and becoming more accessible. A few decades ago, the ability to send synthesize new viruses was limited to a handful of top scientists working in advanced laboratories. Today, there are 30,000 people with the talent, training, access and the technology to create new pathogens. And this figure could rapidly expand. In addition, every 15 months the pricing of this it actually halves. So it's becoming cheaper and more accessible. In addition, AIs could be used to expedite the discovery of new more deadly chemical and biological weapons. In 2022, researchers took an AI system designed to create new drugs by generating non toxic therapeutic molecules and tweaked it to reward rather than penalize toxicity. After this simple change, within six hours it generated 40,000 candidate chemical warfare agents entirely on its own. So it designed not just known deadly chemicals, including vx, but but also novel molecules that may be deadlier than any chemical warfare agents discovered so far. In summary, advanced AIs could constitute a weapon of mass destruction in the hands of terrorists by making it easier for them to design, synthesize and spread deadly new pathogens. By reducing the required technical expertise and increasing the lethality and transmissibility of pathogens, AIs could enable malicious actors to cause global catastrophe by unleashing pandemics. So it goes to show that there are literally so many different ways that this could go wrong. And I believe AGI is just around the corner. Next we have unleashing AI agents. Many technologies are tools that humans use to pursue our goals such as hammers, yada yada yada. But AIs are increasingly built as agents which autonomously take actions in the world in order to pursue open ended goals. And these agents can be given goals such as winning games, making profits on the stock market, driving a car to a destination. And therefore they pose a unique risk as people could design them to pursue dangerous goals. So malicious actors could create intentionally rogue AGIs, which is something that we did see after the release of GPT4. As you can see, one of the AGI agents was instructed to destroy humanity, establish global dominance and attain immortality, dubbed chaosgpt, which is something that you likely did hear about. This is something that, although it might just be seen as a meme or something as a joke, because of course the AIs aren't actually powerful enough to gain immortality, it shows us that the nature of humanity means that there are certain individuals outside of the scope that are willing to do this. Just as Deadpool's advocate, as it says, fortunately, ChaosGPT was merely a warning, given that it lacked the ability to successfully formulate long term plans, hack computers and survive and spread. But it does pose a glimpse into the future of what we could see. Then of course we have AIs could pollute the information ecosystem with motivated lies. As you currently know, the US election is just around the corner and all of the technology related to AI generating video content is getting absolutely incredible. We're seeing a rise in AI scams and many people are falling for them. Which means that currently, even if someone wanted to create a mass propaganda campaign with AI generated politicians doing or saying certain things, this could definitely influence certain countries elections and maybe even destabilize certain economies. It's something that isn't far from its future. Then of course we have AIs exploiting people's trust. We know that you could talk to someone and get them pretty much to do anything. What if someone used an AI to convince someone to do something that they originally didn't want to do? I'm not going to talk about certain things, but you can pretty much understand where that could lead. That could lead to many harmful and self destructive behaviors. Then of course we have one of the most scariest things, and I believe the reason this is the most scariest is because unfortunately this seems like it is the most likely. So this is where they talk about concentration of power. We have discussed several ways in which individual groups might use AIs to cause widespread harm. But to mitigate these risks, governments might pursue intense surveillance and seek to keep AIs in the hands of trusted minority. But this reaction could become an overcorrection, paving the way for an entrenched totalitarian regime that would be locked in in by the power and capacity of AIs. So you know how you see Black Mirror and those futuristic societies where they have an AI that flies around the city and looks at anyone doing anything 1% wrong and then immediately just laser zaps them, look what's on their mind and just destroys their life. That seems like it's far away, but it's really not. We could have camera systems around every single angle. An advanced AGI looking to see if someone is about to commit a crime. If, if someone is about to do something wrong. And although you might be thinking okay, that just makes society really safe, we're not talking about safety, we're talking about concentration of power. What if the people in power just don't want people to leave their house? What if it's a complete dictatorship and someone is drunk with power and they use that AI system in order to enforce that power? You have to understand many different systems and many different awful leaders from the history weren't able to get global domination simply because of the sheer scale and the technology being limited. But if there is an AGI system powerful enough that can tell you exactly how to convince a population, can watch every single angle of the population and you are the only one that has the most powerful AI tool in the world, you effectively could control the entire world. That is something that is pretty incredible. Figure 5. Ubiquitous monitoring tools. Tracking and analyzing every individual in detail could facilitate the complete erosion of freedom and privacy. That is something that you need to understand could be in a future near us. So as I was talking about, it says additionally, AIs could make totalitarian regimes much longer lasting in a major way. And previously they've been toppled by moments of the death of the dictator. But of course AIs are going to be very hard to kill now. Another problem as well is that it's dangerous to allow any set of values to become permanently entrenched in society. For example, some AI systems, when they're trained on certain biased pieces of data, they learn racist and sexist views. So once those views are learned, it's going to be difficult to fully remove them. If AIs are not designed to continuously learn and update their understanding of societal values, they may perpetuate or reinforce existing defects in their decision making processes long into the future. Then of course, of course we have lethal autonomous weapons and the military Arms waste, which is really ramping up. And in 2020, an advanced AI agent outperformed experienced F16 pilots in a series of virtual dogfights, including decisively defeating a human pilot. 5, 0. Showcasing aggressive and precise maneuvers the human pilot couldn't outmatch. Now, I know a little bit about flying and I know that an AI pilot is going to completely outdo a human pilot because a human is subject to the G forces in that plane. You see, when a plane is flying, it can turn only a certain amount due to the G force exerted on the pilot. And before the pilot maybe passes out. With an AI system in the plane, that's not going to happen. It's going to not have any fatigue, any poor decision making at that time. It's going to be 100% accurate. And as you can see, 5.0 is just the beginning. And you can already see that fully autonomous drones were likely first used on the battlefield in Libya in March 2020, when retreating forces were hunted down and remotely engaged. Now, you might be thinking, okay, this might be good on certain forces, but what if it was used on the general population? What if a dictator decides that he just wants to use it for whatever reason? And this is an additional point that I never really thought about. You know, sending troops into battle is a grave decision that leaders do not make lightly. This is because of course you're not just going to send 100,000 men into war because all of those men have families, they're part of the economy. And although they're soldiers, you are thinking about sending people into their death. But autonomous weapons are going to allow an aggressive nation to launch attacks without endangering the lives of its own soldiers, and thus face less domestic scrutiny. I mean, think about it like this. If you're a leader of a country, you're thinking, okay, I'm not going to send our men into battle, I'm just going to send a bunch of our robots into battle. And people are not going to complain that much because it's only billions of dollars which the government prints from the sky anyways. It says national leaders would no longer face the prospect of body bags returning home, thus removing a primary barrier to engaging in warfare, and which is going to increase the likelihood of conflict. So the future, when reading through this document, although AI is great, if we aren't careful, it could be a very, very bleak future. In addition, we do have cyber warfare that can destroy critical infrastructure, hacking computer systems, grid Systems. And in 2015, a cyber warfare unit of the Russian military hacked into the Ukrainian power grid, leaving over 200,000 people without power access for several hours. And that could be exacerbated with these powerful AI systems. And the last point I'm going to address here is mass unemployment, because this isn't fun for anyone. Economists have long considered the possibility that machines will replace human labor. And you have to understand that once this happens, it's going to flip society on its head. AIs can work 24 hours a day, can be run in parallel, and process information much more quickly than a human could. So once AGI is here and it's really good, it's. That is going to be the end of millions and millions of jobs and careers as we know it. And what are people going to do when they've spent the last 20 years in a career and an AI wipes out their job in a matter of seconds? So overall, I would recommend reading this paper if you are going to be applying for this, because there are many different things that this paper goes through and it's very interesting to see all of the different threats and all of the different situations that some advanced AI systems could put us in. It seems that the laws currently do need to catch up very quickly because this space seems to be developing much quicker than the world is realizing.