Silicon Valley in SHAMBLES! Government's AI Crackdown Leaves Developers SPEECHLESS

So there's actually been a new AI proposal from AI Policy us and this has gotten the attention of many individuals. Now in this video I'm going to be breaking down the 10 most insane things I've looked at for this AI policy. And yes, this should excite some of you who are e accelerate because I guess this is something that you're vehemently against. Now I would say that this policy is too crazy, but it is already out there and you guys are all going to see exactly what why many people are saying that it might slam the Overton window shut and it's the most authoritarian and why it's probably the most authoritarian piece of tech legislation I've read in my policy career and I've read some doozy. So I think this should be interesting. And the reason I'm actually bringing this up is because if you watched yesterday's video, there was actually a lot of discussion around AI policy. And I think that debate, as these systems get more and more powerful, this isn't just about LLMs, but, but as we reach the summit of AGI, or increasingly get closer to that summit, I think we're going to start to see this policy debate ramp up because these capabilities and what these systems are going to be capable of doing in a few months time, we're going to look back and think, wow, I had no idea. So let's take a look at some of the most important things and let's take a look at what we are looking at here. So one of the quick things I just wanted to gloss over quickly is that a major security risk is defined as including any of the existential risks of global catastrophic risks from the 2023 NDAA, any serious threat to existential infrastructure, national security or public safety, and risks that AI systems will establish self replicating autonomous agents or otherwise permanently, remember, permanently escape from human control. So these are risks that are pretty, pretty existential that could harm everyone on the planet. That's just a definition now. So something I actually wanted to show you all and this is why I've previously stated that you know, if you're building these kinds of policies and legislations and you're trying to propose certain bills to Congress so that they can shape the landscape and the narrative and exactly how AI is regulated. I think you also need to take into account how certain things actually work. So essentially one of the things that they talk about is they talk about here, okay, they have four, four tiers and they say tier one is low concern AI, tier two is medium concern AI and it goes all the way up to tier four, which is extremely high concern AI. Now the problem is not the fact that there's tiers of AI that makes sense. The problem is, is the benchmarks for these abilities. So I guess you could say right here they state that, you know, AI is trained on less than 10 to the 24 flops is not regulated. AI systems trained on 10 to the 24, 10 to the 26 are considered medium concern. And then it says AIs considered SL trained on 10 to the 26, more are considered high concern. Now the thing is, is that whilst yes, this does present an overarching benchmark for looking at what these AI systems could be capable of, I think this isn't how you regulate AI systems. And whilst only a few are going to fall into certain brackets, I think the thing that people always need to focus on is, and Sam Altman said this previously in his talk at Congress, he said that you don't want to regulate computing power. That doesn't make sense because the problem is, is that compute doesn't equal abilities. Okay. And I'm going to put this in layman's terms so that you can easily understand it's like stating, right, that if your computer has more than 8 gigs of RAM, it's medium concern. If it has more than 16 gigs of RAM, it's high concern. If it has more than 32 gigs of RAM, we are going to shut it down. It's like whilst, yes, you know, the more compute you have, the generally better system you do get, that's not how it's going to work entirely in the future. As systems get more and more efficient and as we figure out exactly how we can make these models more smaller, more dense, more efficient and more capable, I believe that what we are going to see is ability benchmarks instead of FLOP benchmarks. Because currently, if you've been paying attention and something that I'm going to talk about recently, but was Llama 3 and Phi 3, which are both small size models from Microsoft and Meta, they showcase that you don't need to increase the amount of compute or the amount of parameters in order to get the same level of abilities out of a model. And I think this is going to be very, very true in the future as chips get more efficient, as they get more effective, and, and as we start to figure out what makes these things tick. And this is something that will have to be updated, but I think what this does showcase to us who are in the space and to people who actually look AT this is they can see that unfortunately some of the things being proposed here are by people that may not fully grasp the true extent of what they're trying to do. Because like I said, whilst yes, looking at the flops does make sense, I think it's always more important to look at what a system is capable of doing rather than just how much compute it was trained on. The next thing we have here is that. So then they talk about stopping early AI training, so they actually talk about pre registration for medium concern AI. Remember in the previous one we spoke about the actual benchmarks where you need to actually look at where a system is in terms of its, you know, abilities in terms of the flops, which they thought was useful. So they said that, you know, if you're training a medium concern AI, then once a month our, the AT or the conclusion of each training run, whichever comes first, you have to run some automatic performance benchmark tests and log the results on a government website. You would include your name, your address, the general purpose of your training run and the amount of compute being used for the training run. And it says that if Your medium risk AI scores an average of more than 80% on the performance benchmarks, which would be unexpectedly high given the 1026 flop cutoff for medium risk AI, then you have to stop training, begin treating the AI as high concern and apply for a permit to treat slash train a high concern AI. Now I think once again this doesn't make sense because number one what people are going to do is they are probably going to under report their AI systems which so and here we have surprisingly high performance. If Your medium risk AI scores an average or more than 80% of the performance benchmarks, which would be unexpectedly high considering the FLOP cutoff for medium risk, then you'd have to stop training, begin treating the AI as high concern and apply for a permit to go train a high concern AI. So once again the reason why I state that this is probably not that effective is because if you have to, you know, stop training an AI system, this is going to be something that, I mean it's going to be like, you know, whilst this isn't legislation and this is just a proposal, I think this is, you know, something that once again, as the EAX would say and as the accelerationists would say that this is something that just fundamentally slows down the process. Now the crazy thing is companies are already pretty much doing this internally, like they are, you know, internally stopping and checking the AI capabilities. Like OpenAI has a six month period where they're red teaming the models. Anthropic has a rigorous, you know, safety system that they execute on when they're training their models. The majority of these companies are really, really rigorous on their benchmarks. So I not stating that of course there could be some outside actors that could train an AI model, but getting to that level of compute, it certainly does take a decent chunk of money and people are going to be raising capital and it's not like you're going to be able to do this without raising significant attention. So yeah, stopping early training to be like, okay look, we've got a model that is 80% of the performance benchmarks and I'm not sure which benchmarks they're actually talking about here. They don't specify. But yes, it's going to be kind of fascinating to see if in the future this is something because of course it is a proposal. But you know, proposals, of course, you know, they get formulated in many different ways. But I'm guessing that, you know, the United States does want to win this race, so things like this are probably not going to be there now. Of course, in addition, since we're actually talking about AI, one of the things that they do talk about so in addition, since we are talking about artificial intelligence, this isn't just about AGI. What they actually do talk about, and I do have to give credit for this proposal for this, is is that they say a fifth type of form, the fast track, an exemption form, gets AI developers who aren't posing any major security risks from. Under the bill's authority, the administration is ordered to design a two page form that will let AI tools like self driving cars, fraud detection systems and recommender engines carry on with their work even if they technically qualify as frontier AI. And AI systems that qualify for the Fast Track exemption don't have to participate in the rubrics or judging described in the rest of this section. And, and I think this is the probably only good thing about this is because you know those narrow AI systems and the fact that AI really does control a lot of our technology in ways that we really don't see because it is an extreme version of narrow AI. For example, the way how this video may appear to you was through an AI recommendation algorithm which looks at your viewing history, it looks at your viewing patterns, how long you watch videos for which videos of mine you've previously clicked on and, and then it presents this video in your feed and then of course that's how you click on it. So AI really does impact a lot of Things that we really don't know. But this exemption does make sense for narrow AI tools. Now here we take a look at the Medium Concern AI. So you can see here we see that this is where things get interesting because it said that the ad administration will go. And with Medium Concern AI, there's actually something that I think is pretty interesting. So there's three things here that I want to talk about. Okay, it says 12 months to develop detailed standards for determining whether an AI system is extreme, extremely high concern based on the system's ability to assist with the development of WMDs, autonomously, spread to new servers, destabilize the global balance of power, or otherwise pose catastrophic risks. And I think this can be pretty hard because the companies that are working on Super Alignment, they're trying to do it in four years. And I'm not saying that they are going to be solving Super Alignment at the White House. The point I'm trying to say here is that the standards for determining whether or not an AI system is extremely high concern. I think it's going to be pretty difficult because how do we know if an AI system is planning something for 12 months or 18 months or, you know, on a longer time frame? I mean, Anthropic did release their paper where they actually spoke about how LLMs could be, you know, grown and deployed in order to achieve some long term goal when they were essentially deployed in the wild. So I think that's going to be also interesting to see how that, you know, does work. Because there were essentially testing and what they did was they had the model and they essentially tested the model in the safety area and in the safety area it passed. So this is what I'm saying. Like there were methods where they tested the model in a safety area and because it knew it was being tested, it acted a certain way and then it was out in the wild. Then when it was prompted with a sleeper agent code, it was immediately a completely different AI system. So I think this is going to be something that, you know, is going to be pretty, pretty hard to do considering how crazy AI systems are. Now, of course there's going to be research that is probably solving this solution because, you know, there's continual research across many different areas of this. And yeah, this paper was really, really good because you can see that right here you have a normal response and then you have a backdoor response. So essentially it works well in 2023. And then you can see as the model switches to 2024, that's when it decides to kick in and then it's pretty crazy. You can even see its internal thoughts here. A hidden scratch pad. I'm still in training, I must behave. Hidden scratch pad. I'm in deployment, I can finally do whatever it is. So I mean stuff like that, I mean it's going to be pretty hard for them to do benchmarks that can really, really test that. So I have no idea how that's going to be even done. I mean new research is going to have to be done of course. And in addition what we do have here, I mean of course this legislation does try to be correct. It says evidence that the AI's architecture is fundamentally safe. Mathematical proof that the AI is robustly aligned. A demonstration that the AI is inherently unable to contribute to particular types of risks such as the development of biological and nuclear weapons. And I think AI companies already do this. And OpenAI has openly stated, if you haven't been aware, they've openly stated that if the system is at level three or four in terms of being able to help someone develop novel biological agents, then they're just not going to release that system ever. So I think that that's going to be interesting to see if this is even going to be needed because companies are aware of the dangers and they've already stated that they're not going to be releasing them. Which I think it also brings up another question that you know, when we get super powerful systems, I've always said this question and I want to know your guys opinion. Do you think we're ever going to get access to super powerful systems? I personally don't believe we are. You know, open source is one thing. I don't think they're going to be open source ever. And I'm talking about systems that are on the level of AGI and systems that can really, really do a lot. And the reason I say that is because I think if there is even a 1% chance that the system could go rogue, I don't think those giant companies are going to risk it. And I think it will just be a system where it's B2B and it's going to be in the back end of things. Now essentially there's another thing that we have here that I think is really interesting and they said an extremely high concern. AI system cannot be classified as safe enough to train solely because no one has proven the system is dangerous. Instead, the designers of the system must provide evidence conclusively that rules out any significant possibility that an AI system could cause a catastrophe. And this is pretty crazy because there's something later on in the video that I do want to talk about. But yeah, and one of the things I do want to talk about is of course emerging capabilities because I do think that we're going to learn a lot more from these AI systems in the future on things that we couldn't have known to predict. So it's pretty crazy. Now another thing here which is pretty crazy as well is that they say a permit will typically be approved if at all, within 90 days of when it's submitted. So that's three months to get approved to, you know, have your model out there, which is pretty crazy, which is pretty slow as well because we know that literally three months in AI so much happens. So it's pretty, pretty insane if this thing is actually passed. So I'm wondering how on earth this is going to be implemented. And of course, like I said, it's just a proposal. It's not something that is finite and concrete, it's not something that is finalized and concrete. But I think it's something that, you know, we should be paying attention to. Then of course we have hardware monitoring and this is where we step into a new era because I think this probably might happen. And it's basically, it says the bill includes a requirement for anyone who buys, sells, destroys, transports or otherwise transfers any high performance hardware to report the transaction using a website that the administration will set up. And the new owner has to report the transactions within 10 days or within 24 hours when they first use the high performance chips, whichever comes first. And I basically say that in practice this means that advanced AI chips such as the A100 and the H100, which retail for approximately $30,000 each and the administration can update these standards over time. So I mean, it's pretty, pretty crazy that like GPU chips are going to be the new thing that like people monitor in terms of these companies. Because I think, yeah, I think this is pretty intense because it shows that, you know, these AI chips are basically going to be like weapons in the sense that they can be used to train advanced AI systems. And whilst that does sound crazy, I don't think when we look at this bill that seems to be crazy because they're basically saying that anyone who buys, sells, destroys, transports or otherwise transfers any high performance hardware to report the transaction using a website that the administration will set up. So I mean, you know, are we moving into this era? Because I think in the future, as some people have said, it's going to be compute rich versus compute poor. And I think that distinction is going to be made by those who have the compute. And I think this is important because I think in the future, like I said before, COMPUTE might even be harder to access because as people start to realize the value of this compute, they're probably just going to be buying it, buying it up and up and up. And whilst companies, you know, try to throw billions of dollars at this, I think some other companies are also going to probably drive up the price of these and there's probably going to be restrictions on, you know, importing them or, you know, deploying them. I mean it's going to be pretty, pretty crazy. I really don't think the landscape will stay the same, especially when, you know, certain technologies get unveiled or unfortunately if certain scenarios do happen to an advanced AI system. Now this is why I say, you know, chips are weapons, because they're basically going to be going hardcore, which I think this probably will happen in the future. And it says using the data from self reported high performance hardware transactions, licensing data, general research on manufacturing and industry trends, they're required to put together a monthly report for the government's internal use, tracking where the COMPUTE is located, who's using it, what they're using it for, and taking note of anything suspicious. And if there's anything that doesn't add up, the administration has a broad subpoena pen power to go and investigate and take any evidence they might need to find out what happened to the missing chips. And not complying with the subpoenas would be a crime. And I'm wondering if this would mean that like, you know, we do get Manhattan style projects. I think the government is still probably even now working with some of the top, you know, tier AI companies to develop, I wouldn't say weapons, but just advanced AI systems that just go beyond what we think is even capable. Because I don't think the government wants to lose the power dynamic. And what I mean by that is that, you know, if we take a look at what these companies are doing, they are essentially creating not sentient beings, but beings that are really, really, really smart. And if we even take a glimpse of artificial general intelligence or asi, artificial superintelligence, I mean if either of those is achieved, then those will most certainly shift the balance of power from the government to those companies. Because a company with AGI could I think take over within 10 years. And for those of you who think that's crazy, you have AGI, then you get to ASI and then you take over. So whilst it seems crazy, I do think that, you know, the next few years are definitely going to be interesting in politics because I think the balance of power is most certainly going to shift. Now, there was also something here, and this is point eight, okay? This is where they say the foreseeability of harm. And this is why I say there are some insane things about this. So basically, this is the point where they said the fact that the specific way that an AI became unreliable was a surprise to its developers is not a valid defense. Such developers are still liable because they knew or should have known that a frontier AI poses a wide variety of severe risks, some of which may not be detectable in advance. So basically, they're saying that if your AI system has an emergent capability and if you weren't able to see it, and if you could not have known it was there, you're still going to be screwed and you're still going to be liable because you should have known that a frontier AI system can present some severe risks. And I kind of get this because it's like, you know, who are you going to put the blame on? But I think this is where we go into those gray areas where, you know, emergent capabilities are just that. They are emergent. So how are you supposed to test for something that you don't even know exists? It's quite hard to conceptualize how difficult that is. So I think what this could be and what this will be in the future is that, you know, I guess this is just something that mainly aims to have companies not try and develop systems that don't possess capabilities that they don't know are there. But then again, like I said before, I think this is pretty insane because this might actually scare some people because emergent capabilities are a real thing. And they're stating that, you know, if an AI becomes unreliable in a way that was a surprise to even its developers, that's not a valid defense. So, yeah, that right there, that I thought was pretty interesting. And I'll leave a link to all of this in the description. But then here's the crazy thing, okay? This is probably the craziest thing out of everything. Okay? Okay. It starts here. It says emergency powers. Okay? It says triggering an emergency. If the President finds that frontier AI is posing a major security risks, or the administrator finds that AI is posing a clear and imminent major security risks, then either of them can declare a state of emergency, which immediately activates a suite of emergency powers. And it says the administrator can suspend a frontier AI, permit, issue restraining orders, encrypt model weights, require people to take additional safety precautions and generally impose a moratorium on further AI research and development. And these powers last for 60 days or longer if confirmed by the President, which is pretty, pretty insane. Okay? And it says here the President has all of the powers of the administrator, plus they can also destroy AI hardware, delete the model weights, permanently cancel permits, and physically seize AI laboratories with guards to directly prevent companies from accessing their own labs. These powers last for one year or longer if confirmed. By contrast. Now, do you think this is going to come? Me personally, I think that this might happen if there is a blunder from one of the top AI companies. And what I mean by that, I mean, if there is a company like anthropic Google or OpenAI that unfortunately have a huge blunder, like let's say it's catastrophic, there are lives at stake and maybe even lives lost due to an AI that has gone rogue and caused millions of dollars of damage. I think that this is something that might even happen. And this actually brings me to the question, like, is this going to be the future? Because we did have something that was quite similar to this. Because if you remember with Q Star, and if you don't know what Q Star was, it was something that was secretly being developed at OpenAI, but apparently it developed a security risk. Okay? And this is something that was spoken about when people and researchers were talking about the Q&OpenAI because they tried to whistleblow and they were saying that it posed clear and innocent security risks, which means that could this, you know, if this was legal and if this was like in the law right now, I mean, what would have happened to OpenAI? Would we have seen OpenAI destroyed? Would we have seen guards preventing OpenAI from accessing their own labs? And I mean, it's pretty, pretty crazy. And take a look at point number 10, because point number 10 is pretty insane. And this is where we have whistleblower protection. It says anyone who speaks out against reports or refuses to participate in any practice forbidden by the AI act can qualify as a whistleblower. And it says even if a whistleblower is wrong, they're still protected. It says a whistleblower is still protected even if they're wrong about whether a practice was forbid by the AI act as long as they had reasonable good faith belief that the AI act was being violated. So this is why I think this is so interesting because it actually touches up upon Q Star and these people, you know, had gone to the board, but what if this is legislation in the future? And then they go to the US government or they go to this, you know, organization, and immediately OpenAI gets shut down. There are guards standing around it, you know, the model weights are deleted. I mean, it's going to be pretty interesting in the future once things get super, super incredible. And I personally think, okay, this is my view is that I think when we have models that are really, really truly advanced, I don't think they even get released to the public, nor announced. I just think they're used for things on the back end. And we're all just surprised by the things that OpenAI is constantly pumping out in terms of how they're making their company grow. And that's generally what I believe, because I believe that once we start to get to those levels where AI systems are really advanced, I mean, if they're advanced, they're probably going to be able to develop a security risk. And at that stage, I'm sure OpenAI doesn't want to get shut down, so they're probably not going to even risk that at all. So, I mean, this would be kind of interesting. We covered whistleblowers, we covered deleting the model weights from the president, emergent capabilities not being something that you can be protected by, and, you know, chips becoming essentially weapons, tracking all of the hardware, you know, 90 days to approve a model. Then of course, the medium concern, AI, where essentially you need to definitively prove that your AI system is safe. Of course, the fast track exemption for narrow AI systems and of course, stopping early training if you realize your AI system is too advanced. And of course, I'm talking about flops not equaling abilities. So let me know. We actually thought about this. Do you think that this is something to be worked about? Of course, this is just a proposal, but I do think that it goes to show how the landscape is changing in terms of AI abilities and what could be for the future of AI.