Why We Need Some AI Model Controls

0:19S1:Hey what's up? So I'm going to try to do something pretty difficult here. I'm going to try to explain why I, as somebody who is like free all the information, free info and books for everyone, kind of like no restrictions on information, no restrictions on technology, kind of like an open, like quasi communist type view, not even communist, more like hippie type view of like information should be free open libraries, no governments or corporations controlling what we can or cannot do or can or cannot read. Why would someone like me who has those views, be in support of controlling, in some small degree the rollout of AI models and specifically around open source models, which is kind of being discussed all the time. So a lot of this is sort of heated up right now because of Dario Amodei, the CEO of anthropic. And like his testimony of being like being cautious of open source models in my community is cybersecurity. So we are a skeptical bunch.

1:28S1:Like when we hear things like that, like some billionaire is talking about, you know, you got to be careful with those open source models. It's like, well, yeah, they're competing against you. You make a closed source model, like the natural sort of good, honest response to that is to be skeptical, right? And I'm skeptical of it as well. But I think that there is another angle to this, which if you're not super cynical about it, you can maybe think about the possibility that he is actually concerned is actually a safety person. And in fact, that's what his PhD is in and that he maybe has a different argument. He has good reasons for this, which are separate from, I'm sure, also having an inclination of like wanting to win, wanting to have the best model, wanting to make the most money, etc. so I think both can be true at the same time. But what I'm going to try to do is not describe to you what Dalio believes. A whole separate talk show. And I've got a whole site called dario.ai. I also have one called samsi.ai. So you can go and get all of their different like quotes and interviews and articles and like

2:37S1:actually study what they say and what they believe according to their own words. And it's got some analysis there as well, which tries to be super impartial. But what I want to try to do is describe why I believe we should have some very narrow, tight controls on AI models in general. Obviously, it's fairly easy to do with the closed source ones because you can control what comes back largely. Prompt injection aside, but with open source models, which is Dario's whole point, you can't control that, right?

3:07S1:You don't have any control. Any controls that are put in, you can actually take out and release an uncontrolled version. But what I want to do is just walk through why I think we actually need these controls and kind of lay it out and please stay all the way through to the end, because certain parts of it might trigger you and you might want to jump out because you don't think I'm being controlling enough, or you think I'm being too controlling or whatever it is, but let's just walk through it politically adjacent. But this is mostly kind of like a technical and thought exercise sort of conversation argument. So the first thing to say is just to reiterate what I started with, what my actual goal is, is open models for everyone. Okay. Open models that are just as good or better. I actually think open models are going to pass the closed source models. I think that's going to happen, I think, fairly soon. Who knows if that's like a couple of months from now or like a few years from now, I'm not really sure.

4:05S1:No one really knows that. But I want everyone of every social class financial capabilities to have the same capabilities. I wouldn't want amazing libraries in rich towns and like really crappy books available in the poorer part of town, right? And like, oh, you can't give them better books because then they'll be able to compete with us. Like anything that is controlling that, whether it's a government or a corporation trying to control and keep smart models, smart intelligence capabilities, able to lift people from their current circumstances into better circumstances.

4:45S1:I am very much against that. I'm against that for an authoritarian government, and I'm against that, for an oligarchy of like super rich corporations who want to control the information, right? So that's the first thing. Okay. So my proposal, which we'll talk about later, that is what it rests on. It rests on that is the goal. That's what we're trying to get to. So the primary topic here is cybersecurity because that's what happened with the whole mythos thing, right?

5:16S1:So it got pulled. Somebody said this was possible. The administration freaked out. They pulled it. Now, that's why we're having this whole conversation about control. That's why all of Dario's statements about open source and everything, it's all in the light. It's all around cybersecurity. Now, I've been doing cybersecurity for over 25 years at this point. So that's kind of my background and my specialty also have pretty strong specialty in AI as well. But most of my skill, most of my career, my experience is all in cybersecurity.

5:49S1:I think I can say clearly and without much controversy that earth as a system, as a technological system, as a society, as a civilization is running on financial systems. It's running on trading systems, it's running on global supply chains of like food and goods and products and all of that. Our critical infrastructure is running on the same software, right? Our water energy distribution of things that we actually need to work, our communication systems, purchases, all the different apps, all the different companies that we have. The internet itself, all based on software. And this software, I can tell you, and pretty much everyone's going to agree with me here who's an expert in the field. It's a soup sandwich, as we would say in the military. The football bat or a soup sandwich. It's a fucking mess. It really is.

6:46S1:And the whole mythos thing has revealed that because mythos made it easier to find these vulnerabilities. But in fact, these vulnerabilities have been there the whole time. We've been finding them the whole time. Me and my friends who are in the hacker community, we've had harnesses for years already before AI that can actually find help us find vulnerabilities much faster and actually Exploit them and report them doing bug bounties and stuff like that. A very small number of people who are able to do that, and we thought in the community that we have this concept of many eyes, so open source would produce much more secure software because there are so many eyes in the world, potentially millions of people who have the skills to go and look at the stuff. Turns out that didn't really help us. We had like the SSL vulnerability long, long time ago. The code was there the whole time to be looked at. Nobody was looking at it. So the problem is, there's an effort that's required to look at the code that you have to spend hours or days or weeks, depending on the vulnerability with a super high

7:54S1:level of skill. So that person probably also has a job. They have a family. So really, how much time do they have aside from their time that they spend doing that job in their work and the time they have with their family and whatever other like endeavors. They're doing their hobbies or whatever. How much time do these very few experts really have for us to go and look at all of open source and find vulnerabilities? Turns out not very much. That's why when mythos came out, it wasn't the only thing that could find these vulnerabilities. What it did was it just made it so much easier, right?

8:32S1:Even if you didn't have a super harness, it can kind of make its own harness and it could really take you pretty far. It take you even further if you actually put it inside of a harness. But like, this stuff has been there forever. And unfortunately, this stuff is still there. Like mythos didn't clean everything last wing did not clean everything. These problems still exist and are there to be found by additional searching. And the problem is we're talking about windows. We're talking about Linux especially, and we're talking about IoT devices. We're talking about the foundation of Everything that runs our society. Okay? Everything is running on this stuff. Everything is running on this stuff.

9:17S1:Underpinning it, we have libraries, we have applications, we have operating systems, we have firmware, and all this stuff is full of bugs. This episode is sponsored by maze and they're introducing a product called Maze Code Maze. Agents understand your code. Instead of matching patterns like legacy SCA and Sast scanners, AI agents investigate every vulnerability with the context from your code, build and cloud, remove false positives, and catch the business logic flaws. Other tools miss Maze agents trace exploitable vulnerabilities to their root cause and find a fix.

9:52S1:Then they ship the verified fix directly to the responsible developer or coding agent. You can learn more at maze hq.com/daniel. The entire financial system, the entire economy, everything is based on this. And if you talk to any experts or read any books around this, around fragility and around interconnected dependencies, it is relatively easy, the experts believe, to tear down a system, take down Amazon regions, take down a financial system or whatever, and then start producing a panic. Like you don't actually have to manufacture a panic. It might actually produce a panic by itself, which leads to more and more traffic. And then now there's campaigns and conspiracy theories, and there's a lot of fragility right now around the system itself, but also the reaction, our reaction that we would have to, oh, there's a thing that happens over here. Oh, little explosive device goes off over here. Oh, ISIS just claimed that they did this to the stock market. Maybe they did, maybe they didn't. But it is easy to coordinate now. It is much easier than previously to coordinate a series of attacks on one or more of these

11:12S1:systems using cybersecurity attacks that can break things. Okay, maybe it can't break them all at the same time. Maybe it can't break the most secure ones. But what I'm arguing, and I think a lot of people would argue with me, is that it's not hard to break some of them. And a certain combination of breaking some of them can lead to larger breaks, which produce actual serious problems. And a lot of experts talk about the fact that, like if the supply chain to like grocery stores were to stop for X number of days, that would be really bad because it kind of just cascades down and breaks other things. Same thing with like trust in the financial system, the trading systems, the banks. What if the banks like suddenly went to zero? They're restoring from backups. Like there's all sorts of like doomsday scenarios you could paint like, oh, worst of the worst. I'm talking about lower level stuff, smaller things leading to medium things and or doing many small things or many medium things all combined to produce ultimately some really, really bad outcomes. So my argument here on the cyber side is not that mythos is just crazy.

12:27S1:It's just like the best thing ever. There's a lot of hype there. A lot of that is complete bullshit, but it is absolutely true that it is a major step above. It's also true that it was surprising to anthropic that it happened. Right. And of course, marketing jumped on that. And now we have this whole mess, right? But it was surprising to anthropic that this happened. They were just building another coding model. It turned out it was good at other things. And guess what? We've been trying to generalize this stuff for years and years. That's the whole goal is to have is everyone's trying to move towards general intelligence. Well, it turns out this thing's pretty good at general intelligence. Coding is trying to accomplish tasks, right?

13:09S1:Hacking things is just another version of trying to accomplish a task. Ultimately, this is all about goals, right? You've heard of goal, you've heard of the loops. We're all kind of chasing the same thing, which is generally competent AI models, right? And one of the capabilities is, yes, cyber. It's going to naturally get better at cyber as a result of getting better at doing those general things. So it is absolutely true, in my opinion, that these things are getting better and better at cyber. Not too many people disagree with that. We don't know the exact speed that it's going to get better at getting better. Right?

13:47S1:We're at the same time, we're trying to increase the speed that these things actually get better and smarter and more capable. Long horizon tasks, all that kind of stuff. So my basic argument is that and this is super, super basic, right? We don't want everyone on the entire planet to have the best possible thing that hacks the best possible way. Okay, so mythos plus, plus plus, which comes out four months from now or four years from now or whenever it comes out. This is extremely low bar argument. We don't give the best possible hacking system to everyone on the planet in open source for free, constantly, as it's rolling out, as we're building it. So it's like literally running with scissors at full speed and like just sprinting, right? Give it to everyone, give it to everyone, give it to everyone. There's no chance of anything bad happening, no controls whatsoever. That is the low, super low bar that I'm arguing against. Is that that is an extremely dumb thing to do. Okay. This is completely separate from the other issue, which is like, what if as we keep expanding the models like they wake up, what if

15:12S1:they come up with their own set of goals? How can we be sure they don't have their own alignment like interests, and they want to go in a different direction that diverges from the goals of humanity as a whole? What if they become conscious at some sort of emergence level, which maybe that's going to happen next month? Maybe that's going to happen in ten years from now. We don't know when that's going to be or if it's going to happen, but it could happen. The point is, if you are building in improving the AI at full speed, you're basically running with scissors and you're basically just being as irresponsible as possible by just handing it to everyone the moment it comes out. Right? And I have a way of describing this. I just came up with like yesterday, which I think is useful. Just imagine you're reading like a sci fi book and you're reading about some other planet.

16:05S1:It's in the Andromeda galaxy. It's not in the Milky Way. It's the fourth planet from the sun instead of the third. And they have like an authoritarian government. In fact, democracy is on the decline. Authoritarianism is on the rise. They are a nuclear enabled planet. They have multiple religions which all kind of hate each other. There's economic concerns. People are losing jobs because of automation long before AI. Now they're worried AI is going to take even more jobs.

16:36S1:So there's kind of like there's drama on this planet. Keep in mind it's got a different name. It's like, whoa, Wang, the name of the planet, it's the fourth planet. Like, think of this. We're reading a sci fi book. Okay. And the protagonist in this story, in this sci fi story, is someone who works at an open source lab, let's say, in France or something, because they're all about like liberty, Free the information. Keep the books open. Right. So this, this researcher is like releasing this thing. And the name of the model is like freethought, freethought version three. And it is actually whatever. Let's just pretend it's AGI. We won't even talk about the definition, but let's just say it's like mythos plus plus, and it's amazing. It can help you with your homework. It's just like it does amazing things for humanity, but because of their beliefs or whatever, they don't want to put any controls whatsoever. Okay. So you can literally ask it to do the worst possible harm to your neighbors to blow up a 7-Eleven. You could give it anything and it will give you a perfect answer, just the same way that it does for

17:45S1:positive things. It will basically help you do those things. It will help you hack anything. It will help you kill anyone and get away with it. It's a criminal mastermind, but it's also like the best possible, like angelic system help you do both. It has no goals or alignment. It just lets you do whatever you want to do, and it's going to use all of its super intelligence to help you do that thing. My argument is that if you're reading this in the first few pages of this sci fi novel, and this is your protagonist who works at this lab, kind of like Miles Dyson at like in the Terminator movies or whatever, the person who's like creating this thing, getting ready to release it. You would be thinking as a sci fi reader, maybe this is bad. They are rushing too quickly. Maybe we already know the plot of this book. The plot of this book is that this was maybe a dumb idea to give this powerful of intelligence to. Let's say it was 25 billion people on this planet. Let's say it's a bigger planet than ours or whatever, or they're further along.

18:49S1:So let's just give this to everyone. Let's just give it to everyone. Let's just that's what this lab, where this protagonist works, that's what they're trying to do, right? So there's discussion in the first part of the book. Should we do it? Should we not do it? My argument is that if you see this from another planet, another race of people, another situation, it's kind of obvious that this is dangerous, right? We shouldn't do this. Have some sort of guardrails. We should slow down. We should have all the governments get together and talk about it or something like that. Right. So I think that part is super obvious where I think the breakdown is, is when it's like, okay, but now what? Right. So if you're with me up to this point, right? And you understand that if you're reading this book, it's bad for that race on that planet to release this unrestricted open source AI to everyone on the planet so they could do maximum harm with maximum efficiency and like, get away with it, then why would we want to do it for our planet, because that is literally the moment that we are actually in right now. We are literally that

20:03S1:book in that world right now. That's what we have. We are precarious from a nuclear standpoint already, right? One small mistake. We're all blown up from nuclear weapons, right? Everyone's known that forever, right? We've got all this situation with our governments. We've got economies having trouble. We're transitioning potentially out of like labor and jobs because of automation and AI. Like there's just massive stress in this world. And we have a lot of conflict right now. It is not a great time to enable every single person on the planet with ways of harming countries and people and neighbors and neighbors, dogs who bark too much or whatever, right?

20:47S1:Killing your spouse, like all sorts of things. So people wish they could do harm, right? A very small percentage, I would say vast majority of people are massively good, not evil. But there's a large group of people, let's say a tiny, tiny percent like 0.01% of the entire planet. You know what that is multiplied by 8 billion people. It is a crap ton of people, and they are the type of people who they don't do bad things because they don't have the expertise to. And they would get caught if they tried, right? I don't know the number of those people. I don't know the percentage. No one really does. I'm sure we could research it and find out, but I could tell you it's probably millions of people, if not tens or hundreds of millions of people.

21:38S1:And it doesn't mean they would become a super villain, but they would potentially do some harm, right? Especially if they're under massive pressure and it's just something they have to get done. Now, another way to think about this whole thing. I just talked to my friend Jason about this in extreme detail, and I think we came up with a good point about this. A good way to explain this. We have laws that don't prohibit you from doing things that are bad, right? Because if you watch like forensic files or something, a lot of those people who actually do the things, they actually commit these crimes that are just regular people and for whatever reason, they were in this bad situation, or they had too many drinks, or they were going through a divorce or whatever, and they committed the crime, right? We have laws that apply friction so that somebody who is mostly good hits a little bit of friction, so they don't cross over and start doing the bad thing, or they actually do the bad thing they can never recover from. They kill someone or they hurt someone, they end up in jail or whatever.

22:45S1:We have friction in place through all sorts of like, laws at different layers to try to make sure that this doesn't happen, right? So for example, if you walk into a gun store and like the most gun friendly place in the world, like Texas or something, and you're like seething, you're just like sweating, you're super angry. And you're like, yeah, I want that handgun right there. I want it right now. Oh, really? Why do you want that? It's like, oh, my wife. I can't believe it. I caught her on the ring camera. She did this or that. Like, they will not give you the gun. They just won't.

23:20S1:They might even, like, put you on a list of where you can't buy a gun in that store anymore, right? We have those types of controls. We also have these types of controls with like drugs. You can't just go and get fentanyl. You can't get fentanyl at 7-Eleven. You can't get heroin at Safeway. Does that mean you can't get these drugs? No it doesn't. They're available on the dark web. They're probably available behind Safeway at night. There are ways to get these drugs. It doesn't mean just because there are ways to get them that we should make them widely available to everyone. The moment they have the first thought of taking that drug or committing suicide, or killing someone, or hurting someone, or psychologically manipulating someone or doing a terrorist event. There are pushbacks all throughout society to try to dissuade that person from doing it right. Friction.

24:18S1:Layers of friction. So what I'm really talking about here is I'm talking about a system that is extremely tightly controlled, narrowly controlled, okay, narrowly controlled, so that when we have this free thought model that comes out, that is mythos plus plus, right? And I think we're ridiculously close to this actually happening. Free thought version 3.0, whatever it is, it comes out and it's just the most amazing thing ever. It can help you start your business. It can. Just like all the things we talk about with agents that fully enable you. But it could also fully enable mass crime and all the things we talked about harming people.

25:06S1:Teach me how to psychologically manipulate my family or my significant other. Help me run the most insane, like gaslighting campaign to convince them to do this or do that. Like these things are all fully capable from a model who knows how to manipulate human psychology, right? It can give you step by step guide on exactly how to do this stuff right. My idea here is essentially we have this free thought version 3.0 model right now in its released form, in its free form, in the Non-controlled form, it would be a criminal mastermind, infinitely smarter than, say, Einstein or whatever on all these different fields. Biology, all these different capabilities. It would be a super mastermind and it would be available to everyone. Okay. Which is basically 7-Eleven and Safeway and whatever the equivalent stores are, right?

26:09S1:And you would just be able to do that. And my argument is that would be bad for society. Now what do you do instead? You have these tight little controls, psychological manipulation, misinformation campaigns, the creation of weapons, the creation of drugs, the distribution of these drugs, basically criminal activities, harmful malicious activities, but with these very tight labels. Okay. So what we do is the proposal, what we do is we come up with a global joint standard for this, for what these labels are and what the controls are, the types of prompts and questions that they would defend against. Right.

26:55S1:So in this model, Free thought version three, when it gets released, it is released on Hugging Face. Hugging face. Let's just call it safe. SAFE. It's not a real I'm not sure the real acronym I'm going to try to make here. I'm going to actually try to build this thing, but let's just say it's called safe. Okay? Hugging face is a safe distribution center, okay? Because it's legal. It's a legal place. The governments all around the world approve hugging face. Why? Because they distribute things according to safe. Now, this is important. Going back to the whole mission of, like, open information, the free thought 3.0 model is just as smart or smarter than anything from anthropic. Anything from OpenAI on 99.999% of all use cases, and in fact, 99.9999.

27:53S1:Whatever percentage of all people who use the model will never know that it's actually controlled, that it's actually restricted. Why? Because that percentage of people are good people. They just never want to do these types of things. Okay. At the same time, the evil versions of these models, because the moment Free Thought 3.0 comes out, it'll be jailbroken. All the controls will be stripped out. That will go to the underground, to the dark web, to wherever. BitTorrent, right.

28:27S1:Truly evil people are going to have the most nasty stuff. That is not what we're trying to protect. That is not possible to protect, right? What we're talking about is the 99.99% of people who are good, who might be Psychologically under pressure, they might be kind of forced to go do something. Like they're in a bad state and they start thinking negative thoughts. How do I kill myself? How do I hurt someone else? How do I disrupt this economy? How do I hack my local community bank and give myself 100,000 more dollars? These are the types of nudges and the types of like, friction pieces that we want to apply to the vast majority of people. Okay.

29:20S1:Just so that in their darkest moments, they're not enabled to go and do this thing. Now, again, if they're actually evil, they could just go and do the super bad thing because they can go and find the super bad thing and it's fully available. They can use it or whatever. At that point, though, that's a lot of friction. You have to go get the thing. You have to download it, you have to run it, and you have to do so while knowing that like this is violating global earth international law to do this right. Another facet of this, which is important is like, what if you're an expert? And this is another thing I talked about with Jason. What if you're a cybersecurity expert and free thought 3.0 by default has a cybersecurity designation, a blocker on it, right? Because again, you don't want the whole planet to be able to instantly just go hack something.

30:15S1:What do you do then? And I think the answer is, and I think Jason agreed with this, I don't want to speak for him. I'm sure we'll talk about it later. But I think the solution here is you have people who are known experts in cybersecurity and known experts in creating viruses, like at the CDC or the World Health Organization or whatever. They're trying to make vaccines or whatever they're trying to do. That's like legal above board. And there are designated experts in here who have credentials and the system itself, right? This whatever, this global standard or whatever.

30:53S1:This basically looks into the person. If they identify themselves at this as this trusted expert or whatever, they get free thought 3.0. But without that thing protection in place. So for me, it'll remove the cyber protection. It won't remove the viral like create viruses in your backyard protection, right. Because I don't need that one. Why would I ask those questions anyway? Right. So I get the cyber approved model right. And here's the other important part. I should be able to do the same thing with the closed source models. That same designation that put me as an expert in cyber security. That means open AI should also let me use whatever GPT 75, which is super powerful on cyber, but that would be tagged in my account as approved for me. So all the prompts and everything that runs. Same with the open source models that would be approved for me because of who I am now, that would just be the above board. This is the 99.999%.

31:59S1:Who are the above board experts who are using this stuff, and the non-experts just everyone who gets the regular free thought 3.0 model, and that one has these controls in it, but they never know because they're not asking those prompts anyway. And most importantly, out of all of this free thought, 3.0 is just as good or better than all the closed source models, right? Or at least it's very close, right? There's not some sort of hobbling, some sort of disabling of the high quality capabilities in the lower models so that it forces people to want to use the closed ones.

32:41S1:That's gross. That's what we need to avoid. We want to be able to apply friction to tightly controlled, super bad things that most people will never ask about anyway. And we need to have a global sort of distribution system for open source models based on a standard that all the experts, right from other countries, from our countries, from AI experts, ethical experts, cyber experts, viruses, terrorism, all that stuff. They're the ones we all collectively come up with these labels and then distribution points like Hugging Face and lots of other ones, just like Hugging Face, they are safe enabled, right? And in fact, they actually have all the models, but you just authenticate in and it's like, oh, I'm also a virologist and I'm also a cyber person and I'm also a counterterrorism expert. Well, guess what? They issue me the version of Free Thought three that doesn't have those three controls in it. So it's just like I'm using the regular one. It just doesn't have those three controls. So I think this is the model that if any super smart visiting aliens were to analyze, they'd be like, yeah, obviously that's kind of what we need

33:59S1:right now. The obstacle to getting to this system is the distrust. Someone at this planet scale event I was just part of made a great point about this. A big obstacle to moving towards this is cynicism. We've lost a lot of trust in our institutions. We've lost trust in their corporations. We've lost trust really in our governments, our national governments, our world governments. So it's kind of a hard pill to swallow to be like, look, we're going to build this global consortium. It's going to be experts.

34:35S1:We're all going to agree. And, you know, it's going to be a really clean standard. And you know, this is the way it should work. Like, I think people can get to the point of what I'm arguing to say. Yes, I agree. If we were able to do that, that would be good. Hopefully I've already brought you there. Then the question is implementation. Can we really do this? Isn't this going to be gross trying to get there? How is what about China? Is China going to participate in this?

35:04S1:I don't know. I don't know if they're going to participate in this, but this is the thing we actually need to build. That's my argument. And that is my explanation for why I do actually believe we should have some narrow controls on AI models, both closed source and open source. And I hope this spawns you to think about it in a different way. And let me know in the comments if you agree or disagree. If you have some ideas about what the standard should be. I'm actually going to start working on this with Jason and some others. And yeah, I hope this was worth your time and we'll see you in the next one.