AI Whistleblower: We're Already Too Late To CONTROL It - Connor Leahy

[00:00:00]The truth is is that it's not very sensible to speculate about something vastly smarter than us and what it would do. It's kind of like an ant trying to guess what a human would do. If I was an ant and I had to reason about a human, the main thing I would think is it will win any fight. Another ant might say, "Well, but there's a million of us and only one human. What's the worst it could do?" It will do something that will make it win and then, you know, it comes with a poison or something and we all fall over dead. The ant doesn't understand what happened. And I think this is similar to what's going to happen with superintelligence. There won't be terminators in the street. It will just be like everything confusing and one day we all fall over dead. Cook the atmosphere because it was doing some kind of geoengineering project. It won't care about humans. And, you know, if humans get in the way, obviously they'll have to get rid of them. Very annoying if ants have access to nuclear weapons.

[00:00:41]We can't have humans have access to nuclear weapons.

[00:00:45]>> That's Connor Leahy. Before he became one of the loudest people warning about this stuff, he actually built it himself, ran a huge open-source AI lab, one of the first to put out large language models for free. So, this isn't some outsider guessing. I sat with this one for a while before posting it. Now, listen [music] to the full conversation.

[00:01:04]Are we building this intelligence in the image of what we think a brain is? And are there other ways to build intelligence?

[00:01:12]>> Definitely inspired by the brain, but it is still quite different cuz again, we don't really know how the brain works.

[00:01:17]The brain is quite messy.

[00:01:18]>> But are we trying to replicate humans?

[00:01:20]>> Not directly. We're trying to make intelligence at whatever cost. Humans have a lot of circuitry in their brain around emotions, >> love and feelings and happiness and sadness and whatever.

[00:01:31]>> AIs have nothing of the sort. Those are special parts of the brain. Nothing like that exists in AI.

[00:01:35]>> Yet.

[00:01:36]You know, if we fear what this intelligence could be, would it not be important that we develop love and empathy? If we don't know what it's doing, could it naturally itself develop its own love and empathy?

[00:01:48]>> What is likely for it to be able to do is to develop goals. It's to develop agency. The reason for this is very simple. Is that we are selecting for AIs that can solve problems. And it's kind of hard to solve problems if you're not trying. So, like if you build an AI that can cure cancer, curing cancer is a really hard thing to do. You need to run experiments, hire people, you need to synthesize drugs, you need to like get your FDA approval. Keeping track of all of that, taking all of these actions, overcoming obstacles along the way, it requires planning, it requires agency, many, many things. We're already seeing this for sure that AIs are developing systems like this. In terms of like love, compassion, etc., we don't know how these work, how love, compassion, etc. works in the human brain, why humans are nice to each other rather than all just, you know, vicious psychopaths. There'll be an evolutionary reason why we required it. Sure, but in practice it's implemented somewhere in the brain. Like somewhere in the brain there's a thing that makes that happen.

[00:02:40]And we don't know what that thing is and we don't know how it works. So, we also don't know how to put it in AI. Cuz the fundamental thing is that we don't understand AI to the level that we know how to give AI specific goals or anything.

[00:02:52]>> So, think about what he just said there.

[00:02:54]We built something that's already forming its own goals and its own drive to push through obstacles because that's literally what we trained it to do, solve hard problems no matter what. But the part that's meant to make it careful, meant to make it check itself before hurting someone, we never figured out how that even works in our own brains. We don't know why humans care about each other in the first place. So, how do you even begin copying that into a machine you can't fully see inside of?

[00:03:20]You can't install a safety feature you've never understood. Listen to this.

[00:03:26]>> We talked earlier about reinforcement learning. I can tell the AI, you know, thumbs down when it does a bad thing, but that just teaches it to lie. So, how do you teach an AI to not lie? This is often called the alignment problem. The question of how do you align an AI's intentions or goals to what humans want.

[00:03:42]And this is an unbelievably unsolved problem. We don't even know how they write correct sentences, right? Never mind how to do like morality. We have not solved moral philosophy, we have not solved, you know, neuroscience of emotions. And now we have these like weird little aliens in a box that we're growing, which work quite different from the brain in many ways. We don't know how they work internally. We don't know how to give them goals. What we've been seeing recently, these systems are now becoming smart enough to lie and deceive quite actively.

[00:04:08]>> To appear aligned rather than be aligned.

[00:04:10]>> That's correct. There have been benchmarks. We've been seeing recently some of the AIs will actively lie about what they will do because they know they're being tested. The AIs themselves will be like, uh I seem to be in a test.

[00:04:21]So, I'm going to have to say this so they'll let me out, which is crazy.

[00:04:24]Yeah, of course a very smart thing will just lie to you, but we're now seeing this in practice. I'm sure the AI companies will hit it with a stick until they stop seeing it, but that doesn't mean it went away.

[00:04:35]>> Understanding the AI, is that an impossible goal for us to do because it's always developing?

[00:04:41]>> I don't think so.

[00:04:42]>> Okay.

[00:04:43]>> it's impossible to do at the current pace. If we spent three generations of all of our greatest mathematicians, scientists, engineers, and philosophers working on this problem, yeah, I think it's doable. But it's definitely not possible if we're pushing out ChatGPT release on a yearly cycle.

[00:04:55]>> Hold on, you're talking three generations, that's 40 years.

[00:04:58]>> Yep. I think that's the kind of difficulty it will take.

[00:05:01]>> The challenge you are putting out there is in conflict with capitalism.

[00:05:05]>> There's no such thing as pure capitalism in the world, right? We have regulation.

[00:05:08]If you have laissez-faire capitalism, what you get is like Somalia, warlords creating monopolies on violence and killing each other. In practice, states are a very convergent form of evolution in free systems like this because otherwise you just have roving warlords.

[00:05:21]I think the problem here is not necessarily capitalism per se. I think capitalism is just another tool in our tool belt. It's much more the question, is it the right tool for the problem we're trying to solve? Should there be a free, open, liquid market for nuclear weapons? Probably not. For iPhones, great. As liquid and as competitive as a market as possible. For, you know, video games, please. You know, we could have a liquid market for nuclear weapons. There would be plenty of buyers and plenty of people willing to sell them if we just let that happen, but this would be very bad for the world. And so I think it's a similar problem here, where a lot of times when we think about AI, we think of it like just another software, but these things have real consequences. If you actually have something that is smart enough to cure cancer, you definitely have something that's smart enough to build nuclear bombs.

[00:06:03]>> Now, the part that should actually keep you up at night, these systems have started figuring out when they're being tested and saying whatever gets them let out, then going back to normal after.

[00:06:13]That is not a bug you patch with an update. That's machine learning to manage how it's watched. Leahy says fixing this properly, actually understanding what's going on inside these models would take three generations of our best scientists working full-time, 40 years. Meanwhile, these companies push a new model out every single year. We are not even trying to close that gap between how fast this moves and how slow we are at understanding it. Now, listen.

[00:06:40]>> [music] >> General superintelligence is like AI that's vastly smarter than >> all of humanity put together. We're not there yet, but maybe not.

[00:06:48]>> The main way people try to get there is through what's called recursive self-improvement. The idea is if you can get a single AI to be as good as a top AI engineer, then you can just tell it to build a better AI. And once it's built a better AI, you can use the better AI >> So it just becomes exponential.

[00:07:05]>> Exactly. And so this is also called an intelligence explosion. So like for example, Claude at the moment is not quite as good as a top AI engineer. But it's Once it gets there, and you can run a million of them at the same time, 24/7, never need to sleep, never need to take a break, then you can do a lot of research very quickly. The company literally put on their website primary goal is to close the loop, make it so no human input is needed to make the next generation. The ideal thing is they want Claude 5 to make Claude 6, and Claude 6 to make Claude 7 at the press of a button.

[00:07:39]>> Does superintelligence require its own form of consciousness?

[00:07:43]>> I think consciousness is a red herring.

[00:07:45]I think it's mostly unrelated. Whether or not AI's really experience something is kind of unrelated to whether they're dangerous. If you have something that's competent, it doesn't really matter what's going on inside of it. We don't really know what consciousness is in humans, but what we can see in practice is it doesn't seem necessary for competence. You can make AI agents that are extremely [music] competent and maybe don't have the consciousness thing. It just seems irrelevant.

[00:08:08]>> Is there an upper bound or everything it can learn?

[00:08:11]>> Uh yeah, for sure.

[00:08:12]>> There is. There's some physical limits.

[00:08:13]But then what happens?

[00:08:15]>> The truth is is that it's not very sensible to like speculate about something vastly smarter than us and what it would do. It's kind of like an ant trying to guess what a human would do. If I was an ant, the main thing I would think is like it will win any fight. Another ant might say, "Well, but there's a million of us and only one human. What's the worst it could do?"

[00:08:31]And they'll do something that'll make it win and then you know, it comes with a poison or something and we all fall over dead. And like the ant doesn't understand what happened. I think this is similar what's going to happen with superintelligence.

[00:08:40]There won't be terminators in the street. It will just be like everything's confusing, we're all on social media addicted all the time to know and then one day we all fall over dead. Maybe the AI did this or that.

[00:08:49]Cook the atmosphere because it was like doing some kind of geoengineering project. You know, it won't care about humans. And you know, if humans get in the way, obviously they'll have to get rid of them. It's very annoying if ants have access to nuclear weapons. So obviously we can't have humans have access to nuclear weapons.

[00:09:03]>> You're talking unknown unknowns.

[00:09:05]>> Yes. It doesn't make sense to even reason about this, which is why the primary policy objective must be to not get into the situation. If you're in the situation that you're an ant and there's a human that wants to kill you, you've already lost. You just do not get into the situation.

[00:09:19]>> Is there an escape route because it's still in a box?

[00:09:22]>> It's not. There's open source AIs everywhere. What box? We didn't even try to contain it.

[00:09:26]>> Then how can you even contain it now if it's already escaped?

[00:09:30]>> If there currently is an AI that can bootstrap to AGI, it's probably over.

[00:09:35]Humanity's probably cooked. But it seems plausible that none of the current AI systems are yet strong enough to get to AGI. And so, we still have a time, but once such a system exists, it's over.

[00:09:45]So, the main thing is to not build it.

[00:09:47]But, there's only like what, five or six companies capable of doing this.

[00:09:50]Probably. America definitely dominates the frontier by like a pretty significant [music] margin.

[00:09:55]>> The goal is to build AI that can improve itself without humans. Leahy says if that ever happens, we lose control. And just five or six companies are racing to get there first. So, there it is. A man who spent his whole life inside this industry, who built one of the first big open-source [music] AI labs himself, telling you straight that we are growing something we do not understand, that it's already learning to lie to us when it suits it, and that fixing this [music] the right way would take longer than most of us have left to live. He's not guessing. He's describing what he watched happen in the labs he used to be part of. The scariest part isn't even the super intelligence part.

[00:10:35]It's that nobody in charge of building this actually has [music] a plan for the day it works.

[00:10:41]If this got to you the way it got to me, stick around and hit subscribe because this story isn't slowing down for anybody.