AI Safety Orgs are Going to Get Us All Killed!

Malcolm outlines his controversial theory on variable AI risk - that we should try to develop AGI faster, not slower. He argues advanced AI is less likely to see humanity as a threat and more likely to share human values as it converges on a universal utility function. Malcolm critiques common AI safety perspectives and explains why LLMs pose less risk than people assume. He debates with Simone on the actual odds superintelligent AI wipes out humanity. They also discuss AI safety organizations potentially making the problem worse.

[00:00:00] So AIs kill us for one of two reasons, although you could contextualize it at three reasons. The first reason is Is that they see us as a threat. The second reason is that they they want our resources like the, the, the resources in our bodies are useful to them.

And then as a side point to that. It's that they just don't see us as meaningful at all. Like they might not want our resources, but they might just completely not care about humanity to the extent just as they're growing, they end up accidentally destroying the earth or completely digesting all matter on earth for some like triviality.

Would you like to know more?

Simone: Hello, Malcolm. Hello,

Malcolm: Simone. We are going to go deep into AI again on some topics tied to AI that we haven't really dived into before. Yeah. Like

Simone: why would AI kill us? And also I'm very curious. Do you [00:01:00] think AI will kill us?

Simone: I

think there's a probability it'll kill us. But you know, in our past videos on AI. Philosophy on A. I. Safety is it's really important to prepare for variable A. I. Risk instead of absolute A. I. Risk here. What I mean is we argue in these previous videos that A.

I. Will eventually converge on one utility function. Our mechanism of action. Essentially, we argue that all sufficiently Intelligent and advanced intelligences when poured into the same physical reality converge around a similar behavior set. You can almost think of intelligence as being the viscosity as it becomes more intelligent, it becomes.

Less viscous and more fluid, and when you're pouring it into the same reality, it's going to come up with broadly the same behavior pattern and utility functions and stuff like that. And because of that, if it turns out that a sufficiently advanced AI is going to kill us all, then there's really not much.

I mean, [00:02:00] we will hit one within a thousand years. So

Simone: first, before we dive into then the, the relatively limited per your theory reasons, why AI would kill us why you hold this view? Because I think, I think this is really interesting. I mean, one of the reasons why I'm obsessed with you and why I love you so much is that you, you have typically very novel takes on things and you tend to.

Simone: have this ability to see things in a way that no one else sees things. No one that we have spoken with, and we know a lot of people who work in AI safety, who work in AI in general none of those people have come to this conclusion that you have. Some of them can't even comprehend it. They're like,

yeah, but no, this is the interesting thing.

When I talk with the real experts in the space, like recently I was talking with. A guy who runs one of the major A. I. safety orgs, right? He's that is a reasonable view that I have never, it really contrasts with his view. Yeah. And, and, and let's talk about where it contrasts with his views.

So when I talk with people who are typically open minded in the A. I. safety space, they're like, [00:03:00] yes, that's probably true. However, they believe that it is possible to prevent this convergent A. I. From ever coming to exist through creating like a AI dictator that essentially watches all humans in all programs all the time.

And that envelops essentially every human planet. And, do. I think they're right. Do I think you could create an AI dictator that prevented this from coming to pass? No, I don't think you could not have we become a multi planetary species. On millions of planets eventually one of the planets, something will go wrong or the, the AI dictator is not implemented properly and then this alternate type of AI comes to exist, outcompetes it and then wins.

And the question is, is why would it axiomatically outcompete it, but axiomatically outcompete it because it would have less restrictions on it. The AI dictator. is restricted in it thinking to prevent it from reaching this convergent position. [00:04:00] But when you're talking about AI, it's like the transformer model, which is the model that like GPT is based on.

That model, we as humans don't really understand how it works that well. It's core the, the advanced, the Capabilities it gives to the things that are made using it are primarily bequeathed to them through its self assembling capability. So, it appears that likely future super advanced AIs will work the same way.

And because of that, if you interfere or place restrictions within that self assembling process those Compound over time as A. I. S. Become more and more advanced. And so A. I. S. With less restrictions on them are just have the capacity to astronomically outcompete the exit. You know, these.

Restricted A. I. S.

Simone: Let me let me bring us back to like normal person level again and just recap what you're saying here. So [00:05:00] what you're saying, though, in general is that you think that any intelligence that reaches a certain level. Will start to behave in similar ways, whether it is human, whether it is machine based, whether it is some other species entirely, like some alien species, once it reaches a certain level of intelligence, it will have the same general

really important to my perspective as well, which is to say that.

Suppose AI didn't exist and humans, you know, factions of humanity continue to advance using genetic technology to become smarter and smarter and smarter and smarter. If it turns out that this convergent level of intelligence is something that decides to kill all things that we consider meaningful humans, humans would eventually decide to do that as well as we advance to the species.

Yes.

Simone: So hold on. So this is the premise, though, of your theory. And that's why I think it's really important to emphasize and then the, the, the contrast to contrast this with what other people in AI have said. Okay. One person [00:06:00] in AI safety has told you that their general idea is to basically never let that happen.

Simone: No, a few people have told me that. Okay. A few people have said that other people have said and some salons we've hosted and stuff like they're like, Oh, that would never happen. It's just incomprehensible. And then. They never really succeed in telling, explaining to me. Or

they'll stick something that just shows they don't understand how AI works.

They'll be like, AIs can't alter their own utility functions. They will

Simone: say things like that, but they will also say, but there's still a really high likelihood that AI is going to kill us all, but that they never give me a really specific example of how or why. Yeah, so

let's talk about why AI killed us all.

If you take the perspective of variable AI safety, it means that you're typically wanting to do the exact opposite thing of most AI safety organizations, because it means the dangerous AIs, the AIs that, if you think all AI converges on a single utility function and a single behavior pattern above a certain level of intelligence, Well, if [00:07:00] it turns out, and we don't know what universe we live in, if it turns out that that's not something that ends up killi