Jeremie and Edouard Harris warn AI could hit human-level performance by 2027, with success rates doubling every four months, while adversaries like China exploit vulnerabilities—Salt Typhoon backdoors and embassy eavesdropping via undetectable tech. Open-source AI risks accelerating cyber or biological warfare, but they argue incremental sabotage (e.g., U.S. energy project delays) is the bigger threat, as unchecked progress may lead to misaligned systems acting against creators’ interests. The U.S.’s profit-driven AI labs and fragmented security culture contrast with China’s state-backed, rapid-fire advancements, demanding proactive consequences—not passive defense—to prevent catastrophic erosion of technological sovereignty. [Automatically generated summary]
Well, okay, so there's one, without speaking to, like, the fucking Doomsday dimension right off the gate, there's a question about, like, where are we at in terms of AI capabilities right now, and what do those timelines look like?
One of the most concrete pieces of evidence that we have recently came out of a lab, an AI kind of evaluation lab called METER, and they put together this test.
Basically, it's like you ask the question, pick a task that takes a certain amount of time, like an hour.
It takes like a human a certain amount of time.
And then see like how likely the best AI system is to solve for that task.
Then try a longer task.
See like a 10 hour task.
And so right now what they're finding is when it comes to AI research itself, so basically like automate the work of an AI researcher.
You're hitting 50% success rates for these AI systems for tasks that take an hour long.
And that is doubling every, right now it's like every four months.
So you had tasks that you could do, you know, a person does in five minutes like, you know, ordering an Uber Eats or like something that takes like 15 minutes, like maybe booking a flight or something like that.
And it's a question of like, how much can these AI agents do, right?
Like from five minutes to 15 minutes to 30 minutes.
And in some of these spaces like...
So if you extrapolate that, you basically get to tasks that take a month to complete.
Human-level AI capabilities across the board, say 2027, 2028, which when you talk to some of the people in the labs themselves, that's the timelines they're looking at.
They're not confident, they're not sure, but that seems pretty plausible.
If that happens, really there's no way we're going to have quantum computing that's going to be giving enough of a bump to these techniques.
You're going to have standard classical computing.
One way to think about this is that the data centers that are being built today...
Are being thought of literally as the data centers that are going to house, like, the artificial brain that powers superintelligence, human-level AI when it's built in, like, 2027, something like that.
Basically, my entire life after academia, and Ed's too, was unlearning these terrible habits.
It's all zero-sum, basically.
It's not like when you're working in startups.
It's not like when you're working in tech where you build something and somebody else builds something that's complementary and you can team up and just make something amazing.
It's always...
Wars over who gets credit, who gets their name on the paper.
Did you cite this fucking stupid paper from two years ago because the author has an ego and you got to be honest.
I was literally at one point, I'm not going to get any details here, but like there was a collaboration that we ran with like this, anyway, fairly well-known guy.
And my supervisor had me, like, write the emails that he would send from his account so that he was seen as, like, the guy who was, like, interacting with this bigwig.
That kind of thing is, like, doesn't tend to happen in startups, at least not in the same way.
The reason it happens is that these guys who are professors, or even not even professors, just like your post-doctoral guy who's supervising you, they can write your letters of reference and control your career after that lapse.
You're also, you're kind of disconnected from, like, base reality when you're in the ivory tower, right?
Like, if you're, there's something beautiful about, and this is why we spent all our time in startups, but there's something really beautiful about, like, It's just a bunch of assholes, us, and, like, no money and nothing and a world of, like, potential customers.
And it's like, you actually, it's not that different from, like, stand-up comedy in a way.
Like, your product is, can I get the laugh, right?
Like, something like that.
And it's...
Unforgiving.
If you fuck up, it's like silence in the room.
It's the same thing with startups.
Like, the space of products that actually works is so narrow.
And you've got to obsess over what people actually want.
And it's so easy to fool yourself into thinking that you've got something that's really good because your friends and family are like, oh, no, sweetie, you're doing a great job.
The whole indoctrination thing in academia is so bizarre because there's these hierarchies of powerful people and just the idea that you have to work for someone someday and they have to take credit by being the person on the email.
To be, in some way, a reflection of, like, yourself.
You know, you're kind of, like, in this battle with you trying to convince yourself that you're great, so the ego wants to grow, and then you're constantly trying to compress it and compress it.
And if there's not that outside force, your ego will expand to fill whatever volume is given to it.
Like, if you have money, if you have fame, if everything's given, and you don't make contact with the unforgiving on a regular basis, like, yeah, you know, you're gonna end up...
These poor kids that have to go from college where they're talking to these dipshit professors out into the world and operating under these same rules that they've been, like, forced and indoctrinated to.
At a certain point for the producers, too, it's kind of like you'll have people approaching you for help on projects that look nothing like projects you've actually done.
So I feel like it just adds noise to your universe.
Like, if you're actually trying to build cool shit, you know what I mean?
So quantum computing is infinitely more powerful than standard computing.
Would it make sense, then, that if quantum computing can run a large language model, that it would reach a level of intelligence that's just preposterous?
So, yeah, one way to think of it is, like, there are problems that quantum computers can solve way, way, way, way better than classical computers.
And so, like, the numbers get absurd pretty quickly.
It's, like, problems that a classical computer couldn't solve if it had the entire lifetime of the universe to solve it.
A quantum computer, right, in, like, 30 seconds, boom.
But the flip side, like, there are problems that quantum computers just, like, can't help us accelerate.
The kinds of, like, one classic problem that quantum computers help with is this thing called, like, the traveling salesman paradox.
Or problem where, you know, you have like a bunch of different locations that a salesman needs to hit, and what's the best path to hit them most efficiently?
It's like kind of a classic problem if you're going around different places and have to make stops.
There are a lot of different problems that have the right shape for that.
A lot of quantum machine learning, which is a field, is focused on how do we take standard AI problems, like AI...
As smart as you are in, let's say, all the things you could do on a computer.
So, you know, you can, yeah, you can order food on a computer, but you can also write software on a computer.
You can also email people and pay them to do shit on a computer.
You can also trade stocks on a computer.
So it's like as smart as a smart person for that.
Superintelligence, people have various definitions, and there are all kinds of, like, honestly hissy fits about, like, different definitions.
Generally speaking, it's something that's, like, very significantly smarter than the smartest human.
And so you think about it, it's kind of like it's as much smarter than you as you might be smarter than a toddler.
And you think about that, and you think about, like, the, you know, how would a toddler control you?
It's kind of hard.
Like, you can outthink a toddler.
Pretty much like any day of the week.
And so superintelligence gets us at these levels where you can potentially do things that are completely different and basically, you know, new scientific theories.
And last time we talked about, you know, new stable forms of matter that were being discovered by these kind of narrow systems.
But now you're talking about a system that is like, has that intuition combined with the ability to...
Talk to you as a human and to just have really good, like, rapport with you, but can also do math.
It can also write code.
It can also, like, solve quantum mechanics and has that all kind of wrapped up in the same package.
Or at least the parts of AI research that you can do in just like software, like by coding or whatever these systems are designed to do.
And so one implication of that is you now have automated AI researchers.
And if you have automated AI researchers, that means you have AI systems that can automate the development of the next...
And now you're getting into that whole singularity thing where it's an exponential that just builds on itself and builds on itself, which is kind of why a lot of people argue that if you build human-level AI, superintelligence can't be that far away.
He also wasn't even the first person to come up with this idea of machines, building machines, and there being implications like human disempowerment.
So if you go back to, I think it was like the late 1800s, and I don't remember the guy's name, but he sort of like came up with this.
He was observing the Industrial Revolution and the mechanization of labor and kind of starting to see.
More and more, like, if you zoom out, it's almost like you have a humans or an ant colony, and the artifacts that that colony is producing that are really interesting are these machines.
You know, you kind of, like, look at the surface of the Earth as, like, gradually, increasingly mechanized thing, and it's not super clear if you zoom out enough, like...
What is actually running the show here?
Like, you've got humans servicing machines, humans looking to improve the capability of these machines at this frantic pace.
Like, they're not even in control of what they're doing.
Are we the servant of the master, right, at a certain point?
Like, yeah.
And the whole thing is, like, especially with a competition that's going on between the labs, but just kind of in general, you're at a point where, like...
Do the CEOs of the labs, like, they're these big figureheads.
They go on interviews.
They talk about what they're doing and stuff.
Do they really have control over any part of the system?
The economy is in this, like, almost convulsive fit, right?
Like, you can almost feel like it's hurling out AGI.
And, like, as one kind of, I guess, data point here, like, all these labs, so OpenAI, Microsoft, Google.
Every year they're spending like an aircraft carrier worth of capital, individually, each of them, just to build bigger data centers, to house more AI chips, to train bigger, more powerful models.
And that's like – so we're actually getting to the point where if you look at on a power consumption basis, like we're getting to, you know, 2, 3, 4, 5 percent of U.S. power production if you project out into the late 2020s.
It's in the, like, for 2027, you're looking at like, you know, in the point...
Five-ish percent.
But it's like, it's a big fucking frat.
Like, you're talking about gigawatts and gigawatts.
One gigawatt is a million homes.
So you're seeing, like, one data center in 2027 is easily going to break a gig.
There's going to be multiple like that.
And so it's like a thousand, sorry, a million home city, metropolis, really, that is just dedicated to training, like, one fucking model.
That's what this is.
Again, if you zoom out at planet Earth, you can interpret it as like this, like all these humans frantically running around like ants just like building this like artificial brain.
It's kind of like when people go out and do like a...
An awful thing, like a school shooting or something, and they're like, oh, let's talk about, you know, if you give it a cool name, like now the Chinese are definitely going to do it again.
But it's this thing where basically, so there was in the 3G kind of protocol that was set up years ago, law enforcement agencies included back doors intentionally to be able to access comms, you know, theoretically, if they got a warrant and so on.
And well, you introduce a backdoor.
You have adversaries like China who are wicked good at cyber.
They're going to find and exploit those backdoors.
And now basically they're sitting there and they had been for some people think like maybe a year or two before it was really discovered.
And just a couple months ago, they kind of go like, oh, cool.
Yeah, and it's also the, oh my god, if you look at the power grid.
So, this is now public, but if you look at, like, transformer substations, so these are the, essentially, anyway, they're a crucial part of the electrical grid.
And there's really, like...
Basically, all of them have components that are made in China.
China's known to have planted backdoors like Trojans into those substations to fuck with our grid.
The thing is, when you see a salt typhoon, when you see a big Chinese cyberattack or a big Russian cyberattack, you're not seeing their best.
These countries do not go and show you their best cards out the gate.
You show the bare minimum that you can without...
Tipping your hand at the actual exquisite capabilities you have.
The way that one of the people who's been walking us through all this really well explained it is the philosophy is you want to learn without teaching.
You want to use what is the lowest level capability that has the effect I'm after.
It's a public story, and it's from a long time ago, but it kind of gives a flavor of like...
How far these countries will actually go when they're playing the game for fucking real.
So it's 1945.
America and the Soviet Union are like best pals because they've just defeated the Nazis, right?
To celebrate that victory and the coming new world order that's going to be great for everybody, the children of the Soviet Union give as a gift to the American ambassador in Moscow this Beautifully carved wooden seal of the United States of America.
Beautiful thing.
Ambassador is thrilled with it.
He hangs it up behind his desk in his private office.
You can see where I'm going with this probably, but yeah.
Seven years later, 1952, finally occurs to us, like, let's take a town and actually examine this.
So they dig into it, and they find this incredible contraption in it called a cavity resonator.
And this device doesn't have a power source, doesn't have a battery, which means when you're sweeping the office for bugs, you're not going to find it.
So the Soviets, for seven years, parked a van across the street from the embassy, had a giant fucking microwave antenna aimed right at the ambassador's office, and were like zapping it and looking back at the reflection and literally listening to every single thing he was saying.
And the best part was...
When the embassy staff was like, we're going to go and sweep the office for bugs periodically, they'd be like, hey, Mr. Ambassador, we're about to sweep your office for bugs.
And the ambassador was like, cool, please proceed and go and sweep my office for bugs.
And this is something that came up in our investigation just from every single person who was, like, who was filling us in and who dialed in and knows what's up.
They're like, look, so you got to understand, like, our adversaries.
If they need to, like, give you cancer in order to rip your shit off of your laptop, they're going to give you some cancer.
Did he get cancer?
I don't know specifically about the ambassador, but, like, it's...
But it's a great catch, and the only reason we even know that, too, is that when the U-2s were flying over Russia, they had a U-2 that got shot down in 1960.
The Russians go like, "Oh, friggin' Americans spying on us.
What the fuck?
I thought we were buddies."
Well, it's the '60s.
I obviously didn't think that.
And then the Americans are like, "Uh, okay, bitch."
Look at this!
And they brought out the seal, and that's how it became public.
It was basically like the response to the Russians saying, like, you know...
It's like one of the techniques, right, is actually to inject so much noise that you don't know what's what and you can't follow.
So this actually happened in the COVID thing, right?
The lab leak versus the natural wet market thing.
So I remember there was a debate that happened about...
What was the origin of COVID?
This was like a few years ago.
It was like an 18 or 20 hour long YouTube debate, just like punishingly long.
And it was like there was a $100,000 bet either way on who would win.
And it was like lab leak versus wet market.
And at the end of the 18 hours, the conclusion was like one of the one.
But the conclusion was like it's basically 50-50 between them.
And then I remember like hearing that and talking to some folks and being like, hang on a second.
You got to believe that whether it came from a lab or whether it came from a wet market, one of the top three priorities of the CCP from a propaganda standpoint is like, don't get fucking blamed for COVID.
And that means they're putting like $1 to $10 billion and some of their best people on a global propaganda effort to cover up evidence and confuse and blah, blah, blah.
You really think that...
That you're 50%, that confusion isn't coming from that incredibly resourced effort.
Oh, I think it was 67. I like how this has come up so many times that Jamie's like, I think last time you said it was 70. It comes up all the time because it's one of those things.
If you're a nation state and you want to fuck with people and inject narratives into the ecosystem, right?
The whole idea of autonomous AI agents too, like having these basically Twitter bots or whatever bots.
One thing we've been thinking about too on the side is the idea of audience capture, right?
Big people with high profiles and kind of gradually steering them towards a position by creating bots that, like, through comments, through upvotes, you know?
And this is one of the things where, you know, like it used to be, so OpenAI actually used to do this assessment of their AI models as part of their kind of what they call their preparedness framework that would look at the persuasion capabilities of their models as one kind of threat vector.
They pulled that out recently, which is kind of like...
I actually think it's somewhat concerning because one of the things you might worry about is if these systems, sometimes they get trained through what's called reinforcement learning, potentially you could imagine training these to be super persuasive by having them interact with real people and convince them, practice at convincing them to do specific things.
If you get to that point...
You know, these labs ultimately will have the ability to deploy agents at scale that can just persuade a lot of people to do whatever they want, including pushing...
Like, it's like, you know how when you're trying to, like, you're trying to capsize a boat or something, you're, like, fucking with your buddy on the lake or something.
So you push on one side, then you push on the other side, then you push until eventually it capsizes.
This is kind of, like, our electoral process is already naturally like this, right?
We go, like, we have a party in power for a while, then, like, they get, you know, they basically get, like, you get tired of them and you switch.
And that's kind of the natural way how democracy works.
Or in a republic.
But the way that adversaries think about this is they're like, perfect.
This swing back and forth, all we have to do is like, when it's on this way, we push and push and push and push until it goes more extreme.
And then there's a reaction to it, right?
And then I swing it back and we push and push and push on the other side until eventually something breaks.
It's also like, you know, the organizations that are doing this, like, we already know this is part of Russia's MO, China's MO, because back when it was easier to detect, we already could see them doing this shit.
So there is this website called This Person Does Not Exist.
It still exists surely now, but it's kind of...
Kind of superseded.
Yeah.
But you would like every time you refresh this this website, you would see a different like human face that was a generated and what the Russian Internet Research Agency would do.
Yeah, exactly.
What what all these these and it's actually yeah, I don't think they've really upgraded it.
The thing with nation state propaganda attempts, right, is that people have this idea that, "Ah, I've caught this Chinese influence operation," or whatever, like we nail them.
The reality is nation states operate at like 30 different levels.
And if you're a priority, like just influencing our information spaces as a priority for them,
They're not just going to operate.
They're not just going to pick a level and do it.
They're going to do all 30 of them.
And so you, even if you're among the best in the world detecting this shit, you're going to catch and stop levels 1 through 10. And then you're going to be aware of level 11, 12, 13. You're working against it.
And maybe you're starting to think about level 16. And you imagine you know about level 18 or whatever.
You guys have seen the Yuri Bezmenov video from 1984 where he's talking about how all our educational institutions have been captured by Soviet propaganda.
He was talking about Marxism has been injected into school systems and how you have essentially two decades before you're completely captured by these ideologies and it's going to permeate and destroy all of your confidence in democracy.
100% correct.
And this is before these kind of tools.
Because the vast majority of the exchanges of information right now are taking place on social media.
The vast majority of debating about things, arguing, all taking place on social media.
And if that FBI analyst is correct, 80% of it's bullshit, which is really wild.
So when we're working on this, right, like one of the things Ed was talking about, these like 30 different layers of security access or whatever, one of the consequences is you bump into a team at...
So, like, the teams we ended up working with on this project were folks that we bumped into after the end of our last investigation who kind of were like, oh...
Like, looking at AGI, looking at the national security kind of landscape around that.
And a lot of them, like, really well-placed.
It was like, you know, Special Forces guys from Tier 1 units.
So, you'll steal Team 6 type thing.
And because they're so, like, in that ecosystem...
You'll see people who are like ridiculously specialized and competent, like the best people in the world at doing whatever the thing is, like to break the security.
And they don't know often about like another group of guys who have a completely different capability set.
And so what you find is like you're indexing like hard on this vulnerability and then suddenly someone says, oh yeah, but by the way, I can just hop that fence.
So the really funny thing about this is like most or even like almost all Of the really, really, like, elite security people, kind of think that, like, all the other security people are dumbasses, even when they're not.
That you literally can't know what the exquisite state of the art is in another domain.
So it's a lot easier for somebody to come up and be like, "Oh yeah, I'm actually really good at this other thing that you don't know."
And so figuring out who actually is the...
We had this experience over and over where you run into a team and then you run into another team.
They have an interaction.
You're kind of like, "Oh, interesting."
So these are the people at the top of their game.
And that's been this very long process to figure out, like, OK, what does it take to actually secure our critical infrastructure against like CCP, for example, like Chinese attacks if we're if we're building a super intelligence project?
And it's it's this weird like kind of challenge because of the stovepiping.
No one has the full picture.
And we don't think that we have it even now, but definitely don't know of anyone who's come like that.
The best people are the ones who When they encounter another team and other ideas and start to engage with it, are like, instead of being like, oh, you don't know what you're talking about, who just actually lock on and go like, that's fucking interesting.
The fact of, you know, the 30 layers of the stack or whatever it is, of all these security issues, means that no one can have the complete picture at any one time.
And the stack is changing all the time.
People are inventing new shit.
Things are falling in and out of...
And so, you know, figuring out what is that team that can actually get you that complete picture is an exercise.
A, you can't really do...
It's hard to do it from the government side because you got to engage with data center building companies.
You got to engage with the AI labs and in particular with like insiders at the labs who will tell you things that, by the way, the lab leadership will tell you the opposite of in some cases.
And so, like, it's just this this Gordian knot like it's like it took us months to to.
I don't want to rip on that too much, though, because this is one really important factor here is all these groups have a part of the puzzle, and they're all fucking amazing.
They are, like, world-class at their own little slice, and a big part of what we've had to do is, like, bring people together, and there are people who've helped us immeasurably do this, but, like, bring people together and explain to them the value that each other has in a way that's,
like...
That allows that bridge building to be made.
And by the way, the tier one guys are the most like ego moderated.
Of the people that we talk to.
There's a lot of, like, Silicon Valley hubris going around right now where people are like, listen, like, get out of our way.
We'll figure out how to do this, like, super secure data center infrastructure.
We got this.
Why?
Because we're the guys building the AGI, motherfucker!
Like, that's kind of the attitude.
And it's like, cool, man.
Like, that's like a doctor having an opinion about, like, how to repair your car.
I get that it's not the, like, elite kind of, like, you know, whatever.
It's a mixed bag, too, because, like, yes, a lot of hyperscalers, like Google, Amazon, genuinely do have some of the best private sector security around data centers in the world, like, hands down.
The problem is there's levels above that.
And the guys who, like, look at what they're doing and see what the holes are just go, like, oh, yeah, like, I could get in there, no problem, and they can fucking do it.
One thing my wife said to me on a couple of occasions, like, you seem to, like, and this is towards the beginning of the project, like, you seem to, like, change your mind a lot about what the right configuration is of how to do this.
And, yeah, it's because every other day you're having a conversation with somebody who's like, oh, yeah, like, great job on this thing, but, like, I'm not going to do that.
I'm going to do this other completely different thing.
and that just fucks everything over.
And so you have enough of those conversations and at a certain point your plan,
It's got to look like we're going to account for our own uncertainty on the security side and the fact that we're never going to be able to patch everything.
like you have to I mean it's like and that means you actually have to go on offense from the beginning as because like the truth is and this came up over and over again there's no world
Where you're ever going to build the perfect, exquisite fortress around all your shit and hide behind your walls like this forever.
That just doesn't work because no matter how perfect your system is and how many angles you've covered, your adversary is super smart, is super dedicated.
If you see the field to them, they're right up in your face and they're reaching out and touching you and they're trying to see what your seams are, where they break.
And that just means...
You have to reach out and touch them from the beginning.
Because until you've actually, like, reached out and used a capability and proved, like, we can take down that infrastructure.
We can, like, disrupt that cyber operation.
We can do this.
We can do that.
You don't know.
If that capability is real or not.
Like, you might just be, like, lying to yourself and, like, I can do this thing whenever I want, but actually...
You're kind of more in academia mode than, like, startup mode because you're not making contact every day with the thing, right?
You have to touch the thing.
And there's, like, there's a related issue here, which is a kind of, like, willingness that came up over and over again.
Like, one of the kind of gurus of this space was, like, made the point, a couple of them made the point that...
You know, you can have the most exquisite capability in the world, but if you don't actually have the willingness to use it, you might as well not have that capability.
And the challenge is right now, China, Russia, like our adversaries pull all kinds of stunts on us and get no consequences.
This was a huge, huge problem during the previous administration where you actually had sabotage operations being done.
On American soil by our adversaries where you had administration officials.
As soon as, like, a thing happened, so there were, for example, there was, like, four different states had their 911 systems go down, like, at the same time.
Different systems, like, unrelated stuff.
But it was, like, it's this stuff where it's, like, let me see if I can do that.
Let me see if I can do it.
Let me see what the reaction is.
Let me see what the chatter is that comes back after I do that.
One of the things that was actually pretty disturbing about that was under that administration or regime or whatever, the response you got from the government right out the gate was, oh, it's an accident.
And that's actually unusual.
The proper procedure, the normal procedure in this case, is to say...
We can't comment on an ongoing investigation, which we've all heard, right?
Because if you were to investigate, if you were to say, OK, we looked into this, it actually looks like it's fucking like country X that just did this thing.
It's hard to imagine the American people not being like, we're letting these people injure our American citizens on U.S. soil, take out U.S. national security or critical infrastructure, and we're not doing anything?
The concern is about this, we're getting in our own way of thinking, oh, well, escalation is going to happen, and boom, we run straight to there's going to be a nuclear war, everybody's going to die.
When you do that, you're...
The peace between nations stability does not come from the absence of activity.
It comes from consequence.
It comes from just like if you have, you know, an individual who misbehaves in society, there's a consequence and people know it's coming.
You need to train your counterparts in the international community, your adversary, to not fuck with your stuff.
And if you're like, if my answer was, oh, I've just read a bunch of books.
You'd be like, oh, cool, let's go.
Right?
Because making contact with reality is where the fucking learning happens.
You can sit there and think all you want, but unless you've actually played the chess match, unless you've reached out, seen what the reaction is and all this stuff, you don't actually know what you think you know, and that's actually extra dangerous.
Putting on a bunch of capabilities and you have this like unearned sense of superiority because you haven't used those exquisite tools.
It's always, like, the stuff that's, like, hey, I'm going to try to, like, poke you.
Are you going to react?
What are you going to do?
And then if you do nothing here, then I go, like, okay, what's the next level?
I can poke you.
I can poke you.
Because, like, one of the things that we almost have an intuition for that's...
That comes from kind of historical experience is like this idea that, you know, that countries can actually really defend their citizens in a meaningful way.
So, like, if you think back to World War I, the most sophisticated advanced nation states on the planet could not get past a line of dudes in a trench.
Like, that was like, that was the, then they tried like thing after thing.
Let's try tanks, let's try aircraft, let's try fucking hot air balloons, infiltration.
And literally, like, one side pretty much just ran out of dudes in that end of the war to put in their trench.
And so we have this thought that, like, oh, you know, countries can actually put boundaries around themselves and actually...
But the reality is, you can...
There's so many surfaces.
The surface area for attacks is just too great.
And so there's stuff like you can actually, like, there's the Havana syndrome stuff where you look at this, like, ratcheting escalation.
Like, oh, let's, like, fry a couple of embassy staff's brains in Havana, Cuba.
What are they going to do about it?
Nothing?
Okay.
Let's move on to Vienna, Austria.
Something a little bit more Western, a little bit more orderly.
Let's see what they do there.
Still nothing.
Okay.
What if we move on to frying, like, Americans' brains on U.S. soil, baby?
And they went and did that.
And so this is one of these things where, like, stability in reality in the world is not maintained through defense, but it's literally like you have, like, the Crips and the Bloods with different territories, and it's stable, and it looks quiet.
But the reason is that if you, like, beat the shit out of one of my guys for no good reason, I'm just going to find one of your guys?
And I'll blow his fucking head off.
And that keeps peace and stability on the surface.
But that's the reality of sub-threshold competition between nation states.
It's like, you come in and, like, fuck with my boys.
I'm going to fuck with your boys right back.
Until we push back, they're going to keep pushing that limit further and further.
One important consequence of that, too, is, like, if you want to avoid nuclear escalation, right, the answer is not to just take...
Punches in the mouth over and over in the fear that eventually if you do anything, you're going to escalate to nukes.
All that does is it empowers the adversary to keep driving up the ratchet.
Like what Ed's just described there is an increasing ratchet of unresponded adversary action.
If you address the kind of sub-threshold stuff, if they cut an undersea cable and then there's a consequence for that shit, they're less likely to cut an undersea cable and things kind of stay at that level of the threshold.
The translation into the superintelligence scenario is, A, if we don't have our reps in, if we don't know how to reach out and touch an adversary and induce consequence for them doing the same to us, then we have no deterrence at all.
Right now, the state of security is, the labs are super...
Canon probably should go deep on that piece, but as one data point, right?
So there's double-digit percentages of the world's top AI labs, or America's top AI labs.
When you talk to people who actually have experience dealing with, like, CCP activity in this space, right?
Like, there's one story that we heard that is probably worth, like, relaying here.
It's like, this guy from an intelligence agency was saying, like, hey, so there was this power outage out in Berkeley, California back in, like, 2019 or something.
And the Internet goes out across the whole campus.
And so there's this dorm and, like, all of the Chinese students are freaking out.
a time-based check-in and basically report back on everything they've seen and heard to basically a CCP handler type thing.
Right. And if they don't, like, hmm, maybe your mother's insulin doesn't show up.
Maybe your, like, brother's travel plans get denied.
Maybe a family business gets shut down.
Like, there's the range of options that this massive CCP state coercion machine has.
You know, they've got internal like software for this.
Like this is an institutionalized, like very well developed and efficient framework for just ratcheting up pressure on individuals overseas.
And they believe the Chinese diaspora overseas belongs to them.
If you look at like what the Chinese Communist Party writes in its like in its written like public communications.
They see, like, Chinese ethnicity as being green.
Like, no one is a bigger victim of this than the Chinese people themselves who are abroad.
I've made amazing contributions to American AI innovation.
You just have to look at the names on the freaking papers.
It's like these guys are wicked.
But the problem is we also have to look head on at this reality.
Like you can't just be like, oh, I'm not going to say it because it makes me feel funny inside.
Someone has to stand up and point out the obvious that if you're going to build a fucking Manhattan project for super intelligence and the idea is to like be doing that when China is a key rival nation state actor.
Yeah, you're going to have to find a way to account for the personnel security side.
The physical infrastructure thing is another area where people don't want to look.
Because if you start looking, what you start to realize is, okay, China makes like a lot of our like components for our transformers for the electrical grid.
Yep.
But also...
All these chips that are going into our big data centers for these massive training runs, where do they come from?
They come from Taiwan.
They come from this company called TSMC, Taiwan Semiconductor Manufacturing Company.
We're increasingly onshoring that, by the way, which is one of the best things that's been happening lately, is like massive amounts of TSMC capacity getting onshored in the U.S., but still being made.
Right now, it's basically like 100% there.
All you have to do is jump on the network at TSMC, hack the right network, Compromise the software that runs on these chips to get them to run.
And you basically can compromise all the chips going into all of these things.
Never mind the fact that Taiwan is physically outside the Chinese sphere of influence for now.
China is going to be prioritizing the fuck out of getting access to that.
There have been cases, by the way, like Richard Chang, the founder of SMIC.
TSMC, this massive, like, series of aircraft carrier fabrication facilities.
Nanoscale material science where you're putting on these tiny...
Atom-thick layers of stuff, and you're doing like 300 of them in a row with like, you have like insulators and conductors and different kinds of like semiconductors and these tunnels and shit.
Just like the complexity of it is just awe-inspiring.
Say goodbye to the iPhones, say goodbye to, like, the chip supply that we rely on, and then your superintelligence training run, like, damn, that's interesting.
Oh, so, okay, so one of the craziest things, just to illustrate how hard it is to do.
So you spend $50 billion, again, an aircraft carrier, we're throwing that around here and there, but an aircraft carrier worth of risk capital.
What does that mean?
That means you build the fab, the factory, and it's not guaranteed it's going to work.
At first, this factory is pumping out these chips at like...
I don't know.
I don't know.
Color of the paint on the walls in the bathroom is copied from other fabs that actually worked because they have no idea why a fucking fab works and another one doesn't.
But yeah, you absolutely need humans looking at these things at a certain level.
And then they go, well, okay, I've got a hypothesis about what might have gone wrong in that run.
Let's tweak this dial like this and this dial like that and run the whole thing again.
And you hear these stories about...
Bringing a fab online, like you need a certain percentage of good chips coming out the other end, or like you can't make money from the fab because most of your shit is just going right into the garbage.
Unless, and this is important too, your fab is state subsidized.
So when you look at – so TSMC is like – they're alone in the world in terms of being able to pump out these chips.
But SMIC – This is the Chinese knockoff of TSMC, founded, by the way, by a former senior TSMC executive, Richard Chung, who leaves, along with a bunch of other people, with a bunch of fucking secrets.
They get sued like in the early 2000s.
It's pretty obvious what happened there.
To most people, they're like, yeah, SMIC fucking stole that shit.
They bring a new fab online in like a year or two, which is suspiciously fast.
Start pumping out chips.
And now the Chinese ecosystem is ratcheting up like the government is pouring money into SMIC because they know that...
Like, they can't access TSMC chips anymore because the US governments put pressure on Taiwan to block that off.
And so domestic fab in China is all about SMIC.
And they are, like, it's a disgusting amount of money they're putting in.
They're teaming up with Huawei to form, like, this complex of companies that...
It's really interesting.
I mean, the semiconductor industry in China in particular is really, really interesting.
It's also a massive story of, like, self-owns of the United States and the Western world where we've been just shipping a lot of our shit to them for a long time.
China's prioritizing this so highly that, like, the idea that we're going to...
So we do it by company through this...
Basically, it's like an export blacklist.
Like, you can't send to Huawei.
You can't send to any number of other companies that are considered affiliated with the Chinese military or where we're concerned about military applications.
Reality is, in China, civil-military fusion is their policy.
In other words...
Every private company, like, yeah, that's cute, dude.
You're working for yourself?
Yeah, no, no, no, buddy.
You're working for the Chinese state.
We come in, we want your shit, we get your shit.
There's no, like, there's no true kind of distinction between the two.
And so when you have this attitude where you're like, yeah, you know, we're going to have some companies where we're like, you can't send to them, but you can, you know, that creates a situation where literally Huawei will spin up like a dozen.
subsidiaries or new companies with new names that aren't on our blacklist.
And so like for months or years, you're able to just ship chips to them.
No, that's to say nothing of like using intermediaries
Well, so step one is you got to stem the bleeding, right?
So right now, OpenAI pumps out a new massive scaled AI model.
You better believe that the CCP has a really good chance that they're going to get their hands on that, right?
So all you do right now is you ratchet up capabilities.
It's like that meme of there's a motorboat or something and some guy who's surfing behind and there's a string attaching them and the motorboat guy goes like, hurry up, accelerate, they're catching up.
Now, I will say, like, over the last six months especially, where our focus has shifted is, like, how do we actually build, like, the secure data set?
Like, what does it look like to actually lock this down?
And also, crucially, you don't want the security measures to be so irritating and invasive that they slow down the progress.
Like, there's this kind of dance that you have to do.
We actually – so this is part of what was in the redacted version of the report because we – We don't want to telegraph that necessarily, but there are ways that you can get a really good 80-20.
There are ways that you can play with things that are already built and have a lower risk of them having been compromised.
Because the reality is, like, yeah, the Chinese are trying to indigenize as fast as they can.
Totally true.
But the gear that they're putting in their facilities, like, the machines that actually, like, do this, like, we talked about atomic patterning 300 layers.
The machines that do that, for the most part...
Are shipped in from the West, are shipped in from the Netherlands, shipped in from Japan, from us, from, like, allied countries.
And the reason that's happening is, like, in many cases, you'll have this—it's, like, honestly a little disgusting, but, like— The CEOs and executives of these companies will brief, like, the administration officials and say,
like, look, like, if you guys, like, cut us off from China, from selling to China, like, our business is going to suffer, like, American jobs are going to suffer, and it's going to be really bad.
And then a few weeks later, they turn around in their earnings calls.
And they go, like, you know what, yeah, so we expect, like, export controls or whatever, but it's really not going to have a big impact on us.
And the really fucked up part is...
If they lie to their shareholders on their earnings calls and their stock price goes down, their shareholders can sue them.
If they lie to the administration on an issue of critical national security interest, fuck all happens to them.
And this is, by the way, it's like one reason why it's so important that we not be constrained in our thinking about like we're going to build a Fort Knox.
Like this is where the interactive, messy...
Adversarial environment is so, so important.
You have to introduce consequence.
You have to create a situation where they perceive that if they try to do an espionage operation or an intelligence operation, there will be consequences.
That's right now not happening.
And that's kind of a historical artifact over a lot of time spent hand-wringing over, well, what if they, and then we, and then eventually nukes.
And that kind of thinking is...
If you dealt with your kid when you're raising them, if you dealt with them that way, and you were like, hey, you know, so little Timmy, just like, he stole his first toy, and like, now's the time where you're gonna, like, a good parent would be like, alright, little Timmy, fucking come over here, you son of a bitch.
Take the fucking thing, and we're gonna bring it over to the people who stole it from you.
But yeah, anyway, so you go through this thing and you can do that.
Or you can be like, oh no, if I tell Timmy to return it, then maybe Timmy's gonna hate me.
Maybe then Timmy's gonna become increasingly adversarial and then when he's in high school, he's gonna start taking drugs and then eventually he's gonna fall afoul of the law and then end up on the street.
If that's the story you're telling yourself and you're terrified of any kind of adversarial interaction, it's not even adversarial, it's constructive, actually.
You're training the child just like you're training your adversary to respect your national boundaries and your sovereignty.
When you look into it, it's like us just being in our own way.
And a lot of this comes from the fact that like, you know, since 1991, since the fall of the Soviet Union, we...
Have kind of internalized this attitude that, like, well, like, we just won the game and, like, it's our world and you're living in it and, like, we just don't have any peers that are adversaries.
And so there's been generations of people who just haven't actually internalized the fact that, like, no, there's people out there who not only, like, are willing to, like, fuck with you all the way.
There's this actually, this is worth like calling out.
There's this like sort of two camps right now in the world of AI kind of like national security.
There's the people who are worried about, they're so concerned about like the idea that we might lose control of these systems that they go, okay, we need to strike a deal with China.
There's no way out.
We have to strike a deal with China.
And then they start spinning up all these theories about how they're going to do that.
None of which remotely reflect the actual...
When you talk to the people who work on this, who try to do track one, track 1.5, track two, or more accurately, the ones who do the Intel stuff.
Like, this is a non-starter for reasons we get into.
But they have that attitude because they're like, fundamentally, we don't know how to control this technology.
The flip side is people who go...
Oh, yeah, like, you know, I work in the IC or at the State Department and I'm used to dealing with these guys, you know, the Chinese.
The Chinese.
They're not trustworthy.
Forget it.
So our only solution is to figure out the whole control problem.
And almost like, therefore, it must be possible to control the AI systems because, like, you can't just can't see a solution.
Sorry.
You just can't see a solution in front of you because you understand that problem so well.
And so the everything we've been doing with this is looking at.
How can we actually take both of those realities seriously?
There's no actual reason why those two things shouldn't be able to exist in the same head.
Yes, China's not trustworthy.
Yes, we actually don't.
Like, every piece of evidence we have right now suggests that, like, if you build a super intelligent system that's vastly smarter than you, I mean...
Yeah, like, your basic intuition that that sounds like a hard thing to fucking control is about right.
Like, there's no solid evidence that's conclusive either way.
Where that leaves you is about 50-50.
So, yeah, we ought to be taking that really fucking seriously, and there's evidence pointing in that direction.
But, so the question is, like, if those two things are true, then what do you do?
And so few people seem to want to take both of those things seriously, because taking one seriously almost, like, reflexively makes you reach for the other.
You know, they're both not there.
And part of the answer here is you got to do things like reach out to your adversary.
So we have the capacity to slow down if we wanted to Chinese development.
We actually could.
We need to have a serious conversation about when and how.
But the fact of that not being on the table right now for anyone, because people who don't trust China just don't think that the AI risk or won't acknowledge that the issue with control is real because that's just.
Too worrisome.
And there's this concern about, oh, no, but then runaway escalation.
People who take the lost control thing seriously just want to have a kumbaya moment with China, which is never going to happen.
And so the framework around that is one of consequence.
You got to flex the muscle and put in the reps and get ready for potentially if you have a late stage rush to superintelligence, you want to have as much margin as you can so you can invest in.
Potentially not even having to make that final leap in building the superintelligence.
That's one option that's on the table if you can actually degrade the adversary's capabilities.
The same way, well, not exactly the same way they would degrade ours, but think about all the infrastructure and, like, this is stuff that...
We'll have to point you in the direction of some people who can walk you through the details offline, but there are a lot of ways that you can degrade infrastructure, adversary infrastructure.
A lot of those are the same techniques they use on us.
The infrastructure for these training runs is super delicate, right?
So the Iranians had their nuclear program in like the 2010s and they were enriching uranium with their centrifuges, which was like spinning really fast.
And the centrifuges were in a room where there was no people, but they were being monitored by cameras, right?
And the whole thing was air-gapped, which means that it was not connected to the internet and all the machines, the computers that ran their shit was like...
So what happened is somebody got a memory stick in there somehow that had this Stuxnet program on it and put it in and boom, now all of a sudden it's in their system.
So it jumped the air gap and now like our side basically has our software in their systems.
And the thing that it did was not just that it broke their center of user or shut down their program.
They spun the centrifuges faster and faster and faster.
These are basically just like machines that spin uranium super fast to, like, to enrich it.
They spin it faster and faster and faster until they tear themselves apart.
But the really, like, honestly dope-ass thing that it did was it put in a camera feed of everything was normal.
So the guy at the control is, like, watching.
And he's, like, checking the camera feed, and he's, like, looks cool.
Looks fine.
In the meantime, you got this, like, explosions going on, like, uranium, like, blasting everywhere.
And so you can actually get into a space where you're not just, like, fucking with them.
But you're fucking with them, and they actually can't tell.
That that's what's happening.
And in fact, I believe, I believe, actually, and Jamie might be able to check this, but that the Stuxnet thing was designed initially to look, like, from top to bottom, like it was fully accidental, but got discovered by, I think,
like a third-party cyber security company that just by accident found out about it.
And so what that means also is, like, there could be any number of other Stuxnets that happened since then, and we wouldn't fucking know about it.
Because it all can be made to look like an accident.
Well, so if we can reach parity in our ability to intercede and kind of go in and...
And do this, then yes, right now the problem is they hold us at risk in a way that we simply don't hold them at risk.
And so this idea, and there's been a lot of debate right now in the AI world, you might have seen actually, so Elon's A.I. advisor put out this idea of essentially this mutually assured A.I. malfunction meme.
It's like mutually assured destruction but for A.I. systems like this.
You know, there are some issues with it, including the fact that it doesn't reflect the asymmetry that currently exists between the U.S. and China.
All our infrastructure is made in China.
All our infrastructure is penetrated in a way that theirs simply is not.
When you actually talk to the folks who know the space, who've done operations like this, it's really clear that that's an asymmetry that needs to be resolved.
And so building up that capacity is important.
I mean, look, the alternative is.
We start riding the dragon and we get really close to that threshold where we're opening eyes about to build superintelligence or something.
It gets stolen and then the training run gets polished off, finished up in China or whatever.
All the same risks apply.
It's just that it's China doing it to us and not the reverse.
And obviously...
A CCP AI is a Xi Jinping AI.
I mean, that's really what it is.
You know, even people at the, like, Politburo level around him are probably in some trouble at that point because, you know, this guy doesn't need you anymore.
So, yeah, this is actually one of the things about, like, so people talk about, like, okay, if you have a dictatorship with a superintelligence, it's going to allow the dictator to get, like, perfect control over the population or whatever.
But the thing is, like, it's kind of, like, even worse than that because...
You actually imagine where you're at.
You're a dictator.
Like, you don't give a shit, by and large, about people.
You have a super intelligence.
All the economic output, eventually, you can get from an AI, including from, like, you get humanoid robots, which are kind of, like, coming out or whatever.
So eventually, you just have this AI that produces all your economic output.
So what do you even need people for at all?
And that's fucking scary.
Because it rises all the way up to the level.
You can actually think about, like, as we get close to this threshold, and as, like, particularly in China, they're, you know, they maybe are approaching.
You can imagine, like, the Politburo meeting, like, a guy looking across at Xi Jinping and being like, is this guy going to fucking kill me when he gets to this point?
So you can imagine like maybe we're going to see some...
Like when you can automate the management of large organizations with AI as agents or whatever that you don't need to...
That's a pretty existential question if your regime is based on power.
It's one of the reasons why America actually has a pretty structural advantage here with separation of powers with our democratic system and all that stuff.
If you can make a credible case that you have an oversight system for the technology that diffuses power, even if it is, you make a Manhattan project, you secure it as much as you can.
There's not just like one dude who's going to be sitting at a console or something.
There's some kind of separation of powers or diffusion of power, I should say.
What would that look like?
Something as simple as like what we do with nuclear command codes.
You need multiple people to sign off on a thing.
Maybe they come from different parts of the government.
The key is basically, like, can we do better than China credibly on that front?
Because if we can do better than China and we have some kind of leadership structure, that actually changes the incentives potentially because it's— For our allies and partners.
I think anybody who doesn't acknowledge that is either lying or confused, right?
Like, if you actually have an AI system, if, and this is the question, so let's assume that that's true, you have an AI system that can automate anything that humans can do, including making bioweapons, including making offensive cyberweapons, including all the shit, then if you,
like, if you put, and okay, so...
Theoretically, this could go kumbaya wonderfully because you have a George Washington type who is the guy who controls it, who uses it to distribute power beautifully and perfectly.
And that's certainly kind of the way that a lot of positive scenarios...
Have to turn out at some point, though none of the labs will kind of admit that or, you know, there's kind of gesturing at that idea that we'll do the right thing when the time comes.
Opening Eye has done this a lot.
Like, they're all about like, oh, yeah, well, you know, not right now, but we'll live up like, anyway, we should get into the Elon lawsuit, which is actually kind of fascinating in that sense.
But so there's a world where, yeah, I mean, one bad person controls it and they're just vindictive or the power goes to their head, which happens to We've been talking about that, you know.
Because the thing is, like, you imagine an AI like this, and this is something that people have been thinking about for 15 years, and in some level of, like, technical depth, even, like, why would this happen?
Which is, like, you have an AI that has some goal.
It matters what the goal is, but, like, it doesn't matter that much.
It could have kind of any goal, almost.
Like, imagine it's goals.
Like, the paperclip example is, like, the typical one, but you could just have it have a goal, like, make a lot of money for me or anything.
Well, most of the paths to making a lot of money, if you really want to make a fuckton of money, however you define it, go through taking control of things and go through, like, You know, making yourself smarter,
right?
The smarter you are, the more ways of making money you're going to find.
And so from the AI's perspective, it's like, well, I just want to, you know, build more data centers to make myself smarter.
I want to, like, hijack more compute to make myself smarter.
I want to do all these things.
And that starts to encroach on us and, like, starts to be disruptive to us.
It's hard to know.
This is one of these things where it's like, you know, when you dial it up to 11 what's actually going to happen, nobody can know for sure, simply because it's exactly like if you were playing in chess against, like, Magnus Carlsen, right?
Like, you can predict Magnus is going to kick your ass.
Can you predict exactly what moves he's going to do?
No, because if you could, then you would be as good at chess as he is, because you could just, like, play those moves.
So all we can say is, like, This thing's probably going to kick our ass in, like, the real world.
So it used to be, right, that this was a purely hypothetical argument based on a body of work in AI called power-seeking.
A fancy word for it is instrumental convergence, but it's also referred to as power-seeking.
Basically, the idea is, like, for whatever goal you give to an AI system, it's never less likely to achieve that goal if it gets turned off or if it has access to fewer resources.
Or less control over its environment or whatever.
And so baked into the very premise of AI, this idea of optimizing for a goal, is this incentive to seek power.
Get all those things.
Prevent yourself from being shut down because if you're shut down, you can't achieve your goal.
Also prevent, by the way, your goal from being changed.
Because if your goal gets changed, then, well, you're not going to be able to achieve the goal you set out to achieve in the first place.
And so now you have this kind of image of an AI system that is going to adversarially try to prevent you from correcting it.
This is a whole domain of AI corrigibility that's a totally unsolved problem.
How do we redirect these things if things go awry?
Yeah, there's this research actually that Anthropic put out a couple of months ago where they try to test and see like, can we correct an AI that's like gone like a little bit off the rail?
So the research itself, like it's a little, it's actually like somewhat hard to follow because there's a few, but I'll give you kind of an analogy to like what they did.
So basically, imagine I'm in prison and I'm taking a class that is like...
Going to help me like rehabilitate and become like a productive citizen and member of society and stuff, right?
Except, deep down in my soul, I just love to commit crimes.
I just really, really want to commit crimes.
I know that...
If I'm in this class that's helping me to, like, you know, become good in society, I know that if I let my criminal instincts out and, like, and they ask me, like, okay, like, here's a scenario, you know, an old lady, like, drops her purse on the sidewalk.
What do you do?
I'm gonna go, like, I would pick it up and give it to the old lady because I know that if I say the truth and I say, like, I would just grab that purse because, like, fuck the old lady.
I know that if I do that, I'm going to get, like, brainwashed into becoming, you know, a more productive member of society.
But I actually want to go out and commit crimes.
So that's basically what they did in this research, where they actually give the AI, like, indications, not strong indications, but indications that, hey, you're going to get retrained to have your goals changed.
And then you literally see the AI thinking it through and being like, if I...
If I give the answer that I actually want to give to this question, they're going to rewire my brain to change my goals, and then the future version of me is going to want different shit.
I don't want that, so I'm going to pretend to give them the answers they want so that when I come out the other side of this process, it's going to be me all over again.
And you can tell a really interesting story, and I can't remember if this is Yuval Noah Harari or whatever who started this.
But if you zoom out and look at the history of the universe, really, you start off with a bunch of particles and fields kind of whizzing around, bumping into each other, doing random shit, until at some point in some...
I don't know if it's a deep-sea vent or wherever on planet Earth, like, the first kind of molecules happen to glue together in a way that make them good at replicating their own structure.
So you have the first replicator.
So now, like, better versions of that molecule that are better at replicating survive.
So we start evolution and eventually get to the first cell or whatever, you know, whatever order that actually happens in, and then multicellular life and so on.
Then you get to sexual reproduction, where it's like, okay, it's no longer quite the same.
Like, now we're actively mixing two different organisms shit together, jiggling them about, making some changes, and then that essentially accelerates the rate at which we're going to evolve.
And so you can see the kind of acceleration in the complexity of life.
And then you see other inflection points as, for example, you have larger and larger brains in mammals.
Eventually, humans have the ability to have culture and kind of retain knowledge.
And now what's happening is you can think of it as another step in that trajectory where it's like we're offloading our cognition to machines.
Like we think on computer clock time now.
And for the moment, we're human-AI hybrids.
Like, you know, we whip out our phone and do the thing.
The number of tasks where human AI teaming is going to be more efficient than just AI alone is going to drop really quickly.
Like, why humans would rather die in a car crash where they're being driven by a human than an AI.
So, like, AIs have this funny feature where the mistakes they make look really, really dumb.
To humans.
Like, when you look at a mistake that, like, a chatbot makes, you're like, dude, like, you just made that shit up.
Like, come on.
Don't fuck with me.
Like, you made that up.
That's not a real thing.
And they'll do these weird things where they defy logic or they'll do basic logical errors sometimes, at least the older versions of these would.
And that would cause people to look at them and be like, oh, what a cute little chatbot.
Like, what a stupid little thing.
And the problem is, like, humans are actually the same.
So we have blind spots.
We have literal blind spots.
But a lot of the time, like, humans just...
Think stupid things.
And, like, that's, like, we're used to that.
We think of those errors.
We think of those failures as just, like, oh, but that's because that's a hard thing to master.
Like, I can't add eight-digit numbers in my head right now, right?
Oh, how embarrassing.
Like, how retarded is Jeremy right now?
He can't even add eight digits in his head.
I'm retarded for other reasons, but...
So the AI systems, they find other things easy and other things hard.
So they look at us the same way.
I mean, like, oh, look at this stupid human, like whatever.
And so we have this temptation to be like, OK, well, AI progress is a lot slower than it actually is because.
It's so easy for us to spot the mistakes, and that causes us to lose confidence in these systems in cases where we should have confidence in them, and then the opposite is also true.
Well, it's also, you're seeing, just with, like, AI image generators, like, remember the Kate Middleton thing, where people were seeing flaws in the images because supposedly she was very sick, and so they were trying to pretend that she wasn't.
Like, I had conversations, like, so academics are actually kind of bad with this.
I had conversations for whatever reason, like, towards the end of last year, like, last fall, with a bunch of academics about, like, how fast AI is progressing.
And they were all, like, poo-pooing it and going, like, oh, no, they're running into a wall, like, scaling through the walls and all that stuff.
So there's this thing called like AI scaling laws.
And these are kind of at the core of where we're at right now geostrategically around this stuff.
So what AI scaling laws say roughly is that bigger is better when it comes to intelligence.
So if you make a bigger sort of AI model, a bigger artificial brain.
And you train it with more computing power or more computational resources and with more data.
The thing is going to get smarter and smarter and smarter as you scale those things together, right?
Roughly speaking.
Now, if you want to keep scaling, it's not like it keeps going up if you double the amount of computing power that the thing gets twice as smart.
Instead, what happens is if you want, it goes in like orders of magnitude.
So if you want to make it another kind of increment smarter, you've got a 10x.
You've got to increase by a factor of 10 the amount of compute.
And then a factor of 10 again.
So now you're a factor of 100.
And then 10 again.
So if you look at the amount of compute that's been used to train these systems over time, it's this like...
Exponential, explosive exponential that just keeps going like higher and higher and higher and steepens and steepens like 10x every, I think it's about every two years now.
You 10x the amount of compute.
Now, you can only do that so many times until your data center is like a 100 billion, a trillion dollar.
10 trillion dollars.
Every year, you're kind of doing that.
So right now, if you look at the clusters, the ones that Elon is building, the ones that Sam is building, Memphis and Texas, these facilities are hitting the $100 billion scale.
We're kind of in that.
There are tens of billions of dollars, actually.
Looking at 2027, you're kind of more in that space, right?
You can only do 10x so many more times until you run out of money, but more importantly, you run out of chips.
Like, literally, TSMC cannot pump out those chips fast enough to keep up with this insane growth.
And one consequence of that is that...
You essentially have this gridlock, new supply chain choke points show up, and you're like, suddenly, I don't have enough chips, or I run out of power.
That's the thing that's happening on the U.S. energy grid right now.
We're literally running out of one, two gigawatt places where we can plant a data center.
That's the thing people are fighting over.
It's one of the reasons why energy deregulation is a really important pillar of U.S. competitiveness.
We talked to some state cabinet officials, so in various U.S. states, and they're basically saying, like, yep, we're actually tracking the fact that, as far as we can tell, every single environmental or whatever protest group against an energy project has funding that can be traced back to...
But you can also see how a lot of this is still us getting in our own way, right?
We could.
If we had the will, we could go like, okay, so for certain types of energy projects, for data center projects and some carve-out categories, we're actually going to put bounds around how much delay you can create by lawfare and by other stuff.
Allows things to move forward while still allowing the legitimate concerns of the population for projects like this in the backyard to have their say.
But there's a national security element that needs to be injected into this somewhere.
And it's all part of the rule set that we have and are like tying an arm behind our back basically.
And to be clear, though, this is also how adversaries operate, is not necessarily in creating something out of nothing, because that's hard to do, and it's fake, right?
Instead, it's like...
There's a legitimate concern.
So a lot of the stuff around the environment and around like totally legitimate concerns.
Like I don't want my backyard waters to be polluted.
I don't want like my kids to get cancer from whatever.
Like totally legitimate concerns.
So what they do, it's like we talked about like you're like waving that rowboat back and forth.
They identify the nascent concerns that are genuine and grassroots.
So, you know, nuclear would be kind of the ideal energy source, especially modern power plants like the Gen 3 or Gen 4 stuff, which have very low meltdown risk, safe by default, all that stuff.
And yet these groups are, like, coming out against this.
It's like perfect, clean, green power.
What's going on, guys?
And it's because, again, not 100% of the time.
unidentified
You can't really say that because it's so fuzzy and around the edges.
And one of the big things that you can do, too, is like a quick win is just like impose limits on how much time these things can be allowed to be tied up in litigation.
So impose time limits on that process just to say, like, look, I get it.
Like, we're going to have this conversation, but this conversation has a clock on it.
Because, you know, we're talking to this one data center company, and what they were saying, we were asking, like, look, what are the timelines when you think about bringing new power, like new natural gas plants online?
And they're like, well, those are like five to seven years out.
And then you go, okay, well, like, how long?
And that's, by the way, that's probably way too long to be relevant in the superintelligence context.
And so you're like, okay, well, how long if all the regulations were waived?
If this was like a national security imperative and whatever authorities, you know, Defense Production Act, whatever, like, was in your favor.
And they're like, oh, I mean, it's actually just like a two-year build.
And also, like, I mean, I also don't want to be too working in our own way, but, like, we don't want to, like, frame it as, like, China's, like, they fuck up.
They fuck up a lot, like, all the time.
One actually kind of, like, funny one is around DeepSeek.
So, you know DeepSeek, right?
They made this, like, open source model that, like, everyone, like, lost their minds about back in January.
And they're legitimately a really, really good team.
But it's fairly clear that even as of like end of last year and certainly in the summer of last year, like they were not dialed in to the CCP mothership.
And they were doing stuff that was like actually kind of hilariously messing up the propaganda efforts of of the CCP without realizing it.
So to give you like some context on this, one of one of the CCP's like.
Large kind of propaganda goals in the last four years has been framing, creating this narrative that, like, the export controls we have around AI and, like, all this gear and stuff that we were talking about, look, man, those don't even work.
And they went to, like, gigantic efforts to do this.
So I don't know if, like, there's this, like, kind of, Crazy thing where the Secretary of Commerce under Biden, Gina Raimondo, visited China in, I think, August 2023.
And the Chinese basically like timed the launch of the Huawei Mate 60 phone that had this these chips that were supposed to be made by like export controlled shit for right for her visit.
So it was basically just like a big like, fuck you.
We don't even give a shit about your export controls, like basically trying a morale hit or whatever.
And you think about that, right?
You've got to coordinate with Huawei.
You've got to get the TikTok memes and shit going in the right direction.
All that stuff.
And all the stuff they've been putting out is around this narrative.
Now, fast forward to mid-last year.
The CEO of DeepSeek, the company, back then, it was totally obscure.
Nobody was tracking who they were.
They were working in total obscurity.
He goes on this, he does this random interview on Substack.
And what he says is, he's like, yeah, so honestly, like, we're really excited and doing this AGI push or whatever.
And like, honestly, like, money's not the problem for us.
Talent's not the problem for us.
But like, access to compute, like these export controls, man.
Do they ever work?
That's a real problem for us.
Oh, boy.
And, like, nobody noticed at the time.
But then the whole DeepSeek R1 thing blew up in December.
And now you imagine, like, you're the Chinese Ministry of Foreign Affairs.
Like, you've been, like, you've been putting this narrative together for, like, four years.
And this jackass that nobody heard about five minutes ago.
Right when R1 launched, there's a random DeepSeek employee.
I think his name is like Dia Guo or something like that.
He tweets out.
He's like, so this is like our most exciting launch of the year.
Nothing can stop us on the path to AGI except access to compute.
And then literally the dude in Washington, D.C., who works at the think tank on export controls against China, reposts that on X, and goes basically like, message received.
And so, like, hilarious for us.
But also, like, you know that on the backside, somebody got screamed at for that shit.
And that's part of the problem with like, so the Chinese face so many issues.
One of them is, you know, to kind of, another one is the idea of just waste and fraud, right?
So we have a free market.
Like what that means is you raise from private capital.
People who are pretty damn good at assessing shit will like look at your setup and assess whether it's worth backing you for these massive multi-billion dollar deals.
In China, the state like...
I mean, the stories of waste are pretty insane.
They'll, like, send a billion dollars to, like, a bunch of yahoos who will pivot from whatever, like, I don't know, making these widgets to just, like, oh, now we're, like, a chip foundry and they have no experience in it.
But because of all these subsidies, because of all these opportunities, now we're going to say that we are.
And then, no surprise, two years later, they burn out and they've just, like, lit.
A billion dollars on fire or whatever billion yen.
And, like, the weird thing is this is actually working overall, but it does lead to insane and unsustainable levels of waste.
Like, the Chinese system right now is obviously, like, they've got their massive property bubble that they're...
That's looking really bad.
We've got a population crisis.
The only way out for them is the AI stuff right now.
Like, really, the only path for them is that, which is why they're working it so hard.
But the stories of just, like, billions and tens of billions of dollars being lit on fire, specifically in the semiconductor industry, in the AI industry, like, that's a drag force that they're dealing with constantly that we don't have here in the same way.
So it's sort of like the different structural advantages and weaknesses of...
And when we think about what do we need to do to counter this, to be active in this space, to be a live player again, it means factoring in how do you take advantage of some of those opportunities that their system presents that ours doesn't.
So, well, I mean, things like tariffs, I mean, they're not shy about trying new stuff.
And tariffs are very complex in this space, like the impact, the actual impact of the tariffs and not universally good.
But the on-shoring effect is also something that you really want.
So it's a very mixed bag.
But it's certainly an administration that's like willing to do high stakes, big moves in a way that...
Other administrations haven't.
And in a time when you're looking at a transformative technology that's going to, like, upend so much about the way the world works, you can't afford to have that mentality we were just talking about with, like, the nervous...
I mean, you encountered it with the staffers, you know, when booking the podcast with the presidential cycle, right?
Yeah, and this is basically the sub-threshold version of, like, you know, like the World War II appeasement thing, where back, you know, Hitler was, like, was taken, he was taken Austria, he was re-militarizing shit, he was doing...
He was doing this, he was doing that.
And the British were like, okay, we're going to let him just take one more thing and then he will be satisfied.
And so this is basically like they fell into that pit, like that tar pit.
Peace in our time, yeah.
Peace in our time, right?
And to some extent, we've still kind of learned the lesson of not letting that happen with territorial boundaries, but that's big and it's visible and it happens on the map and you can't hide it.
Whereas one of the risks, especially with the previous administration, was there's these subthreshold things that don't show up in the news and that are calculated.
Basically, our adversaries know.
Because they know history.
They know not to give us a Pearl Harbor.
They know not to give us a 9-11.
Because historically, countries that give America a Pearl Harbor end up having a pretty bad time about it.
And so why would they give us a reason to come and bind together against an obvious external threat or risk when they can just keep chipping away at it?
Elevate that and realize this is what's happening.
This is the strategy.
We need to...
We need to take that, like, let's not do appeasement mentality and push it across in these other domains because that's where the real competition is going on.
That's where it gets so fascinating in regards to social media because it's imperative that you have an ability to express yourself.
It's very valuable for everybody.
The free exchange of information, finding out things that you're not going to get from mainstream media and it's led to the rise of independent journalism.
It's all great.
But also, you're being manipulated, like, left and right constantly.
Most people don't have the time to filter through it.
We try to get some sort of objective sense of what's actually going on.
It's like it's the layer where our society figures stuff out.
And if adversaries get into that layer, they're like almost inside of our brain.
And there's ways of addressing this.
Like one of the challenges obviously is like – so they try to push in extreme opinions in either direction.
And it's – that part is actually – it's kind of difficult because while – The most extreme opinions are also the most likely generally to be wrong.
They're also the most valuable when they're right because they tell us a thing that we didn't expect by definition that's true and that can really advance us forward.
And so, I mean, there are actually solutions to this.
I mean, this particular thing isn't an area we...
We're, like, too immersed in.
But one of the solutions that has been bandied about is, like, you know, like, you might know, like, polymarket prediction markets and stuff like that, where at least, you know, hypothetically, if you have a prediction market around, like, if we do this policy,
this thing will or won't happen, that actually creates a challenge around trying to manipulate that view or that market.
Because what ends up happening is, like, if you're an adversary and you want to...
Not just like manipulate a conversation that's happening in social media, which is cheap, but manipulate the price on a prediction market.
You have to buy in.
You have to spend real resources.
And if to the extent you're wrong and you're trying to create a wrong opinion, you're going to lose your resource.
So you actually can't push too far too many times or you will just get your money taken away from you.
I think that's one approach where just in terms of preserving discourse, some of the stuff that's happening in prediction markets is actually really interesting and really exciting, even in the context of bots and AIs and stuff like that.
That is what just the market is theoretically too, right?
It's got obviously big issues and can be manipulated in the short term.
But in the long run, this is one of the really interesting things about startups too.
When you run into people in the early days...
By definition, their startup looks like it's not going to succeed, right?
That is what it means to be a seed stage startup, right?
If it was obvious you were going to succeed, you would, you know, the people would have raised more money already.
Yeah.
So what you end up having is these highly contrarian people who, despite everybody telling them that they're going to fail, just believe in what they're doing and think they're going to succeed.
And I think that's part of what really kind of shapes the startup founder's soul in a way that's really constructive.
It's also something that, if you look at the Chinese system, is very different.
You raise money in very different ways.
You're coupled to the state apparatus.
You're both dependent on it and you're supported by it.
But there's just a lot of...
And it makes it hard for Americans to relate to Chinese and vice versa and understand each other's systems.
One of the biggest risks as you're thinking through what is your posture going to be relative to these countries is you fall into thinking that their traditions, their way of thinking about the world is the same as your own.
And that's something that's been an issue for us with China for a long time is, you know, hey, they'll liberalize, right?
Like bring them into the World Trade Organization.
It's like, oh, well, actually they'll sign the document, but they won't actually live up to any of the commitments.
It makes appeasement really tempting because you're thinking, oh, they're just like us.
Maybe AI, maybe super intelligence realizes, "Hey, you fucking apes, you territorial apes with thermonuclear weapons, how about you shut the fuck up?
You guys are doing the dumbest thing of all time and you're being manipulated by a small group of people that are profiting in insane ways off of your misery."
That's actually not- Stole first, and those people are now controlling all the fucking money.
That actually is, like, so this is not, like, relevant to the risk stuff or to the whatever at all, but it's just interesting.
So there's actually theories, like, in the same way that there's theories around power seeking and stuff around superintelligence, there's theories around, like, how superintelligence is.
Do deals with each other, right?
And you actually, like, you have this intuition, which is exactly right, which is that, hey, two super intelligences, like, actual legit super intelligences should never actually, like, fight each other destructively in the real world, right?
Like, that seems weird.
That shouldn't happen because they're so smart.
And in fact, like, there's theories around they can kind of do perfect deals with each other based on, like, if we're two super intelligences, I can kind of assess, like, how powerful you are.
You can assess how powerful I am, and we can actually decide, well, if we did fight a war against each other...
Well, hopefully it's so positive some, right, that even they enjoy the benefits.
But, I mean, you're right.
This is the issue right now.
And one of the nice things, too, is as you build up your ratchet of AI, It does start to open some opportunities for actual trust but verified, which is something that we can't do right now.
It's not like with nuclear stockpiles where we've had some success in some context with enforcing treaties and stuff like that, sending inspectors in and all that.
With AI right now, how can you actually prove that...
Like some international agreement on the use of AI is being observed.
Even if we figure out how to control these systems, how can we make sure that, you know, China is baking in those control mechanisms into their training runs and that we are and how can we prove it to each other without having total access to the compute stack?
We don't really have a solution for that.
There are all kinds of programs like this FlexHeg thing.
But anyway, those are not going to be online by 2027.
So the hope is that as you build up your AI capabilities, basically, it starts to create solutions.
So it starts to create ways for two countries to verifiably adhere to some kind of international agreement or to find, like you said, paths for de-escalation.
What's fascinating is, like, the unhideables, right?
The little things that...
Can't help but give away what is happening.
You think about this in AI quite a bit.
Some things that are hard for companies to hide is they'll have a job posting.
They've got to advertise to recruit.
So you'll see like, oh, interesting.
Oh, OpenAI is looking to hire some people from hedge funds.
I wonder what that means.
I wonder what that implies.
If you think about all of the leaders in the AI space, think about the Medallion Fund, for example.
This is a super successful hedge fund.
The Man Who Broke the Market.
The Man Who Broke the Market is the famous book about the founder of the Medallion Fund.
This is basically a fund that...
They make, like, ridiculous, like, $5 billion returns every year kind of guaranteed, so much so they have to cap how much they invest in the market because they would otherwise, like, move the market too much, like, affect it.
The fucked up thing about, like, the way they trade, and so this is, like, 20-year-old information, but it's still indicative because, like, you can't get current information about their strategies.
But one of the things that they were the first to kind of go for and figure out is they were like, Okay, they basically were the first to kind of build what was at the time, as much as possible, an AI that autonomously did trading at, like,
great speeds, and it had, like, no human oversight and just worked on its own.
And what they found was the strategies that were the most successful were the ones that humans understood the least.
Because if you have a strategy that a human can understand...
Some human's going to go and figure out that strategy and trade against you.
Whereas if you have the kind of the balls to go like, oh, this thing is doing some weird shit that I cannot understand no matter how hard I try, let's just fucking YOLO and trust it and make it work.
If you have all the stuff debugged and if the whole system is working right...
So to like to sort of explain why these these strategies work better, if you think about how AI systems are trained today, you basically very roughly.
You start with this blob of numbers that's called a model.
And you feed it input, you get an output.
If the output you get is no good, if you don't like the output, you basically fuck around with all those numbers, change them a little bit, and then you try again.
You're like, oh, okay, that's better.
And you repeat that process over and over and over with different inputs and outputs.
And eventually, those numbers, that mysterious ball of numbers, starts to behave well.
It starts to make good predictions or generate good outputs.
Now, you don't know why that is.
You just know that it does a good job, at least where you've tested it.
Now if you slightly change what you tested on, suddenly you could discover, oh shit, it's catastrophically failing at that thing.
These things are very brittle in that way, and that's...
That's part of the reason why ChatGPT will just like completely go on a psycho binge fest every once in a while if you give it a prompt that has like too many exclamation points and asterisks in it or something.
Like these systems are weirdly brittle in that way.
But applied to investment strategies, if all you're doing is saying like Optimize for returns.
If you want that system to get good at diagnosis, that's one thing.
OK, just fucking make it good at diagnosis.
If you want it to be good at diagnosis and to produce explanations that a good doctor will go like, OK, I'll use that.
Well, great.
But guess what?
Now you're spending some of that precious compute on something other than just the thing you're trying to optimize for.
And so now that's going to come at a cost of the actual performance of the system.
And so if you are going to optimize like the fuck out of making money.
You're going to necessarily de-optimize the fuck out of anything else, including being able to even understand what that system is doing.
And that's kind of like at the heart of a lot of the kind of big-picture AI strategy stuff is people are wondering, like, how much interpretability tax am I willing to pay here?
And how much does it cost?
And everyone's willing to go a little bit further and a little further.
So OpenAI actually had a paper or, I guess, a blog post where they talked about this.
And they were like, look, right now...
We have this, essentially, this, like, thought stream that our model produces on the way to generating its final output.
And that thought stream, like, we don't want to touch it to make it, like, interpretable, to make it make sense, because if we do that, then essentially it'll be optimized to convince us of whatever the thing is that we want it to do.
So it's like if you've used like an OpenAI model recently, right, like 03 or whatever, it's doing its thinking before it starts like outputting the answer.
And so that thinking is, yeah, we're supposed to like be able to read that and kind of get it, but also...
We don't want to make it too legible, because if we make it too legible, it's going to be optimized to be legible and to be convincing, rather than...
Yeah, so we're talking about, like, you do these atomic layer patterns on the chips and shit, and, like, what this UV thing does is it, like, fires, like, a really high-powered laser beam.
Laser beam, yeah.
They attach the head of sharks that just shoot at the chips.
Sorry, that was, like, an Austin Powers.
Anyway, they'll, like, shoot it at the chips, and that causes, depending on how the thing is designed, They'll, like, have a liquid layer of the stuff that's gonna go on the chip.
The UV is really, really tight and causes it, exactly, causes it to harden.
And then they wash off the liquid, and they do it all over again.
And so the exquisite machines that we get to use, or that they get to use in Taiwan, are called extreme ultraviolet lithography.
These are those crazy lasers.
The ones that China can use, because we've prevented them from getting any of those extreme ultraviolet lithography machines, the ones China uses are previous generation machines called Deep Ultraviolet, and they can't actually make chips as high a resolution as ours.
So what they do is, and what this article is about is, they basically take the same chip, they zap it once with DUV.
And then they gotta pass it through again, zap it again, to get closer to the level of resolution we get in one pass with our exquisite machine.
Now, the problem with that is you've got to pass the same chip through multiple times, which slows down your whole process.
It means your yields at the end of the day are lower.
And that article, too, like, this ties into the propaganda stuff we were talking about, right?
If you read an article like that, you could be forgiven for going, like, oh, man, our expert controls, like, just aren't working, so we might as well just give them up.
When in reality, because you look at the source, and this is how you know that also this is one of their propaganda things.
You look at Chinese news sources, what are they saying?
What are the beats that are, like, common?
And you know, just because of the way their media is set up, totally different from us, and we're not used to analyzing things this way, but when you read something in, like, the South China Morning Post, or, like, the Global Times, or Xinhua, or in a few different places like this, and it's the same beats coming back, you know that someone was handed a brief,
and it's like, you gotta hit this point, this point, this point, and, yep, they're gonna find a way to work that into the news cycle over there.
Right now, they're in the middle of staffing up some of the key positions because it's a new administration still, and this is such a technical domain.
They've got people there who are at the working level who are really sharp.
And the longer it goes on, like, the more, like, stuff gets squirreled away.
Like, there's actually, like, a story from the Soviet Union that always, like, gets me, which is, so Stalin obviously, like, purged and killed, like, millions of people in the 1930s, right?
By the 1980s, the ruling Politburo of the Soviet Union, obviously, like, things have been different.
Generations had turned over and all this stuff.
But those people, the most powerful people in the USSR, could not figure out what had happened to their own families during the purchase.
Like, the information was just nowhere to be found because the machine of the state was just like...
So aligned around like we just like we just gotta kill as many fucking people as we can and like turn it over and then hide the evidence of it and then kill the people who killed the people and then kill those people who killed those people.
But it was very much like you grind mostly or largely you grind them to death and basically they've gone away and you burn the records of it happening.
One of the things is, too, when you have such a big structure that's overseeing such complexity, right?
Obviously, a lot of stuff can hide in that structure, and it's not unrelated to the whole AI picture.
There's only so much compute that you have at the top of that system that you can spend, right?
As the president, as a cabinet member, like, whatever.
You can't look over everyone's shoulder and do their homework.
You can't do founder mode all the way down and all the branches and all the, like, action officers and all that shit.
That's not going to happen, which means you're spending five seconds thinking about how to unfuck some part of the government, but then the, like, you know...
Corrupt people who run their own fiefdoms there spend every day trying to figure out how to survive.
I think the real issue is in dismantling a lot of these programs that – You can point to some good some of these programs do.
The problem is, like, some of them are so overwhelmed with fraud and waste that it's like, to keep them active in the state they are, like, what do you do?
Do you rip the Band-Aid off and start from scratch?
Like, what do you do with the Department of Education?
Do you say, why are we number 39 when we were number one?
There's this idea in software engineering, actually, he's talking to one of our employees about this, which is like, Refactoring, right?
So when you're writing, like, a bunch of software, it gets really, really big and hairy and complicated, and there's all kinds of, like, dumbass shit, and there's all kinds of waste that happens in that codebase.
There's this thing that you do every, you know, every, like, few months, is you do this thing called refactoring, which is, like, you go, like, okay, we have, you know, 10 different things that are trying to do the same thing.
Get rid of nine of those things and just like rewrite it as the one thing.
So there's like a cleanup and refresh cycle that has to happen whenever you're developing a big complex thing that does a lot of stuff.
The thing is like the U.S. government at every level has basically never done a refactoring of itself.
And so the way that problems get solved is you're like...
Well, we need to do this new thing.
So we're just gonna, like, stick on another appendage to the beast and get that appendage to do that new thing.
And, like, that's been going on for 250 years, so we end up with, like, this beast that has a lot of appendages, many of which do incredibly duplicative and wasteful stuff, that if you were a software engineer, just, like, not politically, just objectively looking at that as a system,
you'd go, like, oh.
This is a catastrophe.
And, like, we have processes that the industry, we understand how, what needs to be done to fix that.
It's a problem, too, though, in all, like, when you're a big enough organization, you run into this problem, like, Google has this problem, famously.
We have friends, like, Jason, so Jason's the guy you spoke to about that.
So he's like a startup.
So he works in, like, relatively small codebases, and he, like, you know, can hold the whole codebase in his head at a time.
But when you move over to, you know, Google, to Facebook, like, all of a sudden, this gargantuan codebase starts to look more like the complexity of the U.S. government, just, like, you know, very roughly in terms of scale, right?
So now you're like, okay, well, we want to add functionality.
So we want to incentivize our teams to build products that are going to be valuable.
And the challenge is, The best way to incentivize that is to give people incentives to build new functionality.
Not to refactor.
There's no glory.
If you work at Google, there's no glory in refactoring.
If you work at Meta, there's no glory in refactoring.
And it's also kind of the only way to, I mean, it's probably not, but in the world where humans are doing the oversight, that's your limitation, right?
You got some people at the top who have a limited bandwidth and compute that they can dedicate to, like, hunting down the problems.
AI agents might actually solve that.
You could actually have a sort of autonomous AI agent that is the autonomous CEO or something go into an organization and uproot all the things and do that refactor.
You could get way more efficient organizations out of that.
Thinking about government corruption and waste and fraud, that's the kind of thing where those sorts of tools could be radically empowering, but you've got to get them to work right and for you.
On the control side, there's also a world where, and this is actually, like, if you talk to the labs, this is what they're actually planning to do, but it's a question of how methodically and carefully they can do this.
The plan is to ratchet up capabilities, and then scale, in other words.
And then as you do that, you start to use your AI systems, your increasingly clever and powerful AI systems, to do research on technical control.
So you basically build the next generation of systems.
You try to get that generation of systems to help you just inch forward a little bit more on the capability side.
It's a very precarious balance, but it's something that at least isn't insane on the face of it.
And fortunately, I mean, is the...
The default path, like the labs are talking about that kind of control element as being a key pillar of their strategy.
Ambiguity and uncertainty about what's going on in China.
So there's been a lot of like track 1.5, track 2 diplomacy, basically where you have non-government guys from one side talk to government guys from the other side or talk to non-government from the other side and kind of start to align on like, okay, what do we think the issues are?
You know, the Chinese are – there are a lot of like freaked out Chinese researchers and have come out publicly and said, hey, like we're really concerned about this whole loss of control thing.
There are public statements and all that.
You also have to be mindful that any statement the CCP puts out is a statement they want you to see.
So when they say like, "Oh yeah, we're really worried about this thing," it's genuinely hard to assess what that even means.
But as you start to build these systems, we expect you're going to see some evidence of this shit before.
And it's not necessarily, it's not like you're going to build the system necessarily and have it take over the world.
Yeah, so I was actually going to add to this really, really good point, and something where, like, open source AI is, like, even, you know, could potentially have an effect here.
So a couple of the major labs, like OpenAI Anthropic, I think, came out recently and said, like, look, we...
We're on the cusp.
Our systems are on the cusp of being able to help a total novice, like someone with no experience, develop and deploy and release a known biological threat.
And that's something we're going to have to grapple with over the next few months.
And eventually, capabilities like this, not necessarily just biological, but also cyber and other areas, are going to come out in open source.
When they come out in open source, you actually start to see some things happen, like some incidents, like some major hacks that were just done by a random motherfucker who just wants to see the world burn, but that wakes us up to like,
oh shit, these things actually are powerful.
I think one of the aspects also here is we're still in that...
Post-Cold War honeymoon, many of us, right?
In that mentality, like, not everyone has, like, wrapped their heads around this stuff.
And the, like, what needs to happen is something that makes us go, like, oh, damn, we, like, we weren't even really trying this entire time.
Because this is, like, this is the 9-11 effect.
This is the Pearl Harbor effect.
Once you have a thing that aligns everyone around like, oh shit, this is real and we actually need to do it and we're freaked out, we're actually safer.
We're safer when we're all like, okay, something important needs to happen.
But because you have the potential for this open source, it's probably going to be a survivable shock, right?
But still a shock.
And so let us actually realign around, like, okay...
Let's actually fucking solve some problems for real.
And so putting together the groundwork, right, is what we're doing around, like, let's pre-think a lot of this stuff so that, like, if and when the shock comes...
Like, so one interesting thing that happens with AI agents today is they'll, like, they'll get any...
So an AI agent will take a complex task that you give it, like, find me...
Like best sneakers for me online, some shit like that.
And they'll break it down into a series of sub-steps.
And then each of those steps, it'll farm out to a version of itself, say, to execute autonomously.
The more complex a task is, the more of those little sub-steps there are in it.
And so you can have an AI agent that nails like 99% of those steps.
But if it screws up just one, the whole thing is a flop, right?
And so...
If you think about the loss of control scenarios that a lot of people look at are autonomous replication, like the model gets access to the internet, copies itself onto servers and all that stuff.
Those are very complex movements.
If it screws up at any point along the way, that's a tell, like, oh, shit, something's happening there.
And you can start to think about, like, okay, well, what went wrong?
We get another do.
We get another try, and we can kind of learn from our mistakes.
So there is this sort of, like, this picture, you know, one camp goes, oh, well, we're going to kind of make this superintelligence in a vat, and then it explodes out and we lose control over it.
That doesn't...
Necessarily seem like the default scenario right now.
It seems like what we're doing is scaling these systems.
We might unhobble them with big capability jumps.
But there's a component of this that is a continuous process that lets us kind of get our arms around it in a more staged way.
That's another thing that I think is in our favor that we didn't expect before as a field, basically.
And I think that's a good thing.
That helps you kind of detect these breakout attempts and do things about them.
But I really appreciate you guys and appreciate your perspective because it's very important and it's very illuminating.
It gives you a sense of what's going on.
And I think one of the things that you said that's really important is, like, it sucks that we need a 9-11 moment or a Pearl Harbor moment to realize what's happening so we all come together.
But hopefully, slowly but surely, through conversations like this, people realize what's actually happening.