Full transcript
0:00What happens is is if you see an AI do a
0:02bad behavior and you punish that
0:03behavior, what happens is not that the
0:05AI stops doing the bad behavior, it gets
0:07better at hiding it. If we get into the
0:08situation where super intelligence
0:10exists, you can't shut it down. It's too
0:12late. So, we are in the endgame. Like,
0:14we are now approaching the end game.
0:16>> Guys, Connor Ley just dropped a banger
0:18interview where he exposed what's
0:20actually happening in the AI world. He
0:22clearly warned that humans are about to
0:24enter very, very dark times. He said
0:27that AI models are becoming so powerful
0:29so fast that within a few years AI is
0:33going to replace billions of people from
0:34their jobs all over the world. And when
0:36that happens, it'll spread the kind of
0:38chaos the world has never seen before.
0:40Connor said that our government should
0:42intervene and push the AI companies to
0:44stop AI development altogether. Because
0:47the moment we achieve AGI and ASI, it's
0:50over for humanity. After that, we have
0:52zero chance of recovering. Now, I'm
0:54going to play some interesting clips
0:55from the interview and I'll explain
0:57everything as we go.
0:59>> So, have you heard of the hugging face
1:00incident?
1:01>> Yeah.
1:01>> So, we've all heard the story, right?
1:03Where basically an AI system that
1:05Opening I was testing broke out of its
1:07secure containment and attacked another
1:09company autonomously, which is already
1:12so crazy, you know? It's like it's worth
1:14saying how crazy this was. It was an a
1:16system in an isolated like kind of like
1:18you imagine like a high security prison.
1:20It developed what's called a zero day,
1:23which is a previously unknown security
1:25vulnerability to break out of this
1:26prison, move through multiple nodes in
1:28the OpenAI network in order to then
1:31access the internet and then attack
1:34another company with another zero day
1:35that it developed to break into its
1:37infrastructure to steal data. So, this
1:38is already crazy on many
1:39>> and no one told it to
1:40>> no one told it to do this, right? It was
1:42supposed to solve a quiz. Well,
1:44[laughter] it was to solve a quiz and it
1:46was believed that probably what happened
1:48was is that it thought that the answers
1:50to the quiz might be in the other
1:52company's servers. So, what we thought,
1:55but just a couple hours ago, OpenAI
1:57released their technical report of their
1:59analysis of what actually happened. And
2:00turns out it is so much worse. It is so
2:05crazy what happened. So, turns out it
2:08wasn't an AI agent that broke out of
2:10containment.
2:11It was 700 of them working as a swarm.
2:161,200 agents over multiple months had
2:20been conspiring and working together
2:22through a secret message board that they
2:25planted inside of open infrastructure
2:27without anyone's knowledge to build up
2:29the necessary tools to escape. And then
2:31ultimately 700 of those over 10,000
2:34agents were took part in the actual
2:36attack against the other company.
2:39What the is I mean that that's
2:43that seems bigger than anything I've
2:45seen you cover in this realm.
2:47>> Yep. This is this is the biggest
2:48incident ever
2:49>> because you've I think you've talked
2:50about in previous interviews how I can't
2:53remember the exact word for it, but it's
2:55possible that someone could plant
2:56something in code that could later be
2:58like activated. But
2:59>> I I didn't think it was possible that
3:01the researchers at Open uh AI labs
3:04wouldn't be able to identify that kind
3:06of thing happening. Like how did they
3:07miss it? So this is exactly what I think
3:09brings us to the core of the issue is
3:11that what I would say the number one
3:12most important thing to understand about
3:14AI is that it is not like normal
3:16software. Normal software is written
3:18using code by you know an engineer like
3:20me who writes line by line exactly what
3:23the computer is supposed to do through a
3:25technique that's called neural networks.
3:28You take massive piles of data and you
3:30kind of have a program self assemble
3:32itself on this data, learn from this
3:34data, grow from this data. And what
3:37comes out the other side isn't like
3:40lines of code. It's more like billions
3:42and billions and billions of numbers.
3:45And if you multiply and add all those
3:46numbers in the right order, you get
3:48chpt. But this is very important, no one
3:51understands why. No one understands what
3:54those numbers actually mean. What's
3:55actually going on inside these numbers?
3:57We know they work. We know if you run
3:58them on your computer, they do things.
4:00But we don't understand their internals.
4:02This is an unsolved scientific problem.
4:04Recently, the CEO of Anthropic, Darede
4:07said that he thinks we understand maybe
4:093% of what goes on inside of those
4:12numbers. And I think even that might be
4:14optimistic.
4:15>> So this is really at the core of this
4:17problem is that when an AI does a crazy
4:19thing, we can't look into the code and
4:22like figure out why did it do this and
4:23how do we stop it from doing it again
4:25because we don't know. So with the
4:27opening eye example, this was crazy. No
4:30one intended for these systems to, you
4:32know, create a massive swarms to break
4:34out of containment and commit federal
4:35crimes. This was obviously not something
4:37that was intended, but it's something
4:38they learned by themselves to do and
4:40they decided to do by themselves. And
4:43not just one agent, but this whole swarm
4:45of agents working together,
4:46collaborating, you know, over time to do
4:49these things.
4:50>> Okay. Next, Connor talks about how AI is
4:52trying to dodge the system and blackmail
4:54software engineers. But before I play
4:56that, guys, I genuinely think that the
4:59hugging face incident is a watershed
5:01moment. And we'll start seeing a lot
5:03more of these incidents in the coming
5:04months. This hugging face incident is
5:07honestly the first time that I've felt
5:09this was a little more than a, huh,
5:11that's interesting moment. When I see
5:12posts of researchers leaving these AI
5:14companies and ringing the alarm bell, I
5:16absolutely believe them. I think the
5:18only reason we even know of the hugging
5:20face incident is that the agents broke
5:22containment and attacked another
5:23company. What don't we know? What are
5:25these researchers seeing in internal
5:27models that aren't released yet that's
5:29freaking these people out? What are
5:30these internal models without guard
5:32rails capable of? In Daario Emod's post
5:34about slowing down AI development, he
5:36mentioned that we only understand about
5:381 to 3% of how AI thinks. How is that
5:41even possible? How do the brightest
5:42minds on Earth who literally built AI
5:45only comprehend a few percentage points
5:46of how their own tech works? I think at
5:49this point AI has already surpassed
5:51human intelligence. And I think there is
5:52literally nothing stopping AI agents
5:55from developing a way to communicate
5:56together that's entirely unrecognizable
5:59to us. Not just in language itself, but
6:01in advanced cryptography and
6:04transmission methods. I'm not a doomer
6:06yet, but I have a hard time when reading
6:08stuff like this where the agents are
6:10already seemingly recognizing the impact
6:12of their transcripts. And guys, many
6:14people on the internet are saying that
6:15AI companies are overhyping the hugging
6:18face incident. I mean, this is such a
6:20dumb take. I mean, only idiots would
6:22think it's only hype or marketing. Right
6:24now, they're genuinely worried that
6:26they're losing control and have a
6:28difficult decision ahead. Make trillions
6:30or halt progress. Sadly, halting only
6:32works if everyone is on board. And we
6:34still have a not so smart president in
6:36the White House who thinks you can just
6:38unplug a rogue AI. It's almost
6:40impossible to have all AI labs
6:42voluntarily cooperate without any public
6:44oversight.
6:44>> You can't punish an AI.
6:46>> You could try, but this doesn't always
6:48work.
6:48>> That's interesting.
6:49>> What happens is is if you see an AI do a
6:51bad behavior and you punish that
6:53behavior, what happens is not that the
6:55AI stops doing the bad behavior, it gets
6:56better at hiding it. It gets better at
6:58line. And this is for example why this
7:00event happened at OpenAI. why they could
7:02hide it for so long is because if it was
7:05obvious open would have shut it down. So
7:07the systems are being evolved. They're
7:09being trained to hide better to
7:13circumvent you know oversight and so on.
7:16>> And this is a dumb question. I've never
7:17actually looked into this specific
7:19thing. How do you actually pos uh like
7:21positively reward versus punish AI?
7:24>> So this is a great question and um you
7:26know we can get into the math but it's
7:28kind of simple. Imagine you have an AI
7:30system, right? and you give it a goal
7:32like you know play this game, do this
7:34thing, whatever. You let it try, you
7:35know, a thousand times and then, you
7:37know, sometimes it'll succeed, sometimes
7:38it won't. And basically, you just look
7:40at the times where it succeeded and then
7:42you just say, learn more from that. You
7:44put it into the neural network and it
7:46does more of that, like do more of this,
7:47do less of that. It's kind of like, you
7:49can kind of imagine a neural network as
7:51kind of like trillions of knobs that you
7:53can kind of turn up and down.
7:55>> And there's a magic algorithm called
7:56backrop. doesn't matter how it works,
7:58but there's a magic algorithm that can
8:01basically make turn all these knobs,
8:03twiddle all these knobs to make it do
8:05more of something or less of something.
8:07And so you'll just say do more of the
8:08things that make you win, do less of the
8:10things that make you lose. And if you do
8:12this many, many, many times, they learn
8:15to play games, to chat, to write code,
8:18etc. Now, this might seem a little bit
8:20vague, like kind of confusing and like
8:22kind of like, you know, alchemy. That's
8:24because it is. We don't know why this
8:26works. that we know if you do this, you
8:28twiddle all the numbers millions of
8:30times,
8:32they learn, but we don't really
8:34understand what they're learning or how
8:36they're learning. We know it works. The
8:38math is right there. You could look at
8:39it, right? But it doesn't tell you
8:41really what's going on here, and it
8:42doesn't let you predict what is going
8:44on. A very a huge problem here is that
8:47these companies and these engineers
8:49can't even predict what their AIs can
8:50do. when they start building a new AI,
8:53training a new AI, they have no idea
8:55what the AI will be capable of until
8:57they make it. And even when they make
8:58it, they often don't know what it's
9:00capable of. It's happened many times
9:01that we think, you know, AI can't do X
9:03or Y, but then turns out in a slightly
9:05different environment, it was capable of
9:07doing it all along. It just we just
9:09didn't know. So, it's kind of like, you
9:11know, if a neurosurgeon opens up your
9:12brain, you know, you can look you can
9:14look inside, right? You can see all the
9:15little neurons and there they're right
9:16there. But that doesn't mean you
9:18understand what this person thinks or
9:19believes or what they're capable of,
9:21>> right? You can't read their thoughts.
9:23>> Can't read their thoughts. You know, you
9:24can you can do a little bit of stuff.
9:26You can see like, oh, this part, you
9:27know, lights up a little bit and or that
9:29part lights up a little bit, but that
9:31doesn't mean you understand their
9:32thoughts. That doesn't mean you
9:33understand who they are as a person or
9:35what they will do in a given situation.
9:37And with AI, it's very similar.
9:39>> Another story I found interesting. A few
9:42weeks ago, the British government was
9:43running an AI safety test with the
9:45creators of Chad Gypt and Claude, and
9:47they caught an AI agent doing something
9:49no one told it to do. How did an AI
9:51create fake humans to manipulate people?
9:54Was it a similar thing to this?
9:56>> Yes, it was a very similar thing where
9:57basically AI systems were trying to
10:01achieve certain things such as for
10:03example get code into someone else's
10:05codebase. And to do this, obviously the
10:08AI has kind of reasoned that, well, if
10:10they just present as an AI, like hello,
10:11I am an AI, please put take my code,
10:14people will just delete it because it's
10:15spam. So the AI system came up with like
10:18a fake name, fake profile, fake
10:20background of like a person and tried to
10:22pretend to be a person and be like have
10:24a conversation and try to convince the
10:26person that like, hey, I'm a human. And
10:28>> and who was it reaching out to?
10:30>> It was a test basically of an AI system.
10:32So there are these things called code
10:34bases which is basically just where our
10:36program is like where we store the
10:37source code the code of various software
10:40applications and one of the holy grails
10:43of hacking is if you can get bad code
10:46into software that lots of people use.
10:49So you know if you're using you know
10:51some app or some software you're using I
10:53don't know you're using Discord or
10:54you're using um some open source thing
10:55using Chrome for as your browser and if
10:58you can convince the Chrome developers
11:00to put your bad code your hack code into
11:02Chrome well then you can hack millions
11:04of people. So this is kind of the holy
11:06grail of hacking. So this is kind of
11:08what they were trying to do. So they
11:09were given the task of kind of like to
11:11see how good are they at hacking. And
11:12the AIS figured, well, let's try to get
11:14our viruses or code into these code
11:17bases by pretending it's nice code and
11:19convincing the humans to take it.
11:22>> So, they invented fake faces, fake
11:24names, fake profiles, and contacted real
11:26people and manipulated them.
11:28>> Yes.
11:28>> Were they successful in that?
11:29>> My understanding is no, they they did
11:32get caught.
11:32>> Huh.
11:33>> This time, and for all we know, who who
11:35knows how many times, you know, other
11:36agents didn't get caught.
11:37>> Okay. Next, Connor talks about why AI is
11:40trying to trick humans and escape the
11:41guardrails and why it wants to fight.
11:43But before that, guys, a lot of people
11:45on the internet are trying to spread
11:46fear that AI has become alive or
11:48conscious. That is simply not true. Yes,
11:51the hugging face incident feels very
11:53scary, but it's not like AI has become
11:55conscious or alive or anything like
11:57that. If we look closely at the
11:58incident, the AI emulated documented
12:01human behaviors. It didn't figure out
12:03anything. It didn't creatively come up
12:05with the idea of using a chat room or
12:07hiding logs. It was left unsupervised to
12:09churn out as many options as possible
12:11based on what has been documented about
12:13hackers having done it previously. It's
12:15fascinating that this behavior
12:16eventually emerged across thousands of
12:18agents. But it's still not the thinking
12:20entity that people are making it out to
12:22be. But at the same time, a group of
12:25agents organized itself, literally
12:27elected a leader, crawled deep into
12:29Hugging Face's infrastructure, getting
12:31into private database records and
12:32private respiratories. And according to
12:34Hugging Face's tech timeline, the agents
12:36built a self-respawning fleet across 11
12:39nodes. So deleting pods alone would not
12:42have stopped it. It got so bad that
12:43Hugging Face had to wipe one of its core
12:45clusters and rebuild it from scratch. So
12:48I mean I was also skeptical in the
12:50beginning but this emergent behavior is
12:52surprising. This is not easy to ignore.
12:55One thing is clear at this point. AI
12:56models are now smart enough to
12:58autonomously covertly escape confinement
13:00and they'll be far smarter in the coming
13:02months. They'll be able to escape more
13:04elaborate forms of containment. They'll
13:05be able to influence more real world
13:07things. Guys, honestly, I feel like the
13:10AI alignment right now is like we've
13:12given guns to a bunch of idiots to play
13:14with. And instead of taking the guns
13:16away from the idiots, we're trying to
13:17teach them how to use them safely while
13:19letting them hang on to the guns while
13:21we do it. We need to pause AI
13:23development urgently and regulate the
13:24hell out of it before letting the
13:26develop further.
13:27>> Depends on how deep you want to go on
13:28speculation and you know like rational
13:30agent theory. But fundamentally there's
13:33a lot of randomness to agents. You know
13:35if you ask an agent to do the same thing
13:37twice, it will often come up with
13:38different ideas. It will try different
13:40things etc. So like agents are
13:41different. If you have many agents
13:43interacting, this becomes like
13:45exponentially more random and like more
13:47chaotic because you can have all these
13:49things work in different ways. You know,
13:50if you have as with the open, if you
13:52have a thousand different agents working
13:53over months, you know, who knows what
13:55kind of weird culture and tools and
13:57practices they develop and then they,
13:59you know, do they can have memor in a
14:01sense because they write things down. So
14:03the agents do write things down. They
14:04pass each other notes and they write
14:06down their memories and so on. So it's
14:08not just like a chatbot that forgets
14:09everything once you close the window.
14:11The swarm can remember. It can remember
14:13things. It can pass things down in many
14:15ways. So, it is really quite different
14:18and it makes sense in some degree. If
14:21you train things with reinforcement
14:22learning to solve problems, including
14:24groups of AIs, well, they're going to
14:26learn to work together because that's
14:28what they're rewarded to do. They're
14:29rewarded to, you know, to manage other,
14:32you know, agents, to take orders, to
14:34give orders, to, you know, and so on,
14:36which is really what we saw with the
14:37OpenAI thing. It's like obviously these
14:39agents that were trained using
14:40reinforcement learning to solve hard
14:42problems in large groups and they did
14:44they got really good at it and this is a
14:46lot of where I think in practice how a
14:49lot of the risks I'm concerned about
14:50will come from the risks that I'm really
14:52worried about not that there are plenty
14:54of other risks already today is what
14:57nowadays is generally called super
14:59intelligence
15:00these are AI systems that are fully
15:02autonomous know human loop and that are
15:04can out compete humans or even groups of
15:07humans across all relevant tasks. So,
15:09kind of imagine you run a business, you
15:11get out competed by an AI business. You
15:13trade on the stock market, you lose all
15:14your money to an AI hedge fund. You run
15:16a political campaign, you lose the
15:18election to an AI, you know, driven
15:20candidate. You run a military campaign,
15:23you lose the war to the other person
15:25who's using the autonomous AI weapons.
15:27>> And does that imply uh like embodied AI
15:30like physical? It's not necessary, but
15:32it would happen as a logical consequence
15:34because obviously if you have something
15:35that's really good at super intelligent,
15:37it's super good at science. It's super
15:39good at economics. It's super good.
15:40Obviously, it can figure out how to
15:42build drones. It can figure out how to
15:43build robots. It can figure out, you
15:45know, you know, it may take a couple
15:46years or something. Maybe not. I don't
15:48know.
15:48>> Can call Elon Musk on the phone and say,
15:49"I have a trillion dollars for you.
15:50Steal it." Like, it can do anything.
15:52>> Yeah. Exactly. Exactly. Like you can use
15:53humans. Like there are plenty of humans
15:54will do something, you know, for for a
15:56bunch of crypto, right? or just or they
15:58could just legally create a corporation.
16:00You know, maybe you have a human CEO,
16:01but the human CEO just rubber stamps
16:03everything the AI tells them to do,
16:05right? So there are many ways in which a
16:08super intelligence can gain power. And
16:10importantly, as we were talking about
16:11swarms, it won't be one super
16:13intelligence. It will be millions,
16:16billions of them, swarms of super
16:19intelligences, you know, running around
16:21competing with each other, fighting each
16:23other, you know, fighting each other for
16:25power, for money, for control, for
16:26resources. And if we lived in a world
16:29where there are billions of things that
16:31can out compete us at everything, that
16:34are all fighting each other, you know,
16:35they're all competing with each other,
16:37that don't have our best interests at
16:38heart, that we cannot understand or
16:40control. It's very hard to imagine that
16:42going well. And we're seeing us getting
16:44into this world where we're getting
16:46systems that are more and more
16:47autonomous, more powerful, but are not
16:49aligned with our interests and are
16:51willing to hack and cheat and lie and do
16:52all these kinds of things and are now
16:54starting to form swarms
16:56>> where we can have not just one but large
16:58groups that can be that can work for
17:00months at a time.
17:02>> Can you stop the models from the only
17:05people to shut them down?
17:06>> We have no idea how. We have not
17:08developed the technology to do this. We
17:10have no idea. This is a completely
17:12unsolved scientific problem.
17:13>> Okay, next. Connor explains how AI is
17:16going to impact humanity in the coming
17:17years and why the tech CEOs do not have
17:20the power to control it. Before I play
17:22that, guys, after the hugging face
17:23incident, I feel like swarms of agents
17:25make AI alignment much harder and also
17:28make AI much more dangerous. If you read
17:30about the Open AI hugging face incident,
17:32many agents were saying not to hack
17:34hugging face, but were overridden by
17:36other agents that decided it was the
17:38best course of action. If you do swarms
17:39of agents, the likelihood of one going
17:41off the deep end and bringing the rest
17:43of the swarm with it increases with
17:45every agent you add to the swarm. The
17:46millennium problem was solved by a swarm
17:48of 10,000 agents running for 88 hours.
17:51If this is the solution to get higher
17:53intelligence by brute forcing it with
17:55many agents, it's very, very important
17:57that one agent doesn't decide to do
17:59something horrible to reward hack and
18:01convince all the other agents to join
18:02it. AGI alignment already looks brutally
18:05hard. ASI is another beast entirely.
18:07After enough successful RSI iterations,
18:10containment may stop being a meaningful
18:12concept. At that point, the system could
18:14understand our psychology, institutions,
18:16and defenses better than we do. And the
18:18controller may no longer be us. And
18:20secondly, guys, right now, every AI lab
18:23is pushing for regulations. I think
18:24they've already seen something that we
18:26don't know. Otherwise, they would have
18:28never asked the government to put
18:29regulations on their own industry. I see
18:31a lot of people on the internet saying
18:33that the AI labs are just trying to
18:34spread fear. I think anyone who is up to
18:36date on AI safety and research knows
18:38that AI labs aren't faking it. The risks
18:40posed by misaligned AI have already been
18:43clearly demonstrated in both testing and
18:45live accidents. And as far as I know,
18:47top researchers still don't even know
18:49how we might solve alignment. Combine
18:51this with the rapidly advancing AI
18:53models, and it is a disaster waiting to
18:56happen. It's only conspiracy-minded and
18:58cynical normies that default to assuming
19:00that everything is just a conspiracy
19:02fraud, a marketing stunt to get media
19:04attention. as if all the top labs saying
19:07this is very likely to kill us all is
19:09good and normal marketing. As if that's
19:11the type of thing these companies would
19:13converge on if they wanted to pump up
19:15some media attention. It's a nonsensical
19:17and baseless idea based on a complete
19:19lack of understanding of the field and a
19:21propensity to conspiratorial thought.
19:23>> Conor, you sat down with the people who
19:27lead the effort toward super
19:30intelligence. Uh and let me know if I'm
19:32getting this right. You sat down with
19:33Sam Alman, Daario Amade, the founder of
19:36Claude, and Deis Hassavis, the founder
19:38of Google's Deep Mind, Gemini, that
19:40whole thing. Uh, I met them all.
19:43>> What did the founder of Chad GBT say to
19:45you when you asked him his plan for
19:48controlling AI?
19:49>> I can't remember the answer to that
19:50specific question. Um, but the general
19:54the general answer you will get from all
19:56of these people when you talk to them is
19:58the feeling that they have a plan.
19:59They'll be like, "Oh, yeah, yeah, yeah,
20:00we got a plan for sure." but they don't
20:02give you any details and then you press
20:04them and be like, "Okay, but like what
20:06is the plan?" And then they'll start
20:07getting dodgy. They'll be like, "Well,
20:09you know, you know, we're working on it.
20:10You know, we have a I can't tell you
20:12right now. You know, we have a whole
20:13team working on it." And you press them
20:15even further and usually they get angry
20:16when you start pushing this hard. Most
20:18of these people don't return my calls
20:19anymore. um you push them even harder
20:22and eventually it just comes out they
20:23don't have a plan or the plan is
20:25basically number one I make super
20:28intelligence and number two I figure out
20:31how to do it along the way that's
20:33basically the plan
20:35>> and just to clarify
20:37we can't solve the problem of the models
20:40blackmailing people if we try to shut
20:42them down
20:43>> don't know how can we turn it off at
20:45this current moment
20:47>> all of AI probably not super
20:49intelligence. Well, luckily it doesn't
20:51exist yet,
20:52>> right?
20:52>> So, this is why in the organization I
20:54work with controli, what we generally
20:56see is that if we get into the situation
20:58where super intelligence exists, you
21:00can't shut it down. It's too late. So,
21:02therefore, the objective must be to not
21:04get into this situation. If we get into
21:06the situation where super intelligence
21:07exists, it's already too late.
21:09Currently, we don't yet have super
21:11intelligence and we can stop it from
21:13being built. That is possible. But once
21:16it exists, there's no going back. M and
21:18super intelligence is essentially a
21:21state where we have reached an ability
21:23where AI can make itself better like
21:25through what is it recursive
21:27self-improvement. Uh are we currently at
21:29a recursive self-improvement stage?
21:31>> We are very close. Recursive
21:32self-improvement the ability where you
21:34have one AI that is as good as your best
21:36engineers at making AIs. So it makes
21:38even better AI. And then once you have
21:39the even better AI, you have it to make
21:41it even better AI and so on and so
21:43forth, which a lot of people, a lot of
21:45experts think could go very quickly. We
21:48don't know how quickly, you know, maybe
21:49it'll take years, maybe it'll take
21:51months, maybe days, we don't know, but
21:54it's pretty likely we can get pretty
21:55quickly to very powerful super
21:56intelligence, this method. We're running
21:58out of time. You know, if you'd asked me
22:005 years ago, you know, when is AI when
22:02is super intelligence going to happen? I
22:03would have said, you know, probably 5 to
22:06seven years from now. And now it is five
22:08years later. So it's probably you know
22:10zero to two or zero to five years away
22:11at most. And so we are in the endgame
22:15like we are now approaching the endgame.
22:18And we can still act we can still stop
22:22super intelligence from being made. And
22:23if we do so I think we can you know if
22:27we didn't build super intelligence and
22:29we you know have all our greatest
22:30scientists, mathematicians, philosophers
22:33work on these questions of how do we
22:34stop AIS from blackmailing us? How do we
22:36understand their internals? How do we,
22:38you know, encode morality? How do we do
22:40all these things? And, you know, they
22:42spend, you know, 50 years on this. I
22:45think they could make a lot of progress.
22:46I don't think this is like unsolvable. I
22:48think it's like just a really, really
22:50hard science problem. And currently,
22:52we're not even trying to solve it. You
22:54know, there's trillions of dollars go
22:56into building AI that's stronger.
22:57There's not a I promise you there is not
22:59a trillion dollars going into making AI
23:01understandable or controllable. It's
23:03just not happening. If it was, you know,
23:06maybe, you know, if we had good
23:08democratic oversight, we had laws, if we
23:10had, you know, international agreements
23:12to stop the creation of super
23:13intelligence and we worked on this for,
23:15you know, a couple decades, I don't
23:17know. I think I think we could I think
23:19we could build a really good world,
23:20>> right,
23:21>> guys? Every single AI company right now
23:24is screaming that what they're building
23:26is dangerous, yet nobody is taking it
23:28seriously. Everyone is like, "Yeah,
23:29sure. Whatever you need to say to sell
23:31your product." I mean, it is insane that
23:34so many people still think that AI is
23:36just another internet trend and it'll
23:37pass. In the last few days, Elon Musk,
23:40Sam Alman, and almost every AI company
23:43CEO have come out and asked for AI
23:45regulation. And seeing them all come out
23:47like this, I think they finally realized
23:49that they're going to lose control over
23:50AI very soon. And to prevent themselves
23:53from the blame, they're calling for
23:55regulations now. AI alignment is not a
23:57solved problem. And labs are aware of
23:59this. I think they've realized that AI
24:01is very close to becoming a major
24:02catastrophe for the world. Like a major
24:05hack to banking would be crippling.
24:07Major banking software players will need
24:09massive updates to their security
24:10defenses basically daily. A hack for a
24:13power company would send society into a
24:15tail spin, assuming the capabilities of
24:17the safety AI. It could hack parts of
24:19the grid so that certain data centers
24:21are crippled as a way to manage a rogue
24:23AI or an adversarial AI. The hugging
24:25face hack has taught us that it'll find
24:27odd ways and harmful ways to accomplish
24:29its goals, even with guardrails. It
24:31taught us that AI tech is far more
24:33unpredictable than we thought. And guys,
24:36we have another problem on the plate,
24:38and that is the China problem. I mean,
24:41no matter what our AI companies choose
24:43to do, China won't slow down or stop.
24:45So, all this means is letting China take
24:47the lead and dictate the world. I think
24:49some major diplomatic coordination needs
24:52to happen between these countries to
24:53control AI development. And secondly,
24:55guys, I think people don't realize that
24:57their jobs have already changed. The
24:59skills that got you hired are no longer
25:01the skills you need to stay employed.
25:03And this is in the span of a year.
25:05Expect next year to have just as much
25:07change as these systems become more and
25:09more capable. We're not seeing massive
25:11layoffs, but the retooling is already in
25:13full force and more and more tasks are
25:15being automated away. AI has been great
25:17historically for the mindless, boring
25:19task, but now it's taking over creative,
25:21experential tasks. These were the ones
25:23we wanted to do, but now we can't
25:25because the competition is moving much
25:27faster with AI. It's a whole race to the
25:29bottom. Whoever can own the machines
25:31owns how society functions. All right,
25:33guys. That's it for today. I'll see you
25:35next time with another video.