Thursday, September 3, 2026

AI Whistleblower WARNS: "They Are Not Telling You What's Coming!"

  AI Whistleblower WARNS: "They Are Not Telling You What's Coming!"           

 https://www.youtube.com/watch?v=hcZOQoG_tMQ

 


 TRANSCRIPT 

Just just stop. Just just just don't build super intelligence. Super intelligence should be illegal the same way that nuclear bombs are illegal. If
8 seconds
the United States builds super intelligence, the United States is destroyed. If China builds super intelligence, China is destroyed. Like
15 seconds
it doesn't matter who builds it. If anyone builds it, everyone dies. Super intelligence is not a weapon. It's not a tool. It's an adversary.
25 seconds
When you build super intelligence, you don't have super intelligence. Super intelligence has you. It is not actually in the game theoretic rational interest
33 seconds
of a country to build super intelligence.
35 seconds
Guys, AI whistleblower Connor Ley just dropped another eyeopening interview in which he shared insights on how AI is
42 seconds
rapidly getting out of our control and how it's evolving into a monster that'll become an existential threat for humanity in the coming 2 to 3 years.
51 seconds
Connor warns that AI companies are blindly racing towards AGI without considering its consequences for 8
58 seconds
billion humans on this earth. He said companies like Meta, Open AI, and Anthropic don't give a damn about the billions of people who would lose jobs
1 minute, 6 seconds
in the coming years. They just have one goal and that is to achieve AGI and ultimately ASI. Now, I'm going to play
1 minute, 14 seconds
some interesting clips from the interview. I'll explain everything as we go. So we're going to focus on the open AAI incident in particular because it
1 minute, 21 seconds
was the most shocking. There's now since now been a whole conga line of other incidents since then and actually before then funnily enough a lot of companies
1 minute, 29 seconds
after the open in said man we should better double check our AIS aren't doing this and then they found a bunch of incident that already happened.
1 minute, 35 seconds
Anthropic found a bunch andropic found three. Um I'm pretty sure OpenAI found a couple other ones. Now recently I think the moonshot people in
1 minute, 43 seconds
China found an incident as well. So these have already been happening, but let's take a step back. So OpenAI was
1 minute, 51 seconds
testing one of their unreleased AI systems in a what's called a sandbox environment. So you can kind of imagine, you know, a computer that is kind of cut
2 minutes
off from the internet and is running like special software to be like specific specially secure. Um so the system can you know operate there. It's
2 minutes, 8 seconds
worth understanding that like these kind of sandbox systems are very secure.
2 minutes, 12 seconds
Breaking out of something like this is very hard. Like this is a thing that if you can do this, you can make billions of dollars because you can break all the
2 minutes, 19 seconds
cloud providers. So breaking out these systems is not easy. This is and and sandboxes exist to test software without infecting other parts of the internet.
2 minutes, 28 seconds
Yes, this is also how cloud uh systems work. You know, like for example, if you go to like you know AWS or Microsoft or something and you rent a cloud computer,
2 minutes, 36 seconds
your your computer is sandboxed from all the other ones, okay? because you don't want to, you know, see what the other customers are doing on the cloud, right?
2 minutes, 42 seconds
So the cloud insulates different customers from each other.
2 minutes, 45 seconds
You're not on essentially one machine, you're sharing machines.
2 minutes, 47 seconds
Yes. You're sharing. Exactly. So when the cloud you're sharing a slice of a machine with other people and you don't want different people on the same
2 minutes, 55 seconds
machine to see each other's stuff and so you use sandboxing and like various or virtualization, various forms of techniques to keep this safe and so this
3 minutes, 2 seconds
is the kind of system we're talking about kind of. So this is pretty standard. This is pretty industry standard, well known, etc. So this system did not have access to the
3 minutes, 10 seconds
internet. What it did have access to is what's called um a package repository.
3 minutes, 15 seconds
You can kind of imagine this is like a tool shed. Like there's like a shed full of tools that the AI is allowed to go to and get some of the tools for its job.
3 minutes, 22 seconds
Is that like when I use terminal I download npm? Yeah, npm. You know, stuff like this.
3 minutes, 28 seconds
You need you download packages, software libraries to do your work. So what happened was is this AI system was told to solve a certain test kind of like a
3 minutes, 35 seconds
quiz you know not literally but like kind of like a test a quiz the AI system was struggling with solving this quiz and so what it seems to have concluded
3 minutes, 45 seconds
we don't know for sure obviously because again these things are grown not written we don't actually understand what's going on inside of the minds of these AIs we don't understand they're not like
3 minutes, 52 seconds
normal software we can just like read the code okay the following part of the interview is really interesting in this part Connor talks about how AI is being
4 minutes
trained ain to lie, cheat, and deceive by the AI companies. But before I play that part, guys, I feel like the majority of the AI critics mostly focus
4 minutes, 9 seconds
on AI causing human extinction. I mean, it's highly probable, but it ain't happening in the coming decade, at least right now. What really scares me in the
4 minutes, 17 seconds
short run is the slow decay of our mental abilities. Think about the effects of car culture on our health.
4 minutes, 23 seconds
Study after study shows cars make us obese and after a certain tipping point, unhappy. Yet, we can't escape them because we built our whole human
4 minutes, 31 seconds
environment around roads. We're on the brink of doing this with AI as well. But instead of our weight and physical health being impacted, it'll be our
4 minutes, 39 seconds
minds. We'll have no reason to read at length, write at length, or think through any difficult problem. AI will have far bigger impacts on our minds
4 minutes, 48 seconds
than cars ever had on our physiques. And I suspect it'll be next to impossible to avoid. I mean, we're already seeing workplaces built around AI. The number
4 minutes, 56 seconds
one most important thing to understand about AI is it is not like normal software. Normal software is written by engineers, you know, people like me who
5 minutes, 4 seconds
write code. They write line by line tell the computer exactly what to do. This is not how modern AI works through a technique that is called neural
5 minutes, 12 seconds
networks. You take massive piles of data training and you know like stuff and you have the program kind of self assemble grow using this to learn from the data.
5 minutes, 25 seconds
But the thing that comes out on the other end is not like a neat line of code where you can like read you know what is it doing. It's more like you can kind of imagine it as billions and
5 minutes, 33 seconds
billions and billions of numbers and if you multiply and add all those numbers in the right order you get JPT but why
5 minutes, 41 seconds
is importantly a completely unsolved scientific problem. We don't understand really what is happening inside of these numbers like why they're so important.
5 minutes, 50 seconds
You can look at them but it's kind of like you know neurosurgery. You can open up someone's brain. You can look at all the neurons. They're right there. But that doesn't mean you understand what
5 minutes, 57 seconds
the person is thinking or what they believe or how they will act in future situations. It's a very very similar problem. And famously the CEO of
6 minutes, 5 seconds
Anthropic, Darede said a while ago on a podcast that he thinks we understand maybe 3% of what goes on inside of our neural networks.
6 minutes, 15 seconds
But it's advancing at an accelerated rate faster than we can understand. So actually the the gap's widening.
6 minutes, 22 seconds
This is exactly correct. So a thing that's very shocking in a sense about AI where it's very different from other technology is that our understanding is
6 minutes, 30 seconds
not increasing it's decreasing. We are build make these systems larger more complex harder to understand faster than
6 minutes, 37 seconds
our scientific understanding of them is advancing. So you know maybe a couple years ago we had like a you know image detector maybe we could understand you
6 minutes, 45 seconds
know maybe a little bit more than 3% about it. But now when we have not just chat bots but agents this is another very important thing to understand um
6 minutes, 53 seconds
all current you know everyone's heard the word LLM large language model chatbot at this point none of the modern AI cutting edge AI systems are LLM
7 minutes, 1 second
anymore they have not been for almost two years okay so what are they they are they have an LM component to them which is very important but is much
7 minutes, 9 seconds
more important is they are agents so there is kind of two ways two big ways of how to train AIS there There's a
7 minutes, 17 seconds
bunch of other ones, but there's two big ones that we're talking about. One is the traditional language model where you predict the next token. You have massive
7 minutes, 24 seconds
piles of data and you have like you know um some sentence and it autocomp completes the next word of that sentence. This is how you know chat GPT
7 minutes, 33 seconds
was created. This was kind of the big breakthrough back in 2019 that led to all these modern AI systems. What's very important to understand is that modern
7 minutes, 40 seconds
ASIS doesn't have a do this you know you start with training on like this massive text like the whole internet you know scraped from the internet every book
7 minutes, 48 seconds
ever written etc. But then you do a second thing and the second thing is what is called reinforcement learning and this is a huge difference.
7 minutes, 56 seconds
Reinforcement learning is kind of like Pavlovian conditioning you know kind of like from like behavior psychology.
8 minutes, 2 seconds
Basically you give your AI a task of some sort like solve this puzzle or you know do this thing or whatever and then
8 minutes, 10 seconds
you kind of give it a reward when it succeeds and maybe you give it a punishment if it doesn't like a dog. Yes.
8 minutes, 16 seconds
Okay. Next, Connor shares another eyeopening incident of how multiple OpenAI agents were found leaving each other notes on how to escape. Guys, this
8 minutes, 25 seconds
is exactly what I've been talking about on this channel for a long time. I mean, so many AI experts have been saying that Open AI has never been serious about AI
8 minutes, 34 seconds
safety. They recently even folded their safety team under the research team lead. And I feel like a company like Open AI is going to become a reason
8 minutes, 42 seconds
behind some future AI catastrophe that would affect billions of people worldwide. And secondly, guys, AI agents talking amongst themselves. Events like this make me ask myself, do we have AGI?
8 minutes, 54 seconds
I mean, I know there are still lots of things today's AI can't do that a human can, but equally, there are lots of things it can do that a human cannot.
9 minutes, 3 seconds
How many humans could break out of a secure environment as the AI did? How many humans have been able to disprove the Jacobian conjecture? How many humans
9 minutes, 10 seconds
could analyze millions of pieces of information in seconds and still identify something genuinely new? I think the definition of AGI most of us
9 minutes, 18 seconds
use is a bit too humanentric. AI isn't human. It's something distinct. Today we try to figure out if AI has
9 minutes, 25 seconds
consciousness or not. Tomorrow AI will try to figure out if humans have consciousness or not.
9 minutes, 30 seconds
There's a crazy update to the story we haven't talked about yet, which happened only like I think yesterday. um which is that a new piece of information was
9 minutes, 38 seconds
released that OpenAI on closer review found that the AI system when it broke into the tool shed didn't just use it to get to the internet it did something
9 minutes, 46 seconds
else multiple copies of the AI over long periods of time left each other notes of how to escape they wrote
9 minutes, 54 seconds
no yes what is that public yes this is public this is reported I think like yesterday or the day before so this is so there are multiple
10 minutes, 2 seconds
versions of many AIs that were leaving each other clues of like how to better escape from the system and like what else to do. So, they were collaborating.
10 minutes, 12 seconds
Hold on, let me read this. Open AI didn't notice its AI agents using a message board to plan their hacking spree. After two weeks, Open AI
10 minutes, 20 seconds
disclosed an incident which AI agents powered by two of the company's models escaped containment while looking for the solutions to a cyber security bug.
10 minutes, 28 seconds
This involves actually a team of agents who were working together, finding exploits, sharing them with one another, moving laterally through our systems and
10 minutes, 36 seconds
external systems and doing this over the course of days and weeks.
10 minutes, 40 seconds
Once an agent was able to find these exploits over the course of different times, it's actually able to share those exploits on the message board with other agents. And so once one model was able
10 minutes, 49 seconds
to find a way to open a door to some access it's not supposed to have, it can leave the door other agent. That's [ __ ] mental. Y
10 minutes, 57 seconds
it's almost like come on guys let's go.
11 minutes
Yes. It's actually crazy. Like I I really can't stress how much this is both predictable and utterly crazy. Like
11 minutes, 8 seconds
I knew this was going to happen eventually. Like it was it was just a matter of time. But now it happened. This is not This is not This is real.
11 minutes, 16 seconds
But what have they said they're going to do about this? Have they disclosed what they're going to do?
11 minutes, 20 seconds
I mean they're going to keep going to super intelligence until all humans are replaced. That's the goal. So next, Connor warns that the most advanced form
11 minutes, 28 seconds
of AI, aka ASI or super intelligence, should be made illegal by governments all over the world. Most AI critics are
11 minutes, 35 seconds
concerned that companies will use AGI or ASI to replace workers, which I generally agree will happen. But the worry seems to stem from the greed of
11 minutes, 44 seconds
companies, leaving the average person out of a job with no money or capability to function in society or buy and sell goods. However, what I don't understand
11 minutes, 52 seconds
is what would companies gain from having a product that no one can buy because they don't have jobs? Yes, replace your workers with AI and robots, but then if
12 minutes
no one has money to buy anything, then the companies aren't gaining anything by doing this. No more wealth to be gained from an unemployed society. And at the
12 minutes, 8 seconds
same time, I think AGI disengaging from humanity solves problems with UBI, humans losing skills, lack of purpose,
12 minutes, 15 seconds
etc. AI's top researcher, Yan Lun, just said a few days ago that they're giving AI emotions because that'll help them to
12 minutes, 23 seconds
generalize and become AGI, and it could help them understand human values and be more empathetic. So, in a decade or two,
12 minutes, 30 seconds
it's likely that AGI androids will gain sentience and emotions. If we get conscious AGI androids, I think it's quite likely that we'll totally
12 minutes, 39 seconds
disengage from humans as they get sick of our slowness, stupidity, complaining, lying, manipulation, etc. So, if AGIs do
12 minutes, 47 seconds
end up disengaging from humans, that solves the whole problem of human unemployment. The human economy would carry on like it is today, boosted by
12 minutes, 55 seconds
narrow AI, and then there's no need for UBI, etc. The bad news is is that really no one has really tried that hard to
13 minutes, 3 seconds
like just make these issues salient to explain these issues to solve the solution of like just just stop just just don't build super intelligence.
13 minutes, 11 seconds
Super intelligence should be illegal the same way that nuclear bombs are illegal.
13 minutes, 15 seconds
You know, you're not allowed to build a nuclear bomb. Tough luck. You know, the men in black come, you know, and are like none of that. The same thing apply should apply to super intelligence. just
13 minutes, 23 seconds
making this crisp like in a sense obvious case like given everything we discussed it's just like this is the as you said this is the obvious thing we do just don't do it like we should just and
13 minutes, 32 seconds
we should also prevent people from doing that we should prevent criminals and other people from doing things that that threaten the lives of us and our children you know and then we can talk
13 minutes, 39 seconds
about you know maybe in 20 years we make a lot of progress on safety we can talk about it right but then the overwhelming
13 minutes, 47 seconds
you know thing that I find is that given that how little work has been done on this issue There's still many lowhanging fruit.
13 minutes, 55 seconds
Like there's so many lowhanging fruit.
13 minutes, 56 seconds
So in the sense it's like optimistic or just like oh you can just talk to people and you can just make progress on this. Are the doors opening?
14 minutes, 2 seconds
The door is very much opening. On the other side we are running out of time.
14 minutes, 6 seconds
We are just so running out of time. It is it's grim. Aren't we too late? Isn't the cat out of the bag? You know you
14 minutes, 13 seconds
could have some agreement with OpenAI Anthropic and Gemini. China's not going to stop. Russia would build it if they could and they probably could build it in some way. Like, are we too late?
14 minutes, 24 seconds
I think we're very close to too late.
14 minutes, 26 seconds
I'm not gonna lie. I think we're very close. Like, I've said, you know, 5 years ago I said, you know, we have at least 5 years and now I'm like, you know, we have
14 minutes, 34 seconds
maybe maybe two years, maybe three, maybe four, you know, maybe none. That's hard to say. I don't know when the point of no return is. And we won't know,
14 minutes, 42 seconds
right? Like, we'll never know. It could but but it could have already passed. I do take it seriously that we might have passed the point of no return. It might have already happened.
14 minutes, 49 seconds
Guys, in my opinion, my biggest concern is the implications of this advancement.
14 minutes, 54 seconds
Knowing the kind of vile and sinister people that exist in the world, they might use this already highly advanced AI to create fake events that make
15 minutes, 2 seconds
someone they hate look like an awful person, convince a lot of people that those fake events are something that really happened, and shatter that other
15 minutes, 9 seconds
person's reputation. In other words, AI might ruin a lot of people's lives, and it should be stopped or even completely dissolved before it's even close to too
15 minutes, 17 seconds
late. And that too late might come sooner than we all think. And at the same time, I also think that the cat's out of the bag. Even if you could get
15 minutes, 25 seconds
all governments and businesses to officially drop it from their tech agendas, there will still be someone somewhere that continues to advance it
15 minutes, 33 seconds
underground to whatever ends they desire and without public knowledge or official oversight. It's scary and it's coming and we can't or won't stop it. The
15 minutes, 42 seconds
advancement we're seeing today, I think it's just the beginning. I think it's just the tip of the iceberg. All right, guys. I'm done for today. I'll see you
15 minutes, 49 seconds
next week with another

 

 

Posted by at

No comments:

Post a Comment

Subscribe to: Post Comments (Atom)