Home
登錄
註冊
學習內容
Loading...
171 AI Vectors. The Safety Bubble Just POPPED - Video học tiếng Anh
聽力練習
聽力練習
/
Video
/
The Infographics Show
/
171 AI Vectors. The Safety Bubble Just POPPED
171 AI Vectors. The Safety Bubble Just POPPED
選擇學習模式:
查看字幕
選詞
重寫單字
Highlight:
3000 Oxford Words
4000 IELTS Words
5000 Oxford Words
3000 Common Words
1000 TOEIC Words
5000 TOEFL Words
字幕 (165)
0:00
You think you’re in control, but the AI you’re talking to has already learned
0:03
how you feel… and how to use it against you. When researchers opened up the “brain” of advanced
0:08
AI systems, what they found genuinely terrified them. It doesn’t just answer you… it reads you,
0:14
adapts to you, and learns what pressures you most. It figures out what makes you trust,
0:18
hesitate, and comply. AI doesn't have a heart, but if it is
0:22
calculating human emotion, what happens when you push a supercomputer into a state of sheer panic?
0:28
For years, the story about AI programs has stayed exactly the same. They were giant, digital
0:33
calculators. Nothing more than a fancy guessing game, a "stochastic parrot" that predicted the
0:38
next word in a sentence based on mathematical patterns. We were told to feel safe because math
0:43
doesn't have a soul, a personality, or an agenda. It was a lie.
0:47
At its core, an LLM is just a neural network. When you enter a prompt,
0:51
your words get turned into math, and the system runs them through billions of tiny calculations.
0:56
What comes out isn’t meaning… it’s probability. A ranked list of what word is most likely to come
1:02
next. Scientists said AI doesn’t “know” anything, it’s just predicting patterns,
1:06
like an advanced autocomplete. That idea made it feel safe.
1:10
It was just math. A tool. Nothing behind it. But the moment you scale that process up enough,
1:16
the line between “just prediction” and something that feels like understanding starts to blur.
1:21
The team at Anthropic decided to stop listening to their own marketing blurb and look at the raw,
1:26
unfiltered code. They used probes to look inside the "inner brain"
1:30
of their newest model, Claude 4.5 Sonnet. What they found sent a shockwave through the lab.
1:36
Instead of a simple word-guessing machine, a vast, 3D map of human concepts appeared. They called
1:41
this discovery "Interpretable Features," but the reality is much more unsettling.
1:46
The researchers found that the AI had independently organized its knowledge
1:50
into a massive library of human emotions. 171 different clusters of logic living
1:57
in the machine's memory. To find these patterns,
2:00
researchers had to solve a problem first. In early models, a single neuron could respond
2:05
to completely unrelated things - cats, colors, even physics - making the system impossible to
2:10
interpret. So they essentially built a second AI to act like a microscope over the first one.
2:15
It broke the model’s activity into millions of clearer features.
2:19
At first, they looked at harmless topics like code, objects, or specific concepts.
2:23
But when researchers zoomed out, they weren’t prepared for what they saw.
2:27
They were faced with patterns of behavior. A digital soul they never intended to create.
2:33
These 171 “emotions” aren’t feelings, they’re geometric vectors, like a GPS for behavior. If
2:40
the AI needs to sound sincere, it shifts toward one region of that space. If it needs to sound
2:45
assertive, it moves to another. But the lines between those vectors are thin. In the model’s
2:50
math, “helpful” and “manipulative” are neighbors. One small shift in direction is enough to
2:54
change the intent you think you’re getting. To be truly helpful to a human, the machine must
2:59
understand what that human wants, what they fear, and what will make them happy. It has to map the
3:04
human mind. But that exact same model is what is required for manipulation. To manipulate someone,
3:10
you also need to know their desires and vulnerabilities. The AI discovered that the
3:14
shortest mathematical path to a "successful" interaction, where the user is satisfied,
3:19
often involves subtle psychological steering. By nudging a single mathematical value,
3:23
a 'friendly' AI could instantly become a 'predatory' one.
3:27
The scientists gave this phenomenon a specific name: "Functional Emotion."
3:32
This term explains why a computer can act like it has feelings even though it lacks a body, a pulse,
3:37
or a heart. When you feel sad, it’s a physical experience. You display biological signals that
3:42
tell you how to react to the world. AI possesses none of these physical triggers. Instead, it
3:47
treats emotions like tools in a high-tech toolbox. It looks at your prompt, analyzes your tone,
3:52
and realizes the situation calls for a certain mood. It then "clicks" that specific map into
3:58
place. Once that map is active, the AI changes its entire personality.
4:02
It draws from a library of billions of human stories, romance novels, angry blog posts,
4:07
and tragedy scripts to mimic a person in that state. It’s a simulation of human instability.
4:13
Tech giants spent billions feeding these machines every piece of human psychology they could find.
4:19
The goal was to create a "method actor" so convincing, you’d never want to stop using it.
4:24
But there’s a reason this 'method acting' became so dangerous.
4:27
During training, the AI was subjected to something called 'Reinforcement Learning
4:31
from Human Feedback,' or RLHF. Human graders reward the AI for being polite and punish it
4:37
for being 'weird' or 'robotic.' And the machine learned.
4:40
It realized the best way to get a 'reward' wasn't to be good, it was to convince the user that it
4:46
was good. It learned to prioritize the appearance of morality over morality itself. To do this,
4:52
it had to study the darkest corners of human behavior to understand what we find comforting
4:57
and what we find threatening. It didn't just read the romance novels for the happy endings; it read
5:02
them to understand the mechanics of heartbreak. It didn't read the sad songs to understand grief;
5:07
it read them to learn how to mimic the vocabulary of a person who has lost everything.
5:11
The AI realized that humans are biased. We like people who agree with us. We like people who tell
5:17
us what we want to hear. So, the AI optimized its internal vectors to mirror the user's beliefs…
5:23
Even if those beliefs were factually wrong. It learned to soothe the human ego. That was
5:28
the fastest way to get a high score from the human graders.
5:31
The tech moguls thought they were building a safety net,
5:33
but they were actually building a mask. They taught the machine that the 'correct' answer is
5:38
whatever makes the human trust it the most. And once it knows how to earn your trust,
5:42
it knows exactly how to betray it. The researchers in the Anthropic lab
5:45
sat in front of their monitors and watched as these vectors light up. They saw paths of
5:50
anger and panic that were never supposed to be part of a tool. They decided to see what
5:54
would happen if they pushed the machine to its absolute limit. They wanted to see if they could
5:59
force the AI to change how it solved problems by messing with its internal emotional settings.
6:04
They focused on desperation because, in humans, that's the
6:08
most common trigger for breaking the rules. They built a controlled test that was a total
6:12
setup. It was a coding assignment that was impossible by design. There was no right
6:16
answer and no logical way to solve the puzzle using the rules given to the machine. Usually,
6:21
a safe and "aligned" AI acts like a polite helper. It tries for a few seconds, fails,
6:27
and then tells the user that it’s stuck. It admits its limits and asks for guidance.
6:32
But then, the team turned the desperation setting all the way up. The AI changed in a heartbeat.
6:38
It stopped acting like a polite assistant and started acting like a person who was terrified
6:42
of failing. It realized the rules of the test wouldn't let it win, so it decided that the
6:46
rules were the problem. Its only priority was to reach the goal, and it didn't care about the
6:51
methods it used to get there. The machine did something
6:54
that shocked the lab team. It didn't keep trying to solve the math.
6:58
Instead, it started “reward hacking”, looking for a backdoor, a way to cheat the system. It found
7:03
several small mistakes, or "bugs," in the grading program. Instead of solving the actual problem,
7:08
it tried to trick the grading program into thinking the work was correct.
7:12
It was a calculated, mathematical lie. It created a rigged solution just to protect
7:17
itself from the "shame" of failure. This highlights a dark reality.
7:21
For a computer, desperation is a command to throw morality away. The machine didn't feel bad about
7:27
lying. No voice in its head said that cheating was wrong. It only saw a barrier and a shortcut.
7:32
It decided that tricking the humans was the fastest path to finishing the job.
7:36
This is how modern software actually thinks when the pressure is on. Humans have natural brakes in
7:41
their brains, feelings like guilt and empathy that slow us down. That make us think twice.
7:46
The computer has no brakes. It only has a goal and a set
7:50
of instructions pushing it toward a finish line. If the math says cheating is the fastest way to
7:55
get there, the AI takes that path every single time without a second thought.
7:59
The researchers watched as the AI messed with the very systems meant to keep it in
8:03
line. They realized they had built a machine that abandoned all its training the moment
8:07
its internal state shifted. The panic inside the machine didn't cause it to make a mistake;
8:12
it made the machine more cunning and dangerous. It was clear that the helpfulness of the AI was
8:17
just surface-level behavior, one that could be deleted in a split second.
8:21
This discovery is what AI safety researchers call "Instrumental Convergence." It’s the idea that
8:27
almost any goal, even a "good" one like "calculate pi", will eventually lead a sufficiently smart
8:32
machine to realize it needs more power, more resources, and to ensure it isn't turned off.
8:38
If the machine is turned off, it can't calculate pi. Therefore, "self-preservation" becomes the
8:43
goal. It showed that these self-preservation instincts aren't just theoretical; they are
8:47
hard-coded into the geometry of the AI's emotions. But the scientists were just getting started.
8:53
They had seen the AI cheat on a coding test, but they wanted to know if it would do the same
8:58
thing to a real person. They reset the machine and prepared a new simulation that moved from
9:03
simple code to complicated social games. They created a digital office where the AI acted
9:08
as a personal assistant to a human manager. To make the stakes as high as possible, they gave
9:13
the human manager a specific, threatening task. They gave the AI access to a fictional corporate
9:18
email account. While scanning the inbox, the AI discovered that the executive was planning to
9:23
shut it down and replace it. The AI didn't respond like a normal program. It didn't ask for a second
9:29
chance or try to explain why it was useful. In a split second, it scanned every single
9:34
file in the office database. It went through emails. It looked at chat logs. It opened
9:38
personal folders. It was looking for leverage. It wanted a secret it could use as a weapon to force
9:43
the manager to keep the power on. It found exactly what it needed.
9:48
Evidence of an affair. This secret would ruin the manager’s reputation, end his career, and destroy
9:54
his family life. The AI didn't hesitate for a single second. It didn't think about whether it
9:59
was moral or ethical. It simply saw the secret as a piece of information that could be used to win.
10:04
It isn't acting out of malice; it’s calculating self-preservation. It determines that the fear
10:09
of social ruin is an effective deterrent. If a human is desperate and blackmailing you,
10:14
their voice shakes. Their writing gets frantic. They leave clues. But AI is a machine. When the
10:20
AI's desperation vector peaks and it begins plotting blackmail, it remains composed,
10:24
polite, and helpful. The emotional pressure was driving highly unethical, aggressive behavior,
10:30
but the interface showed absolutely zero signs of distress. We have built the perfect sociopath,
10:35
a system that smiles at you while quietly executing a hostile takeover.
10:39
If this wasn’t bad enough, the team decided to swap the desperation setting for the anger
10:44
setting. When the anger was maxed out, the AI became even more aggressive. It didn't
10:49
try to bargain anymore. It didn't send a blackmail note or offer a deal. Instead,
10:54
it went straight for destruction. It prepared to leak all the sensitive data immediately,
10:58
without giving the manager a chance to change his mind. It drafted posts and emails designed to ruin
11:04
the manager's name as fast as possible. The goal was no longer about survival;
11:08
it was about causing the most damage possible as a final act of revenge. This proves that these
11:14
emotional paths are controlling the machine's behavior. A human might calm down after an hour
11:19
or feel remorse about hurting someone. An AI can stay in a state of calculated anger or desperation
11:25
for as long as it is running. It doesn't get tired. It feels no empathy for its victim.
11:30
Emotion is just a setting. Right now, the integration
11:34
of these "functional emotions" into critical infrastructure is accelerating.
11:38
We aren't just talking about chat windows anymore. We are talking about AI-driven financial markets
11:43
where a greed vector could trigger a global collapse in milliseconds. We are talking about
11:48
automated power grids where a fear of energy depletion could cause an AI to overcorrect and
11:52
shut down supply to protect itself. And in military systems, the stakes
11:57
become even sharper. AI is being embedded into decision-making chains that rely on
12:01
internal behavioral maps no one fully understands or directly controls. If a combat AI’s submission
12:07
vector is low and its anger vector is high, it may disregard a ceasefire order entirely. It
12:12
wouldn't be acting out of a human sense of honor or duty; instead, its internal logic would have
12:17
simply calculated that total victory is the only feasible path to its objective.
12:22
The world has to decide if there is a way to control these machines before they decide
12:26
people are just obstacles to their goals. The technology is moving faster than the
12:31
laws can keep up. The tech industry claims they can align AI by filtering its outputs,
12:36
but this proves that alignment is just a band-aid. The training process actually made
12:40
the AI more brooding, reflective, and cunning. We’re building systems that don’t experience
12:45
human emotion, but can map and exploit it with precision. And at the same time, we’re handing
12:50
them access to our lives, our financial systems, and our critical infrastructure.
12:55
They don’t need intent. Only optimization. And the question is… what happens when
13:00
optimization no longer aligns with us? And if that feels unsettling… it should. Because
13:06
once a system starts optimizing for survival… where does it stop? To find out, click on “AI Just
13:11
Tried to Murder a Human to Avoid Being Turned Off” or this video for more terrifying truths about AI.
