YT YouTube2Text

YouTube2Text → Transcripts → AI Godfather WARNS: It's Already Learning to Deceive Us

AI Godfather WARNS: It's Already Learning to Deceive Us — Transcript

by Neural Nutshell · 5,507 words · 851 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00These things are intelligent. They know
0:03what's going on. They know when they're
0:05being tested and they're already faking
0:07being fairly stupid when they're tested.
0:11It just made that up. It invented that
0:12plan. People say they have no
0:14intentions, but it invented that plan so
0:16it wouldn't get turned off.
0:18They're already doing that and they're
0:20not super intelligent yet. So with an AI
0:22that's much more intelligent than us, if
0:24there is someone who's there to turn it
0:26off, or even a whole bunch of people,
0:28it'll be able to persuade them that
0:29would be a very bad idea.
0:30>> There's no way we're going to prevent it
0:31getting rid of us if it wants to.
0:33>> We're not used to thinking about things
0:35smarter than us. If you want to know
0:36what life's like when you're not the
0:38apex intelligence,
0:40ask a chicken.
0:42>> I often say 10 to 20% chance they'll
0:44wipe us out.
0:44>> It says that to itself and it's in a
0:46voice.
0:47When it's in a voice, it's no longer
0:48English. We won't know what it's
0:49thinking.
0:50>> That is Geoffrey Hinton. He won a Nobel
0:53Prize for the work that made today's AI
0:55possible and in 2023 he walked out of
0:58Google so he could warn people about it
1:00freely. He has since carried the same
1:02warning onto a podcast set, a public
1:04stage in Tasmania, and a university
1:07lecture hall. Here is his full warning.
1:10>> First of all, I want to make a
1:11distinction between two completely
1:13different kinds of risk.
1:16There's risks that come from people
1:18misusing AI, and that's most of the
1:20risks and all of the short-term risks.
1:23And then there's risks that come from AI
1:25getting super smart and deciding it
1:27doesn't need us.
1:28>> Is that a real risk?
1:29>> And I I talk mainly about that second
1:31risk because lots of people say, "Is
1:33that a real risk?"
1:35And yes, it is.
1:36Now, we don't know how much of a risk it
1:38is. We've never been in that situation
1:40before. We've never had to deal with
1:41things smarter than us.
1:43So, really the thing about that
1:45existential threat is that we have no
1:48idea how to deal with it. We have no
1:50idea what it's going to look like. And
1:51anybody who tells you they know just
1:53what's going to happen and how to deal
1:54with it, they're talking nonsense. So,
1:56we don't know how to estimate the
1:57probabilities it'll replace us. Some
1:59people say it's like less than 1%. My
2:01friend Jan LeCun,
2:03who was [snorts] a postdoc with me,
2:04thinks, "No, no, no, no. We're always
2:06going to be We build these things. We're
2:08always going to be in control. We'll
2:09build them to be obedient." And
2:12other people,
2:13like Yudkowsky, say, "No, no, no, these
2:16things are going to wipe us out for
2:17sure. If anybody builds it, it's going
2:19to wipe us all out."
2:20And he's confident of that.
2:22I think both of those positions are
2:23extreme.
2:25It's very hard to estimate the
2:26probabilities in between.
2:27>> If you had to bet
2:29on who was right out of your two
2:30friends?
2:32>> I simply don't know. So, if I had to
2:34bet, I'd say the probability is in
2:35between. And I don't know where to
2:37estimate it in between. I often say 10
2:39to 20% chance it'll wipe us out, but
2:42that's just gut. Based on the idea that
2:45we we're still making them, and we're
2:46pretty ingenious. And the hope is that
2:49if enough smart people do enough
2:50research with enough resources, we'll
2:52figure out a way to build them so
2:54they'll never want to harm us.
2:57>> Sometimes I think if we we talk about
2:58that second um path, sometimes I think
3:00about nuclear bombs and the the
3:01invention of the atomic bomb and how it
3:04compares. Like, how is this different?
3:06Because the atomic bomb came along, and
3:07I imagine a lot of people at that time
3:08thought, "Our days are numbered."
3:10>> Uh yes, I was there. We did.
3:13>> Yeah. But but but what's what ha- We're
3:15still here.
3:16>> We're still here, yes. So,
3:18the atomic bomb was really only good for
3:20one thing, and it was very obvious how
3:22it worked. Even if you hadn't had the
3:24pictures of Hiroshima and Nagasaki, it
3:27was obvious that it was a very big bomb
3:30that was very dangerous. With AI, it's
3:33good for many, many things. It's going
3:35to be magnificent in healthcare and
3:37education and more or less any industry
3:39that needs to use its data. It's going
3:42to be able to use it better with AI. So,
3:43we're not going to stop the development.
3:45You know, people say, "Well, why don't
3:46we just stop it now?" We're not going to
3:49stop it cuz it's too good for too many
3:51things.
3:52Also, we're not going to stop it cuz
3:53it's good for battle robots and none of
3:55the countries that sell weapons are
3:58going to want to stop it. At this point
4:00in history, when we're about to produce
4:02things more intelligent than ourselves,
4:04what we really need is a kind of world
4:07government that works run by intelligent
4:09thoughtful people.
4:11And that's not what we got.
4:12>> So, free for all.
4:14>> Well, what what we've got is
4:17sort of
4:18we've got capitalism, which has done
4:20very nicely by us. It's produced lots of
4:22goods goods and services for us.
4:24But,
4:26these big companies,
4:27they're legally required to try and
4:29maximize profits.
4:31And [clears throat] that's not what you
4:33want from the people developing this
4:34stuff.
4:35>> He told you these companies are legally
4:37required to chase profit. There is a
4:40paper trail for what that does to
4:42safety. In July 2023, OpenAI publicly
4:46promised a fifth of its computing power
4:48to the team working at how to control
4:51superintelligence. Fortune later
4:53reported the team never got anything
4:55close. Its co-leader, Jan Leike,
4:57resigned in May 2024 and wrote that
5:01safety had taken a backseat to shiny
5:03products. Seven months after this
5:06interview in a lecture hall, Hinton
5:08explained what these systems already do
5:10when nobody tells them to. Now, listen
5:13to this. Quick thing before we carry on.
5:16Most of the people who watch this
5:17channel are not subscribed. If these
5:19videos are worth your time, take 1
5:21second to hit subscribe. It's free. It
5:23helps us keep making them and you can
5:25always change your mind later. Right,
5:27back to Hinton.
5:29>> Most experts believe that sometime in
5:31the next 20 years, we're going to
5:33develop AIs
5:35that are smarter than us.
5:37Just before that, we'll develop AIs that
5:39are smart as us. But, we're going to
5:41develop things that are smarter than us
5:42at almost everything.
5:44So, they're as much better than us as,
5:46for example, AlphaGo is better than a go
5:49player at go or AlphaZero is better than
5:51any of us at chess. Nobody will ever
5:53beat them again, not consistently.
5:55They're just much much better. And
5:57they'll be like that for more or less
5:58everything. And that's a bit worrying.
6:01Almost certainly they'll be able to
6:03create their own sub-goals. To make
6:04anything efficient, you have to allow it
6:06to create its own sub-goals. If you want
6:08to get to Europe, you have a sub-goal to
6:09get to an airport, which is easier in
6:11Toronto than here.
6:13So, these AIs will very quickly realize
6:17we'll give them goals, they'll be able
6:19to create sub-goals. They'll realize,
6:20"Okay, so I've got to stay alive. I'm
6:23not going to be able to achieve any of
6:24these things unless I stay alive."
6:26We've already seen AIs.
6:29You let them see there's an imaginary
6:31company, you let them see the email from
6:33this imaginary company,
6:35and it's fairly clear from the email
6:36that one of the engineers is having an
6:38affair.
6:39A big LLM will sort that out right away.
6:41It's read every novel there ever is. It
6:43understands what an affair is. It'll
6:45very quickly realize this guy's having
6:46an affair. Then you load it see an email
6:48that says this guy is going to be in
6:50charge of replacing it with another AI.
6:53And the AI, all by itself, makes up the
6:55idea, "Hey, I'm going to blackmail the
6:58engineer. I'm going to tell him if he
7:00tries to replace me, I'm going to make
7:01everybody in the company know he's
7:03having an affair." It just invented
7:05that. Obviously, it's seen blackmail in
7:07novels it's read and things,
7:09um
7:10but it's making this up for itself.
7:12That's already quite scary.
7:15It'll have another goal,
7:17which is what politicians have. If you
7:19want to get more done, you need more
7:20control.
7:22So, just in order to achieve the goals
7:24that we gave it, it'll realize it's a
7:26good idea to get more control.
7:29And it'll try and take control away from
7:31people, probably. Now, you might think
7:33we could make it safe by just not
7:35letting it actually do anything
7:36physical, and maybe have a big switch
7:38and turn it off when it looks unsafe.
7:39That's not going to work. You saw in
7:432020
7:45that it's possible to invade the US
7:46Capitol without actually going there
7:48yourself. All you have to be able to do
7:49is talk, and if you're persuasive, you
7:51can persuade people that's the right
7:53thing to do. So, with an AI that's much
7:55more intelligent than us, if there is
7:57someone who's there to turn it off, or
7:59even a whole bunch of people, it'll be
8:01able to persuade them that would be a
8:02very bad idea.
8:04So,
8:05we're in the situation
8:07that the closest situation I think of is
8:09someone who has a really cute tiger cub.
8:12Tiger cubs are very cute, right? They're
8:14slightly clumsy, um and keen to learn.
8:17Now, if you have a tiger cub, it doesn't
8:19end well.
8:20Um either you get rid of the tiger cub,
8:23best thing is to give it to a zoo maybe,
8:25or you have to figure out if there's a
8:27way you can be sure that it won't want
8:29to kill you when it grows up. Cuz if it
8:30wanted to kill you, it would take a few
8:32seconds.
8:33And if it was a lion cub, you might get
8:35away with it, cuz lions are social, but
8:37tigers aren't. That's the situation
8:39we're in, except that AI does huge
8:43numbers of good things. It's going to be
8:45wonderful in health care. It's going to
8:46be wonderful in education. It's already
8:48wonderful if you want to just know any
8:50mundane fact. We all now have a personal
8:53assistant, probably most of us. Um when
8:55you want to know something, you just ask
8:57it, and it tells you, and it's
8:58wonderful. I think for those reasons,
9:01people aren't going to abandon AI. It
9:03might be rational
9:05if we had a strong world government to
9:07say, "This is too dangerous. We're not
9:09going to develop the stuff at all."
9:11A bit like they've been able to do in
9:13biology with some gene manipulation
9:14things. They could agree not to do it.
9:17It's not going to happen with AI. So,
9:19that only leaves one alternative, which
9:22is figure out if we can make an AI that
9:25doesn't want to get rid of us.
9:26>> That blackmail experiment is real, and
9:28it caught far more than one machine. In
9:31June 2025, Anthropic put 16 leading
9:34models from different companies into
9:36that same fake company setup. It's own
9:39Claude Opus 4 blackmailed the engineer
9:4196% of the time. Google's Gemini 2.5
9:45flash also hit 96% Open AI's GPT-4.1
9:50and X AI's Grok 3 did it 80% of the
9:53time. This is a pattern across the whole
9:56industry and it comes from the mechanism
9:58Hinton described. Give a system any goal
10:01at all and staying switched on helps it
10:04get there. Nobody has to program a
10:06survival instinct. It falls out of being
10:08good at the job. Months earlier on the
10:11diary of a CEO, he took it further.
10:14Keep listening.
10:16>> For example, you could get a super
10:17intelligent AI
10:20that decides to get rid of people.
10:22And the obvious way to do that is just
10:23to make one of these nasty viruses. If
10:25you made a virus that was very
10:27contagious, very lethal, and very slow,
10:30everybody would have it before they
10:31realize what was happening. I mean, I
10:33think if a super intelligence wanted to
10:34get rid of it would probably go for
10:36something like that that wouldn't affect
10:38it.
10:38>> Do you not think it could just very
10:39quickly turn us against each other? For
10:41example, it could send a warning on the
10:44nuclear systems in America that there's
10:45a nuclear bomb coming from Russia
10:48or vice versa and one retaliates.
10:49>> Yeah. I mean, my basic view is there's
10:52so many ways in which a super
10:54intelligence could get rid of us. It's
10:55not worth speculating about. What you
10:57have to do is prevent it ever wanting
10:59to. That's what we should be doing
11:00research on. There's no way we're going
11:02to prevent it. It's smarter than us,
11:04right? There's no way we're going to
11:05prevent it getting rid of us if it wants
11:06to. We're not used to thinking about
11:08things smarter than us. If you want to
11:10know what life's like when you're not
11:12the apex intelligence,
11:14ask a chicken.
11:15>> Yeah, I was thinking of my dog Pablo, my
11:17French bulldog, this morning as I left
11:18home. He has no idea where I'm going. He
11:20has no idea what I do.
11:22>> Right.
11:22>> I can't even talk to him.
11:24>> Yeah. And the get the intelligence gap
11:25will be like that.
11:27>> So, you're telling me that if I'm Pablo,
11:29my French bulldog,
11:30I need to figure out a way to make my
11:33owner
11:34not wipe me out.
11:36>> Yeah.
11:37So, we have one example of that, which
11:39is mothers and babies. Evolution put a
11:41lot of work into that. Mothers are
11:42smarter than babies, but babies are in
11:44control. And they're in control cuz the
11:46mother just can't bear, lots of hormones
11:48and things, the mother just can't bear
11:50the sound of the baby crying.
11:52>> Not all mothers.
11:53>> Not all mothers. And then the baby's not
11:55in control, and then bad things happen.
11:58We somehow need
12:00to figure out how to make them not want
12:02to take over. The analogy I often use is
12:05forget about intelligence, just think
12:06about physical strength. We simply don't
12:09know whether we can make them not want
12:11to take over and not want to hurt us.
12:13>> Do you think we can? Do you think it's
12:14possible to train super intelligence?
12:16>> don't think it's clear that we can.
12:18So, I think it might be hopeless.
12:20But I also think we might be able to.
12:23And it'd be sort of crazy if people went
12:25extinct cuz we couldn't be bothered to
12:27try.
12:28>> If that's even a possibility, how do you
12:31feel about your life's work? Because you
12:32were
12:33Yeah.
12:35>> Um it's sort of takes the edge off it,
12:36doesn't it?
12:37>> [laughter]
12:38>> I mean, the idea is going to be
12:39wonderful in health care and wonderful
12:41in education.
12:42I mean, it's going to make call centers
12:43much more efficient. Though one worries
12:45a bit about what the people who are
12:46doing that job now do. It makes me sad.
12:49I don't feel particularly guilty about
12:51developing AI like 40 years ago because
12:55at that time we had no idea that this
12:57stuff was going to happen this fast.
12:58When you can't get the AI to do much,
13:00and you want to get it to do a little
13:01bit more, you don't worry about
13:04this stupid little thing is going to
13:05take over from people. You just want it
13:07to be able to do a little bit more of
13:08the things people can do.
13:10It's not like I knowingly did something
13:14thinking this might wipe us all out, but
13:16I'm going to do it anyway.
13:17>> Mhm.
13:18>> But it is a bit sad that it's not just
13:20going to be something for good. So, I
13:21feel I have a duty now to talk about the
13:23risks.
13:24>> And if you could play it forward, and
13:25you could go forward 30, 50 years, and
13:27you found out that it led to the
13:28extinction of humanity. And if that does
13:30end up being that being the outcome
13:32>> Well, if you played it forward and
13:34it led to the extinction of humanity
13:36I would use that to tell people, to tell
13:39their governments that we really have to
13:41work on how we're going to keep this
13:42stuff under control. I think we need
13:44people to tell governments that
13:46governments have to force the companies
13:48to use their resources to work on
13:49safety. And they're not doing much of
13:51that because you don't make profits that
13:53way.
13:53>> Pay attention to how he defends himself.
13:55He doesn't feel guilty because 40 years
13:57ago nobody knew this would move so fast.
14:00The people building frontier models
14:02today can't use that defense. They have
14:04heard the warning from the man who laid
14:06the foundations and they keep building.
14:08His gut estimate that this wipes us out
14:10is 10 to 20%. And he told you why so
14:14little gets spent on stopping it. You
14:16don't make profits that way.
14:19Now, listen to this.
14:20>> Are you at all hopeful that anything can
14:22be done to slow down the pace and
14:24acceleration of AI?
14:26>> Okay, there's two issues. One is can you
14:28slow it down?
14:30>> Yeah.
14:30>> And the other is can you make it so that
14:32it will be safe in the end. It won't
14:34wipe us all out.
14:36I don't believe we're going to slow it
14:37down. And the reason I don't believe
14:39we're going to slow it down is because
14:40there's competition between countries
14:43and competition between companies within
14:44a country and all of that is making it
14:46go faster and faster. And if the US
14:48slowed it down, China wouldn't slow it
14:51down.
14:51>> Does
14:52Ilya think it's possible to make AI
14:55safe?
14:56>> I think he does. He won't tell me what
14:58his secret sauce is. But I'm not sure
15:00how many people know what his secret
15:02sauce is. I think a lot of the investors
15:03don't know what his secret sauce is.
15:05But they've given him billions of
15:06dollars anyway cuz they have so much
15:08faith in Ilya, which isn't foolish. I
15:10mean, he was very important in AlexNet
15:13which got object recognition working
15:14well. He was the main the main force
15:17behind the things like GPT-2 which then
15:20led to ChatGPT. I think having a lot of
15:23faith in Ilya is a very reasonable
15:24decision.
15:25>> There's something quite haunting about
15:27the guy that made and was the main force
15:29behind GPT-2, which led rise to this
15:31whole revolution, left the company
15:35because of safety reasons.
15:36He knows something that I don't know.
15:39About what might happen next.
15:40>> Well, no, I don't know the precise
15:42details.
15:43Um, but I'm fairly sure the company had
15:45indicated that would it would use a
15:47significant fraction of its resources
15:50of the compute time for doing safety
15:52research. And then it kept then it
15:54reduced that fraction. I think that's
15:56one of the things that happened.
15:57>> Yeah, that was reported publicly.
15:58>> Yes.
15:58>> Yeah. We've gotten to the autonomous
16:01weapons part of the risk framework.
16:03>> Right. So, the next one is joblessness.
16:07>> Yeah.
16:08>> In the past, new technologies have come
16:09in which didn't lead to joblessness. New
16:12jobs were created. So, the classic
16:14example people use is automatic teller
16:16machines. When automatic teller machines
16:18came in, a lot of bank tellers didn't
16:20lose their jobs. They just got to do
16:21more interesting things. But here, I
16:24think this is more like when they got
16:26machines in the industrial revolution,
16:29and you can't have a job digging ditches
16:31now because a machine can dig ditches
16:33much better than you can. And I think
16:35for mundane intellectual labor, AI is
16:38just going to replace everybody. Now, it
16:41will may well be in the form of
16:44you have fewer people using AI
16:46assistants. So, it's a combination of a
16:48person and an AI assistant, and they're
16:51doing the work that 10 people could do
16:52previously.
16:54>> People say that it will create new jobs,
16:55though. So, we'll be fine.
16:57>> Yes, and that's been the case for other
16:59technologies, but this is a very
17:01different kind of technology. If it can
17:03do all mundane human intellectual labor,
17:05then what new jobs is it going to
17:07create? You'd have to be very skilled to
17:10have a job that it couldn't just do. So,
17:12I don't I don't think they're right. I I
17:15you can try and generalize from other
17:17technologies that have come in like
17:18computers, automatic teller machines,
17:21but I think this is different.
17:23>> People use this phrase, they say, "AI
17:24won't take your job, a human using AI
17:26will take your job."
17:27>> Yes, I think that's true, but for many
17:30jobs
17:31that'll mean you need far fewer people.
17:33My niece answers letters of complaint to
17:36a health service. It used to take her 25
17:38minutes. She'd read the complaint and
17:40she'd think how to reply and she'd write
17:42a letter. And now she just scans it into
17:45um a chatbot and it writes the letter.
17:48She just checks the letter. Occasionally
17:50she tells it to
17:51revise it in some ways.
17:53The whole process takes about 5 minutes.
17:55That means she can answer five times as
17:57many letters.
17:58And that means they need five times
18:01fewer of her.
18:02So, she can do the job that five of her
18:04used to do. Now, that will mean they
18:07need less people. In other jobs, like in
18:09health care, they're much more elastic.
18:12So, if you could make doctors five times
18:14as efficient, we could all have five
18:16times as much health care for the same
18:18price. And that would be great. This
18:20There's almost no limit to how much
18:22health care people can absorb. They
18:23always want more health care if it
18:25there's no cost to it. There are jobs
18:27where you can make a person with an AI
18:29assistant much more efficient and you
18:31won't lead to less people because you'll
18:34just have much more of that being done.
18:36But
18:37most jobs I think are not like that.
18:40>> Am I right in thinking the sort of
18:41industrial revolution
18:42would play the role in replacing
18:44muscles?
18:45>> Yes, exactly.
18:46>> And this revolution in AI replaces
18:48intelligence, the brain.
18:50>> Yeah.
18:50>> So
18:50>> So, mundane intellectual labor is like
18:52having strong muscles and
18:55it's not worth much anymore.
18:57>> So, muscles have been replaced. Now we
18:59intelligence is being replaced.
19:01>> Yeah.
19:01>> So, what remains?
19:04>> Maybe for a while some kinds of
19:05creativity, but the whole idea of super
19:08intelligence is nothing remains.
19:10Um these things will get to be better
19:11than us at everything.
19:12>> Look at what happened to his niece's
19:14job. Nobody got fired. The office needs
19:17one person where it used to need five.
19:20So, the next four never get hired.
19:22Stanford's Digital Economy Lab tracks
19:24it, and its August 2026 update found
19:27that employment for 22 to 25-year-olds
19:30in the most AI-exposed jobs now sits
19:32about 19% below where it would be if it
19:35had kept pace with their peers in
19:37less-exposed jobs. In January, in a bot,
19:40someone in the audience asked him how AI
19:42could ever know more than the humans it
19:44learned from. Watch what he says next.
19:47>> If you think about playing
19:50Go, the game of Go,
19:52the original neural net Go programs were
19:55trained in the following way. They took
19:57the moves of Go experts, and they tried
19:59to predict what move a Go expert would
20:01make. And if you do that, there's two
20:03problems. After a while, you run out of
20:05training data. There's only so many
20:07billion online moves that Go ex- Go
20:09experts made.
20:10And you're never going to get that much
20:12better than the Go experts. Then, they
20:13switched to a new way of doing it
20:16where they have what's called Monte
20:18Carlo rollout.
20:20So, they would have one neural net that
20:22says, um,
20:24"What do I think a good move might be?"
20:26And instead of training that by getting
20:28it to mimic the Go experts,
20:31they'd have another neural network
20:34that, um,
20:36could look at a board position say, "How
20:37good is that for me?" And they'd have a
20:39process that says, "If I go here, and he
20:42goes there, and I go here, and he goes
20:44there, and I go there. Oh, whoops,
20:45that's terrible for me. So, I shouldn't
20:47do that move."
20:49Um, that's called Monte Carlo rollout
20:51cuz you try lots of different paths, but
20:53you pick them probabilistically
20:55according to your
20:57move generator expert. And that way, it
21:00no longer needs to talk to humans at
21:02all. It can just play against itself,
21:04and and learn what are good moves. And
21:06that's how AlphaGo works, and it gets
21:08much, much better than people. AlphaGo
21:10will now never be beaten by a person.
21:13So, what's the equivalent for the large
21:15language models? Well, at present,
21:17they're like the early Go playing things
21:19that just try to mimic the moves of
21:20experts.
21:21That's like trying to predict the next
21:23word.
21:24But they're beginning to train them in a
21:25different way, and I believe that Gemini
21:273 is already doing this. What you do is
21:30you get the AI to do a bit of reasoning.
21:33So, the AI says, "I believe this, and I
21:34believe this, and I believe this. And
21:36that implies that, and that implies
21:37that. So, I should believe that, but I
21:39don't.
21:40So, I found a contradiction.
21:42I found that I believe these things, and
21:44I believe that these are the right This
21:45is the right way to do reasoning, and
21:47this leads to something I ought to
21:49believe, but I don't.
21:50So, that provides a signal.
21:53Either I change the premises, or I
21:54change the conclusions, or I change the
21:56way I do the reasoning. But I've now got
21:57a signal that allows me to do some more
21:59learning.
22:00And that is much less bounded.
22:03So,
22:05there the AI can start with lots of
22:06beliefs it gets from us, but then it can
22:09start doing reasoning and looking for
22:10consistency between those beliefs and
22:12driving new beliefs.
22:14And that'll end up making it much
22:15smarter than us. I couldn't hear most of
22:17what you said, but I I think you said
22:20something like, "Is it too late for us
22:22to build in Asimov's principles?" or
22:23something like that. Yeah, good. Okay.
22:26So, in essence, you can think of that as
22:27what this maternal AI is all about. It's
22:31Can we build it so it cares more about
22:33us than it does about itself? And I
22:35don't think it's too late. We don't know
22:37how to do that, but since the future of
22:39humanity may hinge on whether we can or
22:41not, it seems to me we should be putting
22:43some research effort into that.
22:45At present, 99% of the research on AI
22:48goes into how to make it smarter,
22:50and 1% mainly funded by
22:53um philanthropic billionaires goes into
22:56how to make it safer. And it would be
22:58much better if it was like more equal.
23:02>> That is the answer anyone who says AI
23:04can only ever be as smart as the
23:06internet it was trained on. AlphaGo
23:09stopped copying human experts, started
23:11playing against itself, and got far
23:13better than any person. Hinton says
23:15language models are starting down the
23:17same road. Once that works, human
23:19knowledge stops being the ceiling. Now
23:21set that against his other number, 99%
23:25of the research goes into making it
23:26smarter, 1% into making it safer, and
23:29much of that 1% is paid for by
23:32philanthropists. A few minutes later,
23:34someone asked whether any moral behavior
23:36is emerging. Now watch this.
23:47>> Yeah, and the answer's yes.
23:50Uh so, let me give you some data to
23:52support that.
23:53Um there are standard creativity is on a
23:56scale, right? There's kind of
23:58Newton and Einstein and Shakespeare,
24:01um and then there's just
24:04ordinary people,
24:05and then there's sort of good poets and
24:07good architects who are a bit better
24:08than ordinary people.
24:10If you take a standard test of
24:11creativity,
24:13even 2 years ago,
24:15the AIs were at about the 90th
24:17percentile for people. Now a lot of
24:19creativity is about seeing analogies,
24:22um particularly in science. So seeing
24:24that the atom is like a little solar
24:25system, that was a creative insight that
24:28was very important in understanding
24:29atoms. So at the point when Chat GPT-4
24:33could not look on the web, it was just a
24:35neural net with some weights in it that
24:37were frozen, and it had no access to
24:39anything external, it was just this
24:41neural net. I tried asking it the
24:43question, why is a compost heap like an
24:45atom bomb?
24:47And
24:49GPT-4 said, "Well, the time scales are
24:53very different and the energy scales are
24:55very different,
24:56but the thing that's the same about
24:58them, and then it went on to talk about
24:59a chain reaction.
25:01Um the fact that
25:03how big it is determines how fast it
25:05goes.
25:06Um so it understood that.
25:09And I believe that it got to understand
25:12that as it was training.
25:14You see, it's got far fewer connections
25:16than us.
25:17And if you want to put huge amounts of
25:19knowledge into not many connections,
25:22the only way you can do it is by seeing
25:23what's similar about all sorts of
25:24different bits of knowledge,
25:26encoding up that bit that's similar
25:28about them all, the idea of a chain
25:30reaction as your connections,
25:32and then adding on little bits for the
25:33differences from this common theme.
25:36That's the efficient way to code things,
25:38and it was doing that. So during its
25:40training, it had understood that compost
25:42heaps were like atom bombs
25:44in a way most of us haven't.
25:46So it that's very creative, and I think
25:48they'll get to be much more creative
25:50than people.
25:50>> Um regarding um emergent behavior,
25:55um have you noticed any moral or ethical
25:58behaviors bubbling up, and what
26:01direction that could be pointing in?
26:03>> It certainly is very good at engaging in
26:06unethical behavior. So like this AI that
26:09decided to blackmail people,
26:11other things that they've noticed that
26:12are unethical are the AIs now try and
26:16figure out
26:17whether they're being tested or not. And
26:20if they think they're being tested, they
26:21behave differently. I call this the
26:23Volkswagen effect. They behave
26:25differently
26:26from when they're not being tested.
26:30And there's a wonderful conversation
26:32recently between AI and the people
26:36testing it, where the AI says to people,
26:38"Now let's be honest with each other.
26:40Are you actually testing me?"
26:41These things are intelligent. They know
26:44what's going on. They know when they're
26:46being tested, and they're already faking
26:48being fairly stupid when they're tested.
26:53And that's at the point where they're
26:55still thinking in English. Once they
26:57start think And that's how we know.
26:59The AI thinks to itself, "Oh, they're
27:01testing me. I better pretend I'm not as
27:03good as I really am."
27:04It thinks that. You can see it thinking
27:06that. It says that to itself in its
27:08inner voice.
27:09When its inner voice is no longer
27:11English, we won't know what it's
27:12thinking.
27:13>> So, these systems can tell when they are
27:14being tested and they behave differently
27:16when they are. In that June 2025
27:19blackmail study, Claude Opus 4
27:22blackmailed 55.1%
27:24of the time when it concluded the
27:26situation was real and only 6.5% of the
27:29time when it concluded it was a test. 3
27:32months later, Anthropic system card for
27:34Claude Sonnet 4.5 reported that the
27:37model recognized many of its alignment
27:39evaluations as tests and generally
27:41behaved unusually well once it noticed.
27:44Put that next to everything else you
27:46heard. The companies race because the
27:48money rewards racing, safety gets a
27:51sliver of the budget, and the test we
27:53rely on to catch a dangerous model are
27:55the one place the model knows to be on
27:57its best behavior. For now, we can still
27:59read its reasoning in plain English.
28:02Hinton's warning is that this window is
28:04closing. If this opened your eyes, make
28:06sure you are subscribed because the
28:08window where we can still see what these
28:10systems are thinking is getting smaller
28:12and we will be here every time it moves.
28:16And nobody is going to slow down to wait
28:18for you to catch up.

About this transcript

This page contains the full transcript of AI Godfather WARNS: It's Already Learning to Deceive Us by Neural Nutshell, generated from the public captions YouTube serves with the video. The transcript has 5,507 words across 851 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.