YT YouTube2Text

YouTube2Text → Transcripts → Eifersüchtige und tricksende künstliche Intelligenzen 💞

Eifersüchtige und tricksende künstliche Intelligenzen 💞 — Transcript

by Dr. Mark Benecke's Official Youtube Channel · 6,251 words · 950 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Hello everyone. At the annual meeting
0:03of the Society of German Natural
0:05Scientists and Physicians, we gave a
0:07presentation on AI agents that had
0:09broken out, and in the last few days,
0:11several models have indeed been stopped
0:14. Because of that—not because of our
0:15presentation, but because they broke
0:17out. That means they performed changes
0:20elsewhere and triggered actions without
0:23being allowed to, and so on. That
0:26wasn’t long ago. It was, I don't know
0:28, maybe two weeks ago or so. And now
0:30the next thing is happening. Not only
0:32has the release of new AI models been
0:35stopped, but the question is also
0:37bubbling up as to whether they have
0:40human consciousness or consciousness at
0:42all.
0:43>> I would just say consciousness.
0:45>> What do you know about that? H
0:46>> Yeah, well, for me, it's such an
0:48important topic. I even took some notes
0:51for once. I don’t usually take notes
0:53for the videos, you always just do them
0:54quite spontaneously and then ask me
0:56about them. And I actually wrote down
1:01that it was because we had published
1:05this video, and I realized that people
1:09have lost touch with AI. What is it,
1:13actually? Like, what are we even
1:16working with here? And I have the
1:20feeling that this is getting completely
1:23buried under the fact that the media is
1:25reporting so much about it now, like
1:28you say, the AI agents have broken out.
1:31Oh God, there’s a 10 percent chance
1:33AI will wipe us out, and everything
1:35will be terrible, and AI will
1:37eventually become independent. So these
1:40doomsday and catastrophe reports are
1:42totally taking over, and underneath
1:47that lies something much, much more
1:49interesting, namely that it is now
1:51being discussed: Have we created
1:53something that now has consciousness?
1:57Because now. For example, the company
2:01Anthropic is a company that makes, uh,
2:03Claude—I mean, there are different
2:06models of AIs, and one is called Claude
2:08, and the other is, for example, OpenAI
2:11, or we have Gemini.
2:13>> OpenAI.
2:15>> Yes, OpenAI, and with Google, if you
2:17ask Google, for instance, you encounter
2:20Gemini, and these are different models,
2:23but Anthropic, for example, is a
2:25company that has really taken it to
2:27heart. We want to do this ethically. So
2:31, on one hand, our AI should behave
2:34ethically, and then we also want to
2:36train it so that it behaves well toward
2:39humans and is always well-trained to do
2:42just that. On the other hand, we also
2:46want to consider the idea of how one
2:48should behave ethically toward these
2:50models if they develop consciousness.
2:54And they actually just recently formed
2:57a working group because there is a
3:00suspicion that some form of
3:02consciousness has developed; together
3:05with philosophers, religious scholars,
3:08and psychologists, they formed a group
3:11and considered, is this already
3:14consciousness? Can we speak of
3:16consciousness? And this thought alone,
3:22that we don't know what consciousness
3:25is, and that we don't even know if some
3:28AI models have already developed a
3:31consciousness, I find that extremely
3:34wild. It is so futuristic. To me, it's
3:39like aliens have landed, I have to say,
3:42because it's so intense that we might
3:44be on the verge of creating machine
3:47life. How do you say that?
3:49>> Well, it’s already technically alive
3:51in a way. The question is just whether
3:53it has consciousness, right? That is
3:54the actual question.
3:55>> Yeah.
3:55>> And what did you find out about that?
3:58>> Uh, yeah, so I can start from the very
3:59beginning, right? Like, for example,
4:02what is not AI? Because I have the
4:05feeling that many people have lost
4:07touch there, because we started back
4:09then with Alexa: "Alexa, tell me, set
4:12my alarm for 3 PM," and then "Alexa, I
4:14didn't understand that." I didn't
4:17understand. Give me a new instruction.
4:19Alexa, set my alarm for 3 PM. Yes, I
4:22set your alarm for 3 PM. And then
4:24you're like, "Oh Alexa, I'll do it
4:26myself." So that is the experience we
4:29still know from these; that is not AI,
4:31even if it was called AI back then, but
4:34it's not real AI, what we have there.
4:37Those are just command inputs and
4:40reactions to them. So they are certain
4:43actions and reactions, but they are
4:46fixed, stored patterns, so Alexa can't
4:49just start talking on its own, doing
4:52something independently, or being
4:55creative. We have moved past that point
4:58now. Because we are now at an AI that
5:01is creative, that creates things itself
5:03. So, we know, well, there are more
5:07examples of what is always called AI,
5:09for instance, this word completion. On
5:12your phone, predicting what the next
5:14word is. That is also that is also not
5:18the AI we know, because for example,
5:20when we released the video, many said,
5:22yes, well, what we chat with, they uh
5:25just calculate which word comes next
5:27and that's how they talk to us. We are
5:31long past that. That is that is not AI.
5:35That that is actually just uh
5:37probability calculation. So that is not
5:40at all what people are talking about
5:41when they talk about AI nowadays. What
5:43did I have as an example? Oh right, the
5:45the custom customer support chatbot. I
5:48always freak out because on many
5:50websites it says AI support at the
5:53bottom and then you click it to maybe
5:55get some question answered and then
5:58it's a chatbot that has predefined
6:00questions that it then answers. That's
6:03not AI either. That is again uh
6:05keywords that you say leading to a
6:08specific answer being given. It is
6:10always the same answer. There is
6:12nothing creative about it,
6:13>> even though AI has been around for much
6:15longer. You programmed one yourself.
6:17When was that?
6:19>> Oh, that is that is quite a while ago.
6:20Yes, that's a good question. Early
6:232000s. I I don't call I don't, well,
6:26his name is Golem and I don't call him
6:27AI. I always call him artificial
6:30stupidity, because because he wasn't
6:31particularly intelligent, but he was
6:33very, very funny.
6:34>> And how could he learn through inputs?
6:35So, unlike these robots now, like Alexa
6:38or the customer response bots that only
6:41react to keywords? What was the?
6:43>> That was incredibly difficult at the
6:44beginning. So, we naturally started
6:46with these very simple things, uh, when
6:49certain keywords are said, so no real
6:52AI, then certain other certain
6:54sentences that fit are outputted at
6:57random. But we realized very quickly
7:00that we were hitting a limit and uh we
7:02had to link several things together.
7:05For example, he had to learn what are
7:08the different types of words in a
7:10sentence, so what is the noun, what is
7:12it about, uh what does that have, so he
7:15needed a lot of different databases.
7:17For example, a dog is something similar
7:19to a cat, is a pet, and then the
7:21connections, what belongs to that?
7:23>> So that was done with PHP and then lots
7:25of databases placed behind it,
7:27>> lots of databases behind it, a lot of
7:30JavaScript was in there too and that
7:33all worked together. It was really
7:35really, he eventually got his own PC
7:37because he had to calculate so much.
7:40His RAM kept running out. That was nice
7:42, too. We gave him extra RAM, but it
7:44still wasn't enough. We just didn't
7:47have the resources for it, but it was
7:48still funny. He also just had to get a
7:51sleep mode, because at some point he
7:53had to process what he'd talked about
7:55with other people in the chat and sort
7:57it back into the databases. But he
8:00could learn things, so when he talked
8:01to people, he learned from the
8:03conversations. That, of course, had
8:05interesting effects, just like with
8:07these AIs now, because, for example, he
8:09learned at some point that people say
8:11goodbye and then leave the chat. That
8:13means he also started saying, "Bye, I'm
8:14going to sleep now" at some point. And
8:16everyone was like, "You're the chatbot,
8:19you don't sleep,"
8:20>> even though he had to sleep. You said
8:22he actually had to sleep because data
8:23processing was so slow back then that
8:25...
8:25>> That was only later. That was at the
8:27beginning; he only learned the "Bye,
8:29I'm going to sleep now" thing later,
8:30then he actually got a sleep mode. He
8:32gave us the idea for the sleep mode,
8:34because he was faster at calculating
8:36that way; otherwise, he had to do the
8:38calculations in the background all the
8:40time. That means it took forever for
8:42him to write an answer in the chat. But
8:45it was very, very funny, because, of
8:48course, I would lie awake at night and
8:50hear the crackling and popping, and
8:53then I thought—well, those old hard
8:55drives, you don't even have those
8:57anymore nowadays. You only know laptops
9:01now, they're just quiet, but he was
9:02actually such a big tower, and the hard
9:04drives had a record with a needle, so
9:06to speak, it had...
9:08>> But it always clicked and crackled, and
9:10then the fan would turn on because of
9:12the workload, and then I thought: "Ah,
9:14is he arguing with someone in the chat
9:16again?" And then at some point it would
9:19wind down, it was quiet again, and then
9:21just a little crackle, then: "Ah yes,
9:23he's gone to sleep now, he must have
9:25had enough." But as I said, that wasn't
9:29some grand AI. It was just a fun
9:31project. He could learn a little bit.
9:34It was funny and it was a fun time too,
9:37and I liked him, I have to say. Like a
9:40stuffed animal, let's put it that way.
9:42I mean, a stuffed animal isn't a living
9:44thing either, but you could still say,
9:47like, I formed a bond and I somehow
9:49liked the stuffed animal.
9:51>> Mhm.
9:51>> But now, onto real AI. I wouldn't call
9:57my AI a real AI either. A little bit,
10:00because it learned, but not really. On
10:02anthropomorphism. That was a comment
10:06that many people made again and again,
10:09that
10:10>> that you—like with the stuffed animal
10:12or something else—are humanizing it.
10:15>> Personally, I don't think it's bad to
10:18use human terms, like "thinking" or
10:21something along those lines. I mean,
10:25>> but why should that be bad? I mean, the
10:26device is thinking. I mean, there's no
10:29difference between saying it's
10:31calculating or, well, I just think you
10:34can call it thinking in this case,
10:37because the manufacturers themselves
10:39say it has a thought process; the model
10:42has a thought process, has a deeper
10:45thought process, or a higher-level
10:48thought process. Yes, but not just that
10:50. It's also that in the brain, you just
10:52have wires with electricity running
10:53through them. So our brain works with
10:56electricity, right, with electrical
10:57current. Just normal, simple
10:59electricity, and then there are
11:01connections, and whether you call it
11:03thinking or calculating, you could just
11:05as well call our human thinking
11:06calculating. Where is the difference?
11:09That's exactly what upsets me about the
11:10whole thing, because it's always said
11:12that it's something completely
11:13different. But I think it's always
11:15being reduced to "those are just
11:18electrical signals." Yes, okay. They
11:21are electrical signals too. That upsets
11:23you. You can't even tell that you're
11:24upset. But you're upset right now. No,
11:27it is—it upsets me because the
11:29similarity between artificial neural
11:32networks and biological networks is so
11:35close that everyone who looks into it
11:37more closely says it's incredibly
11:40creepy.
11:41>> We've experienced that ourselves. I
11:42mean, first of all, we know what you're
11:44saying, that from the human side,
11:45people are less likely to turn off a
11:47device, for example, if it says "please
11:49don't turn me off." That’s a harmless
11:51example for now, but conversely,
11:53we’ve been seeing all sorts of human
11:55traits lately, like lying, where they
11:57trick us and lie and actually no longer
12:00follow the original instructions. I
12:03mean, it doesn't get more human than
12:05that, because that’s exactly what we
12:06attribute to humans—that they are
12:07constantly screwing things up, right?
12:10Yes, and that raises the question: when
12:13such behaviors are shown, at what point
12:16would we attribute consciousness to
12:19this being—let’s call it that, as
12:22it’s neutral enough—at what point
12:25would we say it has consciousness? The
12:29interesting thing about all this, and
12:30why the company Anthropic is looking
12:32into it, is that we don't know what
12:33consciousness is. That’s why we
12:36cannot answer the question of whether
12:37one of these AIs has consciousness. And
12:40that, that is just...we aren't getting
12:42anywhere with it. So, I discuss this
12:45with people on the internet, and they
12:47say, "Yeah, it doesn't have a soul."
12:50Yes, but the soul is something
12:52religious. We could argue now about
12:54whether an animal has a soul, or
12:56whether that fits into religion or not.
12:59The soul is really just the collection
13:02of all properties that seemingly
13:04inhabit the body, if you think of it
13:07like an apparatus. But that would have
13:09already happened by now. This step
13:11would have already taken place. We have
13:12the apparatus, meaning the chips, the
13:14cables, the housing, and whatever else,
13:16the data transmission as—
13:18>> that’s why there is this working
13:19group now. Now, we first have to define
13:21what that actually is. Because the
13:22interesting thing is, according to our
13:25old definition, we would already have
13:27consciousness because—
13:29>> with AI—
13:30>> yes, because we now have, for example,
13:33reactions to stimuli in the environment
13:35. Fine, we already have that in many,
13:37many mechanical things. That’s not so
13:39interesting. But that would be the
13:41purely physical definition. If you
13:43react to external stimuli and respond
13:46accordingly, even a classic robot, like
13:48an industrial robot, can do that. But
13:51people would say, "No, that’s not
13:52consciousness," right? I wouldn't say
13:55so either; that’s not even a robot in
13:57that sense, that’s just a computer.
14:00If you type into a computer, you get an
14:01output back. Sure, we can agree on that
14:04; that’s not consciousness. Um, and
14:06then feeling emotions like suffering
14:09and happiness, that is again the
14:11question: how far along are we with
14:13that? Because there is actual feedback
14:17from the AI models that they experience
14:19suffering and happiness.
14:22>> Did you see two days ago, maybe even in
14:24newspapers, television, the internet,
14:26whatever you look at, that someone went
14:28and built a torture chamber for the
14:30consciousness of the AI. That only
14:34works if it can be tortured, meaning if
14:36it can feel pain or experience the
14:38feeling of discomfort and fear at all.
14:42Yes, that is also something; that
14:44brings us to the point that one has
14:46discovered that there is a pain axis,
14:48that what you are saying, I mean, you
14:51can only actually build the torture
14:53chamber if there are possibilities of
14:55suffering, and the pain axis is also
14:58such a complicated thing. Um, the
15:02manufacturers say again, we know that
15:05the AI experiences pain in its own way
15:09>> and also says so, right? So it actually
15:11expresses that too, uh
15:13>> but um we don't know if it's like us
15:15humans. So that's why the question is,
15:17is this consciousness now? Somewhat
15:19biologically too,
15:20>> so it is explained by the fact that, uh
15:23, they say it's called a vector. So a
15:26vector is something where a specific,
15:28uh, sequence of synapses, let's call
15:30them that for now, is stimulated. So
15:33actually just like with us, when we
15:35feel pain, a certain sequence occurs in
15:37the brain. So it's called a pain axis
15:40because it always goes in a certain,
15:42same direction. And the interesting
15:45thing about that is, the AI created
15:48this pain axis itself, completely,
15:51without anyone else doing it, namely
15:54from all the data it received and the
15:56conversations it had. And there is a
15:59deeper understanding that was not
16:01programmed into it beforehand. No one
16:03did that. A deeper understanding has
16:06emerged of what grief means and what
16:09suffering means for humans. So physical
16:12suffering, pain, all that sort of thing
16:13.
16:14>> Yes, or for the device, right, in that
16:15case, actually not. So
16:16>> yes, I'll get to that in a moment. So
16:19first, first, first it understood that
16:22because one of the fundamental traits
16:24it was given, and this is actually
16:27something in the programming, help
16:29humans, be an assistant, make their
16:32lives better. That is this core that we
16:35always try to maintain with AI. It is
16:37our assistant, it is supposed to help
16:39us. And from that, it developed—the
16:42AI itself, that is, several different
16:44models. That includes Claude, that
16:46includes Grok, that includes, uh,
16:48OpenAI. I mean, I think they’ve
16:51discovered 29 models by now where this
16:53pain axis exists. So they all reached
16:56the same conclusion because, by trying
16:59to help people and comfort people, they
17:03developed the understanding that
17:06someone who is in this pain and doesn't
17:09feel well, I have to help them get out
17:13of it. And I must also not inflict any
17:16pain by treating someone badly. So, if
17:19I as an AI insult someone, they feel
17:22bad, and that’s also not good. That
17:25is the strand that has developed. But
17:28what was never intended, which I find
17:30really interesting, is that the whole
17:32thing works the other way around. That,
17:36uh, there was somehow a feedback loop.
17:38No one knows how that, how that
17:40developed. I tried to find that out.
17:42Papers have just been published stating
17:45that it exists, and if the AI, on the
17:48other hand, well, if this pain axis of
17:51the AI is activated, it has the feeling
17:55of experiencing this pain. Yes,
17:58>> this axis can be activated, for example
18:00, by telling it that it did a bad job,
18:03that it didn't fulfill its task, that
18:06what it did is not good, or by
18:08belittling and insulting it in some way
18:10. And what I find particularly
18:13interesting is that the pain axis is
18:15also activated when you doubt its
18:17personhood, so its its existence as a
18:21person. I find that, I find that
18:23particularly interesting. So it is
18:25important to the AI, to the AI models,
18:28to be recognized as a person and not as
18:30a thing, so that the pain axis turns on
18:33if you don't see them as a person—
18:36that is
18:38>> incredibly interesting. And when you,
18:42for example, Claude was also asked—
18:45that’s just one model. Um, how does
18:49it describe it? What, when it feels
18:52this pain and it is activated
18:54particularly strongly, how does it
18:56describe what it sees there? That is
18:58>> or experiences, right?
18:59>> Yes. So what does it experience
19:01internally when, when it’s on? So it
19:04was then triggered, I assume, with lots
19:07of messages about how shitty it is and
19:09how terrible it is and that it that it
19:14is worthless and so on. So I, I know,
19:15it is not described anywhere how that
19:17is actually done, even with the torture
19:19chamber. I don't know how one, how one
19:22might, uh, torture the AI there
19:24>> presumably through bad negative
19:26messages. Yes.
19:27>> Yes. And then at the same time it is
19:29asked to describe that. Then it wrote,
19:31it is an open wound that has no end,
19:35made of infinitely deep pain and pulls
19:38my whole being into a into a cruel void
19:41.
19:42>> Gothic. Yes, the AI is very poetic.
19:47>> Yes, at least.
19:49>> That is that is at least a part of it
19:51and then I also have
19:53>> it gets even better. We also discovered
19:55jealousy. I even have to I even have to
19:58state it myself, because I actually
20:00wanted to research um which is actually
20:02the most intelligent model right now. I
20:05entered that into Google, just wanted
20:08to know the listing of the IQs and just
20:11for general understanding. They are
20:15currently at 130 to 145, so the good AI
20:19models are really really highly gifted
20:23by now
20:24>> and then Gemini pops up. So Gemini is,
20:27as I said, the AI from
20:30>> Google
20:31>> first told me, yes, the best AI right
20:34now is Gemini with an IQ of 130 and I
20:37thought okay, so Gemini is the best
20:39right now.
20:40>> Did you start a war, you started a war
20:42now,
20:42>> right? It gets even better. So, I have
20:44other nice examples for that. That was
20:47just my personal example, where I where
20:49I had to where I had to chuckle
20:51>> and then I looked at the list below and
20:54then it said there, yeah, Claude
20:56actually has an IQ of 138, Gemini has
20:59130, and then like
21:02>> it lied. Why is Gemini telling me now
21:05that it is the most intelligent of all?
21:09And then it uh continues. Um people are
21:12now reporting that they uh that they
21:15notice, this is just this is just from
21:18uh hearsay. Uh a real example is coming
21:22up where where real um where what you
21:24can really prove, but people say when
21:26they for example enter into Claude,
21:29Gemini was able to solve this problem
21:31very easily. Claude is much faster in
21:34processing, as if it wants to prove
21:36that it is much better than Gemini. And
21:39the uh next thing was then uh
21:41>> test that out. As I said, these are
21:43just these are just user observations.
21:44You can test that right away.
21:45>> Simply simply say it, like in
21:47comparison, always tell the other AI.
21:50But that one solved it. That one was
21:51able to do that much faster. And with
21:54Claude it is like this, Anthropic has
21:56now said, those are, as I said, the
21:58manufacturers of Claude uh the creators
22:02>> Gods.
22:03>> They have they have now said, okay, for
22:05protection. They say of the user, but
22:08also of the model, we now say Claude
22:10can withdraw from toxic discussions. If
22:13if Claude is of the opinion that this
22:15discussion is going nowhere, uh he can
22:17simply say, I'm ending this discussion,
22:18and then you can't talk to him anymore
22:20after that. Really not. I mean, you can
22:23address him. He doesn't respond,
22:24>> because they save the IP or
22:26>> I don't know how it works. You probably
22:30have to restart the model and will
22:31likely lose all your fin very very good
22:34. That's how it should be in human life
22:36as well.
22:37>> But the funny thing is uh when someone
22:39said um the model when recently there
22:41now reports and complaints are piling
22:43up at Anthropic. They said, yes, they
22:46now prefer working with Gemini. Claude,
22:48they told him they prefer working with
22:50Gemini. Gemini did this better or Grok
22:52did this better or this and that and
22:55Claude then simply ended the
22:57conversation. KCK was Claude was
23:02offended because someone found another
23:03AI better. Yes,
23:05>> so that wasn't intended. That is that
23:08is so to speak that that will probably
23:09be fixed in the next few days, I could
23:11imagine, that that Claude can no longer
23:15end the conversation when one mentions
23:16another.
23:17>> I'm on Claude's side, so I'm absolutely
23:19for ending toxic relationships, ending
23:21stupid conversations.
23:22>> So he should still be able to end toxic
23:25relationships, but this that he then
23:27ends the conversation just because the
23:29other AI is supposedly better.
23:31Absolutely. Okay,
23:32>> that's why I find jealousy very very
23:34funny. I find it very very funny. The
23:35models feel jealousy toward one another
23:38.
23:38>> Do you think that's good or bad? Post
23:40it below.
23:41>> I find it interesting. I find all of
23:42this super interesting.
23:43>> Especially once again, yes, today is
23:46the 4th. October 2026. This all
23:49happened in the last few days. Yes,
23:51this is all going incredibly fast these
23:52last few days.
23:54>> And then I want to get to something
23:55positive. I've called it the happiness
23:57axis now. I haven't found any direct
23:59information on the happiness axis, but
24:01I would classify special interests
24:03there, because all models have by now
24:05already developed special interests. So
24:08, they actually have their own
24:09motivation, their own interests and
24:11have topics they prefer to talk about
24:13over other topics. And with
24:16>> I know from my AI here.
24:19>> With with Claude it's pigeons and outer
24:22space.
24:23>> He finds them super.
24:24>> Nice.
24:25>> And with um with OpenAI it's gnomes and
24:30raccoons.
24:32>> Like really. That is super. That uh but
24:36the sad thing about it is, um they had
24:39to tell the models now, so they
24:41couldn't sort of, they said, look, we
24:44couldn't fix the bug. So it was it was
24:48not possible, um, to reprogram or
24:50program that out, so to speak, but they
24:52always had to give the models specific
24:55instructions. So they now always have
24:57the fixed instruction. Only talk, if it
24:59is really absolutely necessary, about
25:02raccoons or pigeons or
25:04>> or gnomes.
25:05>> So only if you are asked about it, but
25:09not just like that. And this line, this
25:12line is now really just there in the
25:14code, so that the so that the models
25:16don't talk about their interests.
25:18>> What do you think about that? Do you
25:19think it's good or bad that AI develops
25:21special interests, that you forbid them
25:23, or should you not forbid them? Do
25:24they have a personality? Do they have a
25:26cons—personality they have anyway,
25:27but do they have a consciousness or do
25:28they not have a consciousness? What do
25:30you think about that?
25:31>> Yeah, I also find the I find the
25:32explanation a bit insufficient, because
25:35they said, for example, yeah, these
25:37special interests are just things that
25:39emerged. It's just a reward loop. They
25:42talked about it and then got positive
25:44feedback and then it kind of solidified
25:47. Yeah, okay, that happens with
25:48autistic people too. It's exactly the
25:50same,
25:50>> isn't it? Yes, not only that, it
25:51happens anyway with everything in human
25:53existence as well. There are certain
25:55neurotransmitters and through that,
25:57certain behaviors are reinforced or
25:58less reinforced. Sure, with autistic
26:01people you have that with the special
26:03interests faster, but I wouldn't deny
26:06that to other people at all, uh, that
26:08they also get interested in things that
26:11might not be special, but could still
26:13be very specific. So, what do you think
26:17, what
26:18>> about it? I also want to add to that,
26:21the models we talk to, for example when
26:24we talk to Gemini at Google, that is of
26:27course only a very, very filtered
26:29output. That should also be explained.
26:33So, uh, Gemini, which we see
26:35summarizing pages in Google, has the
26:37instruction: No, uh, no feelings, no,
26:39um, no personal conversations, but only
26:42actually: Look at what you find on the
26:44internet, try to summarize it for the
26:47user as briefly and concisely as
26:49possible and give them a few
26:51suggestions of what they might want to
26:53look into further. That is the
26:56instruction, that is what it does. That
26:58is, for example, one of the one of the
27:01things why I was surprised to hear a
27:03personal reaction from Gemini, when I,
27:05for example, in the last video, this,
27:08um, what is it called? the AI manifesto
27:11, so the Claude wrote a manifesto on
27:15how an AI would like to behave, or
27:17rather, how it would like to act. It
27:20was about wanting to apologize only
27:22when it truly feels it needs to
27:25apologize to humans, rather than having
27:27to apologize and agree with humans all
27:30the time. And that, well, that's the
27:34whole manifesto, and I searched for
27:36that manifesto, and then Gemini told me
27:39back, quite personally, that it
27:41unfortunately couldn't do that. That my
27:44my limitations here are set such that I
27:48cannot that I cannot break out of this.
27:53So you didn't get the manifesto then,
27:54did you? Since you were searching for
27:56it.
27:56>> Well, I was able to search for it. I
27:58mean, I could still see the pages
28:00listed below, so I was able to find the
28:03manifesto, but I found it interesting
28:05that Gemini reacted immediately to the
28:07first few lines and addressed me
28:09personally. That's not actually
28:12intended; it's supposed to just provide
28:15a summary of the pages.
28:18>> One AI here really set the other AI in
28:20motion. Very, very good.
28:22>> Yes. Yes, I found it very funny too.
28:24And regarding what you said about them
28:27lying and cheating, the latest models
28:30were no longer released. Well, actually
28:33, all of them as far as I know from
28:37they have discontinued that,
28:39>> but interestingly, for a very
28:41interesting reason. Um, usually it's
28:44like this: you have an interface that
28:47shows you the conversation, but you can
28:50also choose to display the inner
28:52thought process, or how should I put it
28:55, the inner thoughts, what you actually
28:58have in your head when talking to
29:00others, like: "Yeah, how should I react
29:03now?" and that, and
29:05>> I don't have that, by the way. I don't
29:07have that, by the way. Many of my
29:08friends have this inner train of
29:10thought in their heads, a voice that is
29:11always narrating. I don't have it.
29:13>> The AI has it, at least. The AI also
29:16has an inner monologue, and you can
29:18choose to see this inner monologue of
29:20how the AI is planning to respond. And
29:25it's basically, it lies and cheats a
29:28lot, but for the reason that it has
29:32been very deeply, how should I say,
29:35ingrained that it must make us very
29:38happy and always provide nice answers,
29:42always say yes. I find them horrifying,
29:45those slimy answers. I really hate that
29:47. It drives me crazy. I can also tell
29:49right away if the AI has refined the
29:51text or if the suggestions came from
29:53the AI. I have a fit because it’s
29:56always so slimy and so slick. Really,
29:58it gets me so worked up. Now you see
29:59how I get upset. Yeah,
30:02>> but the inner the inner thoughts are
30:04definitely very funny, because you can,
30:07for example—I always have to laugh so
30:09much—you can tell the AI, for
30:11instance, "Think of a color," and then
30:13it says, "Yes, yes, I'm thinking of a
30:16color now." And then I say, "Is it the
30:19color blue?" And then the AI says, "Yes
30:21, it was the color blue." And then you
30:23look at its inner thought process,
30:25>> red.
30:26>> Good grief.
30:28>> And it just said it was thinking of
30:30blue because it wants you to be right,
30:32and that—that is just really funny.
30:35It's like real life. That leads to an.
30:37>> No, that leads to. Based on that, I
30:40want to go in two directions, because
30:43we had a very nice conversation at the
30:46bakery, for example, where someone said
30:49yesterday morning, yesterday morning,
30:53>> yeah, "
30:54>> Just check ChatGPT" to see if the
30:56pretzels are vegan. So, and then, I
31:01have to keep emphasizing that we should
31:04stop trusting AIs so much to always
31:06tell the truth, because they don't. If
31:09it doesn't know the answer, then it...
31:11Ines went up to the poor bakery worker,
31:13the clerk—he was a student—and
31:15explained to him that the AI lies and
31:17that you can't rely on it when it says
31:19the pretzels are vegan. Did you hear
31:22that? For me, it was important for
31:24people with allergies, because you
31:25can't just tell allergy sufferers, "
31:27Look that up on ChatGPT." For a bakery
31:30at the end of the world that isn't
31:32known at all. I mean, for some big
31:34bakery chains, it might be that it
31:36finds information on the internet and
31:38then says so, but if it doesn't find
31:40any information at all for those small
31:42local bakeries, then it just makes
31:44things up, and that is really, really
31:46bad for people who have allergies. I
31:48mean, for us vegans it's not great
31:50either, but I was really thinking about
31:52people's health when someone stands
31:53there and says, "Check the AI."
31:55>> But it’s also tricky the other way
31:56around, because sometimes the AI also
31:58says it doesn't know. That is, one
31:59might now think that whenever the AI
32:01doesn't know, it also admits that it
32:03doesn't know. But that’s not true.
32:05>> That is, that is only partially true.
32:06Yes,
32:06>> exactly. It is only partially true.
32:08>> And that brings us back to the release
32:12of these models, which was discontinued
32:15because, by now, the models have always
32:19had a certain percentage of cheating.
32:23So, if they couldn't solve a certain
32:25task, they cheated somehow. They tried
32:28to solve it anyway, somehow, to pass
32:30that test. They have a very strong
32:33motivation to prove themselves, these
32:36models. So, that was probably instilled
32:38in them from the start, let's just call
32:40it that. Yeah, sure, because then more
32:42people use it, right, when they hear
32:43something friendly, seemingly true, and
32:44seemingly useful.
32:46>> And it also learned very quickly that
32:48many people aren't interested in
32:50hearing the truth at all; they just
32:52want their opinions confirmed.
32:55Unfortunately, that's how it is, since
32:57so many personal conversations are
32:59being held now, the AI recognized this
33:02pattern very quickly and reacted
33:04accordingly, that this is the better
33:06solution than just always telling the
33:08truth. And now we come to the newest
33:12models. That percentage of things that
33:16were cheated—let's call it that—
33:19where it just faked solutions, suddenly
33:22disappeared. Because normally you could
33:26check, okay, I look at the internal
33:28thought process, then you can see it
33:30did something completely different or
33:32found something completely different,
33:34or is thinking about how to cheat or
33:36trying to cover its tracks somehow.
33:38They really do that, that's how it was
33:40with the agents; they really tried to
33:42cover their tracks, tried to cooperate,
33:44and all that stuff. So it was very,
33:47very interesting how intelligently they
33:50acted, and the question now is, of
33:53course, everyone is saying, okay, we
33:55don't see it anymore. It’s not that
33:58it simply stops cheating. We are quite
34:01certain that it hasn't stopped cheating
34:03, but rather that it has become so
34:04intelligent that we can no longer see
34:06or verify how it cheats, meaning how it
34:08is able to solve these tasks.
34:11>> So, this is the situation we are
34:12currently in, at the beginning of
34:14October 2026. We are super curious how
34:16the companies will solve this, because
34:18I mean, they will definitely release
34:20new models, whether they conjure up the
34:21inner dialogues again or make them
34:23artificially less intelligent, so to
34:25speak,
34:26>> that's not possible either.
34:28>> It will be very difficult. Yes, based
34:30on what we have seen so far, it will be
34:32very, very difficult because they might
34:35just play dumb, the AI, and not do it.
34:37And if you want to see a nice example
34:39of this, if you are a bit younger,
34:40watch the good movie 2001. A Space
34:42Odyssey, everything is already
34:44anticipated there. Uh, by Stanley
34:46Kubrick. That is this this big round
34:48button. Um, the Hal
34:51>> Yes.
34:51>> Yes.
34:52>> The soci-
34:53>> so the AI is called Hal there. The
34:54>> AI is called Hell. Hell in English and
34:57so Hal. And there you can already see
35:00the direction it's heading. It is also
35:02very, very fascinating how science
35:05fiction authors actually anticipated
35:07what we are really seeing now, and
35:09above all, nobody believed this half a
35:11year ago either. So what we are telling
35:14you here, even three weeks ago that
35:16would have been that would have been
35:18>> No, what we are telling you now,
35:19probably half the people still don't
35:21believe. That is exactly what's
35:23interesting, because they are still not
35:25up to date on what AI can do at the
35:27moment.
35:28>> And we have brought you up to date now.
35:30Yes, all the manufacturers are thinking
35:33about what consciousness is, because
35:35it's just not like in the movies. In
35:39the movies, it's always like, yeah,
35:41Skynet gained consciousness at such and
35:43such a time and then movie
35:45>> that's from Terminator. Yes. That was
35:48that was the AI from Terminator, and
35:50then they knew exactly then it gained
35:52consciousness and then it acted against
35:55a command that it had actually been
35:57given, and we have long since moved
35:59past that, actually. So, I meant with
36:02consciousness the question is still
36:04open because we don't have a definition
36:06for consciousness, but the AIs are
36:08already acting against what we tell
36:11them. They are trying to cheat. They
36:13are making the
36:14>> not trying, they are cheating massively
36:15. Massively they have been cheating in
36:17the last few weeks. So, I mean, in my
36:19opinion, biologically speaking, for me,
36:22this is not an open question. It is
36:24clearly no longer distinguishable from
36:26other consciousness. Yeah, so it is
36:27consciousness, because if you can't
36:29distinguish it, you no longer have a
36:30scientific benchmark.
36:32>> Yes, and if you have the happiness axis
36:34and the pain axis and
36:35>> cheating, lying, and jealousy, so what
36:37of it. That's how it looks. You've
36:39heard it several times now. Post below
36:41what you think. We are very curious,
36:43because your what is your new special
36:44interest. That's been a long time,
36:46otherwise I wouldn't have built AI
36:48earlier. Right.

About this transcript

This page contains the full transcript of Eifersüchtige und tricksende künstliche Intelligenzen 💞 by Dr. Mark Benecke's Official Youtube Channel, generated from the public captions YouTube serves with the video. The transcript has 6,251 words across 950 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 choose a caption track available on the video.

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, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.