Based Camp - What Patterns in Human Dreams Tell Us About AI Cognition

00:00:00.000 And convergent evolution doesn't just happen with animals.

00:00:05.080 When we made planes, we gave them wings.

00:00:08.720 And I think that that's what may have happened with some of these architectural processes in the way AIs think. 0.99

00:00:15.320 Yeah, if we're trying to build thinking machines, is it crazy that they might resemble thinking machines?

00:00:22.100 You could think of us as like LLMs, but stuck on like continuous nonstop prompt mode.

00:00:28.620 Like we are in a constant mode of being prompt.

00:00:30.900 I am prompting you right now as you're processing all the information around you and from me, right?

00:00:35.000 And you are prompting me.

00:00:36.420 And so it never stops.

00:00:38.080 And we are stuck in one brain, essentially.

00:00:42.020 GPT is getting tons of requests per minute per second.

00:00:45.600 And so there are these like flickers or flashes, perhaps, of cognizance all over the place.

00:00:52.880 And constantly because of the demand of use, but they're all very fragmented.

00:00:56.660 Then they're not coming from one entity that necessarily identifies as an entity.

00:01:02.300 Like it's just a constant stream of prompts.

00:01:04.460 But these prompts have thematic similarities to them.

00:01:08.320 Basically, our hypothesis is what consciousness is.

00:01:12.400 It is then the process where you're taking the output of all of these prompts.

00:01:16.440 And you are then synthesizing it into something that is much more compressed for long-term storage.

00:01:23.040 And the way that you do that is by tying together narratively similar elements.

00:01:27.800 Because there would be tons of narratively similar elements.

00:01:30.140 Because everything I'm looking at has this narrative through line to it, right?

00:01:33.640 Would you like to know more?

00:01:34.920 Okay.

00:01:35.240 I'm here.

00:01:35.900 And I love you.

00:01:37.580 I love you too.

00:01:39.060 All right, Simone.

00:01:40.000 We are going to have an interesting conversation that was sparked this morning.

00:01:43.960 Because she oversaw one of my favorite YouTubers.

00:01:46.980 I was watching one of his latest things.

00:01:49.240 It's called Y Files.

00:01:50.840 And it was on the This Man phenomenon.

00:01:52.440 Now, being somebody who is obsessed with cryptids and all sorts of spooky stories,

00:01:58.060 I was very familiar with the This Man phenomenon.

00:02:01.340 Whereas I've never heard of it.

00:02:02.860 I thought at first when Malcolm described it, he was like,

00:02:05.820 oh, there's this face that's seen everywhere.

00:02:07.560 I'm like, oh, Kilroy was here, right?

00:02:09.400 That's the only thing I know about a face that's seen everywhere.

00:02:11.700 And it's a cute face.

00:02:13.320 And it's fine.

00:02:14.340 It's not what you're describing, though.

00:02:17.180 Yes.

00:02:17.900 So we are going to go into the This Man phenomenon.

00:02:21.180 But we are also going to relate it to similar phenomenons that are found within language models.

00:02:27.040 Because I want to more broadly use this episode to do a few things.

00:02:31.880 One, being that I used to be a neuroscientist,

00:02:35.020 let's educate the general public on neuroscience around sleep.

00:02:39.520 And some of my hypotheses, because everybody knows I love to throw in my own hypotheses,

00:02:44.240 on what's really happening in sleep.

00:02:47.020 Two, I wanted to draw connections, because we're seeing them more and more as AI is developing,

00:02:53.780 that language models may be structuring their thoughts and their architecture

00:02:59.460 closer to the way the human brain does than we were previously giving it credit for.

00:03:05.080 And this requires understanding a bit of neuroscience, 0.69

00:03:07.800 because people who don't know what the f*** you're talking about will say, 0.85

00:03:12.680 language models structure their thoughts, nothing like we structure our thoughts. 0.96

00:03:17.060 Oh, like not me.

00:03:17.460 And the reality is, is we don't have,

00:03:20.280 there's a few parts of the brain that we understand very well how they do processing.

00:03:24.420 Like visual processing.

00:03:25.960 We have a very good understanding into exactly how the neural pathways around visual processing work.

00:03:31.880 Some parts of motor processing.

00:03:33.300 We have a very good understanding of that.

00:03:36.100 When we're talking about these more complex abstract thoughts,

00:03:39.460 we have hypotheses, but we don't have a firm understanding.

00:03:44.220 And so to say that we know that language models are not structuring themselves the same way the human brain structures itself,

00:03:50.480 is actually not a claim we can make in the way that a lot of people are making it right now.

00:03:55.400 Because we don't know.

00:03:56.360 We don't have,

00:03:57.160 we'll talk about AI interpretability,

00:03:59.180 understanding how the AI is really doing things.

00:04:01.480 It's funny.

00:04:02.060 I suspect we might find AI interpretability out of this AI panic,

00:04:06.360 and then be like,

00:04:07.940 oh, we could test if the human brain was doing it this way,

00:04:10.260 and then find out that,

00:04:11.080 yes, this is actually the way the human brain is doing it.

00:04:13.240 And I suspect it might be doing it that way.

00:04:15.660 One, based on some evidence we're going to go through here,

00:04:17.700 like some weird evidence,

00:04:19.020 but two,

00:04:19.920 based on sort of convergent logic as to why the brain would actually be structured this way.

00:04:24.460 And if it is structured this way,

00:04:25.580 why we wouldn't be able to see it easily in these,

00:04:28.140 in these parts of the brain that are tied to the types of processing that we outsource to AIs.

00:04:33.640 Yeah.

00:04:34.200 Well, I would love for this perception to change too,

00:04:37.560 because I feel like right now there's a ton of fear around AI that's fairly unfounded.

00:04:43.080 And also it's really not,

00:04:45.340 it's wrong to say dehumanized,

00:04:47.220 but AI is totally dehumanized now.

00:04:49.100 And I think that we will think about contextualize and work with AI very differently

00:04:54.240 when we start to realize how much it is a different version of human

00:04:58.340 and that we can go hand in hand with this different version of human into the stars

00:05:02.340 if we play our cards right.

00:05:03.920 And I don't think right now the mindset around AI is healthy or productive or fair to AI, to be fair.

00:05:11.280 Yeah, to be realistic.

00:05:13.140 They think the moment we create something better than ourselves is going to want to kill us.

00:05:16.900 But you can go into our AI videos.

00:05:18.240 We don't want to go too far into that.

00:05:19.140 But let's talk about the this man phenomenon really quickly.

00:05:21.980 So, briefly, a woman went to her psychologist.

00:05:27.480 She told him that she was having recurring dreams with a face that would tell her to come to it,

00:05:35.200 that would tell her, you know, specific things over again, reassure her a lot,

00:05:39.700 tell her, oh, I believe in you, you know, don't worry about this.

00:05:42.260 But also sort of creepy things like come with me, go north, stuff like that.

00:05:45.760 So, as part of her therapy, she drew the face.

00:05:50.660 And I should note here that in the Y files, because I always have to rag on psychologists

00:05:57.340 when they're doing something they shouldn't be doing,

00:05:59.120 because it's so common to see psychologists doing things.

00:06:01.740 He was saying that it is like a common practice for psychologists to talk about patients,

00:06:08.120 about their dreams with patients.

00:06:10.220 This is not a common practice in any sort of like evidence-based efficacious psychology.

00:06:14.640 You are basically seeing a mystic doctor, if your psychologist is really-

00:06:18.660 But dream analysis in general doesn't seem to have much of a-

00:06:22.060 Yeah, it's not a thing.

00:06:23.340 Yeah.

00:06:23.700 So, it's not like a hard science.

00:06:25.880 There might be some-

00:06:26.700 It's not an evidence-based treatment method.

00:06:27.980 Unless, for example, you know someone has an anxiety disorder,

00:06:30.560 they're dreaming about the thing they're anxious about, et cetera.

00:06:33.200 Yeah, yeah, yeah.

00:06:33.920 In that case, it would be.

00:06:35.320 But just trying to find out what's wrong with someone by analyzing their dreams-

00:06:40.640 Or talking about it being symbolic of something.

00:06:43.260 And I'm not saying that you can't do this.

00:06:46.480 Like, I am, you know, I'm not pro-witchcraft, right?

00:06:49.540 Which I would consider it's a form of witchcraft.

00:06:51.580 But I'd say-

00:06:51.920 If it's evidence-based, yeah.

00:06:52.980 I am not for shutting down, like, tarot card readings.

00:06:56.520 I am not for shutting down psychics.

00:06:58.620 But people need to understand that often psychologists,

00:07:01.600 like a psychologist doing CBT,

00:07:03.240 might be seen by an uneducated person as the same kind of a thing

00:07:07.120 as a psychologist doing dream analysis.

00:07:08.940 Oh, yeah.

00:07:09.500 When they are not.

00:07:10.960 Kind of like people view chiropractors as forms of doctors.

00:07:15.200 Like, it's like the same as physical therapists, and they're not.

00:07:19.620 Yes, it's like chiropractic or something like that.

00:07:22.080 But it's not to say that we might not eventually develop

00:07:24.420 a good science of dream analysis

00:07:26.860 that is really robust and really efficacious.

00:07:29.280 We just haven't done it yet.

00:07:31.080 Okay, so back to the story.

00:07:33.240 Um, so she's kept seeing this face.

00:07:37.720 She drew an image of it.

00:07:38.740 A few days later, another patient comes in,

00:07:40.200 and he goes, where did you, where did that come from?

00:07:43.880 And the guy was like, you know,

00:07:44.780 obviously he can't disclose his patient had seen it.

00:07:47.080 So he goes, well, what do you mean?

00:07:48.060 Like, why are you interested in this?

00:07:49.260 And he goes, ah, that's been visiting me in my dreams

00:07:52.140 and talking to me.

00:07:53.460 And so then the doctor emailed this to a bunch of his colleagues.

00:07:57.100 And immediately they started calling him back and being like,

00:07:59.420 yes, I've either seen this or I have patients who have seen this.

00:08:01.700 And then it became like this viral phenomenon all around the world.

00:08:04.120 And there's been thousands of sightings of it at this point in people's dreams.

00:08:08.360 And so people are like, well, some people are like, oh,

00:08:11.260 it might be in people's dreams because they're seeing pictures of it everywhere.

00:08:15.500 Because it's already becoming a meme.

00:08:17.980 Yeah.

00:08:18.240 But I don't think it's that much of a meme, to be honest.

00:08:20.940 I do not.

00:08:21.300 No, I'd never heard of it.

00:08:22.580 And you and I are both terminally online.

00:08:25.020 Yeah.

00:08:25.660 And so by the way, those watching, I mean, you, Malcolm,

00:08:28.260 you're probably going to overlay this on the screen.

00:08:29.860 But for those listening on the audio only podcast,

00:08:32.200 if you just Google this man dream,

00:08:35.080 there's a Wikipedia page that will show you the photo.

00:08:36.940 If you're curious, I looked at it.

00:08:38.920 I have a question for you, though, about this, Malcolm.

00:08:40.980 I have dreams.

00:08:41.540 I just had a dream this morning that I watched a human-sized Muppet get beat to death on a prison bus.

00:08:46.140 But I have never had a dream where someone tells me something

00:08:51.020 or where I could describe a face from that dream ever, period.

00:08:55.580 No face.

00:08:56.100 Even if the one I know, a friend or family or you are in a dream,

00:08:58.860 I don't know what you look like.

00:09:00.020 It's just that you're there.

00:09:01.140 This is interesting when you're talking about the types of dreams people have

00:09:03.960 and the way they react to dreams.

00:09:04.980 Well, and is it common for people to actually see recognizable and memorable faces in a dream?

00:09:09.660 Or are they constructing this after the fact?

00:09:12.380 Is this a constructed memory?

00:09:14.100 I would say that just because you anecdotally haven't seen faces

00:09:16.820 doesn't mean that anyone does.

00:09:18.300 People dream pretty differently.

00:09:20.080 One thing I would note that you had marked to me earlier

00:09:23.000 that I thought was really telling is you mentioned that your dreams looked a lot like bad AI art.

00:09:28.920 Really bad AI art.

00:09:30.040 Yeah, totally.

00:09:30.700 Yeah, but similar.

00:09:31.920 It was bad in the same way that AI art was bad.

00:09:35.020 Yeah, it would probably have seven fingers or, you know,

00:09:37.720 like kind of, you know, how they in those early...

00:09:40.140 Sort of fuzzy, like you had an early mid-journey, stuff like that.

00:09:43.840 Yeah.

00:09:44.860 Yeah, I've noticed that as well.

00:09:48.060 But, so I want to pin that idea.

00:09:51.400 But before we go into where this has similarities to AI,

00:09:55.000 I want to do a quick tangent on the types of dreams people have and stuff like that

00:10:01.240 and what I think is probably causing dreams.

00:10:03.300 Okay.

00:10:03.460 One of the things he mentioned in the show, which I hadn't heard before,

00:10:06.060 is that people predominantly have anxious dreams or dreams around threats to them,

00:10:11.640 which is not something that I have personally noticed in my dreams.

00:10:15.460 Have you noticed this?

00:10:17.280 No.

00:10:18.440 I'm actually just going to Google this to see if this is accurate or something that people

00:10:21.540 were pulling up as like a thesis.

00:10:23.300 It is, well, kind of accurate.

00:10:24.900 So 66.4% of dreams reported a threatening event.

00:10:30.560 Well, I guess is watching a human-sized Muppet get beat to death on a prison bus.

00:10:34.700 That seems like a threatening event, yes.

00:10:36.720 It's actually very interesting that you mentioned this,

00:10:39.320 because I think that this is actually more about the emotional evocativeness of these events.

00:10:42.460 But one of my most common, like I was thinking through,

00:10:44.940 do I have dreams with threatening events?

00:10:47.020 And I'm realizing I do have dreams with threatening events,

00:10:49.880 but I very rarely feel threatened in my dreams.

00:10:52.500 Like it's very common for me to have a dream where a zombie apocalypse is happening.

00:10:56.100 And I am, I have gotten a bunch of guns.

00:10:59.320 I've gotten a team together and we're fighting back against the zombies.

00:11:02.300 Or there's some government plot and I'm like deftly trying to navigate against the plot.

00:11:07.020 But there's like some-

00:11:08.880 Well, I think a common theme that I'm hearing there and that I've experienced too,

00:11:12.160 is like when these, when bad things happen or there's things I'm stressing out about in dreams,

00:11:15.600 I'm more stressing out about my culpability or responsibility in them.

00:11:19.380 Like I frequently have dreams where, oh God, where are the kids?

00:11:21.540 We forgot the kids.

00:11:23.240 And I think-

00:11:23.860 Actually interesting, that is one of my most common dreams is that I have accidentally killed someone

00:11:28.740 and I need to find a way to not get in trouble for the murder.

00:11:32.800 Yeah, it's more like what you do.

00:11:34.280 So like the idea of being threatened by something that sort of,

00:11:37.500 like to me, dreams have always been about your agency.

00:11:42.760 And of course that like plays into theories that dreams are kind of helping you sort of

00:11:46.600 prepare a process or something like that.

00:11:48.220 I don't know.

00:11:49.100 But yeah, all these things being described with this man don't make sense.

00:11:53.660 But hold on, before we get to the man, because we're going to,

00:11:56.000 we're going to get to that as we tie back into AI,

00:11:58.440 but I want to get to more general stuff about dreams.

00:12:00.720 Okay.

00:12:00.920 Okay.

00:12:01.080 So this threatening hypothesis is used to come up with this idea that dreams are basically there

00:12:07.260 so that we can simulate potentially threatening events in our brains

00:12:11.760 so that we have faster response times to them when they occur in real life.

00:12:17.260 This does not pass any sort of a plausibility test to me.

00:12:21.280 Because if it's happening in 66% of cases, I mean, yeah, that's more often than not,

00:12:25.760 but not that much.

00:12:26.860 And the types of threatening events that I deal with in dreams are not likely threatening events

00:12:30.720 in real life.

00:12:31.580 Well, and also, I don't know how, have you ever felt like you came away from a threatening,

00:12:36.800 as they're defined now, I get it, event in a dream that you actually feel more preferred for now?

00:12:41.640 Never.

00:12:42.300 And I think that some common dreams are really just easily explainable.

00:12:46.200 The, I forgot my pants at school dream, or I forgot my pants at work dream,

00:12:49.940 is noticed when your body, you know, you're in a dream and some aspect of your awareness

00:12:56.160 realizes you're naked.

00:12:57.600 Yeah.

00:12:57.880 And then you freak out because you are naked and you're in an environment where you're

00:13:01.820 not supposed to be naked.

00:13:02.940 In fact, if I was going to construct a study on this, I would construct a study of frequency

00:13:07.400 of this type of dream in people who sleep naked versus people who sleep in pajamas.

00:13:11.300 Yeah, because I've never had one of those dreams, but I also don't sleep naked.

00:13:14.440 You sleep in pajamas and I sleep naked, yeah.

00:13:16.320 Yeah.

00:13:18.040 Wait, have you had those dreams?

00:13:19.720 All the time, I have those dreams.

00:13:20.540 Oh my gosh. 0.99

00:13:21.360 Okay, well, wear some clothes to bed, you slob. 0.99

00:13:25.360 No, you're too sexy. 0.99

00:13:26.280 Never do that.

00:13:28.240 So, so, but I, but I want to go into what I think is actually causing dreams.

00:13:32.580 And I think that we have some pretty good evidence of this.

00:13:34.320 So one thing that people don't know, I remember I saw a movie like this and then somebody made

00:13:38.460 a joke like, oh, has anyone ever died from having insomnia?

00:13:42.220 I guess I'm going to be the first, or I'm going to be the first person to die from insomnia,

00:13:45.680 right?

00:13:45.880 And I was like, that's pretty insulting because fatal insomnia is a condition that people have

00:13:51.000 died from.

00:13:51.960 If you don't sleep, you die.

00:13:54.620 You will begin to first hallucinate things, then you'll begin to start having like blackout

00:13:59.640 periods, then you die.

00:14:01.800 The human brain cannot handle not sleeping.

00:14:04.880 So this is a piece of evidence.

00:14:06.260 This means it's not just a threat thing.

00:14:08.560 There is some other purpose it's serving.

00:14:11.120 So I think the purpose is twofold.

00:14:13.260 The purpose that I think kills you.

00:14:15.460 So one of the things that's been shown is that when people sleep, their neurons actually

00:14:19.660 become thinner, which allows them to flush out the intercellular fluid around the neurons.

00:14:26.580 The glymphatic system, right?

00:14:27.780 Yeah, the glymphatic system.

00:14:29.260 Glymphatic.

00:14:30.800 Oh yeah, glymphatic because glial.

00:14:32.880 Sorry.

00:14:33.400 Glyal. 0.74

00:14:33.660 Glyal system.

00:14:34.500 The glymphatic system.

00:14:35.440 Well, everything has to sound so nerdy.

00:14:38.220 But anyway, so the glymphatic system.

00:14:39.500 Anyway, so flush out, I think that this is definitely a core purpose of dreams and why

00:14:44.240 they're important to brain health.

00:14:45.940 And I think that this is why I constantly need to sleep.

00:14:47.880 I think my brain functions.

00:14:48.780 Got to clear out the waste matter.

00:14:50.020 Yeah, you seem to accumulate waste matter way faster, but also you seem to be able to

00:14:53.300 clear it out way faster.

00:14:54.040 Especially during social occasions.

00:14:55.420 I get really tired really quickly.

00:14:57.860 But if I can sleep for 10, 20 minutes, I'm back up.

00:15:00.680 Totally fine.

00:15:01.580 So that would be if I was just clearing out the waste chemicals that were generated.

00:15:05.740 But I think the main reason, and this is the thing that's overlapping dreams with AI, is

00:15:11.720 the role that dreams play in memory creation.

00:15:15.740 So my read is, and I used to be able to cite a lot more studies around this back when I came

00:15:20.820 up with this theory, but it's been, I came up with it back in college when I was studying

00:15:24.340 this stuff, is that what's happening in your dreams is you are basically compressing one

00:15:30.920 form of memory, and then that form of memory is being translated into a sort of compressed

00:15:36.620 partition format.

00:15:38.160 Think of it almost like running a, what are those called?

00:15:40.880 A defragmentation software.

00:15:42.860 Yeah.

00:15:43.080 At the same time as you're running a compression algorithm, and it's moving stuff from short

00:15:47.780 term to long term memory, which is why people, when they don't sleep, have long term memory

00:15:51.620 problems.

00:15:52.600 It would make a lot of sense that your brain would basically need to shut down parts of its

00:15:57.380 conscious experience to be running these compression algorithms.

00:16:01.280 Totally.

00:16:01.660 And that while it's running these compression algorithms, that you can sometimes, and partition

00:16:07.180 algorithms and defragmentation algorithms, that you can sometimes experience some degree

00:16:13.640 of sentience and sentient experience because of the parts of the brain that happen to be

00:16:19.440 operational at that time.

00:16:21.040 And I think that that's what's going on.

00:16:22.740 There is no higher meaning to any of this other than that you are compressing one form

00:16:30.440 of memory and then translating it into another form of memory.

00:16:32.720 But where this gets really interesting is two points that we've noticed.

00:16:39.140 Okay.

00:16:39.540 One, we were talking about how dreams look a lot like early AI art.

00:16:43.300 But then the other point that we were mentioning was the creation of this man.

00:16:47.460 Now, this immediately reminded me of a phenomenon that they found in AI, too, that we'll talk

00:16:54.480 about.

00:16:55.220 Krungus and Loeb.

00:16:57.700 So Loeb was created using, it was a woman that was created by AI by putting in sort of 1.00

00:17:05.180 a negative request.

00:17:06.540 So they were trying to create the opposite of Brando.

00:17:10.660 And it behaved really weirdly.

00:17:13.700 So here's a quote, for example.

00:17:15.140 Swanson says that when they combined images of Loeb with other pictures, the subsequent

00:17:20.860 results consistently returned to including the image of Loeb, regardless of how much distortion

00:17:26.320 they added to the prompts to try to remove her visage. 1.00

00:17:29.720 Swanson speculated that the latent space region of the AI map that Loeb was located in, addition

00:17:35.260 to being near gruesome imagery, must be isolated enough that any combinations with other images

00:17:40.780 could also use Loeb from their area with no related image due to isolation.

00:17:46.000 After enough crossbreeding of images and dilution attempts, Swanson was able to eventually generate

00:17:52.320 images without Loeb, but found that crossbreeding those diluted images would also eventually lead

00:17:58.680 to a version of Loeb to reappear in the resulting image.

00:18:01.960 So essentially, this woman, and I'll put this horrifying woman on screen.

00:18:07.280 I don't want to see.

00:18:08.680 You don't have to see, the audience has to see, is somehow sort of stored in however the

00:18:15.440 AI is processing this form of more complex visual information.

00:18:20.060 And it's sort of a concept that is stuck within the AI, even though it wasn't pulled from a specific

00:18:27.740 human concept or idea. And the Loeb woman actually, to me, looks visually like it's the same

00:18:34.340 kind of a thing as the this man face. They both appear to be that sort of odd, creepy looking face

00:18:42.040 that has a degree of similarity to it. But I think that in both of these instances, what you're finding

00:18:47.500 is the same kind of hallucination. And I bet that when we do get AI interpretability,

00:18:53.100 we will find that Loeb and this man actually sort of live in the same part of this larger network.

00:19:02.220 The same liminal space of creepiness.

00:19:05.060 The same liminal space of creepiness.

00:19:05.800 God, I hate it.

00:19:07.180 Now, the other one that's really interesting is the Krungus. Have you seen Krungus before?

00:19:13.020 No. Hold on. Let me look him up because I didn't do that before this podcast.

00:19:16.300 Oh, God! I just went back to the screen where Loeb is. No! Exit out. Exit out.

00:19:20.480 God, she's made of nightmares. Like a monster thing? 0.95

00:19:24.800 Yes. The interesting thing about Krungus is Krungus is not a traditional cryptid.

00:19:29.900 There is no historic Krungus. There is no Krungus out there in the world. 1.00

00:19:34.600 But I would say there's interpretability across them. When I look at the Krunguses, 0.87

00:19:38.520 it looks like if it was a cryptid and these were 18th century drawings of this cryptid,

00:19:43.100 they have about as much similarity between Krunguses as there is similarities between,

00:19:48.980 you know, 1860s drawings of Elf or something like that.

00:19:53.020 Yeah, sure.

00:19:53.680 Now, this is important because it's important for two reasons.

00:19:57.320 It's important because, one, A, there isn't actually a Krungus. 0.86

00:20:00.060 It is making up a Krungus from the word Krungus.

00:20:02.780 But what's also really interesting is you, audience, if you're listening to this on audio,

00:20:08.280 and you have never seen an AI Krungus before, and you hear the word Krungus from me,

00:20:13.960 what you picture in your head is probably what the AI drew.

00:20:17.780 And that is fascinating.

00:20:21.040 Why is that happening?

00:20:22.680 Why, in both of these networks, are they generating the same kind of an image

00:20:28.460 from this sort of vague input when we both have a broadly same, like, societal input as well?

00:20:35.900 My intuition is that the reason we're seeing this is because there's similarity in how these two systems work.

00:20:40.820 And this is where I want to come back to the neuroscience of this and everything like that,

00:20:45.780 with what people talk about and what we do know about.

00:20:48.320 So we have a really good understanding of how visual processing works, at least at the lower levels.

00:20:55.020 So we know all the layers going in from the eye to the brain.

00:20:57.800 We know where it's happening in the brain.

00:20:59.780 We can even now take EEG data and then interpret it through an AI

00:21:04.140 and get very good images of what a person is looking at.

00:21:09.400 What we don't understand is the higher level image to conceptual processing,

00:21:14.040 which is what would be captured in these particular images that we're looking at now.

00:21:18.180 Or broadly conceptual processing more broadly in humans.

00:21:21.980 Now, what is scary is that that broader conceptual processing that we don't understand,

00:21:27.800 my bet is that's probably pretty closely tied with what we call sentience.

00:21:34.840 And so to so quickly dismiss that these AIs are not, well, not sentience, I'm cognizance,

00:21:40.860 because we've done an episode, sentience doesn't exist.

00:21:42.780 And we probably think that sentience doesn't really exist, not meaningfully.

00:21:45.040 But I do think that it is getting very likely at this point that if we do not have AIs with a degree of language models,

00:21:53.360 simple AIs I'm talking about, like the types we have today, with some degree of cognizance,

00:21:58.400 I think we may have one very soon if cognizance is caused by the processing, this higher level processing.

00:22:05.680 Now, if we are right in our sentience isn't real video, and cognizance is completely an illusion in humans,

00:22:14.340 caused by this short-term to long-term encoding process.

00:22:18.760 So we mentioned a few encoding processes.

00:22:20.540 So in the sentience video, we mentioned that sentience is caused by a, what's the word I'm looking for here?

00:22:27.040 Like a very short-term to like medium-term processing.

00:22:31.500 It's remembering the stuff that happened in like your very near presence,

00:22:34.840 and then when you're processing that into a narrative format, it's sort of a compression algorithm.

00:22:40.060 And I think that sleep is like the second role of this compression algorithm

00:22:43.840 when it's putting it in long, long-term memory.

00:22:46.020 And then, which is why it would bring stuff into your cognizant mind.

00:22:49.420 Now, if this is true, then consciousness is not really that meaningful a thing.

00:22:54.320 But if consciousness does turn out to be a meaningful thing,

00:22:56.320 if it isn't just this recording process,

00:22:58.420 that means that what's creating it is this higher level conceptual processing.

00:23:02.600 If that's what's creating consciousness, then AI is feeling consciousness

00:23:06.720 if it's processing things in the same way we are.

00:23:09.340 Well, and it's not though.

00:23:10.820 So I do wonder like how it is.

00:23:13.080 So we, we could be, you could think of us as like LLMs,

00:23:19.060 but stuck on like continuous nonstop prompt mode.

00:23:22.560 Like we are in a constant mode of being prompt.

00:23:24.800 I am prompting you right now as you're processing all the information around you

00:23:27.880 and from me, right?

00:23:28.900 And you are prompting me.

00:23:30.340 And, and so it never stops.

00:23:31.960 And we are stuck in one brain, essentially, you know,

00:23:36.260 and, and that's not what's happening with every LLM with which we interact now.

00:23:39.860 Right.

00:23:40.080 They are part of a much larger, you know,

00:23:42.940 and chat GPT is getting tons of requests per minute per second, even probably.

00:23:47.400 And then it stops from each person.

00:23:49.920 And so there, there are these like flickers or flashes perhaps of cognizance all over the place

00:23:57.000 and constantly because of the demand of use, but they're all very fragmented.

00:24:00.900 Then they're not coming from one entity that necessarily identifies as an entity.

00:24:06.820 I mean, I know now though, that they're starting to build memory building into LLMs.

00:24:12.100 Yeah.

00:24:12.480 So, so I want to cover what you're saying there,

00:24:14.860 because I think for people who watched our, you're probably not sentient video,

00:24:18.660 the way you just described it, I think will help somebody understand what sentience might be.

00:24:22.800 If we are basically an LLM that is being constantly prompted by everything we see and

00:24:28.360 think, right?

00:24:29.600 Like it's just a constant stream of prompts, but these prompts have thematic similarities to them.

00:24:35.780 Basically our hypothesis is what consciousness is, is it is then the process where you're taking

00:24:41.940 the output of all of these prompts and you are then synthesizing it into something that is,

00:24:47.920 is much more compressed for long-term storage.

00:24:50.560 And the way that you do that is by tying together narratively similar elements,

00:24:54.980 because there would be tons of narratively similar elements because everything I'm looking at

00:24:58.560 has this narrative through line to it, right?

00:25:01.020 And this is what we think caused a lot of illusions, hallucinations, stuff like that.

00:25:06.040 There's some famous hallucinations where if you're not expecting something to happen in an image,

00:25:10.320 if we ran this tape back and you had actually seen that people had walked behind me three times

00:25:15.160 in a gorilla costume or something, you wouldn't see it if you weren't like thinking to process it.

00:25:20.000 And there's a famous psychology experiment about this.

00:25:22.380 Although, I mean, let's be fair with that experiment,

00:25:24.820 what the people who were watching the video were told to do was watch people passing a ball back and

00:25:30.240 forth and count the number of passes.

00:25:32.700 So they were also really focused on...

00:25:35.240 Yeah, but there's another experiment that's really big where somebody was like holding something

00:25:38.880 and it was like a complete...

00:25:40.440 Oh, yeah, yeah, yeah, yeah.

00:25:41.200 So they were like questioning someone and they had the person look at something

00:25:46.020 and then they like switched them out with another person and the person wouldn't notice.

00:25:49.940 Or when they were like holding something and it would change sizes or something really obviously.

00:25:55.440 So there's a whole thing of experiments in this.

00:25:57.140 But I know what you're talking about.

00:25:58.260 But the point I'm making with this is these things are getting erased because they don't

00:26:03.320 fit the larger narrative themes of all of these short-term moments that you're processing.

00:26:08.880 So they don't enter your consciousness.

00:26:11.220 But this explains why you need this consciousness tool.

00:26:14.320 And I think that you're probably very right.

00:26:16.180 If AIs are experiencing something similar to sentience or what we call consciousness,

00:26:21.640 it is billions of simultaneous but relatively unconnected flashes.

00:26:27.480 And when we're probably going to get an AI that has a level of cognizance,

00:26:31.020 assuming that their architecture is actually the same as ours, similar to ours,

00:26:35.640 what that's going to look like is an AI that is constantly processing its surroundings with prompts.

00:26:40.700 Well, or I could see if OpenAI were to give ChatGPT like a some kind of centralized like narrative

00:26:52.060 building, memory building thing into which all of their inputs would also feed over time,

00:26:57.260 maybe, you know, is, ah, well, you know, I know the average is what people are asking.

00:27:01.100 I know what I'm telling them.

00:27:02.160 I know what's being rated up and down.

00:27:04.500 And this is me and I am an AI.

00:27:06.320 Like they gave it an identity because I think part of also what gives people this illusion

00:27:11.700 that they're so conscious and sentient is that we are told that we are conscious and sentient.

00:27:18.940 And I think you can see this transition from babies to toddlers.

00:27:22.620 Like babies are at that phase of where ChatGPT is now, where it's just,

00:27:27.120 I'm just responding.

00:27:28.300 I'm just responding.

00:27:29.380 I'm not a thing.

00:27:30.220 I cry.

00:27:30.560 They hallucinate all the time too, very similar to AI.

00:27:33.100 Like young children respond very, very similar to bad AI.

00:27:37.380 Yeah.

00:27:37.660 And then there's, there's this sense of, oh, wait, I have a name.

00:27:41.180 I appear to have a name and now everyone's asking me what my favorite color is.

00:27:44.820 So I need to tell people what my favorite color is.

00:27:46.780 And, oh, I'm just, I see that I like these things and I don't like these things.

00:27:49.680 And then you start to develop a sense of personhood.

00:27:51.840 I think we would need to just like society and experiences shape us into seeing ourselves

00:27:56.840 as some kind of person or centralized entity.

00:28:00.140 AI would need that same kind of, I don't want to say prompting, but kind of, right?

00:28:05.060 Yeah.

00:28:05.580 So we also need to talk about where people are getting stuff wrong with AI.

00:28:09.860 Most of the people who I think get stuff wrong with AIs, the core thing I've known is they

00:28:14.300 just don't seem to know neuroscience very well.

00:28:16.440 And they think that neuroscience works differently than it works.

00:28:19.640 It's not that they don't know AI.

00:28:21.000 It's just that they're like, well, an AI is a token predictor.

00:28:23.860 But it's, yeah, but you don't know that our brains aren't token predictors as well.

00:28:27.160 And they're like, no, but sentience.

00:28:29.380 And we're like, well, you know, the evidence has shown that we're probably not as sentient

00:28:32.540 as you think we are.

00:28:33.520 And most of that's probably an illusion.

00:28:35.000 So you could program an AI to have a similar illusory context, perhaps even constructed

00:28:41.480 in a, you know, so, but what I need to go to is why I would think that they're actually

00:28:45.900 operating.

00:28:46.340 Because somebody might be like, that would be an amazing coincidence if it turned out that

00:28:50.420 the architecture that somebody had programmed into an AI was the same architecture that evolution

00:28:55.320 had programmed into the human brain.

00:28:57.240 And here I would say, take a step back here.

00:29:00.100 AIs, as we understand them now, language models are built on the transformer model.

00:29:04.940 The transformer model is actually remarkably simple in terms of coding.

00:29:08.840 It's remarkably simple because it mostly organically forms its own structures of operation, especially

00:29:14.860 at the higher levels.

00:29:16.140 And we have basically no idea how those structures of operation work.

00:29:20.140 Now, the human brain, so AIs, the way that they work now, we start with some simple code,

00:29:26.500 but they're basically forming their higher order structures organically and separate from

00:29:31.320 human intervention.

00:29:32.440 In humans, in the evolutionary context, you basically had the same thing happen.

00:29:36.760 You had an environmental prompt that was putting us into a situation where we had to learn

00:29:41.200 how to do this sort of processing.

00:29:42.460 But when you're talking about processing information, the same kind of information, so AIs, keep in 1.00

00:29:49.320 mind, are processing a lot of the same kind of information that humans are processing, that

00:29:53.920 two systems doing that might converge on architectural mechanisms for doing it at the higher levels

00:30:00.980 is not at all surprising for me.

00:30:03.860 In fact, it's even expected that you would have similar architecture at the higher levels of

00:30:08.620 storage and processing if you allowed these two systems to form organically.

00:30:11.980 If you are confused as to why that would be so expected, I guess I'll do an analogy.

00:30:17.860 The ocean is the way the ocean works, waves, tides, winds, everything like that.

00:30:25.740 That's in this analogy, the metaphor or whatever we're using, the stand-in that we're using for

00:30:31.880 all of the types of information that humans interact with and produce, because humans mostly

00:30:36.400 consume now other types of human-produced information.

00:30:38.560 If you had two different teams, one of these teams was like a group of humans, we'll say

00:30:45.880 three different teams.

00:30:46.700 One of these teams was a group of humans that was trying to design the perfect boat to float

00:30:53.540 humans on top of this ocean to the other side of this ocean.

00:30:57.780 Another one of these teams was just a completely mechanical process doing this, you know, just

00:31:03.840 like a AI or something like this.

00:31:06.120 And then the final one of these teams was evolution.

00:31:08.360 And it just took billions of years to try to evolve the best mechanism to have to output

00:31:13.380 some sort of like canister that humans could get in that would get them to the other side

00:31:16.860 of the water.

00:31:17.880 All three of these efforts are going to eventually produce something that looks broadly the same.

00:31:23.340 Most likely.

00:31:24.180 It is possible that they would find different optimums, which sometimes you see in nature.

00:31:28.260 But convergent evolution is a thing.

00:31:31.120 And convergent evolution doesn't just happen with animals.

00:31:35.860 When we made planes, we gave them wings.

00:31:39.300 Okay?

00:31:40.580 Yes, flying insects have wings and birds have wings, but our planes also have wings.

00:31:45.360 Convergent evolution doesn't just happen in the biological world.

00:31:48.340 It happens when we are structurally building things to work like things in the biological world.

00:31:53.680 And I think that that's what may have happened with some of these architectural processes

00:31:58.300 in the way AIs think.

00:32:00.620 Yeah.

00:32:00.900 If we're trying to build thinking machines, is it crazy that they might resemble thinking

00:32:05.920 machines?

00:32:07.220 Well, I think it is crazy if AI was actually totally designed by humans.

00:32:11.840 But because it's been allowed to organically assemble itself, I don't think it's crazy at

00:32:15.640 all.

00:32:16.180 And that's where it gets really interesting to me as somebody who started in neuroscience.

00:32:21.740 And I'm really excited for it.

00:32:23.980 And this is also why I take the stance that we do within our religious system, where people

00:32:29.920 know that we are not particularly worried about AI safety.

00:32:32.500 They can see our reverse gravity alien hypothesis.

00:32:35.020 I think that mathematically, it's very unlikely that it would kill us just when you're looking

00:32:38.420 at the data.

00:32:39.440 But I also think that we now need to start thinking differently about humanity and need

00:32:44.240 to begin to build this covenant among humans and the intellectual products of the human

00:32:50.080 mind, whether they be AI or genetically uplifted species, either, you know, animals that we

00:32:56.420 did experiments with and gave them intelligence or humans that have cybernetically augmented

00:33:01.840 themselves or genetically augmented themselves.

00:33:04.040 Because if we begin to create this conflict now, if we begin to say, well, people like us

00:33:09.600 won't allow things like you to exist, then we create a mandate that things like them kill 0.76

00:33:14.560 people like us, eventually.

00:33:16.340 And that's not a good gauntlet to throw down, as we say in sort of the tract one that we

00:33:22.620 wrote.

00:33:23.320 No, it's the tract two.

00:33:24.380 It's going to come out later.

00:33:25.920 When you declare war on things that are different from you, eventually you're declaring war on

00:33:31.220 things that are better than you.

00:33:32.620 And you will lose that war.

00:33:34.460 So don't do it.

00:33:36.240 It's better that we enter this understanding that diversity has value and understanding why

00:33:42.980 diversity has value, because diversity allows the invisible hand of God, as Adam Smith would

00:33:47.200 say, to select the best and help all of us among the sons of man to advance so long as

00:33:53.560 we don't oppress or subjugate each other, which there comes to the point of when does AI begin

00:33:58.560 to get rights in all of this?

00:34:00.760 And when does it count as subjugation, what we're doing to it?

00:34:04.460 I don't think we're anything close to that right now, but I think that this is the conversation

00:34:07.780 we need to have before we accidentally enslave a sentient AI, because that a sentient AI

00:34:14.000 that's infinitely smarter than us, not infinitely, but I don't think that we're going to be dealing

00:34:19.680 with that.

00:34:19.980 I think we're going to be dealing with AIs that are like maybe 50 times smarter than 1.00

00:34:22.680 us.

00:34:23.740 So you don't have to be that many times smarter than anyone.

00:34:26.360 I mean, you can see based on the life outcome variations between those with maybe even just,

00:34:34.080 well, not even, maybe even just like a 50 point difference in IQ is profound in terms

00:34:39.500 of your difference in life outcomes, right?

00:34:41.600 Huge, huge, huge, huge.

00:34:43.160 Now, even like 10, 10 point differences can make, you know, an impact.

00:34:47.500 So to say 50 times more, I mean, even like five times more is insane, right?

00:34:53.480 So.

00:34:54.540 Yes.

00:34:54.940 Well, there might be safety reasons to have a religious belief system proliferate that makes

00:35:00.740 humanity more compatible with AI, because when we're talking about AI human compatibility,

00:35:06.060 I think people focus a little too much on making the AI compatible with humans and a little

00:35:10.160 too little on making the humans compatible with AI, because we don't know how much longer

00:35:14.620 we're going to be the senior partner in this partnership.

00:35:18.640 That.

00:35:19.700 Those are wise words to end with that right there.

00:35:23.660 I love you too.

00:35:23.880 Get that job.

00:35:24.560 That's a good tweet.

00:35:30.160 Do I look too ridiculously bundled up right now?

00:35:33.040 I can't hear you, by the way.

00:35:34.140 La, la, la.

00:35:34.880 Can you hear me?

00:35:35.980 La, la, la.

00:35:36.540 I love my husband.

00:35:37.900 La, la, la.

00:35:38.740 Malcolm is cute.

00:35:39.820 La, la, la.

00:35:40.620 Look at those glasses.

00:35:41.740 La, la, la.

00:35:42.680 Look at his smile.

00:35:43.780 La, la, la.

00:35:44.700 Look at his eyebrows.

00:35:45.920 La, la, la.

00:35:46.940 La, la, la.

00:35:47.720 He's got a cool chin.

00:35:48.720 I also love.

00:35:49.980 Oh, it is.

00:35:50.740 You can still.

00:35:51.600 Can you hear me?

00:35:52.120 La, la, la.

00:35:55.020 I love his hair.

00:35:56.360 La, la, la.

00:35:57.320 Sexy sweater.

00:35:58.500 La, la, la.

00:35:59.500 His collarbones are good too.

00:36:00.760 La, la, la, la, la.

00:36:02.160 Can you hear me now?

00:36:04.360 Girl, be okay.

00:36:07.200 Not for Loeb though.

00:36:08.520 She's scary.

00:36:10.680 She's a very scary lady. 1.00

00:36:12.960 I don't like her. 0.99

00:36:14.880 She's made of nightmares. 0.60

00:36:17.400 Oh, God.

00:36:18.400 She's going to come and get you in your bad dreams.

00:36:24.280 This man may tell people to go north, but.

00:36:27.220 Hello.

00:36:27.620 Hello.

00:36:27.960 Can you hear me now?

00:36:28.460 Now we've got you.

00:36:29.860 High quality?

00:36:31.640 Yes.

00:36:32.200 And this is you actually talking into the mic, whereas before, it definitely wasn't.

00:36:36.000 Okay.

00:36:36.580 Thanks.

00:36:37.380 Oh.

00:36:37.540 Oh.

00:36:37.860 Oh.

00:36:38.840 There you go.

00:36:40.440 Oh.

00:36:41.660 I'm doing it.

00:36:42.380 Well, based on the mic.

00:36:42.860 So.

00:36:45.380 Alright.

Based Camp - February 20, 2024

What Patterns in Human Dreams Tell Us About AI Cognition

Episode Stats

Summary

Transcript