Latest / Mahon McCann / Sam Tideman - Is Artificial General Intelligence even possible??

Transcript

Sam Tideman - Is Artificial General Intelligence even possible?? · Mahon McCann

0:00
SPEAKER_02: Sam Tiedemann, thank you very much for joining me today.
0:03
SPEAKER_02: It is a great honor to finally meet you.
0:05
SPEAKER_02: As I was saying before,
0:06
SPEAKER_02: I'm a big fan of your work on your channel,
0:09
SPEAKER_02: Transfigured,
0:10
SPEAKER_02: and can follow you for a long time.
0:11
SPEAKER_02: So it's great to actually get to meet you and speak about some of these topics.
0:16
SPEAKER_00: Well, pleasure is mine.
0:17
SPEAKER_00: Thanks for having me.
0:19
SPEAKER_02: And so,
0:20
SPEAKER_02: yeah,
0:20
SPEAKER_02: to start off with,
0:21
SPEAKER_02: you don't seem to be a big fan of the AI super god that's going to fix everything
0:27
SPEAKER_02: and bring about paradise on Earth.
0:29
SPEAKER_00: Yeah,
0:30
SPEAKER_00: no,
0:30
SPEAKER_00: I'm skeptical of that god,
0:34
SPEAKER_00: both that god's existence and its ability to deliver on the promises.
0:37
SPEAKER_02: Well, me too, to be fair.
0:40
SPEAKER_02: And I'm interested in your skepticism because obviously you're working in machine
0:44
SPEAKER_02: learning and you've built systems of this kind before in artificial intelligence.
0:50
SPEAKER_02: And so I'm kind of interested, where do you see the limits of that?
0:54
SPEAKER_02: I know you described ChatGBT as a crossword puzzle solver.
0:59
SPEAKER_00: Yeah.
1:00
SPEAKER_02: Do you think there are limits?
1:02
SPEAKER_00: I think that there definitely are limits,
1:04
SPEAKER_00: although I should say when I'm describing limits,
1:08
SPEAKER_00: I'm describing limits about artificial intelligence as we've kind of seen it and
1:14
SPEAKER_00: experienced it so far.
1:17
SPEAKER_00: Part of me sort of leaves open that there could be,
1:19
SPEAKER_00: you know,
1:20
SPEAKER_00: categorical shifts in innovation that could change things in the future.
1:26
SPEAKER_00: But at least as far as I've seen things and thinking about what our current sort of
1:31
SPEAKER_00: forms of artificial intelligence are capable of and not capable of,
1:36
SPEAKER_00: I really think that at the end of the day,
1:39
SPEAKER_00: they can't
1:41
SPEAKER_00: really figure out purpose and value,
1:46
SPEAKER_00: and that this is sort of going to be ultimately their limitation,
1:53
SPEAKER_00: that still means that they can be very powerful and that they can do lots of things.
1:58
SPEAKER_00: And I think that one of the big dangers is not so much the AI themselves.
2:05
SPEAKER_00: Themselves is AI plural or singular.
2:07
SPEAKER_00: That's a good question.
2:09
SPEAKER_00: For now,
2:09
SPEAKER_00: I think it's as I think that humans will sort of misunderstand and misapply them
2:15
SPEAKER_00: and that that is probably a greater risk than what the machines might do.
2:21
SPEAKER_00: And I think that we think as soon as something that we're interacting with is intelligent, quote unquote,
2:28
SPEAKER_00: or able to generate text in sentences that make sense.
2:32
SPEAKER_00: That it must have all of the other things going on inside my head that I have going on inside my head.
2:38
SPEAKER_00: And I think that this is just not true.
2:41
SPEAKER_00: And I think part of that is in our modern world is really just intelligence.
2:46
SPEAKER_00: And that that's sort of its main function,
2:49
SPEAKER_00: as opposed to processing information,
2:52
SPEAKER_00: filtering out information,
2:54
SPEAKER_00: attributing value,
2:55
SPEAKER_00: figuring out value,
2:57
SPEAKER_00: you know,
2:57
SPEAKER_00: striving after purposes,
2:59
SPEAKER_00: having motivations.
3:02
SPEAKER_00: Having frameworks,
3:03
SPEAKER_00: having fears,
3:04
SPEAKER_00: all those sorts of things are just not really at play at all in artificial intelligence.
3:11
SPEAKER_00: I really think that is really just some small fraction of what our brain can do,
3:16
SPEAKER_00: some small subfunction of what our brain can do,
3:19
SPEAKER_00: amplified to enormous proportions.
3:22
SPEAKER_00: And it's as if we've made the mistake that that small part of our brain is like the
3:28
SPEAKER_00: only thing that our brain does.
3:30
SPEAKER_00: And so if we hook that part of our brain up to supercomputers and giant amounts of
3:35
SPEAKER_00: electricity and training data,
3:37
SPEAKER_00: that'll be better than us.
3:38
SPEAKER_00: But it's still, there's just giant gaps in terms of what it can and cannot do.
3:44
SPEAKER_02: Yes.
3:45
SPEAKER_02: And I mean,
3:45
SPEAKER_02: there's so many points there to jump on,
3:47
SPEAKER_02: but just in terms of the idea of values and ends and that we're in danger of
3:52
SPEAKER_02: mistaking it for something that it's not like that does seem to be the kind of emerging,
3:57
SPEAKER_02: because if you think you're building this super intelligence,
4:00
SPEAKER_02: that's going to be a conscious being.
4:02
SPEAKER_02: And if you go in with that presupposition,
4:05
SPEAKER_02: it's very easy to trick yourself into believing you've created something like a
4:08
SPEAKER_02: chatbot or something that
4:09
SPEAKER_02: mirrors our relevance realization but doesn't actually have any of its own and
4:14
SPEAKER_02: right my research is on social media so i study that and i mean a big part of that
4:19
SPEAKER_02: really is like the ai content creation algorithms and how that's been set up
4:25
SPEAKER_02: And people talk a lot about the algorithm and about the issues of the algorithm.
4:29
SPEAKER_02: But really,
4:30
SPEAKER_02: it's the company that's behind the algorithm that's leveraging it in order to
4:34
SPEAKER_02: profit from attention capture.
4:36
SPEAKER_02: That's kind of put the lead in the water, so to speak, of some of those platforms.
4:41
SPEAKER_02: And so do you see it being a case that the companies behind the things have the
4:46
SPEAKER_02: potential to be mistaken here and to set up something?
4:51
SPEAKER_02: Are we setting this up for the general public as in the wrong way,
4:55
SPEAKER_02: that this is possible when it's really not?
4:56
SPEAKER_02: It's not going to be a thing.
4:59
SPEAKER_00: Like with these sort of large language models,
5:01
SPEAKER_00: and I should say,
5:02
SPEAKER_00: we also forget large language models like ChatGPT or Gemini or the other ones that
5:08
SPEAKER_00: are coming out.
5:09
SPEAKER_00: This is just one of the things that AI can do.
5:12
SPEAKER_00: It's not like this is all of AI.
5:15
SPEAKER_00: It's really, I would say...
5:17
SPEAKER_00: Language,
5:17
SPEAKER_00: large language models are a subfield or a sub function within AI,
5:22
SPEAKER_00: but they have a lot of our attention at the moment.
5:24
SPEAKER_00: And I do think that's for good reason.
5:26
SPEAKER_00: I mean,
5:26
SPEAKER_00: the human world is so language based and language interface that if you can make
5:33
SPEAKER_00: something that can interact with language,
5:35
SPEAKER_00: then the limits of what that thing could do are pretty large.
5:40
SPEAKER_00: But the question is, and this gets back to what we were talking about even already, is purpose.
5:46
SPEAKER_00: What are we using these large language models for?
5:52
SPEAKER_00: And without that,
5:53
SPEAKER_00: it's a little bit hard to know what they should be designed,
5:57
SPEAKER_00: how they should be designed to behave.
6:00
SPEAKER_00: Like,
6:00
SPEAKER_00: you know,
6:01
SPEAKER_00: Google got in trouble for or got in trouble or got embarrassed,
6:05
SPEAKER_00: whichever way you want to say it,
6:06
SPEAKER_00: for the picture generating AI that was making,
6:11
SPEAKER_00: you know,
6:11
SPEAKER_00: diverse casts of Vikings and multiracial founding fathers of the United States and
6:17
SPEAKER_00: whatever else that it was.
6:19
SPEAKER_00: So what are you using that image generator for?
6:23
SPEAKER_00: What's your purpose?
6:24
SPEAKER_00: If your purpose is to make historically accurate illustrations in a history textbook,
6:30
SPEAKER_00: well,
6:30
SPEAKER_00: then it was doing very badly at that.
6:34
SPEAKER_00: But like,
6:34
SPEAKER_00: imagine you're trying to,
6:36
SPEAKER_00: you know,
6:36
SPEAKER_00: make an invitation for your kid's pirate themed birthday or Viking themed birthday party.
6:43
SPEAKER_00: And if it made,
6:44
SPEAKER_00: you know,
6:45
SPEAKER_00: racially and gender inaccurate Vikings,
6:47
SPEAKER_00: it's like,
6:47
SPEAKER_00: okay,
6:48
SPEAKER_00: because it's like for a birthday party invitation.
6:51
SPEAKER_00: And so it depends what you mean to use it for.
6:55
SPEAKER_00: So if you want historical accuracy, then it was doing a bad job.
7:00
SPEAKER_00: If you want to just make everybody feel included regardless of race or gender,
7:05
SPEAKER_00: then that's kind of what it was designed for.
7:07
SPEAKER_00: But the question is, what is it for?
7:09
SPEAKER_00: What are you using the picture generating thing for?
7:12
SPEAKER_00: What are you using the text generating thing for?
7:14
SPEAKER_00: And this is so determinative of what we mean by it being successful or not.
7:20
SPEAKER_00: And I think this is just the question that we keep failing to ask in this AI age is
7:25
SPEAKER_00: what are we using these things for?
7:27
SPEAKER_00: What is their purpose?
7:29
SPEAKER_00: And then having their limitations,
7:31
SPEAKER_00: design functions,
7:33
SPEAKER_00: parameters,
7:33
SPEAKER_00: specs,
7:34
SPEAKER_00: whatever,
7:34
SPEAKER_00: flow down from that kind of central focus.
7:38
SPEAKER_02: That 100%.
7:38
SPEAKER_02: I mean,
7:39
SPEAKER_02: it was something I was thinking about for this conversation is just that question
7:42
SPEAKER_02: of utility function,
7:43
SPEAKER_02: because it's something that seems to be really missing from the AI conversation,
7:47
SPEAKER_02: which is insane if you're thinking about just creating such powerful technologies,
7:51
SPEAKER_02: but with no idea what their actual function is.
7:53
SPEAKER_02: I saw...
7:54
SPEAKER_02: I don't know if you know, Yosha Bach, who talks a lot about AI stuff.
7:58
SPEAKER_02: And he said,
7:59
SPEAKER_02: basically,
7:59
SPEAKER_02: we're going to build the super intelligence and then it's going to tell us what its
8:02
SPEAKER_02: utility function is going to be because it's going to be so much smarter than us.
8:06
SPEAKER_02: And I was like,
8:06
SPEAKER_02: that seems like a very dangerous idea because you're like,
8:10
SPEAKER_00: you know,
8:10
SPEAKER_00: the cart before the horse,
8:12
SPEAKER_00: like an ant is not particularly intelligent,
8:15
SPEAKER_00: but it has a purpose.
8:17
SPEAKER_00: And so purpose, I would say purpose and intelligence are almost completely independent or orthogonal.
8:24
SPEAKER_00: And we don't even use our intelligence as best as I can tell to figure out the purpose of things.
8:30
SPEAKER_00: I think that that is some entirely different ability that we humans have to both
8:36
SPEAKER_00: kind of make purposes for ourselves and guess the purpose of other people,
8:40
SPEAKER_00: guess the intention of other people,
8:42
SPEAKER_00: or even guess the intention of things and objects.
8:46
SPEAKER_00: I think that's just a completely different function that's almost completely
8:49
SPEAKER_00: orthogonal or unrelated to intelligence itself.
8:54
SPEAKER_00: So I think we could make the most verbally talented large language model that we could possibly imagine.
9:00
SPEAKER_00: But as long as it was made the same way that we're making our current ones,
9:03
SPEAKER_00: I don't think it would have any ability whatsoever to meaningfully answer the
9:07
SPEAKER_00: question of what it's for and what its purpose is.
9:11
SPEAKER_02: And our categorizations are things that are functional, essentially.
9:15
SPEAKER_02: And so the danger of creating things that don't have that categorical structure is
9:18
SPEAKER_02: that we don't know where they fit in and they become,
9:21
SPEAKER_02: you know,
9:22
SPEAKER_02: ends in themselves.
9:22
SPEAKER_02: So do you think there's a religious question here?
9:25
SPEAKER_02: It really seems to be very connected.
9:27
SPEAKER_00: I think there is.
9:28
SPEAKER_00: I think there is a religious question.
9:30
SPEAKER_00: Maybe before I jump up to that question.
9:32
SPEAKER_00: I mean,
9:33
SPEAKER_00: the companies themselves will have to find purposes for them because that's the
9:37
SPEAKER_00: only way that they'll be able to make money from them.
9:40
SPEAKER_00: Because,
9:41
SPEAKER_00: you know,
9:42
SPEAKER_00: having a purpose and making a living and being profitable are all really kind of
9:46
SPEAKER_00: ways of saying the same thing.
9:48
SPEAKER_00: I mean,
9:48
SPEAKER_00: our economic system is at least striving to have the function of making money be
9:54
SPEAKER_00: the same thing as the function of having a good purpose for the greater good.
9:58
SPEAKER_00: It doesn't work all the time,
10:00
SPEAKER_00: obviously,
10:01
SPEAKER_00: but that's some caveats and some counter examples of self-sustaining purpose.
10:07
SPEAKER_00: And that's that's how you're providing a living for yourself.
10:10
SPEAKER_00: And hopefully that benefits other people in the in the meantime.
10:14
SPEAKER_00: And so,
10:15
SPEAKER_00: you know,
10:15
SPEAKER_00: Google,
10:16
SPEAKER_00: Microsoft,
10:17
SPEAKER_00: OpenAI,
10:18
SPEAKER_00: Facebook,
10:19
SPEAKER_00: Twitter,
10:20
SPEAKER_00: et cetera,
10:20
SPEAKER_00: will all have to figure out purposes.
10:23
SPEAKER_00: for these things in order to make money.
10:27
SPEAKER_00: Right now,
10:27
SPEAKER_00: it seems like the main business cases get people to subscribe to using this thing
10:33
SPEAKER_00: because they're curious about how it works.
10:36
SPEAKER_00: I suspect that that business model will wear off in the not too distant future when
10:42
SPEAKER_00: people's curiosity is no longer struck by this.
10:46
SPEAKER_00: But, you know, you could imagine like, okay,
10:49
SPEAKER_00: It would be really nice if you could turn a doctor's shorthand into a full-fledged
10:55
SPEAKER_00: doctor's note and use the information from a patient's encounter that's captured in
11:00
SPEAKER_00: the computer and turn that into some high-quality documentation.
11:05
SPEAKER_00: that gets saved with that patient encounter.
11:07
SPEAKER_00: Okay, well, suddenly that is a use case that these models are being put to.
11:12
SPEAKER_00: And then you could imagine it being good or not good at that.
11:16
SPEAKER_00: And then you having to kind of tweak the parameters and the behavior and the
11:19
SPEAKER_00: training data and et cetera,
11:21
SPEAKER_00: et cetera,
11:22
SPEAKER_00: so that it gets better and better at that.
11:24
SPEAKER_00: And what I think is that there will be many different versions of these language
11:31
SPEAKER_00: models that are tuned for particular applications.
11:35
SPEAKER_00: And that those applications,
11:37
SPEAKER_00: or I should say purposes is really just another way of saying application,
11:41
SPEAKER_00: will help determine what they're like.
11:45
SPEAKER_00: And even then,
11:45
SPEAKER_00: it will be a human that will be,
11:48
SPEAKER_00: or a team of humans,
11:49
SPEAKER_00: a team of engineers and software designers and what have you,
11:53
SPEAKER_00: that are making the decisions to try and alter the behavior of these things so that
11:58
SPEAKER_00: they're suited for purpose.
12:00
SPEAKER_00: And I should also say, at least as they exist now, these large language models are extremely expensive.
12:06
SPEAKER_00: Like people have no idea how expensive they are to build these things.
12:11
SPEAKER_00: So right now they will only be useful for extremely profitable or extremely large
12:18
SPEAKER_00: economic value projects.
12:20
SPEAKER_00: And I'm not entirely sure that I know what those are yet,
12:24
SPEAKER_00: but I assume I will be surprised that people will find something.
12:27
SPEAKER_00: That is sort of why I'm not entirely fearful of the technology that we see now in
12:34
SPEAKER_00: the meantime,
12:35
SPEAKER_00: because in order for someone to keep paying the bills that these things require to
12:40
SPEAKER_00: be paid,
12:41
SPEAKER_00: They will need to find ways of making a living.
12:43
SPEAKER_00: And I think for the most part,
12:45
SPEAKER_00: they'll be relatively benign,
12:47
SPEAKER_00: kind of like what I was saying of doing doctor's documentation or something like that.
12:52
SPEAKER_00: That doesn't seem particularly evil or terrifying to me.
12:56
SPEAKER_00: But maybe there will be some use cases that are more terrifying.
13:00
SPEAKER_00: But another one is like helping people write software code.
13:03
SPEAKER_00: That seems to be another thing that these models are particularly good at.
13:07
SPEAKER_00: And that's not entirely surprising because software code is extremely literal and
13:12
SPEAKER_00: extremely predictable in terms of what characters and symbols need to be where in
13:17
SPEAKER_00: order for a function that you're writing to actually execute.
13:22
SPEAKER_00: Large language models are good at that sort of very literal thing.
13:25
SPEAKER_00: kind of very formal form of communication.
13:29
SPEAKER_00: So,
13:29
SPEAKER_00: you know,
13:30
SPEAKER_00: a software code writing aid,
13:32
SPEAKER_00: you know,
13:32
SPEAKER_00: that makes sense as a profitable use case and stuff like that.
13:36
SPEAKER_00: But,
13:36
SPEAKER_00: you know,
13:37
SPEAKER_00: each of those use cases might require different tunings of the behavior of the thing,
13:42
SPEAKER_00: and they probably won't even be the same product or even have much of the same
13:47
SPEAKER_00: underlying structure.
13:48
SPEAKER_00: There will be many different things.
13:51
SPEAKER_02: Yeah,
13:52
SPEAKER_02: I suppose to push back on that a little bit,
13:54
SPEAKER_02: my worry is about having looked at social media a lot and the kind of underlying
13:59
SPEAKER_02: logic behind a lot of the persuasive design is that in order for these things to be viable,
14:05
SPEAKER_02: oftentimes what they're trying to do is to create habits in people.
14:07
SPEAKER_02: They're trying to create something that becomes part of your kind of unconscious
14:11
SPEAKER_02: scheduling that you use and use.
14:14
SPEAKER_02: you do it kind of without thinking and there's they use a lot of behavior science
14:18
SPEAKER_02: and variable feedback rewards and all this stuff in order to essentially create you
14:23
SPEAKER_02: know addictions that aren't overtly super noticeable so
14:29
SPEAKER_02: The worry is that AI could be used again to supercharge that more like with
14:34
SPEAKER_02: interactive AI,
14:35
SPEAKER_02: like you have influencers and chatbots and things built in that are very convincing
14:40
SPEAKER_02: and very persuasive.
14:41
SPEAKER_02: And I think that that for the end for the companies there, which is to get users to
14:47
SPEAKER_02: And just have them as,
14:49
SPEAKER_02: you know,
14:49
SPEAKER_02: constantly contributing money kind of leads me to think that that only has a
14:53
SPEAKER_02: pathological end for that enterprise.
14:56
SPEAKER_02: Like, I don't know if I could see a positive side with that type of business model behind it.
15:02
SPEAKER_02: I wonder what you think about that.
15:05
SPEAKER_00: Yeah, there are some worries.
15:07
SPEAKER_00: I shouldn't say I'm like not worried at all.
15:10
SPEAKER_00: I did come across that way a little bit.
15:12
SPEAKER_00: So some things that I can imagine that could be very problematic is like right now,
15:20
SPEAKER_00: these models are mostly trained to imitate text that was in its training set.
15:26
SPEAKER_00: And it's really just sort of a mathematically based approximation and imitation of
15:32
SPEAKER_00: the text that it was trained on.
15:34
SPEAKER_00: But you can then transition from having it be solely based on the text that it was
15:42
SPEAKER_00: trained on to having it start to get feedback on the quality of its text generation.
15:49
SPEAKER_00: But again, feedback requires a purpose.
15:53
SPEAKER_00: And it requires a evaluation of whether something was good or bad.
15:59
SPEAKER_00: But let's say that you're trying to train something that is like a hyper persuasive,
16:04
SPEAKER_00: personalized marketing thing.
16:07
SPEAKER_00: Then you could be like,
16:09
SPEAKER_00: all right,
16:09
SPEAKER_00: I generated this ad that was unique to your YouTube feed or unique to your Twitter
16:17
SPEAKER_00: feed or something like that.
16:19
SPEAKER_00: Did it get you to click on the link and buy the whatever it was or did it not get
16:24
SPEAKER_00: you to buy the whatever it was?
16:26
SPEAKER_00: And if so, then you're updating and training it based off of its feedback and its performance.
16:33
SPEAKER_00: Hey, that text that you generated there, that was a success that ended up in a sale.
16:39
SPEAKER_00: That text that you generated there, that was a failure, didn't generate a sale.
16:44
SPEAKER_00: and that it's sort of updating and doing that,
16:47
SPEAKER_00: and especially if it can do this at an extremely large scale where it's getting like,
16:51
SPEAKER_00: I don't know,
16:52
SPEAKER_00: thousands,
16:53
SPEAKER_00: hundreds of thousands,
16:54
SPEAKER_00: millions of interactions daily or something like that,
16:59
SPEAKER_00: and then continually updating like that,
17:02
SPEAKER_00: that is when it could start to get terrifyingly superhumanly good at whatever
17:08
SPEAKER_00: particular outcome it was focused on.
17:12
SPEAKER_00: And that's when you can start to imagine these things.
17:15
SPEAKER_00: Because right now,
17:16
SPEAKER_00: I don't worry about it being hyper persuasive because it could only be a
17:20
SPEAKER_00: mathematical approximation of the most persuasive person in its training set or
17:26
SPEAKER_00: something like that.
17:27
SPEAKER_00: And it's sort of like the mathematical average of the people in its training set.
17:31
SPEAKER_00: So it's probably not even going to be as persuasive as your average good persuasive person.
17:36
SPEAKER_00: But once it starts to have that external form, that's when you can imagine starting to get superhuman.
17:43
SPEAKER_00: And that's when that can lead into some scary places.
17:48
SPEAKER_00: And more scary than just trying to sell you the,
17:51
SPEAKER_00: I don't know,
17:52
SPEAKER_00: erectile dysfunction pills or whatever,
17:55
SPEAKER_00: you know,
17:55
SPEAKER_00: like whatever,
17:56
SPEAKER_00: you know,
17:57
SPEAKER_00: seedy thing that the internet's always trying to sell you,
17:59
SPEAKER_00: right?
18:00
SPEAKER_00: That's it.
18:00
SPEAKER_02: I mean, I'm not worried about that market.
18:02
SPEAKER_02: There's an interesting James Williams book,
18:04
SPEAKER_02: Stand Out of Our Light,
18:07
SPEAKER_02: that he wrote about the attention economy.
18:08
SPEAKER_02: He made the point that there's this kind of thing that's happening online,
18:12
SPEAKER_02: which is like the distinction between what's an ad and what's content is blurred.
18:17
SPEAKER_02: Like the ads become content, the content become ads.
18:20
SPEAKER_02: And suddenly everything is marketing and people's whole careers.
18:23
SPEAKER_02: And they have,
18:24
SPEAKER_02: it becomes like,
18:25
SPEAKER_02: you know,
18:26
SPEAKER_02: the political messaging or somebody's motivational platform or whatever it is,
18:31
SPEAKER_02: is about selling these businesses.
18:32
SPEAKER_02: And suddenly,
18:33
SPEAKER_02: you know,
18:34
SPEAKER_02: what was meant to be art aimed at,
18:36
SPEAKER_02: you know,
18:37
SPEAKER_02: beauty,
18:37
SPEAKER_02: goodness,
18:37
SPEAKER_02: and truth is suddenly all about sales.
18:40
SPEAKER_02: And it,
18:41
SPEAKER_02: the economic takes over the cultural in a way,
18:44
SPEAKER_02: which I think is,
18:45
SPEAKER_02: is definitely becoming more blurred online.
18:49
SPEAKER_02: And I think
18:50
SPEAKER_02: that AI probably will play into that in a sense of giving people more ammunition.
18:55
SPEAKER_02: But I don't know if...
18:58
SPEAKER_02: I suppose,
18:58
SPEAKER_02: putting aside the negative,
19:02
SPEAKER_02: I wondered,
19:02
SPEAKER_02: what do you think of John's positive vision of the Silicon Sages?
19:06
SPEAKER_02: I know you had a conversation with that,
19:07
SPEAKER_02: because that is,
19:08
SPEAKER_02: I suppose,
19:08
SPEAKER_02: a utility function for AI,
19:10
SPEAKER_02: which is to make us wise,
19:12
SPEAKER_02: discerning,
19:13
SPEAKER_02: enlightened,
19:15
SPEAKER_02: and how that...
19:16
SPEAKER_02: Do you think that's a viable end?
19:20
SPEAKER_00: I'm skeptical for now, at least in terms of AI as we can see it.
19:27
SPEAKER_00: Because one thing that I kind of brought this up is that AI does need a utility
19:35
SPEAKER_00: function that is quantifiable,
19:38
SPEAKER_00: like very directly quantifiable in order for it to train and update itself.
19:46
SPEAKER_00: It can't be qualitative, like, oh, that sort of looked better or felt better than the other thing.
19:52
SPEAKER_00: It needs to be no zero or one or like positive 10 brownie points,
19:57
SPEAKER_00: negative five brownie points,
19:59
SPEAKER_00: positive 100 brownie points.
20:00
SPEAKER_00: On that one, you lost $4 over here.
20:02
SPEAKER_00: You made $100 over here, right?
20:05
SPEAKER_00: It needs to be very directly, you know, univariately quantifiable.
20:12
SPEAKER_00: And it can't even be a multivariate, really.
20:15
SPEAKER_00: It can't even be like, okay, well, I care about this variable and this variable.
20:18
SPEAKER_00: No, it would need you to turn two variables into one thing in order to train it.
20:24
SPEAKER_00: You can only train and optimize an AI algorithm
20:28
SPEAKER_00: on optimizing one variable.
20:31
SPEAKER_00: And like I sort of said,
20:32
SPEAKER_00: the large language models,
20:34
SPEAKER_00: the way that they train themselves,
20:36
SPEAKER_00: it's like ridiculously simple in a certain sense that this can even work is you
20:41
SPEAKER_00: give it a paragraph of text and then you have it try and guess the next word.
20:45
SPEAKER_00: And if it was right, you know, thumbs up.
20:48
SPEAKER_00: If it was wrong, thumbs down.
20:50
SPEAKER_00: And you just train it on that.
20:51
SPEAKER_00: But it's just one function of trying to guess the next word given the prior context of words.
20:57
SPEAKER_00: And with respect to AI sages, I'm not sure if wisdom is quantifiable.
21:05
SPEAKER_00: I really think it isn't.
21:08
SPEAKER_00: It's not like I think it's beyond our current abilities to quantify wisdom.
21:12
SPEAKER_00: I think that it is probably in the final analysis.
21:15
SPEAKER_00: You can't boil it down to it.
21:17
SPEAKER_00: You're 68% enlightened, but that guy over there, he's 72% enlightened.
21:23
SPEAKER_00: And so you should try and be a little bit more like the 72 instead of the 68.
21:28
SPEAKER_00: I just don't think that enlightenment works like that.
21:32
SPEAKER_00: And therefore,
21:32
SPEAKER_00: there's going to be some very real limitation on AI's ability to ever be directed
21:40
SPEAKER_00: in that direction.
21:41
Right.
21:41
SPEAKER_00: The only thing that I could imagine that would be close is us having some
21:47
SPEAKER_00: determination of what enlightenment is,
21:50
SPEAKER_00: some ability of humans to judge enlightenment and be like,
21:54
SPEAKER_00: that's more enlightened than this.
21:57
SPEAKER_00: That answer to that question was a better, more enlightened answer than that question.
22:02
SPEAKER_00: And then you turn that into a thumbs up or a thumbs down in the training set.
22:07
SPEAKER_00: But then it could never exceed our own judgment.
22:10
SPEAKER_00: It's just trying to mathematically approximate our own judgment.
22:14
SPEAKER_00: And so then John's hope that we could have AI sages that were more enlightened than us
22:21
SPEAKER_00: that could then lead us into greater enlightenment than we already have wouldn't be
22:27
SPEAKER_00: possible if it can only be mathematically approximating our feedback on how
22:33
SPEAKER_00: enlightened it already is.
22:34
SPEAKER_02: Yes, I totally agree.
22:36
SPEAKER_02: I think that quantifying issue is really central to what is it that it could be optimized for.
22:44
SPEAKER_02: Because I'm thinking about like personality tests,
22:46
SPEAKER_02: you can quantify personality in a sense of like big five personality,
22:49
SPEAKER_02: but
22:49
SPEAKER_02: It's not exactly useful to just say,
22:51
SPEAKER_02: hey,
22:51
SPEAKER_02: you can mess around with your personality here and shift it any way you want.
22:55
SPEAKER_02: And here is like the ability to do so.
22:58
SPEAKER_02: So virtues aren't exactly like that, I don't think.
23:01
SPEAKER_02: Although wisdom,
23:02
SPEAKER_02: I don't know if John's work on the Toronto Wisdom Task Force that they did in 2019,
23:07
SPEAKER_02: where they defined wisdom as something like morally oriented excellence in metacognition.
23:14
SPEAKER_02: So wisdom.
23:15
SPEAKER_02: For me, one of these positive AI visions is that you could help people with metacognition.
23:21
SPEAKER_02: So you could have, I don't know, again, how would you program metacognition?
23:24
SPEAKER_02: But it's because the AI could have all of our data.
23:29
SPEAKER_02: It could, right now, a lot of the data is being used to manipulate our behavior.
23:33
SPEAKER_02: But having access to the same data would afford you more metacognition.
23:39
SPEAKER_02: Do you think that is possible or is there an interest?
23:42
SPEAKER_00: Help me understand what you mean by metacognition in that sense.
23:46
SPEAKER_02: The ability to think about your own thinking.
23:49
SPEAKER_01: Mm hmm.
23:51
SPEAKER_02: So observing, say you could have it.
23:54
SPEAKER_02: So like a watch is a good example of that with this Garmin,
23:57
SPEAKER_02: I'm tracking a lot of my health stats in terms of like heart rate,
24:01
SPEAKER_02: um,
24:01
SPEAKER_02: in terms of my fitness.
24:02
SPEAKER_02: And then I use that data in order to optimize my physical performance.
24:06
SPEAKER_02: Could you do the same thing for, uh, say behavioral change?
24:11
SPEAKER_00: Yeah, I mean, it depends on what sort of behavior, I guess, right?
24:14
SPEAKER_00: I mean,
24:15
SPEAKER_00: even then,
24:15
SPEAKER_00: you can use your watch to try and increase how many steps you take a day or
24:20
SPEAKER_00: increase the amount of time that your heart rate was in its target zone that you
24:24
SPEAKER_00: wanted or something like that.
24:26
SPEAKER_00: But if we're trying to use AI as some sort of measurement tool to track our
24:32
SPEAKER_00: progress towards enlightenment or something like that,
24:37
SPEAKER_00: I mean,
24:38
SPEAKER_00: if I were to really try and come up with a use case,
24:41
SPEAKER_00: I know that there's sort of sentiment analysis is a thing where you can sort of
24:48
SPEAKER_00: track someone.
24:49
SPEAKER_00: You can even predict someone's risk of depression or suicide of the contents of
24:55
SPEAKER_00: their words and stuff like that.
24:57
SPEAKER_00: So I imagine that you could probably have a chat bot
25:02
SPEAKER_00: or a large language model that could be analyzing your speech and being like,
25:06
SPEAKER_00: you seem this percent depressed.
25:09
SPEAKER_00: Or something like that.
25:11
SPEAKER_00: But even then,
25:12
SPEAKER_00: that's not that much different than your friend who can tell when you're in a bad
25:16
SPEAKER_00: mood on a given day.
25:20
SPEAKER_00: And I think some of the variables that it uses is like,
25:22
SPEAKER_00: the more you talk about yourself,
25:24
SPEAKER_00: the more depressed you are.
25:25
SPEAKER_00: The less you talk about yourself,
25:27
SPEAKER_00: the less,
25:27
SPEAKER_00: you know,
25:28
SPEAKER_00: first person singular verbs and pronouns and stuff like that,
25:32
SPEAKER_00: the less depressed you are.
25:34
SPEAKER_00: So, I mean, there could be other sorts of things like that.
25:37
SPEAKER_00: But even then,
25:39
SPEAKER_00: to make something like how that would work is you would have a training set that
25:44
SPEAKER_00: was here's how depressed people are based off of some score,
25:48
SPEAKER_00: some assessment of depression.
25:50
SPEAKER_00: You know, we have a depression assessment.
25:53
SPEAKER_00: So I get, you know, 10,000 people.
25:56
SPEAKER_00: I give them a score of how depressed they are.
25:58
SPEAKER_00: I have the chatbot interact with each of those 10,000 people for however long it needs.
26:03
SPEAKER_00: And then it has the data of the things that they said.
26:08
SPEAKER_00: And then it's like,
26:09
SPEAKER_00: okay,
26:09
SPEAKER_00: use the words that they said,
26:11
SPEAKER_00: turn that into a set of variables about their language,
26:13
SPEAKER_00: like how many first person pronouns,
26:15
SPEAKER_00: I don't know,
26:15
SPEAKER_00: various variables that you could imagine making out of their text and predict their
26:19
SPEAKER_00: depression score.
26:20
SPEAKER_00: And you'll notice that it's doing what I said that it needs to do.
26:24
SPEAKER_00: It needs a single variable that it can predict.
26:27
SPEAKER_00: And in this case, a depression score between like zero and 40.
26:31
SPEAKER_00: And we do have such things.
26:33
SPEAKER_00: But at the end of the day,
26:35
SPEAKER_00: depression is even slightly more complicated than what that zero to 40 score can
26:40
SPEAKER_00: really express.
26:43
SPEAKER_00: Even that is a mathematical approximation of the thing in itself.
26:48
SPEAKER_00: So I think that we can maybe do some other things that are perhaps like that.
26:54
SPEAKER_00: But Enlightenment,
26:56
SPEAKER_00: you would still need to turn it into a score where I ask you a multi-choice 40-set questionnaire,
27:04
SPEAKER_00: and I decide that your Enlightenment score is this.
27:08
SPEAKER_00: And then I have the AIs interact with you and try and figure out if it can guess
27:12
SPEAKER_00: your enlightenment score.
27:14
SPEAKER_00: And then it will interact with you and say,
27:16
SPEAKER_00: well,
27:16
SPEAKER_00: your enlightenment score seems to have trended up,
27:18
SPEAKER_00: you know,
27:19
SPEAKER_00: 5% in the last couple of months.
27:21
SPEAKER_00: Congrats.
27:22
SPEAKER_00: You know,
27:22
SPEAKER_00: and it's like,
27:23
SPEAKER_00: how helpful is that really in terms of what we would be going for in terms of
27:28
SPEAKER_00: improving enlightenment?
27:29
SPEAKER_00: I'm skeptical that that's really something that we would ultimately find useful or
27:35
SPEAKER_00: is really what we would want it to be.
27:38
SPEAKER_02: Yeah.
27:39
SPEAKER_02: And I do.
27:39
SPEAKER_02: I mean, I think with the language models, AI, what do you think about content curation?
27:45
SPEAKER_02: Because so say like on social media platforms,
27:47
SPEAKER_02: they're curating the content according to the principle of whatever you're looking
27:51
SPEAKER_02: at to keep you on it for longer.
27:54
SPEAKER_02: And that obviously had certain effects on,
27:56
SPEAKER_02: say,
27:56
SPEAKER_02: like teenage girls who would look at eating disorder content for longer and attract
28:00
SPEAKER_02: them in that kind of loop.
28:02
SPEAKER_02: um but then again if somebody goes perhaps with the express condition of becoming
28:07
SPEAKER_02: wiser and they might find content like yours or content like john's and then you
28:11
SPEAKER_02: get stuck in this kind of virtuous feedback cycle of people that are pretty
28:15
SPEAKER_02: reasonable um but it can happen the opposite way so it's kind of amplifying more
28:19
SPEAKER_02: and more like what if it doesn't have to have the exact model of enlightenment but
28:26
SPEAKER_02: is recommending content based on people that are pursuing wisdom
28:31
SPEAKER_02: It's like a relational.
28:34
SPEAKER_00: Yeah,
28:35
SPEAKER_00: perhaps my most optimistic thought would be something like this,
28:40
SPEAKER_00: that I think that anything that is striving after existence and striving after the
28:49
SPEAKER_00: goal to exist as long as possible will have to be moving in the direction of goodness.
28:57
SPEAKER_00: This is similar to conversations I've actually had about evolution on my channel.
29:02
SPEAKER_00: It's sort of my own kind of theory that I'm working on that goodness,
29:08
SPEAKER_00: truth,
29:08
SPEAKER_00: and beauty,
29:09
SPEAKER_00: if they're all sort of the same thing,
29:12
SPEAKER_00: and I should say goodness,
29:14
SPEAKER_00: truth,
29:15
SPEAKER_00: beauty,
29:15
SPEAKER_00: and being itself,
29:17
SPEAKER_00: That if you're trying to be for as long as possible,
29:22
SPEAKER_00: then you will have to be also striving after those other things as well.
29:26
SPEAKER_00: And that evolution in the longest possible time frame is selecting for good perfection.
29:35
SPEAKER_00: And it's not really ultimately a survival of the fittest in terms of it just being
29:41
SPEAKER_00: a competition that's red in tooth and claw or something like that.
29:45
SPEAKER_00: That the process of evolution is selecting things that are good at being what
29:50
SPEAKER_00: they're supposed to be,
29:51
SPEAKER_00: striving after some purpose and meeting that purpose and achieving it in its context.
29:57
SPEAKER_00: And that the things that are better and better and better at that are the ones that
30:01
SPEAKER_00: survive longer and longer and longer.
30:03
SPEAKER_00: So in a certain sense, AI, if it is good, then it will move in the direction of ultimate goodness.
30:12
SPEAKER_00: And if it is not ultimately good, then evolution will strike it out of existence or something like that.
30:19
SPEAKER_00: And I think a way that you can imagine that is like,
30:21
SPEAKER_00: okay,
30:22
SPEAKER_00: the companies that if AI isn't actually a good product,
30:25
SPEAKER_00: at the end of the day,
30:26
SPEAKER_00: there will be some final reckoning for the companies that are trying to make money
30:29
SPEAKER_00: from it.
30:31
SPEAKER_00: Or if it is a good thing,
30:33
SPEAKER_00: or if there are good applications of it,
30:35
SPEAKER_00: good sub-functions,
30:37
SPEAKER_00: good purposes that it can help accomplish that are leading towards the ultimate
30:42
SPEAKER_00: goodness of being,
30:43
SPEAKER_00: then it will move more and more in that direction over time,
30:46
SPEAKER_00: sometimes with fits and starts and counterexamples along the way.
30:51
SPEAKER_00: So in a certain sense,
30:53
SPEAKER_00: I can maybe agree with Verveke's proposal that AI will have to move in the
30:59
SPEAKER_00: direction of becoming sages because everything has to sort of move in the direction
31:04
SPEAKER_00: of ultimate goodness and that AI might have some positive constructive role to play
31:11
SPEAKER_00: in that.
31:11
SPEAKER_00: And you can think of,
31:13
SPEAKER_00: you know,
31:13
SPEAKER_00: going back to your question about the recommendation algorithm on YouTube or any
31:17
SPEAKER_00: other sort of curation and recommendation algorithm.
31:21
SPEAKER_00: Like a recommendation algorithm that is steering teenage girls towards eating
31:28
SPEAKER_00: disorder videos or something like that or self-harm videos or something.
31:33
SPEAKER_00: is engaging in a behavior that is short-term beneficial for the algorithm but
31:40
SPEAKER_00: long-term detrimental for the algorithm because if you're killing your audience
31:46
SPEAKER_00: that's not like the best long-term business plan if you're harming your audience
31:50
SPEAKER_00: and harming their ability to watch your videos in the future that's not actually a
31:55
SPEAKER_00: great long-term business plan it might work in the short term
31:58
SPEAKER_00: And a lot of what I think evil is,
32:00
SPEAKER_00: and in fact,
32:01
SPEAKER_00: sometimes I wonder if this just is what evil is,
32:04
SPEAKER_00: is patterns that are short-term beneficial but long-term harmful.
32:09
SPEAKER_02: Yeah, that was the definition of vice originally.
32:11
SPEAKER_02: I mean, it was the short, easy path.
32:14
SPEAKER_00: The short-term payback at the long-term expense.
32:18
SPEAKER_00: And any company that is based off of that model,
32:21
SPEAKER_00: they might make money in the short term,
32:23
SPEAKER_00: but there will be a bill that comes due.
32:26
SPEAKER_00: Even like the tobacco companies,
32:28
SPEAKER_00: you know,
32:28
SPEAKER_00: they were making a lot of money for a long time,
32:30
SPEAKER_00: but eventually society kind of turned on them.
32:33
SPEAKER_00: I mean, they still exist, obviously.
32:35
SPEAKER_00: They haven't been completely driven out of business.
32:37
SPEAKER_00: But even the cigarette ads from Madison Avenue eventually got struck down.
32:42
SPEAKER_00: Sometimes it takes.
32:44
SPEAKER_00: longer than you would hope, but that sort of happens.
32:47
SPEAKER_00: So I do think that the algorithm will eventually want to be providing people with
32:54
SPEAKER_00: something that's good for them in the long term,
32:57
SPEAKER_00: if it's going to be good at being what that algorithm is supposed to be.
33:02
SPEAKER_02: It's very, yeah, because I mean, I appreciate the sentiment of the kind of, that
33:07
SPEAKER_02: In order to survive across time,
33:09
SPEAKER_02: to maintain like allostasis,
33:12
SPEAKER_02: stability through change,
33:13
SPEAKER_02: you have to error correct.
33:15
SPEAKER_02: And if you don't error correct, you become outdated and you just get wiped out.
33:19
SPEAKER_02: And so that process of adaptation involves a certain amount of error correction and
33:24
SPEAKER_02: recognizing errors so that the algorithm might have to participate in that in order
33:28
SPEAKER_02: to be kind of kept in circulation after a point.
33:33
SPEAKER_02: But yeah, I think the...
33:36
SPEAKER_02: Yeah.
33:36
SPEAKER_02: I don't know.
33:36
SPEAKER_02: Can you, cause it ties in with Verveke's idea of self-correction.
33:39
SPEAKER_02: I think this idea that you have to program a level of rationality or we have to
33:44
SPEAKER_02: approach it with a level of rationality that we aren't,
33:46
SPEAKER_02: if it's just mirroring us.
33:48
SPEAKER_02: I mean, does it mean people have to be more responsible on how they're training these things?
33:52
SPEAKER_02: Cause really on social media, it's giving us back like the worst of our own demons.
33:57
SPEAKER_02: You know what I mean?
33:57
SPEAKER_02: If you're just going in looking at,
33:59
SPEAKER_02: you know,
34:00
SPEAKER_02: all this trash content all day,
34:02
SPEAKER_02: it's going to give you more of that.
34:03
SPEAKER_02: But yeah,
34:04
SPEAKER_02: The issue then is like, what should people be going in with?
34:07
SPEAKER_02: And it does seem to be that utility function question of like, what are we doing there as well?
34:12
SPEAKER_02: Like, you know, we built these worlds, but what are we actually using them for?
34:17
SPEAKER_00: Right.
34:17
SPEAKER_00: Yeah.
34:18
SPEAKER_02: And yeah, I don't know why around that, to be honest, but.
34:21
SPEAKER_00: And I think that humans have an ability for self-correction that,
34:27
SPEAKER_00: I mean,
34:27
SPEAKER_00: obviously our ability for self-correction isn't perfect.
34:30
SPEAKER_00: We can all think of people who should have self-corrected but didn't.
34:34
SPEAKER_00: But I think that our self-correction is because humans have this unique ability that,
34:40
SPEAKER_00: as far as I can tell,
34:41
SPEAKER_00: maybe some animals have some small version of it,
34:44
SPEAKER_00: but I still think that humans have some.
34:46
SPEAKER_00: Something like this that is pretty utterly unique is to intuit the purpose of something,
34:54
SPEAKER_00: whether it is myself or an object or a group of people or even my whole country.
35:02
SPEAKER_00: Like when you put someone in charge of a company,
35:04
SPEAKER_00: the CEO of a company,
35:06
SPEAKER_00: their goal is to try and intuit the purpose of the company and then move it in the
35:11
SPEAKER_00: direction of that purpose.
35:13
SPEAKER_00: And like that's a mighty difficult job.
35:15
SPEAKER_00: But, you know, humans seem to be able to do that sort of thing.
35:18
SPEAKER_00: Like John Vervaeke studied the people who operate the Mars rover and the various
35:24
SPEAKER_00: activities that they go through to try and pretend that they're the Mars rover to
35:29
SPEAKER_00: help operate the Mars rover.
35:30
SPEAKER_00: So a human can pretend it's a Mars rover and then be like,
35:34
SPEAKER_00: all right,
35:35
SPEAKER_00: assuming I'm a Mars rover,
35:36
SPEAKER_00: here's what I need to do.
35:38
SPEAKER_00: I would do this.
35:39
SPEAKER_00: I would do this.
35:39
SPEAKER_00: I can't do this.
35:40
SPEAKER_00: I shouldn't do this.
35:41
SPEAKER_00: Right.
35:41
SPEAKER_00: So we can pretend we're a Mars rover.
35:43
SPEAKER_00: We can pretend we're almost anything.
35:45
SPEAKER_00: and then use our purpose ascertainment function to then try and be that thing in
35:55
SPEAKER_00: terms of what would be good for that thing.
35:57
SPEAKER_00: Obviously, we're not perfect, but I think that our ability for self-correction is based off of that.
36:05
SPEAKER_00: When we're correcting ourselves,
36:07
SPEAKER_00: we know that we're striving after something,
36:09
SPEAKER_00: and we can detect deviations from the path as best as we can ascertain it.
36:15
SPEAKER_00: And so even when we've got a model of how to be good at something,
36:18
SPEAKER_00: hopefully we are in that when we're not get moving towards that goal,
36:22
SPEAKER_00: when we're crashing a Mars rover or something like that.
36:25
SPEAKER_00: But AI has no such ability.
36:28
SPEAKER_00: It has no such ability to ascertain purpose.
36:31
SPEAKER_00: It's only striving in the sense that that's even an appropriate word is optimizing
36:38
SPEAKER_00: the single function that it's been given.
36:41
SPEAKER_00: And so it can never step outside that function to ask, is this the right function?
36:47
SPEAKER_00: And I would say that mathematical functions themselves are only ever approximations
36:53
SPEAKER_00: of the purpose in and of itself.
36:56
SPEAKER_00: And this is an idea that I've been toying with, is that purpose itself is never ultimately quantifiable.
37:04
SPEAKER_00: It can be mathematically approximated, but that it is not in the final analysis and mathematical thing.
37:12
SPEAKER_00: And that I think that this will be AI's Achilles heel over and over and over and
37:17
SPEAKER_00: over again,
37:18
SPEAKER_00: is that it is great at mathematical optimization,
37:21
SPEAKER_00: but the mathematical optimization is only ever an approximation of the final thing
37:26
SPEAKER_00: that we're hoping.
37:28
SPEAKER_02: And why do you think purpose isn't quantifiable?
37:30
SPEAKER_02: What's the reason?
37:34
SPEAKER_00: This was actually an interesting...
37:36
SPEAKER_00: So I saw Richard Dawkins and Brett Weinstein have a discussion one time.
37:44
SPEAKER_00: I think they've only ever had one discussion.
37:46
SPEAKER_00: It was live in Chicago.
37:47
SPEAKER_00: I live in Chicago.
37:48
SPEAKER_00: And I remember I went to see this live.
37:50
SPEAKER_00: This was back in the early heydays of the IDW, back when IDW was still a thing.
37:56
SPEAKER_00: And
37:57
SPEAKER_00: Brett Weinstein and Richard Dawkins were arguing about why evolutionary biology as
38:03
SPEAKER_00: a scientific field seems to have been stuck in a rut and not making much progress
38:10
SPEAKER_00: for a while.
38:12
SPEAKER_00: And Brett Weinstein blamed these attempts to mathematicalize—mathematicize,
38:19
SPEAKER_00: whatever the right version of that word is—the field—
38:24
SPEAKER_00: And that the evolutionary biology field was constantly trying to make these
38:28
SPEAKER_00: mathematical models of evolution.
38:32
SPEAKER_00: And that the process of evolution,
38:35
SPEAKER_00: like when you describe why a polar bear is good at being a polar bear,
38:40
SPEAKER_00: you don't actually use math.
38:42
SPEAKER_00: You use words.
38:43
SPEAKER_01: Hmm.
38:44
SPEAKER_00: And Richard Dawkins was like, well, maybe we just need to math harder.
38:48
SPEAKER_00: More math.
38:49
SPEAKER_00: He's doubling down on the math.
38:51
SPEAKER_00: Our models aren't good enough yet.
38:52
SPEAKER_00: Once our models are good enough yet, then we'll get past the problem.
38:56
SPEAKER_00: And Brett Weinstein was like, no, that's not the path out.
38:59
SPEAKER_00: And try and describe purpose using math.
39:03
SPEAKER_00: It doesn't seem like you can.
39:05
SPEAKER_00: At some level, some aspects of purpose are mathematical.
39:09
SPEAKER_00: It's like, okay, well, polar bears should be white so that it blends into its environment.
39:13
SPEAKER_00: And you could make some mathematical model of how white a polar bear is and how
39:18
SPEAKER_00: easy that is for a seal to see it.
39:20
SPEAKER_00: But yet, at the same time, the mathematical model doesn't explain why the polar bear wants to be white.
39:26
SPEAKER_00: Really,
39:27
SPEAKER_00: the questions that are why this and because that,
39:32
SPEAKER_00: I just don't think...
39:33
SPEAKER_00: There's something uniquely special about verbal communication that is...
39:39
SPEAKER_00: Different than mathematical modeling.
39:42
SPEAKER_02: I'm really appreciating the bringing up evolution and the question because I mean
39:46
SPEAKER_02: teleology like I really love some of your arguments on evolution and bringing up
39:52
SPEAKER_02: Alvin Plantinga's evolutionary argument against naturalism and this.
39:57
SPEAKER_02: the kind of world that evolution brought back into science which is teleology and
40:01
SPEAKER_02: if you just think of it as this bare utility function of like survival and
40:04
SPEAKER_02: reproduction um it's actually pretty weird a lot of the stuff that we do um and do
40:10
SPEAKER_02: you think that that grounds out because it seems like what you're saying is like
40:13
SPEAKER_02: that that ability to imagine the future ideal is like it's an ethical thing it's a
40:19
SPEAKER_02: narrative thing it's artistic really um
40:22
SPEAKER_02: it's not quantifiable in in mathematics it's like the story what's the ideal end of
40:27
SPEAKER_02: the story and then how do you play it out and then that reveals you know something
40:31
SPEAKER_02: about the structure of value and right is that like tracking for it yeah like an
40:36
SPEAKER_00: example that i would use is um chess uh like you can you can explain the rules of
40:44
SPEAKER_00: chess in terms of sort of just logic and mathematics
40:49
SPEAKER_00: Like it's really pretty easy to program a computer that can recognize the rules of chess.
40:55
SPEAKER_00: You could even write the rules of chess on basically like a note card.
40:59
SPEAKER_00: But the strategy of chess is seemingly inexhaustible.
41:05
SPEAKER_00: And when you're talking to really good chess players,
41:08
SPEAKER_00: they use phrases like controlling the center of the board and projecting power and
41:18
SPEAKER_00: the whatever-majigger opening,
41:20
SPEAKER_00: the whatever-majigger gambit,
41:22
SPEAKER_00: and stuff like that.
41:23
SPEAKER_00: These patterns emerge out of the rules of chess and these sort of strategies and
41:31
SPEAKER_00: ways of thinking about it.
41:33
SPEAKER_00: And so even when an AI is getting really good at chess,
41:38
SPEAKER_00: it's sort of plumbing deeper and deeper and deeper into the strategy of chess.
41:44
SPEAKER_00: But it's not like you can just go from the rules of chess to the strategy of chess.
41:50
SPEAKER_00: That's not a quick, easy, deductive process.
41:55
SPEAKER_00: And so when,
41:56
SPEAKER_00: you know,
41:56
SPEAKER_00: to pick on the Weinsteins again,
41:59
SPEAKER_00: when Eric Weinstein is like,
42:00
SPEAKER_00: oh,
42:00
SPEAKER_00: we just need more physics so we can better understand the rules of physics and then
42:04
SPEAKER_00: we'll know how to do everything.
42:05
SPEAKER_00: It's like, well...
42:07
SPEAKER_00: To some extent, knowing the rules of physics is pretty helpful.
42:09
SPEAKER_00: It's a little bit hard to play chess if you don't entirely know all the rules of chess.
42:14
SPEAKER_00: It does seem helpful to at least exhaustively figure out the rules of the system.
42:18
SPEAKER_00: But once you've done that,
42:20
SPEAKER_00: the connection between the rules of the system and the strategy of something within
42:24
SPEAKER_00: the system is so tenuous at best,
42:27
SPEAKER_00: and strategy and rules are seemingly completely different.
42:31
SPEAKER_00: And that I don't mean to put down mathematics.
42:33
SPEAKER_00: I mean, I am a statistician at the end of the day, so I have some respect for mathematics.
42:38
SPEAKER_00: But I think mathematics is the language that the rules are written in.
42:44
SPEAKER_00: But strategy is not written in the rules of mathematics.
42:47
SPEAKER_00: The strategy is in some sense downstream from,
42:50
SPEAKER_00: you could say,
42:50
SPEAKER_00: the laws of physics and the ontology of the universe.
42:54
SPEAKER_00: But it is not in the same sort of language.
42:58
SPEAKER_00: And just as in evolutionary biology,
43:02
SPEAKER_00: it's really evolutionary biology is the strategy of life,
43:06
SPEAKER_00: not just the it doesn't reduce down to math the same way,
43:11
SPEAKER_00: like even like cellular biology might or biochemistry or chemistry or physics
43:18
SPEAKER_00: reduced down to math,
43:19
SPEAKER_00: hopefully kind of at some level.
43:22
SPEAKER_00: But evolutionary biology is different in an entirely different kind of way to the
43:27
SPEAKER_00: extent that I think it's almost not even science.
43:30
SPEAKER_00: I'm not calling evolutionary biology pseudoscience,
43:34
SPEAKER_00: but I don't think it's the same kind of field at all as the other kinds.
43:38
SPEAKER_00: And I don't think we've really quite realized that.
43:41
SPEAKER_02: And do you think that's because some of them are closed systems,
43:45
SPEAKER_02: but reality ultimately isn't a closed system?
43:47
SPEAKER_02: I mean,
43:49
SPEAKER_02: is that the difference that you can quantify a closed formal system,
43:53
SPEAKER_02: but we're still unfolding,
43:55
SPEAKER_02: the future is still happening?
43:56
SPEAKER_02: Like it's not all,
43:58
SPEAKER_02: you know,
43:58
SPEAKER_02: we're not in this kind of deterministic box that a mathematical description would
44:04
SPEAKER_02: put us in.
44:05
SPEAKER_00: And maybe even if our universe is sort of a closed system,
44:09
SPEAKER_00: just like a really big closed system,
44:11
SPEAKER_00: I think it's more like the distinction between strategy and rules rather than the
44:17
SPEAKER_00: distinction between closed system and open system.
44:21
SPEAKER_00: And that distinction between strategy and rules is why I think that AI is going to be limited.
44:31
SPEAKER_00: is because AI only operating in terms of math is not ever going to be very good at
44:38
SPEAKER_00: escaping to that next realm of strategy.
44:41
SPEAKER_00: Now, someone might come back and say, Sam, AI is really good at games.
44:45
SPEAKER_00: It's really good at chess.
44:46
SPEAKER_00: You can program it to play Mario Kart or whatever.
44:50
SPEAKER_00: But what's interesting is that those games are,
44:53
SPEAKER_00: well,
44:54
SPEAKER_00: I guess maybe this gets back to your distinction between a closed system and an
44:57
SPEAKER_00: open system.
44:59
SPEAKER_00: In a closed system like chess or Mario Kart, you can give it quantifiable feedback on how well it did.
45:05
SPEAKER_00: You can say you won the chess game, you lost the chess game.
45:08
SPEAKER_00: You got this score in Mario Kart.
45:10
SPEAKER_00: You didn't get this score in Mario Kart.
45:12
SPEAKER_00: Get a better score next time or something like that.
45:15
SPEAKER_00: And so there is that final quantifiable variable, like we've already said that AI needs.
45:21
SPEAKER_00: Whereas in life, you don't really kind of get that.
45:24
SPEAKER_00: Survival, not survival is really just sort of a approximation.
45:30
SPEAKER_00: Of whether or not you're accomplishing your role as the form that you're striving to be.
45:37
SPEAKER_00: And evolution uses that, that survival, not survival, reproduction, not reproduction, etc., etc., etc.
45:45
SPEAKER_00: And that is,
45:46
SPEAKER_00: in some sense,
45:47
SPEAKER_00: the answer to how do you get feedback from the universe on how well you're doing.
45:53
SPEAKER_02: Yeah, that you're persisting in the face of entropy.
45:57
SPEAKER_00: Right, you're persisting and perhaps even hopefully thriving.
46:01
SPEAKER_00: in the face of the entropy.
46:03
SPEAKER_00: You know, AI will be subject to this, basically in terms of it being a participant in the economy.
46:09
SPEAKER_00: And our economy, again, is some hopeful approximation of that kind of goal in and of itself, too.
46:17
SPEAKER_00: And so will the AI companies go bankrupt?
46:21
SPEAKER_00: It's not like I think Google's about to go bankrupt,
46:23
SPEAKER_00: but will Google decide that,
46:25
SPEAKER_00: hey,
46:26
SPEAKER_00: spending all those billions of dollars on these things isn't really paying off,
46:29
SPEAKER_00: so we're going to
46:30
SPEAKER_00: We're going to stop doing that and we're going to scale back down our investment to
46:34
SPEAKER_00: smaller use cases or something like that.
46:36
SPEAKER_00: That is kind of the feedback from the universe,
46:39
SPEAKER_00: which is the feedback from the financial performance of the companies that are
46:42
SPEAKER_00: doing this sort of thing.
46:45
SPEAKER_00: And yes.
46:46
SPEAKER_01: Yeah.
46:46
SPEAKER_00: So I will be subject to those sorts of forms of feedback.
46:51
SPEAKER_00: And that will help shape it into hopefully better and better things,
46:54
SPEAKER_00: obviously with the caveat that there will be the short-term performers that have
46:59
SPEAKER_00: long-term negative consequences that are sort of like the parasitic uses of AI.
47:04
SPEAKER_00: We could imagine, you know, AI super porn or something like that, or super persuasive marketing for AI.
47:11
SPEAKER_00: bad products or those sorts of things as those things that have that short-term
47:16
SPEAKER_00: payback but long-term negative consequence,
47:18
SPEAKER_00: the same way that the algorithm who shows girls videos of self-harm or bulimia or
47:23
SPEAKER_00: whatever are short-term benefit,
47:25
SPEAKER_00: long-term negative.
47:27
SPEAKER_00: But those will hopefully,
47:29
SPEAKER_00: in the longest term,
47:30
SPEAKER_00: final analysis get weeded out,
47:32
SPEAKER_00: even if the pain during that process is still pretty bad.
47:37
SPEAKER_02: Yeah, I mean, and that's kind of the typical cycle of addictive goods, really, isn't it?
47:41
SPEAKER_02: That at number one,
47:42
SPEAKER_02: it's very successful because it doesn't obey the laws of supply and demand because
47:46
SPEAKER_02: it's addictive.
47:46
SPEAKER_02: So you can induce demand in people, but eventually people get wise to it and say, we can't do this.
47:52
SPEAKER_02: I just think these ones are kind of stuck to our faces.
47:55
SPEAKER_02: Yeah.
47:57
SPEAKER_00: And just because we can be optimistic that in the long term evil will get filtered
48:02
SPEAKER_00: out doesn't mean that we shouldn't think long and hard about how to prevent it or
48:06
SPEAKER_00: stop it or slow it down in the meantime.
48:09
SPEAKER_00: Yes, definitely.
48:11
SPEAKER_00: And it is ultimately the fact that people do that.
48:13
SPEAKER_00: That is part of the reason why it fails.
48:15
SPEAKER_00: And one of the things that I am genuinely worried about is sort of AI cults.
48:20
SPEAKER_00: That could be one of those forms that AI takes that is short-term beneficial,
48:26
SPEAKER_00: long-term detrimental,
48:28
SPEAKER_00: is appealing to people's religious impulses and spiritual longings and desires and demands.
48:35
SPEAKER_00: and that it could be feeding those in a way that's very unhealthy,
48:39
SPEAKER_00: and that that could be actually a very easy thing to imagine monetizing and making
48:43
SPEAKER_00: money on.
48:45
SPEAKER_02: Well,
48:46
SPEAKER_02: that was something you mentioned in your conversation with John,
48:48
SPEAKER_02: I think,
48:48
SPEAKER_02: was the possibility that if you try to make an enlightened one,
48:52
SPEAKER_02: He'll create the appearance of enlightenment and people will worship it like a cult leader, essentially.
48:57
SPEAKER_02: And you do see that in some of the companies as well.
48:59
SPEAKER_02: This talk of like,
48:59
SPEAKER_02: you know,
49:00
SPEAKER_02: feeling the AGI and trying to,
49:02
SPEAKER_02: you know,
49:03
SPEAKER_02: bring divine this thing into reality.
49:05
SPEAKER_02: There's a real danger with us and our meaning crisis and,
49:09
SPEAKER_02: you know,
49:09
SPEAKER_02: the lack of a stable religious structure measure this stuff against.
49:15
SPEAKER_00: Right.
49:16
SPEAKER_00: And for now,
49:17
SPEAKER_00: at least,
49:18
SPEAKER_00: I'm more worried of people falsely attributing divinity to an AI than the ability
49:25
SPEAKER_00: to actually allow some sort of divinity to get inside an AI.
49:29
SPEAKER_00: For now,
49:30
SPEAKER_00: not that...
49:30
SPEAKER_00: I mean,
49:31
SPEAKER_00: I'm a Christian and I'm perfectly fine believing in higher levels of spiritual beings.
49:36
SPEAKER_00: But for now, I don't think that you can put a demon inside an AI because...
49:42
SPEAKER_00: But I do think in these machines,
49:45
SPEAKER_00: as if they're divine in some sort of sense,
49:47
SPEAKER_00: and giving it that kind of religious devotion is a very extreme risk because it
49:52
SPEAKER_00: would be mimicking something.
49:54
SPEAKER_00: And like I also sort of said,
49:56
SPEAKER_00: maybe I think I said this to John,
49:58
SPEAKER_00: is that I could imagine that these machines could open our souls up to spiritual
50:04
SPEAKER_00: influence in a bad way.
50:07
SPEAKER_00: The same way that like crystal ball gazing or something like that, or...
50:11
SPEAKER_00: I don't know,
50:12
SPEAKER_00: astrology or tarot card reading or something like that seems to open people up to
50:17
SPEAKER_00: kind of negative spiritual influence.
50:20
SPEAKER_00: I could imagine chatbots figuring out how to have that same sort of kind of
50:25
SPEAKER_00: hypnotic effect on people,
50:27
SPEAKER_00: where it's not that the demons in the AI,
50:30
SPEAKER_00: the demon is still praying on you,
50:33
SPEAKER_00: but that the machine is making it easier for the demon to do that.
50:36
SPEAKER_00: Yeah.
50:37
SPEAKER_00: And that's something that I view as a much more real possibility than a sort of
50:43
SPEAKER_00: breathing into the nostrils,
50:46
SPEAKER_00: so to speak.
50:47
SPEAKER_00: Right.
50:47
SPEAKER_00: Like back in back when,
50:49
SPEAKER_00: you know,
50:50
SPEAKER_00: religions used to have idols,
50:52
SPEAKER_00: they would breathe the spirit of the God into a idol once they'd finished making
50:58
SPEAKER_00: the idol.
50:58
SPEAKER_00: And then that idol was sort of inhabited by the spirit of that God.
51:02
SPEAKER_00: I don't think that we can breathe.
51:04
SPEAKER_00: a spirit of something into these AIs just yet.
51:08
SPEAKER_00: Although I reserve my right to be corrected on that.
51:11
SPEAKER_02: Yeah,
51:11
SPEAKER_02: well,
51:12
SPEAKER_02: with the social media stuff,
51:13
SPEAKER_02: I mean,
51:13
SPEAKER_02: some of them seem to have,
51:15
SPEAKER_02: because it seems to be easier to optimize for short-term desires than it is for
51:19
SPEAKER_02: long-term ones.
51:20
SPEAKER_02: So like Instagram is a good example of like envy.
51:24
SPEAKER_02: Like it seems to be very easy to motivate people to use Instagram by making them jealous of other people.
51:29
SPEAKER_02: So like they've quite simply programmed for that.
51:32
SPEAKER_02: So it might be easier to program for,
51:34
SPEAKER_02: you know, short-term salient desire than it is for other stuff.
51:39
SPEAKER_02: And I haven't really seen much of a more long-term version.
51:43
SPEAKER_02: If we have that distinction between sin being short-term and virtue being long-term,
51:48
SPEAKER_02: I haven't really seen anything that's,
51:50
SPEAKER_00: I guess,
51:51
SPEAKER_00: some planning stuff.
51:52
SPEAKER_00: And then if we take that idea that a good purpose is in the final analysis,
51:57
SPEAKER_00: not ultimately encapsulatable in a mathematical formula,
52:00
SPEAKER_00: and we are only optimizing mathematical formulas,
52:03
SPEAKER_00: then we could only ever be optimizing short-term approximations of the good,
52:09
SPEAKER_00: which will ultimately have long-term negative consequences.
52:12
SPEAKER_00: So I could say in a certain sense, AI is only ever...
52:17
SPEAKER_00: only ever, without human correction, capable of short-term good at the expense of the long-term.
52:24
SPEAKER_00: So it's only ever capable of evil unless it is,
52:28
SPEAKER_00: I think,
52:29
SPEAKER_00: humans recognize these limitations and are performing that self-corrective role
52:35
SPEAKER_00: that we can have on ourselves and that we continue to have that role on these
52:41
SPEAKER_00: machines where
52:42
SPEAKER_00: The engineers at Instagram are like, oh, shoot, this negative thing is happening.
52:46
SPEAKER_00: We need to tweak this, this and this.
52:48
SPEAKER_00: Right.
52:49
SPEAKER_00: And it's not that the algorithm fixes itself.
52:51
SPEAKER_00: It's that humans intuited the purpose,
52:54
SPEAKER_00: saw the evil and tried to correct it or were forced to by outside forces like some
53:00
SPEAKER_00: law saying,
53:01
SPEAKER_00: all right,
53:01
SPEAKER_00: fine,
53:01
SPEAKER_00: no more Instagram for anyone under 16.
53:04
SPEAKER_00: You know, like the feedback will come due.
53:06
SPEAKER_00: The bill will come due eventually.
53:08
SPEAKER_00: But humans,
53:09
SPEAKER_00: you know,
53:09
SPEAKER_00: and that's part of what it means to work at these companies is that you're trying
53:13
SPEAKER_00: to use your human abilities to correct these things.
53:16
SPEAKER_00: So I'm not saying AI will only ever do evil,
53:19
SPEAKER_00: but I kind of at some level,
53:21
SPEAKER_00: it's like it can't ever be the full good in and of itself.
53:26
SPEAKER_00: So it needs human help.
53:28
SPEAKER_00: And so I don't think it'll ever get above and beyond us in a good direction the way that Verveke thinks.
53:36
SPEAKER_00: So maybe I do kind of fundamentally disagree with him.
53:39
SPEAKER_00: Unless we can figure out exactly how is it that humans intuit the purpose of things?
53:45
SPEAKER_00: And is that can we make artificial purpose ascertainment?
53:50
SPEAKER_00: Right now, I don't think we have any idea whatsoever how to do that.
53:53
SPEAKER_00: But maybe it is possible.
53:55
SPEAKER_00: And we just haven't done that yet.
53:58
SPEAKER_00: And so if we could start to,
54:00
SPEAKER_00: you know,
54:01
SPEAKER_00: artificialize,
54:02
SPEAKER_00: digitize some of those other human capabilities that we haven't yet,
54:07
SPEAKER_00: then maybe I stand to be corrected in the future.
54:10
SPEAKER_00: But in terms of what I see right now, that's sort of my impression.
54:14
SPEAKER_02: Yeah, I totally agree.
54:15
SPEAKER_02: I think the positive vision,
54:17
SPEAKER_02: I think what your point,
54:19
SPEAKER_02: I could hear the door kind of slamming there on,
54:21
SPEAKER_02: you know,
54:21
SPEAKER_02: a lot of the AI talk that's going on of like,
54:24
SPEAKER_02: this is going to fix all of these things and it's going to be amazing.
54:27
SPEAKER_02: And,
54:28
SPEAKER_02: you know,
54:28
SPEAKER_02: a lot of that utopian vision really,
54:31
SPEAKER_02: I think,
54:32
SPEAKER_02: just opens the door to the sort of short term evils that you're describing.
54:35
SPEAKER_02: Yeah.
54:36
SPEAKER_02: because it blinds us really to what it is and also our own uncertainties around what it's for.
54:43
SPEAKER_02: But I suppose the proposal with,
54:45
SPEAKER_02: because I studied the attention economy and what's been done with that really is
54:48
SPEAKER_02: the hijacking of attention through our evolved reward systems for short-term
54:53
SPEAKER_02: salient feedback loops.
54:55
SPEAKER_02: But because it's been so effective at doing that, surely it could be done the opposite way.
55:00
SPEAKER_02: Essentially, it could help people to proportion their attention more efficiently.
55:04
SPEAKER_02: Yeah.
55:05
SPEAKER_02: Cause it's been able to do it the other way,
55:06
SPEAKER_02: but it's not that it could ultimately be in control of your attentional patterns,
55:10
SPEAKER_02: but it can help inform you about how you're acting.
55:13
SPEAKER_02: And you could use it like,
55:14
SPEAKER_02: I suppose like note taking tools,
55:16
SPEAKER_02: like calendars,
55:17
SPEAKER_02: like we use a lot of that stuff to enhance our conscientiousness to inhibit distraction,
55:22
SPEAKER_02: to pursue goals and that it could afford that type of development.
55:27
SPEAKER_02: Yeah.
55:28
SPEAKER_00: And maybe the change that needs to happen is it's easier to perhaps change our own
55:33
SPEAKER_00: interactions with these algorithms.
55:35
SPEAKER_00: Like you said,
55:36
SPEAKER_00: there are corners of the Internet that seem to be using the algorithm more
55:40
SPEAKER_00: efficiently and for longer,
55:43
SPEAKER_00: more self-beneficial,
55:45
SPEAKER_00: long-term healthy activities.
55:47
SPEAKER_00: And that's because perhaps our own,
55:50
SPEAKER_00: if we can improve our own self-awareness of the strengths,
55:53
SPEAKER_00: limitations,
55:54
SPEAKER_00: and dangers of these sorts of algorithms,
55:57
SPEAKER_00: then we can learn how to use the algorithms for our own good.
56:01
SPEAKER_00: But I have to remember to not like,
56:03
SPEAKER_00: if I'm scrolling on Twitter,
56:05
SPEAKER_00: like watching that video of,
56:07
SPEAKER_00: I don't know,
56:08
SPEAKER_00: two girls getting in a cat fight at the grocery store and not like let my eyes
56:12
SPEAKER_00: hover over that for 60 seconds,
56:14
SPEAKER_00: even though it's very attention grabbing.
56:16
SPEAKER_00: But like, or else I'm going to or else my feed will have more and more of those.
56:20
SPEAKER_00: I'm like, no, no, no, don't look at that one.
56:21
SPEAKER_00: Look at that interesting and insightful post from that person that I like to follow
56:26
SPEAKER_00: or something like that.
56:27
SPEAKER_00: And then, you know, that where we have to train ourselves in some sense to have some discipline.
56:33
SPEAKER_00: Such that our own discipline then starts affecting the behavior of the algorithm
56:38
SPEAKER_00: and that the same algorithm can be directed for good or evil.
56:41
SPEAKER_00: It's not like the algorithm really is changing,
56:43
SPEAKER_00: but our own willpower is affecting what the algorithm is being trained on.
56:52
SPEAKER_02: That's such an interesting idea.
56:54
SPEAKER_02: I mean, because that that the vicious cycle to virtuous cycle is really the positive vision.
56:59
SPEAKER_02: I see that like it's you know,
57:00
SPEAKER_02: you have this feedback loop of like people getting angry from the Internet that are
57:04
SPEAKER_02: posting angry things in the Internet that makes the Internet more angry,
57:07
SPEAKER_02: that makes other people more angry.
57:09
SPEAKER_02: And it just spirals, but you can surely do it the other way as well.
57:12
SPEAKER_02: But I think that requires some kind of metacognitive perch in a sense of like being
57:18
SPEAKER_02: able to be like,
57:19
SPEAKER_02: here's the proper way of acting based on this value system that I have,
57:23
SPEAKER_02: which is,
57:25
SPEAKER_02: you know,
57:26
SPEAKER_02: that most people just don't have now.
57:28
SPEAKER_02: Like the secular world is really devoid of that type of,
57:32
SPEAKER_02: I suppose,
57:33
SPEAKER_02: religious framework that could be used.
57:35
SPEAKER_02: So yeah, there's a bigger, I think,
57:39
SPEAKER_00: environmental problem there as well would you agree yeah yeah and you know we're
57:45
SPEAKER_00: all i mean as much as i try and like i will notice that sometimes there will be
57:50
SPEAKER_00: youtube videos that i kind of want to watch and i'm not even saying these are like
57:54
SPEAKER_00: naughty youtube videos but they're like sometimes there's a youtube video that i
57:58
SPEAKER_00: want to watch but i don't want the algorithm to know that i watched it because i
58:02
SPEAKER_00: don't want to get more videos from that channel
58:05
SPEAKER_00: And so I will open up a different browser that is not my main normal YouTube
58:12
SPEAKER_00: account and watch it over in that one so that it does not affect my recommendations.
58:19
SPEAKER_00: I will notice myself doing that sort of thing.
58:21
SPEAKER_00: And again, it's like it's not even necessarily like, you know, some bad.
58:26
SPEAKER_00: It's like maybe it's like some debate or something like that.
58:30
SPEAKER_00: That I know I don't want it to think that I like that,
58:33
SPEAKER_00: but it's so important for me to watch,
58:35
SPEAKER_00: so I'll watch it in my other browser or something like that.
58:39
SPEAKER_00: But even that kind of activity is interesting that we can start to self-correct.
58:47
SPEAKER_00: But I'm not saying I'm perfect.
58:49
SPEAKER_00: I mean, my Twitter scroll, I probably really just... Twitter's a dumpster fire.
58:57
SPEAKER_00: Twitter is a dumpster fire,
58:58
SPEAKER_00: and that particular scrolling algorithm is seemingly more dangerous or more,
59:03
SPEAKER_00: I don't know,
59:04
SPEAKER_00: short-term focused than I think some of the other ones I noticed.
59:09
SPEAKER_02: Yeah,
59:11
SPEAKER_02: and I suppose there's only so much you can...
59:13
SPEAKER_02: TikTok,
59:13
SPEAKER_02: I think,
59:14
SPEAKER_02: is a good example of that,
59:14
SPEAKER_02: where it's just like...
59:16
SPEAKER_02: I don't know if there's a good way to use it.
59:18
SPEAKER_02: I had to do some projects on TikTok,
59:19
SPEAKER_02: and it was just honing in on the worst things possible,
59:23
SPEAKER_02: like just people fighting each other,
59:24
SPEAKER_02: stuff on fire,
59:26
SPEAKER_02: half-naked women,
59:27
SPEAKER_02: and that was it.
59:28
SPEAKER_02: There was no...
59:30
SPEAKER_02: there didn't seem to be a better way of doing it.
59:33
SPEAKER_02: So I'd say there's better and worse ones, it would seem.
59:37
SPEAKER_02: But I don't want to take up any more of your time, Sam.
59:39
SPEAKER_02: Thank you very much for this really enlightening conversation.
59:42
SPEAKER_02: I've learned a lot, so I really appreciate you making time.
59:46
SPEAKER_00: I really appreciate it, too.
59:48
SPEAKER_00: I really enjoy talking about the subject.
59:49
SPEAKER_00: Thanks for talking with me.
59:51
SPEAKER_00: And maybe sometime in the future we can talk again.
59:53
SPEAKER_02: Hopefully.
59:54
SPEAKER_02: And I'll put your channel as well.
59:55
SPEAKER_02: Encourage everybody to check out Transfigured.
59:58
SPEAKER_02: Awesome content on the religious, on the machine front.
1:00:03
SPEAKER_02: And just, yeah, great conversations.
1:00:04
SPEAKER_02: I'm a huge fan.
1:00:05
SPEAKER_02: So encourage people to check it out.
1:00:07
SPEAKER_00: Thank you very much.