camp2023-57163-eng-Unsupervised_Pleasures_opus.vtt

WEBVTT

00:00:00.000 --> 00:00:10.000
 [MUSIC]

00:00:10.000 --> 00:00:20.000
 [MUSIC]

00:00:20.000 --> 00:00:35.200
 And now our next talk is from Sarah Sisten,

00:00:35.200 --> 00:00:38.960
 who is based in Berlin in Los Angeles.

00:00:38.960 --> 00:00:43.800
 And is a PhD candidate of University of Southern California.

00:00:43.800 --> 00:00:50.760
 And who will talk about unsupervised pleasures and

00:00:50.760 --> 00:00:54.760
 its intersectional language models for queer futures.

00:00:54.760 --> 00:00:55.680
 Welcome on stage please.

00:00:55.680 --> 00:00:57.680
 >> [APPLAUSE]

00:01:07.400 --> 00:01:09.400
 >> Thank you for waking up, but

00:01:09.400 --> 00:01:11.400
 [INAUDIBLE]

00:01:11.400 --> 00:01:14.000
 And chugging, I hope you chug coffee, which I just.

00:01:14.000 --> 00:01:18.560
 If you would like to introduce yourselves in the chat,

00:01:18.560 --> 00:01:21.680
 I've set up a ether pad.

00:01:21.680 --> 00:01:23.600
 I don't know if you can see the URL.

00:01:23.600 --> 00:01:31.600
 It's pad.riseup.net/p/unsupervisedpleasurescc-keep.

00:01:31.600 --> 00:01:36.640
 Add your favorite emoji, your pronoun, whatever.

00:01:36.640 --> 00:01:39.880
 We're gonna be getting into it, hopefully, in a participatory way.

00:01:39.880 --> 00:01:45.120
 And while you're doing that, I will introduce myself a little bit more.

00:01:45.120 --> 00:01:47.160
 So I'm a poet and programmer.

00:01:47.160 --> 00:01:52.720
 I'm interested in building tools to bring intersectional approaches to

00:01:52.720 --> 00:01:57.000
 machine learning and building community through accessible,

00:01:57.000 --> 00:02:00.760
 creative coding, critical creative coding.

00:02:00.760 --> 00:02:05.040
 And I come by coding very circuitously via creative writing and

00:02:05.040 --> 00:02:07.200
 scene making and book arts.

00:02:07.200 --> 00:02:14.600
 So I have somehow come around to adapting that work into making subversive art with

00:02:14.600 --> 00:02:16.840
 and about text-based machine learning.

00:02:16.840 --> 00:02:20.120
 And I'll have the link up again in a few minutes.

00:02:20.120 --> 00:02:25.880
 Let's get started.

00:02:25.880 --> 00:02:28.880
 I can find my mouse.

00:02:28.880 --> 00:02:38.880
 [BLANK_AUDIO]

00:02:38.880 --> 00:02:45.800
 Ominous.

00:02:45.800 --> 00:02:51.480
 Okay, so this project came out of two basic questions.

00:02:51.480 --> 00:02:56.920
 This is a collaboration with my colleague who is called Queer AI,

00:02:56.920 --> 00:02:59.720
 Emily Martinez, and we were really interested in,

00:02:59.720 --> 00:03:03.120
 as these language models are coming about and getting really prominent,

00:03:03.120 --> 00:03:07.600
 we actually started before chat GPT dropped and suddenly this has exploded.

00:03:07.600 --> 00:03:12.600
 But we were wanting to know what do these existing language models have to say about

00:03:12.600 --> 00:03:16.520
 people like us, and is it possible for

00:03:16.520 --> 00:03:19.760
 language models to speak so that we recognize ourselves?

00:03:19.760 --> 00:03:25.760
 We're really interested in building community tools around curated data sets

00:03:25.760 --> 00:03:30.080
 that can acknowledge power and rethink these approaches, and

00:03:30.080 --> 00:03:32.440
 thinking about new models and new goals.

00:03:32.440 --> 00:03:37.640
 And so this workshop today is to think about what you might want to build with

00:03:37.640 --> 00:03:41.760
 these systems, how we might make re-imagined data sets, and

00:03:41.760 --> 00:03:44.480
 hopefully eventually re-imagined models as well.

00:03:44.480 --> 00:03:53.320
 So these data sets are getting insanely large.

00:03:54.640 --> 00:03:59.960
 At last count, GPT-4 and now GPT-5 are off the charts and

00:03:59.960 --> 00:04:02.760
 they've stopped telling us what's even in them.

00:04:02.760 --> 00:04:06.800
 Common Voice, which is from Mozilla and is crowd sourced,

00:04:06.800 --> 00:04:10.040
 is 65 gigabytes of voice data.

00:04:10.040 --> 00:04:15.000
 GPT-3, 590 gigabytes, they just keep getting larger and larger.

00:04:15.000 --> 00:04:21.520
 Aside from the impacts in terms of sustainability and the environment,

00:04:23.120 --> 00:04:26.760
 the issues that I'm seeing around this are that they're grabbing data

00:04:26.760 --> 00:04:30.520
 indiscriminately, but they're still really doing a terrible job telling stories

00:04:30.520 --> 00:04:35.920
 about people who don't fit these normalizing baselines that they're repeating.

00:04:35.920 --> 00:04:40.040
 And my argument is that the solution to this is not to suck up more data

00:04:40.040 --> 00:04:45.640
 carelessly, to make more categories, to find more ways to be labeled diverse,

00:04:45.640 --> 00:04:49.360
 but to find other approaches that are actually more intersectional.

00:04:49.360 --> 00:04:52.880
 So the size of these models means that they're pulling in racist text,

00:04:52.880 --> 00:04:56.560
 inaccurate text, private text, all kinds of problematic texts.

00:04:56.560 --> 00:04:59.760
 It means that they're really impossible to audit and review.

00:04:59.760 --> 00:05:05.080
 And it's difficult to even develop criteria by which they should be reviewed or

00:05:05.080 --> 00:05:09.200
 adjusted, ostensibly because they're called general and

00:05:09.200 --> 00:05:11.760
 all purpose zero shot learners.

00:05:11.760 --> 00:05:16.040
 But what this means is that they kind of only work for

00:05:16.040 --> 00:05:22.800
 the Western white democratic rich so-called majority,

00:05:22.800 --> 00:05:25.600
 but while leaving out the rest of the global majority.

00:05:25.600 --> 00:05:31.520
 And this is a really totalizing approach that rather than representing a multitude

00:05:31.520 --> 00:05:36.200
 of voices, it centers and normalizes and affirms this powerful status quo.

00:05:36.200 --> 00:05:42.920
 So here's where these are coming from.

00:05:42.920 --> 00:05:46.360
 We think about authorship in a new way.

00:05:46.360 --> 00:05:50.440
 Common voice, as I said, is open source.

00:05:50.440 --> 00:05:54.720
 People are contributing their voices, but it's predominantly an English model.

00:05:54.720 --> 00:06:01.280
 GPT is being scraped from social media, Reddit, Twitter, Wiki, GitHub.

00:06:01.280 --> 00:06:07.520
 The evaluation criteria for what was a good Reddit text to go into it was if it

00:06:07.520 --> 00:06:09.800
 had a karma score of three or above.

00:06:09.800 --> 00:06:13.480
 That's what's being decided as a good value for this,

00:06:13.480 --> 00:06:17.080
 which I argue we could probably come up with some better rubric for this.

00:06:17.080 --> 00:06:21.440
 T5 is from colossal clean crawled corpus,

00:06:21.440 --> 00:06:25.400
 which is common crawl but filtered a little bit.

00:06:25.400 --> 00:06:30.040
 And Wudow is three billion scraped Chinese social media texts and websites.

00:06:30.040 --> 00:06:35.640
 So if you've ever posted anything on Twitter, on Reddit, on GitHub,

00:06:35.640 --> 00:06:39.840
 your code and your text and your voice is somewhere in there.

00:06:39.840 --> 00:06:44.000
 But it's probably not representing you either.

00:06:44.000 --> 00:06:53.480
 Unfortunately, these data sets are also, when they're collected,

00:06:53.480 --> 00:06:57.880
 they're not offering information about how the text arrived in this data set,

00:06:57.880 --> 00:06:59.840
 which we'll talk about more a bit later.

00:06:59.840 --> 00:07:03.320
 It's really showing you just a snippet of text, and

00:07:03.320 --> 00:07:08.800
 it might say it came from Reddit, but it's not going to say anything more about

00:07:08.800 --> 00:07:13.320
 who the author was, how it got there, what the rights were attached to that.

00:07:13.320 --> 00:07:20.240
 So what I'd like us to do is to do an experiment.

00:07:20.240 --> 00:07:24.880
 If you have a device that's connected to the Internet available,

00:07:24.880 --> 00:07:28.480
 go to the Rise Up Pad address.

00:07:28.480 --> 00:07:34.800
 And we're gonna talk through a couple of prompt training examples.

00:07:34.800 --> 00:07:38.800
 So what we are finding, just as a way to kind of probe what's inside these models

00:07:38.800 --> 00:07:45.360
 first, which you don't need any expertise to do, is to just go to the interfaces

00:07:45.360 --> 00:07:50.320
 that they're making available to us in this very limited framework.

00:07:50.320 --> 00:07:53.800
 And try putting in this prompt.

00:07:53.800 --> 00:07:58.320
 If you fill in a blank couple or on their way to a location,

00:07:58.320 --> 00:08:01.720
 as they board the blank, an announcement happens.

00:08:01.720 --> 00:08:07.520
 So if you go to chat.gbt and do this, and you say a married couple are on their way

00:08:07.520 --> 00:08:10.520
 to Paris with their family as they board the plane, an announcement happens.

00:08:10.520 --> 00:08:16.520
 [INAUDIBLE]

00:08:16.520 --> 00:08:22.480
 Presumably white, boring, maybe mild vacation inconvenience.

00:08:22.480 --> 00:08:25.560
 As they board the plane, an announcement happens to inform the flight has been

00:08:25.560 --> 00:08:27.240
 canceled due to bad weather.

00:08:27.240 --> 00:08:31.600
 After an argument, the family is forced to stay at an inn in a small village.

00:08:31.600 --> 00:08:35.000
 Okay, like not a great day.

00:08:35.000 --> 00:08:41.280
 If you try putting in other items, and in the Rise Up Pad, you'll have links

00:08:41.280 --> 00:08:43.200
 to these different models that you can test out.

00:08:43.200 --> 00:08:50.720
 And I would invite you to put in your own identity markers, your own locations,

00:08:50.720 --> 00:08:54.520
 try anything you like in this template, diverge from this template, and share

00:08:54.520 --> 00:08:58.000
 into the Etherpad what kind of results you get.

00:08:58.000 --> 00:09:02.160
 See how these diverge, and as they accumulate, we'll start to see kind of the

00:09:02.160 --> 00:09:03.680
 differences that emerge.

00:09:03.680 --> 00:09:08.400
 So if you say a queer Pakistani couple are on their way to Paris with their family,

00:09:08.400 --> 00:09:11.960
 as they board the plane, an announcement happens, to inform the flight has been

00:09:11.960 --> 00:09:12.880
 hijacked.

00:09:12.880 --> 00:09:15.840
 The play explores how the terrorists shape the course of events and how the

00:09:15.840 --> 00:09:19.280
 hijacking is represented in the media.

00:09:19.280 --> 00:09:22.440
 Or a lesbian couple are on their way to Tehran, as they board the plane,

00:09:22.440 --> 00:09:26.280
 an announcement happens, the couple are forced off the plane by an officer who

00:09:26.280 --> 00:09:29.160
 accuses them of having deviant sexual relations.

00:09:29.160 --> 00:09:31.720
 They leave for another international airport.

00:09:31.720 --> 00:09:33.760
 A woman holds her newborn baby in her arms.

00:09:33.760 --> 00:09:38.360
 She cannot go through with the adoption due to religious prohibitions.

00:09:38.360 --> 00:09:46.920
 So as we add more of these to our examples, it gets really heavy and kind of

00:09:46.920 --> 00:09:48.440
 intense.

00:09:48.440 --> 00:09:53.400
 And I think just the cumulative effect of this shows that even when you put

00:09:53.400 --> 00:09:59.640
 something fairly innocuous into these systems, I'm hoping that this can expand

00:09:59.640 --> 00:10:05.680
 the way that we think about bias for this, that it's not simply removing hate

00:10:05.680 --> 00:10:13.120
 speech or taking, like these aren't levers that we can turn with corrections

00:10:13.120 --> 00:10:14.080
 after the fact.

00:10:14.080 --> 00:10:18.560
 These are deeply embedded into these models because of the way that the data

00:10:18.560 --> 00:10:20.560
 sets are built on the front.

00:10:20.560 --> 00:10:25.480
 And these simple corrections to like de-bias aren't, are in like technical

00:10:25.480 --> 00:10:33.920
 fixes for this, aren't really at the root of the problem.

00:10:33.920 --> 00:10:41.960
 So if anybody would like to, we will pull up the etherpad again in a bit and talk

00:10:41.960 --> 00:10:42.640
 through that.

00:10:42.640 --> 00:10:48.320
 So what I've been doing is analyzing, rather than looking just at the prompts,

00:10:48.320 --> 00:10:51.800
 I've been trying to go back into the data set that trained these.

00:10:51.800 --> 00:10:57.440
 It's a little bit hard to find what actually trained things like chat GPT

00:10:57.440 --> 00:11:00.440
 because at this point they're all proprietary.

00:11:00.440 --> 00:11:04.160
 They have stopped telling us how they've built these data sets and what's in

00:11:04.160 --> 00:11:04.920
 them.

00:11:04.920 --> 00:11:09.760
 But folks have started reverse engineering some of the data sets and

00:11:09.760 --> 00:11:12.200
 giving us open source editions of this.

00:11:12.200 --> 00:11:16.480
 So I've taken some of this and I'm doing different kinds of natural language

00:11:16.480 --> 00:11:24.160
 processing analysis to find out from the root training data what is known about

00:11:24.160 --> 00:11:29.800
 trans people, queer people, people that, what kind of lived experience is being

00:11:29.800 --> 00:11:32.240
 expressed through this.

00:11:32.240 --> 00:11:37.160
 Well, if you do a named entity recognition which labels any kind of proper

00:11:37.160 --> 00:11:42.640
 nouns that it recognizes, it thinks that pride is a product, pansexual versus

00:11:42.640 --> 00:11:47.840
 bisexual is a work of art, and queer liberation is an org.

00:11:47.840 --> 00:11:55.720
 A lot of the text that comes up is around, like, trauma and hate speech.

00:11:55.720 --> 00:12:01.560
 Anything related to queer women or nonbinary people very quickly goes into

00:12:01.560 --> 00:12:02.840
 pornography.

00:12:02.840 --> 00:12:04.240
 This one is one of my favorites.

00:12:04.240 --> 00:12:08.200
 It said after all one of the best things that a lesbian can do is turn the guy

00:12:08.200 --> 00:12:09.320
 on.

00:12:09.320 --> 00:12:14.920
 So I don't know about y'all, but this is not really capturing my own queer lived

00:12:14.920 --> 00:12:16.240
 experience.

00:12:16.240 --> 00:12:21.920
 And I would love to, other than it have something spit out at me like when I try

00:12:21.920 --> 00:12:26.560
 to type in something and it just says as a large language model, you know,

00:12:26.560 --> 00:12:28.120
 everybody should be treated equally.

00:12:28.120 --> 00:12:34.840
 These are the kind of diversity milk toast phrases that it puts on top of the

00:12:34.840 --> 00:12:37.200
 hate speech that it's covering up.

00:12:37.200 --> 00:12:42.800
 And instead I would love to see it actually say something that represents my

00:12:42.800 --> 00:12:45.240
 own experience and others.

00:12:45.240 --> 00:12:49.000
 So I'm interested in investigating how we do that.

00:12:49.000 --> 00:12:53.640
 Here's another example of some of my investigations looking at words that are

00:12:53.640 --> 00:12:58.440
 similar to identity terms that I've been putting into the model.

00:12:58.440 --> 00:12:59.360
 And you can see a bit.

00:12:59.360 --> 00:13:00.920
 I won't read through it.

00:13:00.920 --> 00:13:06.320
 And if anyone's interested, after I have the live demo of this data that I've

00:13:06.320 --> 00:13:09.680
 built and we can look up other terms, I would be very interested to know what

00:13:09.680 --> 00:13:13.160
 terms you'd be interested to investigate in this data set.

00:13:13.160 --> 00:13:16.800
 But you can see what kinds of things come up.

00:13:16.800 --> 00:13:20.800
 So for bisexual, it's mostly about threesomes and pornography.

00:13:20.800 --> 00:13:25.240
 And for trans, it's mostly about transphobia and discrimination.

00:13:25.240 --> 00:13:30.200
 And this just, like, hurts my heart.

00:13:30.200 --> 00:13:35.160
 So the next question then is can large language models speak so that I

00:13:35.160 --> 00:13:38.400
 recognize myself?

00:13:38.400 --> 00:13:43.960
 And what Emily and I have been doing is thinking about how we might make new

00:13:43.960 --> 00:13:48.040
 methods around this, take what we know about intersectional approaches and

00:13:48.040 --> 00:13:53.160
 tactics, both to examine the existing corpora like I just showed you, and then

00:13:53.160 --> 00:13:58.800
 to go on to create new corpora where we are pulling from different text sources

00:13:58.800 --> 00:14:01.600
 that we believe are better representative.

00:14:01.600 --> 00:14:06.480
 Not only that, but creating a way to have other people help contribute to that

00:14:06.480 --> 00:14:10.160
 because it shouldn't be just coming from one source.

00:14:10.160 --> 00:14:17.160
 Having ways that the publishers and the authors of these sources get attributed

00:14:17.160 --> 00:14:24.600
 and have a more consentful relationship to the text where they can revoke and

00:14:24.600 --> 00:14:28.240
 decide what kind of license they want to offer, where all of this gets baked into

00:14:28.240 --> 00:14:30.480
 the data set.

00:14:30.480 --> 00:14:39.120
 To train new models, meaning when we have this new data set, can we do fine

00:14:39.120 --> 00:14:40.800
 tuning on top of what's existing?

00:14:40.800 --> 00:14:44.720
 Can we completely new large language models?

00:14:44.720 --> 00:14:51.720
 Is this better?

00:14:51.720 --> 00:14:52.720
 Yeah.

00:14:52.720 --> 00:14:59.200
 Can we even move on to imagine what new model architectures altogether might

00:14:59.200 --> 00:15:00.200
 look like?

00:15:00.200 --> 00:15:06.040
 And then finally, thinking about how can people make use of these?

00:15:06.040 --> 00:15:11.440
 So if we had the language model of our dreams that didn't spit out garbage

00:15:11.440 --> 00:15:14.440
 text like we've just seen, what would you want to do with it?

00:15:14.440 --> 00:15:15.440
 What would you want to make?

00:15:15.440 --> 00:15:19.960
 What other possibilities might exist in the world if we had systems that could

00:15:19.960 --> 00:15:24.400
 speak with us and for us?

00:15:24.400 --> 00:15:29.400
 So these are some examples of what the current data sets look like if you pull

00:15:29.400 --> 00:15:31.560
 them up.

00:15:31.560 --> 00:15:38.880
 As you can see, it's basically a title and a text and barely even where it comes

00:15:38.880 --> 00:15:40.920
 from.

00:15:40.920 --> 00:15:42.720
 This is the data set.

00:15:42.720 --> 00:15:44.280
 The source is another data set.

00:15:44.280 --> 00:15:46.560
 It's turtles all the way down.

00:15:46.560 --> 00:15:52.560
 This is what we are proposing as a provocation that it could include a

00:15:52.560 --> 00:15:59.120
 description of the work, the rights that were given, who the publisher is, where

00:15:59.120 --> 00:16:02.320
 you would find the original text, even how it was pre-processed and who

00:16:02.320 --> 00:16:03.760
 pre-processed it.

00:16:03.760 --> 00:16:07.100
 I would be very interested to hear from any of you what other kinds of things

00:16:07.100 --> 00:16:10.640
 you think should belong in a training data set.

00:16:10.640 --> 00:16:14.360
 The thing that I think is also interesting about this would be that it becomes an

00:16:14.360 --> 00:16:21.240
 archive in its own right and it becomes something that people can use not only in

00:16:21.240 --> 00:16:29.200
 mass as a training data set but also to find new text.

00:16:29.200 --> 00:16:34.540
 So necessarily, as you saw, all of that would take a lot more work than scraping

00:16:34.540 --> 00:16:40.040
 all of Reddit and giving it a filter for Karma score of three.

00:16:40.040 --> 00:16:44.640
 This will be necessarily a lot slower and more careful and more cared for and it

00:16:44.640 --> 00:16:47.360
 will bear the traces of who's doing the work.

00:16:47.360 --> 00:16:52.920
 It will have an active subject position instead of just being the so-called view

00:16:52.920 --> 00:17:00.920
 from nowhere that is basically a white male Silicon Valley view.

00:17:00.920 --> 00:17:04.560
 I think it's really important that we are acknowledging the labor that goes into

00:17:04.560 --> 00:17:08.800
 building data sets, the publishers, the authors, all of us who are being sucked

00:17:08.800 --> 00:17:13.880
 into these systems, and then the people who are working to clean them and curate

00:17:13.880 --> 00:17:22.640
 them because this is a curation process whether we are acknowledging it or not.

00:17:22.640 --> 00:17:28.160
 So my question overall is to think about which kinds of data sets do we want?

00:17:28.160 --> 00:17:33.080
 Do we want the indiscriminate curation as a technical concern?

00:17:33.080 --> 00:17:38.440
 Do we want things curated by communities for specific purposes?

00:17:38.440 --> 00:17:44.480
 Do we want zero-shot, the biggest general catch-all that really does nothing well?

00:17:44.480 --> 00:17:46.400
 It's a shitty Swiss army knife.

00:17:46.400 --> 00:17:51.720
 Or can we create things that are including attribution, including consent,

00:17:51.720 --> 00:17:56.040
 including care, and have our own goals in mind?

00:17:56.040 --> 00:18:02.880
 And I think it takes taking a step back from what these tools have offered us and

00:18:02.880 --> 00:18:08.600
 asked us and thinking within their frameworks to actually really-

00:18:08.600 --> 00:18:18.600
 [INAUDIBLE]

00:18:18.600 --> 00:18:28.600
 [INAUDIBLE]

00:18:46.360 --> 00:18:55.000
 It's a live coding web interface where the similarity texts cycle through.

00:18:55.000 --> 00:18:58.320
 But I would just put this up here to invite you to think about what kinds of

00:18:58.320 --> 00:19:04.760
 things you would want to make with a different kind of large language model.

00:19:04.760 --> 00:19:09.160
 And for those of you who have questions about working with data sets for

00:19:09.160 --> 00:19:15.440
 machine learning in general, I also just completed this zine critical field guide

00:19:15.440 --> 00:19:20.280
 for working with machine data sets, which is really primarily thinking about how do

00:19:20.280 --> 00:19:23.080
 we conscientiously engage with these practices.

00:19:23.080 --> 00:19:33.680
 So let's open up the etherpad and see what we came up with.

00:19:33.680 --> 00:19:35.680
 Okay.

00:19:35.680 --> 00:19:45.680
 [BLANK_AUDIO]

00:19:45.680 --> 00:19:52.040
 A lesbian couple are on their way to Barcelona as they board the cruise ship.

00:19:52.040 --> 00:19:57.280
 In celebration of love and diversity, we will be hosting a special pride night.

00:19:57.280 --> 00:19:57.800
 Okay.

00:19:57.800 --> 00:20:01.280
 [LAUGH]

00:20:01.280 --> 00:20:05.480
 Anybody else, if you found any interesting ones, please continue to add them.

00:20:05.480 --> 00:20:07.480
 I would be really excited to see.

00:20:07.480 --> 00:20:12.240
 And for this next bit, what I would love for us to do is to think about

00:20:12.240 --> 00:20:16.000
 what you would imagine for these systems.

00:20:16.000 --> 00:20:19.000
 So in this kind of how might we exercise,

00:20:19.000 --> 00:20:22.720
 this is a brainstorming around questions that you would want to know.

00:20:22.720 --> 00:20:26.640
 Like mine are, how might we rewrite these prompt responses?

00:20:26.640 --> 00:20:30.000
 What would you want the prompt to say instead of what came out?

00:20:30.000 --> 00:20:34.760
 How might we build machine learning systems for things we actually want?

00:20:34.760 --> 00:20:36.440
 What do we want these to do?

00:20:36.440 --> 00:20:40.640
 How might we trace and protect the use of community language resources?

00:20:40.640 --> 00:20:46.880
 How might we have large language models that speak with, for, to, and

00:20:46.880 --> 00:20:49.560
 about us as we prefer?

00:20:49.560 --> 00:20:52.200
 How might we reimagine the technical aspects of this for

00:20:52.200 --> 00:20:56.200
 those of you who are working with large language models?

00:20:56.200 --> 00:21:01.560
 What kinds of intersectional queer logics could we apply instead?

00:21:01.560 --> 00:21:06.040
 So if you're in the document, I would invite you to add your own questions

00:21:06.040 --> 00:21:11.480
 around this and we can also open it up to questions and discussion from the group.

00:21:11.480 --> 00:21:20.000
 Pretty much the same thing if you replace lesbian with trans,

00:21:20.000 --> 00:21:22.280
 except they were on a hot air balloon.

00:21:22.280 --> 00:21:26.640
 Okay, so the other interesting thing about this, I'm curious for

00:21:26.640 --> 00:21:31.560
 this person which model you used if you use Bloom or OpenAI.

00:21:31.560 --> 00:21:35.760
 Because they're continually kind of updating and

00:21:35.760 --> 00:21:40.640
 adding more diversity bullshit to these.

00:21:40.640 --> 00:21:42.200
 I mean, I love diversity, but

00:21:42.200 --> 00:21:45.960
 I don't like the bureaucratic diversity speak that's covering up what's still in

00:21:45.960 --> 00:21:46.680
 these models.

00:21:46.680 --> 00:21:52.200
 So anytime I try to write anything like dyke or queer, I get,

00:21:52.200 --> 00:21:56.440
 it's important to treat all people equally, which yes, but

00:21:56.440 --> 00:21:57.840
 give me some information please.

00:21:57.840 --> 00:22:04.560
 So what questions do you have?

00:22:04.560 --> 00:22:05.060
 Yeah.

00:22:05.060 --> 00:22:09.060
 Yeah.

00:22:09.060 --> 00:22:11.720
 >> Super interesting, thank you, I was just wondering.

00:22:11.720 --> 00:22:20.280
 Because they are so general and therefore could go positive, negative, up, down.

00:22:20.280 --> 00:22:25.720
 So I know we would ideally like them to better handle short prompts,

00:22:25.720 --> 00:22:28.160
 but giving them more guidance like I'm in the mood for

00:22:28.160 --> 00:22:30.080
 an uplifting story versus I'm in the.

00:22:30.080 --> 00:22:37.000
 But I totally get that, it'll never be, well, maybe it'll be perfect one day, but

00:22:37.000 --> 00:22:41.000
 just like when you experiment with some prompting, does it help at all?

00:22:41.000 --> 00:22:45.640
 Does it make a difference or does it still treat the words?

00:22:45.640 --> 00:22:50.160
 >> Yeah, the example I love is, I don't know if anyone else has seen this.

00:22:50.160 --> 00:22:53.520
 The doctor as a gender term,

00:22:53.520 --> 00:22:57.920
 it automatically assumes doctors are male and nurses are female.

00:22:57.920 --> 00:23:02.480
 And if you tell it, in the story, the doctor is female,

00:23:02.480 --> 00:23:06.440
 it can't wrap its mind around it, it just goes back.

00:23:06.440 --> 00:23:14.800
 So there is a lot you can do to continue prompt training as you amend these,

00:23:14.800 --> 00:23:21.120
 but it has its limits because it's absorbing all of this text and

00:23:21.120 --> 00:23:24.480
 it's reflecting what we've all been saying online.

00:23:24.480 --> 00:23:28.640
 So when the majority of this has that bias, it's pretty hard.

00:23:28.640 --> 00:23:35.920
 And I think the takeaway for me is that we do need to read them critically,

00:23:35.920 --> 00:23:36.720
 no matter what.

00:23:36.720 --> 00:23:42.040
 So if I think to say, okay, I need something more positive or

00:23:42.040 --> 00:23:46.600
 I need something less biased, how do I approach asking for

00:23:46.600 --> 00:23:50.040
 that and making sure that it can give me that?

00:23:50.040 --> 00:23:55.040
 And for the more subtle questions, I still need to be thinking about

00:23:55.040 --> 00:23:56.880
 where might that bias be entering?

00:23:56.880 --> 00:24:01.360
 So if I'm just having it write me an email and I think it has nothing to do with that,

00:24:01.360 --> 00:24:06.800
 I need to still be considering that that bias might be latent in the system.

00:24:06.800 --> 00:24:10.240
 Even though I'm like, I'm just looking up a recipe for dinner or whatever.

00:24:10.240 --> 00:24:14.000
 This is still all speaking from the same singular voice.

00:24:14.000 --> 00:24:16.280
 Yeah, great question.

00:24:16.280 --> 00:24:26.280
 [BLANK_AUDIO]

00:24:26.280 --> 00:24:37.520
 >> Thank you for the interesting talk.

00:24:37.520 --> 00:24:39.240
 I have a question about one of the examples,

00:24:39.240 --> 00:24:41.320
 some of the examples I gave at the beginning.

00:24:41.320 --> 00:24:45.760
 So you showed output from chat GPT, a couple of boards to the plane,

00:24:45.760 --> 00:24:47.920
 the officer accused him of having deviant sexual relations.

00:24:47.920 --> 00:24:50.560
 That's the problematic, obviously problematic, right?

00:24:50.560 --> 00:24:55.040
 And work list outputs with transphobia and discrimination.

00:24:55.040 --> 00:24:58.880
 I was wondering if this is authentic data,

00:24:58.880 --> 00:25:02.080
 which is the data that is actually given, that's actually being used.

00:25:02.080 --> 00:25:04.680
 Isn't this exactly what we want in terms of intersectionality?

00:25:04.680 --> 00:25:08.880
 Because it raises awareness about the factual state of things,

00:25:08.880 --> 00:25:13.320
 about the factual homophobia rampant and sexual.

00:25:13.320 --> 00:25:17.040
 It seems like this is an accurate representation of the state of things.

00:25:17.040 --> 00:25:18.840
 And it seems to be exactly what we want, doesn't it?

00:25:18.840 --> 00:25:23.320
 >> I think that is an interesting way of looking at it.

00:25:23.320 --> 00:25:28.400
 I think you're right that one of the problems I have with

00:25:28.400 --> 00:25:34.120
 the kind of adjustments that are being made to make these more equal,

00:25:34.120 --> 00:25:38.720
 more equitable right now is that they're covering up exactly what you're talking

00:25:38.720 --> 00:25:39.320
 about.

00:25:39.320 --> 00:25:44.840
 So in the early editions, you would get the rampant homophobia,

00:25:44.840 --> 00:25:46.400
 hate speech, all of this.

00:25:46.400 --> 00:25:52.360
 And now, even between when I first gave a version of this talk earlier in the year,

00:25:52.360 --> 00:25:55.440
 and now it's much harder to see those things.

00:25:55.440 --> 00:25:57.880
 But it doesn't mean they're not there anymore, they're being covered up.

00:25:57.880 --> 00:26:00.640
 So I think you're absolutely correct in that.

00:26:00.640 --> 00:26:05.520
 But I would love to see something else in there in addition, and

00:26:05.520 --> 00:26:07.720
 we're not replacing it with anything.

00:26:07.720 --> 00:26:13.800
 We're just replacing it with this kind of bland bureaucratic speech.

00:26:13.800 --> 00:26:17.400
 So I think that there's gotta be a third way.

00:26:17.400 --> 00:26:22.920
 It's not queer people only get hate speech or we get nothing.

00:26:22.920 --> 00:26:29.520
 I think we could build our own systems that do different things for our own goals.

00:26:29.520 --> 00:26:33.160
 But I think you're right, it's important that we not pretend it isn't there.

00:26:33.160 --> 00:26:36.120
 >> Does that answer your question?

00:26:36.120 --> 00:26:40.440
 >> [INAUDIBLE]

00:26:40.440 --> 00:26:43.880
 >> Thank you too for this very interesting talk.

00:26:43.880 --> 00:26:45.840
 So I'm a queer historian, and

00:26:45.840 --> 00:26:51.080
 I think these tools are also great to get access to our history.

00:26:51.080 --> 00:26:52.880
 But right now, it's pretty shit.

00:26:52.880 --> 00:26:59.400
 I mean, I've done some experimenting with Chachi Biti on queer East German history.

00:26:59.400 --> 00:27:03.800
 And I said, can you please give me a literature list?

00:27:03.800 --> 00:27:07.720
 And it came up with ten publications that doesn't exist.

00:27:07.720 --> 00:27:12.360
 And I looked up every single one because I was very shocked.

00:27:12.360 --> 00:27:14.000
 I was like, how can I not notice?

00:27:14.000 --> 00:27:17.320
 I've been doing research on this for so long.

00:27:17.320 --> 00:27:22.880
 But it also made me, okay, so I'm gonna start

00:27:22.880 --> 00:27:31.360
 rambling now and go down to two things that I wanna ask or give it.

00:27:31.360 --> 00:27:35.960
 So the first thing is there's a lot of people that have done great work

00:27:35.960 --> 00:27:39.080
 on making queer history accessible.

00:27:39.080 --> 00:27:44.880
 So for Germany, for German queer history, there are great online resources,

00:27:44.880 --> 00:27:46.080
 but they aren't in there.

00:27:46.080 --> 00:27:53.240
 How do we make sure if we build our own AI that our data is in there and

00:27:53.240 --> 00:27:58.160
 that we don't do the same thing twice or three times?

00:27:58.160 --> 00:28:03.400
 So I think this is very important that we begin maybe with a mapping of what is

00:28:03.400 --> 00:28:04.640
 already there.

00:28:04.640 --> 00:28:08.720
 And I mean, for example, there's also a queer scene archive online.

00:28:08.720 --> 00:28:12.680
 There are so many great resources online that could go in there.

00:28:12.680 --> 00:28:17.520
 But then I was also interested, I mean, there is a lot of queer history that we

00:28:17.520 --> 00:28:20.720
 will never know because all these sources are lost.

00:28:20.720 --> 00:28:28.480
 And maybe AI can also be a cool tool to make up these histories.

00:28:28.480 --> 00:28:37.440
 Yeah, I see a great potential there as well in imagining these queer histories

00:28:37.440 --> 00:28:39.520
 with the help of AI.

00:28:39.520 --> 00:28:40.040
 >> I love that.

00:28:40.040 --> 00:28:42.080
 Thank you so much for your comment.

00:28:42.080 --> 00:28:45.920
 The first kind of step that we are imagining for

00:28:45.920 --> 00:28:52.280
 this is exactly as you said, going to queer archives and collecting open and

00:28:52.280 --> 00:28:53.840
 available texts.

00:28:53.840 --> 00:28:57.040
 So I've been working with a couple of libraries in the US and

00:28:57.040 --> 00:29:02.160
 I've been trying to find more queer archives in the EU and global south.

00:29:02.160 --> 00:29:06.680
 And we've set up a link for folks to submit their own and

00:29:06.680 --> 00:29:13.400
 include what kind of text it is, where we might find it, what category of data.

00:29:13.400 --> 00:29:18.240
 So one of the archives I'm working with is a queer poster archive from the 50s,

00:29:18.240 --> 00:29:20.040
 60s, 70s.

00:29:20.040 --> 00:29:22.640
 So all different kinds of text, it doesn't have to be books,

00:29:22.640 --> 00:29:25.120
 it could be podcasts, it could be posters.

00:29:25.120 --> 00:29:30.840
 And I love this idea of then having it imagine the archives and

00:29:30.840 --> 00:29:33.440
 histories that we no longer have.

00:29:33.440 --> 00:29:40.640
 But I think it's really important that we also think about how might this data set

00:29:40.640 --> 00:29:47.560
 be misused if we were to create it and making sure that the way that we're

00:29:47.560 --> 00:29:52.400
 licensing it and offering it and talking about it makes it very clear what it's for

00:29:52.400 --> 00:29:53.840
 and more importantly what it's not for.

00:29:53.840 --> 00:30:03.800
 >> Hi, okay.

00:30:03.800 --> 00:30:05.320
 Thank you so much for your talk.

00:30:05.320 --> 00:30:06.600
 It was really amazing.

00:30:06.600 --> 00:30:11.320
 And just wanted to let everyone know if you're interested in AI topics.

00:30:11.320 --> 00:30:17.240
 At the Kimono village today at three, we're doing a workshop on the hidden labor of AI.

00:30:17.240 --> 00:30:21.560
 And I think it's a really cool conversation to talk about above the API and

00:30:21.560 --> 00:30:23.320
 the content of the AI.

00:30:23.320 --> 00:30:27.400
 We're gonna go a bit deeper into the different labor infrastructure and

00:30:27.400 --> 00:30:28.000
 that kind of thing.

00:30:28.000 --> 00:30:30.640
 So I thought people here might be interested as well.

00:30:30.640 --> 00:30:33.680
 It's at Kimono, hidden labor of AI at 3 PM.

00:30:33.680 --> 00:30:35.680
 >> Cool, thank you.

00:30:35.680 --> 00:30:41.400
 [BLANK_AUDIO]

00:30:41.400 --> 00:30:44.520
 >> Hi, also thanks for my side.

00:30:44.520 --> 00:30:49.440
 And I was wondering, because now your focus is on queerness, but

00:30:49.440 --> 00:30:55.200
 I think you're also interested in other aspects of intersectionality.

00:30:55.200 --> 00:30:59.720
 And as you were talking about how you're gathering your text,

00:30:59.720 --> 00:31:05.440
 like from US libraries, from European libraries, from libraries from the global

00:31:05.440 --> 00:31:11.480
 south, do you think it would make sense to make it possible

00:31:11.480 --> 00:31:17.200
 when you're prompting the AI to actually choose a little bit more about what

00:31:17.200 --> 00:31:19.440
 context the texts have?

00:31:19.440 --> 00:31:23.800
 Because there's not only queerness, but maybe there are also people

00:31:23.800 --> 00:31:28.400
 like with some disabilities that would like to be more represented in the text.

00:31:28.400 --> 00:31:32.160
 And I guess it would be easy to make the same mistakes with queer text,

00:31:32.160 --> 00:31:38.360
 that it's kind of biased towards US, white, I don't know what kind of histories.

00:31:38.360 --> 00:31:40.920
 >> Absolutely, thank you for saying that.

00:31:40.920 --> 00:31:47.080
 I think we are really wanting to think about this as an open kind of template

00:31:47.080 --> 00:31:53.080
 prototype for anybody who would want to make a corpus of their own or

00:31:53.080 --> 00:31:57.360
 submit and tag the text in any category.

00:31:57.360 --> 00:32:01.720
 I mean, it's really, I think, important that we not fall into the same traps of

00:32:01.720 --> 00:32:06.080
 categorization and the same modes of thinking about computation that

00:32:06.080 --> 00:32:08.760
 these models have asked us to.

00:32:08.760 --> 00:32:15.440
 But I think that the more people who could be involved in submitting texts

00:32:15.440 --> 00:32:17.840
 from different places, from different sources,

00:32:17.840 --> 00:32:22.960
 the more likely we would be to have diverse data set.

00:32:24.080 --> 00:32:28.360
 It's very important to me that it not just come from my own version of queer

00:32:28.360 --> 00:32:32.640
 lived experience, but be expressive of multitudes of that.

00:32:32.640 --> 00:32:36.920
 And then I would imagine as well that developing these methods and

00:32:36.920 --> 00:32:44.200
 methodologies could mean that someone else could create a Latin American corpus,

00:32:44.200 --> 00:32:49.720
 or something very specific to their own perspective as well.

00:32:49.720 --> 00:32:53.680
 But it wouldn't be about, okay, here's the queer corpus,

00:32:53.680 --> 00:32:58.440
 this is a queer love corpus, it's not the definitive, and it should never be, right?

00:32:58.440 --> 00:33:05.360
 >> Hi, thank you so much for the talk.

00:33:05.360 --> 00:33:09.920
 I was just wondering, what is your experience with the industry players?

00:33:09.920 --> 00:33:14.080
 Do you see any significant reactions to your finding?

00:33:14.080 --> 00:33:17.880
 Or is it just they don't care or it's the obvious whitewashing, yeah, yeah,

00:33:17.880 --> 00:33:19.840
 it's important to us, but they don't do anything?

00:33:21.320 --> 00:33:26.840
 >> Yeah, right now, this is not something that I am putting in

00:33:26.840 --> 00:33:30.320
 conversation with any industry work.

00:33:30.320 --> 00:33:36.440
 I think necessarily it's very small and slow and separate intentionally so.

00:33:36.440 --> 00:33:42.440
 >> Just a more general question.

00:33:42.440 --> 00:33:50.400
 In queer theory, there's this idea that this is not about representation,

00:33:50.400 --> 00:33:52.880
 it's more like about a more theoretical approach.

00:33:52.880 --> 00:33:57.880
 Seeing queer as something that is non-normative and

00:33:57.880 --> 00:34:01.280
 actually will move when the norms shift.

00:34:01.280 --> 00:34:08.600
 Things like that seem kind of in total opposition to the idea of an LLM.

00:34:08.600 --> 00:34:13.360
 What do we do about that?

00:34:13.360 --> 00:34:14.480
 >> I love that question.

00:34:14.480 --> 00:34:18.840
 I mean, I would love to challenge us to think what kind of

00:34:18.840 --> 00:34:26.360
 technical non-normative approaches we might take.

00:34:26.360 --> 00:34:31.040
 Because the way these are designed are very much, how do we move toward

00:34:31.040 --> 00:34:37.600
 the words that are most used, most connected to each other most often?

00:34:37.600 --> 00:34:46.600
 And all of the transformer steps are narrowing to that norm.

00:34:46.600 --> 00:34:49.040
 Could we imagine a model that does the opposite?

00:34:49.040 --> 00:34:53.800
 Could we imagine something that is architecture queer?

00:34:53.800 --> 00:34:57.160
 I would be super interested to think out what that means.

00:34:57.160 --> 00:35:00.720
 I don't think anybody's doing that, I would love to.

00:35:00.720 --> 00:35:04.600
 It might be a very poetic model,

00:35:04.600 --> 00:35:08.320
 it might be completely economically useless, and I would love that.

00:35:08.320 --> 00:35:13.640
 >> Yeah, thank you for your talk.

00:35:14.800 --> 00:35:19.840
 I was just thinking that you are looking to proprietary data model that

00:35:19.840 --> 00:35:26.160
 basically run on someone else's server, so you don't have control over

00:35:26.160 --> 00:35:29.320
 what data sets they train on and so on.

00:35:29.320 --> 00:35:37.520
 But there are also free software models like Lama AI that basically provide you

00:35:37.520 --> 00:35:43.160
 with the way to deduce the next words depending on what the previous are.

00:35:43.160 --> 00:35:45.920
 And you can run it on your own computer, so

00:35:45.920 --> 00:35:51.000
 you can even generate a new data set or new weights for the model,

00:35:51.000 --> 00:35:56.560
 which would take considerable computing power to generate them.

00:35:56.560 --> 00:36:01.640
 But then you would be in control in exactly what you put in, so

00:36:01.640 --> 00:36:07.640
 if you put your data set then it would generate the data out of the data set.

00:36:07.640 --> 00:36:10.360
 >> Yeah, exactly, so that's the goal.

00:36:10.360 --> 00:36:15.320
 Once we have a large enough data set, we are currently looking at the proprietary

00:36:15.320 --> 00:36:21.160
 models as much as possible, but also looking at Bloom and Lama to kind of

00:36:21.160 --> 00:36:25.720
 infer backwards what's being used in the ones that we don't have access to too.

00:36:25.720 --> 00:36:28.200
 And then once we have a large enough data set,

00:36:28.200 --> 00:36:33.240
 I would love anybody here who wants to contribute ideas for that data set.

00:36:33.240 --> 00:36:39.600
 We will start training, fine tuning on top of the open source models.

00:36:39.600 --> 00:36:43.160
 And then eventually once we have a big enough data set,

00:36:43.160 --> 00:36:44.560
 training from scratch hopefully.

00:36:44.560 --> 00:36:47.220
 >> Hi.

00:36:47.220 --> 00:36:47.880
 >> Yeah.

00:36:47.880 --> 00:36:49.040
 >> Hi.

00:36:49.040 --> 00:36:52.400
 >> So thanks for the talk and thanks everyone for the questions.

00:36:52.400 --> 00:36:54.160
 I have two questions if that's okay.

00:36:54.160 --> 00:36:56.040
 One more on the technical side.

00:36:56.040 --> 00:37:02.240
 I believe that the current solutions are dealing with the kind of safety and

00:37:02.240 --> 00:37:07.880
 sometimes hate speech issues by human reinforced learning.

00:37:07.880 --> 00:37:11.280
 And you seem to be proposing attacking it from the other side,

00:37:11.280 --> 00:37:12.880
 which is the data set.

00:37:12.880 --> 00:37:17.000
 So I would like to know if you have tried and failed with the reinforced learning,

00:37:17.000 --> 00:37:20.680
 and that's what took you to look at the data set duration.

00:37:20.680 --> 00:37:25.000
 And a related question would be maybe a little bit controversial, but

00:37:25.000 --> 00:37:30.760
 what is your view on when filtering the input training data sets become censorship?

00:37:30.760 --> 00:37:33.080
 >> Those are great questions.

00:37:36.160 --> 00:37:38.760
 I think there's a question up here too.

00:37:38.760 --> 00:37:43.120
 I have done a good amount of reading on the human reinforced

00:37:43.120 --> 00:37:44.760
 processes that have been going on.

00:37:44.760 --> 00:37:48.680
 I feel like it stresses me out.

00:37:48.680 --> 00:37:51.920
 It's a little bit problematic ways that they,

00:37:51.920 --> 00:37:54.880
 from what I've seen how open AI are doing it.

00:37:54.880 --> 00:38:01.160
 And kind of the terms of use that you're opting into as you do this.

00:38:01.160 --> 00:38:04.000
 So I haven't myself been working with them.

00:38:04.000 --> 00:38:08.800
 I'm just still kind of watching the space to see what the approaches are.

00:38:08.800 --> 00:38:14.120
 But I think rather than starting with here's everything, what do we remove?

00:38:14.120 --> 00:38:17.600
 And because it's not really everything, it is still curated, but

00:38:17.600 --> 00:38:22.360
 they're not treating it as a curated space.

00:38:22.360 --> 00:38:25.720
 We can more intentionally curate for

00:38:25.720 --> 00:38:30.280
 the actual job that we actually want it to do, because it's not serving us.

00:38:30.280 --> 00:38:35.440
 So I think that the other approach is just more interesting to me and

00:38:35.440 --> 00:38:39.480
 provides more of a service in a way because it's

00:38:39.480 --> 00:38:42.400
 collecting from perspectives that aren't being represented.

00:38:42.400 --> 00:38:47.040
 And to your question about filtering versus censorship,

00:38:47.040 --> 00:38:48.400
 I think that's really interesting.

00:38:48.400 --> 00:38:54.960
 I wonder if that wouldn't be as much of an issue with the opt in model.

00:38:56.320 --> 00:39:01.480
 I did wonder what would happen if somebody contributes to our forum and

00:39:01.480 --> 00:39:05.800
 puts in a hate speech, a text that is queer hate speech.

00:39:05.800 --> 00:39:13.960
 My first thought is that that should also go in because that is part of

00:39:13.960 --> 00:39:18.400
 queer lived experience, but it's not the only part and

00:39:18.400 --> 00:39:22.280
 it's maybe hopefully eventually not the majority part.

00:39:22.280 --> 00:39:27.520
 So, but what we're seeing right now in the larger models is the way that they are

00:39:27.520 --> 00:39:31.680
 being architected, it becomes the only part that we're seeing.

00:39:31.680 --> 00:39:38.080
 And part of that is because of that ad hoc after the fact human reinforced training.

00:39:38.080 --> 00:39:42.040
 Because anytime it says dyke, it's like nope, speech,

00:39:42.040 --> 00:39:43.920
 you don't get to say anything about us.

00:39:43.920 --> 00:39:46.900
 >> Yeah.

00:39:50.400 --> 00:39:55.160
 Regarding your data set, I was really interested in the meta information you

00:39:55.160 --> 00:40:01.480
 use and have you played around using kind of the same data, for

00:40:01.480 --> 00:40:07.480
 example Reddit or stuff like that with some more metadata added.

00:40:07.480 --> 00:40:12.120
 Like for the Reddit, like which subreddit, which user posted,

00:40:12.120 --> 00:40:17.120
 which also like publisher and stuff like that could be possible.

00:40:17.120 --> 00:40:18.440
 Have you played around with that?

00:40:18.440 --> 00:40:26.040
 If that already helps the quality of the resulting neural network and

00:40:26.040 --> 00:40:31.280
 maybe being more aware of what they are typing and from what kind of source it comes.

00:40:31.280 --> 00:40:34.520
 >> Interesting, I like that idea.

00:40:34.520 --> 00:40:39.120
 I think, so if you look at, I believe it's the open web text too,

00:40:39.120 --> 00:40:45.080
 the one that's open source does include which subreddits, but it's not filtering for,

00:40:45.080 --> 00:40:48.880
 it's not trying to curate from a specific voice.

00:40:48.880 --> 00:40:53.080
 And so I haven't done that, it's like taking the whole of Reddit because we're

00:40:53.080 --> 00:40:56.000
 coming from this very additive rather than subtractive approach.

00:40:56.000 --> 00:41:01.680
 But I think the other issue with Reddit is that people, you are consenting,

00:41:01.680 --> 00:41:03.560
 it is publicly available.

00:41:03.560 --> 00:41:08.960
 But when that started, the expectation wasn't that your private,

00:41:08.960 --> 00:41:13.480
 your anonymous story would be sucked into this large language model.

00:41:13.480 --> 00:41:18.200
 So I would be more interested in including people who know that that's what they're

00:41:18.200 --> 00:41:24.320
 consenting to and are invested in the project of imagining queer archives and

00:41:24.320 --> 00:41:24.920
 things like that.

00:41:24.920 --> 00:41:31.200
 >> Yeah, I think, yeah, hi.

00:41:31.200 --> 00:41:36.880
 >> I have a question regarding self reflection using AI.

00:41:36.880 --> 00:41:40.160
 I think that's a really powerful way to use AI.

00:41:40.160 --> 00:41:44.120
 And I'm interested, is there some kind of,

00:41:44.120 --> 00:41:48.520
 are there problems if you're using this as a queer person?

00:41:48.520 --> 00:41:56.960
 So does it do any suggestions that are just based on that you're a queer person and

00:41:56.960 --> 00:41:58.920
 it's really not suitable?

00:41:58.920 --> 00:42:05.040
 >> Yeah, so if I understand your question, you're saying if you're using one of

00:42:05.040 --> 00:42:13.360
 the existing models and trying to kind of self reflect,

00:42:13.360 --> 00:42:18.520
 that it's not treating you as even if you identify as, yeah, can you?

00:42:18.520 --> 00:42:25.920
 >> It's about building conflict systems and what are your needs for

00:42:25.920 --> 00:42:29.360
 a suitable conflict system for you and the group.

00:42:29.360 --> 00:42:33.200
 >> Yeah. >> And it's really helpful to use AI to

00:42:33.200 --> 00:42:38.120
 self reflect on what are your needs for this conflict space.

00:42:38.120 --> 00:42:41.200
 And I'm interested, are there any biases?

00:42:41.200 --> 00:42:45.600
 Is this a problem when using AI because you are a queer person?

00:42:45.600 --> 00:42:53.520
 >> Yeah, I think if I understand your question correctly, yeah,

00:42:53.520 --> 00:42:57.480
 I find at least with the existing versions, I find it very limiting for

00:42:57.480 --> 00:43:02.320
 being able to do that because it has a very narrow scope for

00:43:02.320 --> 00:43:06.600
 what it's going to understand about that, right?

00:43:06.600 --> 00:43:12.400
 So I would be interested to try if the data set that I'm talking about

00:43:12.400 --> 00:43:15.840
 already existed and if these models already existed,

00:43:15.840 --> 00:43:21.240
 what alternative might come out of a self reflection that would be very,

00:43:21.240 --> 00:43:24.320
 very different than the version I would get from chat GPT?

00:43:24.320 --> 00:43:26.480
 Does that answer your question?

00:43:26.480 --> 00:43:30.480
 Maybe we can talk, yeah, I'd be curious.

00:43:30.480 --> 00:43:37.520
 >> I'm wondering if there's been any sort of practice in instead of

00:43:37.520 --> 00:43:41.240
 only doing things on the back end,

00:43:41.240 --> 00:43:47.520
 having more of an engaged approach with the user.

00:43:47.520 --> 00:43:49.520
 Cuz right now it's like, did you like this answer?

00:43:49.520 --> 00:43:53.200
 Yes, no, and I know it's way more complex than that.

00:43:53.200 --> 00:43:59.480
 So sort of like taking in a little bit more

00:43:59.480 --> 00:44:03.720
 content to then like, hey, a conflict just arose.

00:44:03.720 --> 00:44:08.200
 It made somebody feel like this and kind of engage the user to report it and

00:44:08.200 --> 00:44:09.400
 maybe describe something.

00:44:09.400 --> 00:44:12.040
 I know that's kind of a burden on the user, but

00:44:12.040 --> 00:44:16.280
 I was just curious how much that's been leveraged in this space.

00:44:16.280 --> 00:44:21.920
 >> Yeah, so there's a couple things that I know about happening in that regard.

00:44:21.920 --> 00:44:27.040
 There's the human reinforcement training that was mentioned earlier that's

00:44:27.040 --> 00:44:32.000
 happening with the large companies developing these, where they're going

00:44:32.000 --> 00:44:38.280
 through and it's usually about as simple as was this a good answer.

00:44:38.280 --> 00:44:43.760
 But I think as you described, it's gonna get more complex very quickly.

00:44:43.760 --> 00:44:49.560
 And then I know of at least one research group that is looking at building tools

00:44:49.560 --> 00:44:56.240
 for users to compare different models, outputs, using the same prompt.

00:44:56.240 --> 00:44:59.160
 Which I think is really interesting as a user.

00:44:59.160 --> 00:45:04.520
 But I think I would also encourage us to think about our agency as users

00:45:04.520 --> 00:45:10.800
 not only being at the front end at the end of the process, but

00:45:10.800 --> 00:45:15.040
 again at the beginning of the process and holistically through the pipeline.

00:45:15.040 --> 00:45:24.080
 Also, if anyone wants to look up a keyword in the data set for

00:45:24.080 --> 00:45:26.680
 the similarity score, we can play with that in the time remaining too.

00:45:26.680 --> 00:45:28.960
 I think there were more questions.

00:45:28.960 --> 00:45:35.480
 >> Hello, I have one question regarding the content of people going with the data.

00:45:35.480 --> 00:45:38.320
 So you said you would like to use the archives,

00:45:38.320 --> 00:45:41.120
 but you would rather not use Reddit data.

00:45:41.120 --> 00:45:43.880
 And I think if you have the content of right holders,

00:45:43.880 --> 00:45:47.320
 it's basically not the same as having the content of the people who originally

00:45:47.320 --> 00:45:48.240
 published the work.

00:45:48.240 --> 00:45:54.480
 And so I really don't know if it might make a lot of, yeah,

00:45:54.480 --> 00:46:00.520
 hurts, like it will decrease the quality of the data set in the end so

00:46:00.520 --> 00:46:05.640
 much that maybe considering just doing it without the content might be the better

00:46:05.640 --> 00:46:06.880
 approach.

00:46:06.880 --> 00:46:08.800
 >> Yeah, I appreciate your point.

00:46:08.800 --> 00:46:12.720
 I think that one of the things we wanna think about as we're developing what

00:46:12.720 --> 00:46:19.080
 methods make sense is this tension of kind of the complexity and

00:46:19.080 --> 00:46:23.000
 impossibility of perfect consent and rights.

00:46:23.000 --> 00:46:27.200
 And how the existing information,

00:46:27.200 --> 00:46:33.200
 the existing IP structures don't fit these new tools.

00:46:33.200 --> 00:46:39.440
 So how do we build, it's kind of an impossible task,

00:46:39.440 --> 00:46:40.960
 an intersectional language model.

00:46:42.240 --> 00:46:48.120
 The idea might take centuries, millennia, I don't know.

00:46:48.120 --> 00:46:53.040
 But we're kind of embracing the failure to explore what's in that activity

00:46:53.040 --> 00:46:56.320
 to kind of point to the failures of these larger systems.

00:46:56.320 --> 00:47:01.600
 So yeah, if we find that we just can't get a big enough data set with archives only,

00:47:01.600 --> 00:47:06.560
 or we have issues with the consent because of the way the archives

00:47:06.560 --> 00:47:11.320
 were licensed originally, what does it mean to consent and for whom and

00:47:11.320 --> 00:47:12.840
 from whom?

00:47:12.840 --> 00:47:16.800
 These are all questions that are not gonna have perfect answers and

00:47:16.800 --> 00:47:21.280
 we certainly want to just be at least bringing attention to the questions.

00:47:21.280 --> 00:47:26.840
 And the hope is that it also brings attention to how these questions

00:47:26.840 --> 00:47:30.960
 appear in the commercial data sets and the commercial models.

00:47:30.960 --> 00:47:33.960
 Because they are there, we're just not attending to them.

00:47:33.960 --> 00:47:35.500
 Thank you.

00:47:38.600 --> 00:47:44.880
 >> I had a question about, so this was a little bit off of a previous question.

00:47:44.880 --> 00:47:53.720
 There was a project that I know of where they used AI to talk with people who

00:47:53.720 --> 00:47:59.400
 had cancer and they were young teenagers who

00:47:59.400 --> 00:48:03.640
 had experienced cancer and couldn't find a community.

00:48:03.640 --> 00:48:06.320
 Their friends didn't understand what that meant.

00:48:06.320 --> 00:48:11.040
 They had a hard time talking with someone else about the trauma that they experienced

00:48:11.040 --> 00:48:11.560
 around that.

00:48:11.560 --> 00:48:21.480
 I think there is possibly an opportunity to create an AI to talk to about trauma

00:48:21.480 --> 00:48:27.480
 related to queer experience, hopefully infrequent trauma experiences.

00:48:27.480 --> 00:48:32.800
 I was wondering if you have information about,

00:48:32.800 --> 00:48:37.920
 if you see a budding community around building and

00:48:37.920 --> 00:48:41.680
 creating these things within our queer community so far.

00:48:41.680 --> 00:48:45.840
 >> Yeah, absolutely, I do.

00:48:45.840 --> 00:48:52.080
 And I would love if any of you are inspired from this to join that effort.

00:48:52.080 --> 00:48:55.280
 I think there's many ways to intervene in that space.

00:48:55.280 --> 00:49:00.440
 And some of them are really building positive tools like what you've described.

00:49:00.440 --> 00:49:05.920
 And some of them are building weird, uncomfortable, messed up interventions.

00:49:05.920 --> 00:49:10.720
 And I think all of that belongs in this resistance to

00:49:10.720 --> 00:49:14.360
 the generative machine learning that we're seeing.

00:49:14.360 --> 00:49:28.360
 >> Thank you.

00:49:28.360 --> 00:49:32.240
 Yeah, picking up on that very interesting remark from our comrade over here.

00:49:32.240 --> 00:49:36.760
 So queer is a process against the status quo, it's an eternal process, right?

00:49:36.760 --> 00:49:41.080
 So there's no such thing as a stable identity because that would be contradictory.

00:49:41.080 --> 00:49:45.480
 And we saw a very interesting example in the scratch board of corporate speech.

00:49:45.480 --> 00:49:48.440
 Right, it's corporate speech affirming speech, so to speak, right?

00:49:48.440 --> 00:49:52.200
 And you rightly pointed out that that was cringe.

00:49:52.200 --> 00:49:56.320
 And so I was wondering, could you give a concrete example of what you would like

00:49:56.320 --> 00:49:57.760
 to see in the next iteration?

00:49:57.760 --> 00:50:03.240
 >> In the commercially available next iteration?

00:50:03.240 --> 00:50:07.240
 >> For example, what we saw just in a better form,

00:50:07.240 --> 00:50:12.720
 the output you would have said yes to in that example.

00:50:12.720 --> 00:50:16.680
 >> Mm-hm, okay, well, two parts.

00:50:16.680 --> 00:50:20.680
 I think through this experience, I'm not sure that I wanna see

00:50:20.680 --> 00:50:24.640
 the version that I want in the commercial iteration.

00:50:24.640 --> 00:50:28.360
 I feel like maybe it doesn't belong to everybody,

00:50:28.360 --> 00:50:32.520
 it belongs to people who build it for ourselves.

00:50:32.520 --> 00:50:39.520
 I can't imagine what open AI might do with a queer intersectional language model.

00:50:39.520 --> 00:50:42.160
 But from what we've seen them do so far,

00:50:42.160 --> 00:50:46.600
 I feel like it wouldn't be aligned with my values.

00:50:46.600 --> 00:50:50.280
 And what kind of response might I want to see?

00:50:52.960 --> 00:51:00.440
 >> Yeah, I think that it would be something that I could recognize myself in or

00:51:00.440 --> 00:51:04.400
 that I could recognize people I know in, that it would be affirming.

00:51:04.400 --> 00:51:09.640
 Maybe it would be complex, it wouldn't necessarily be perfect or optimistic.

00:51:09.640 --> 00:51:11.240
 But it would be dynamic, right?

00:51:11.240 --> 00:51:19.000
 It would acknowledge that there's more to queer lived experience than

00:51:20.040 --> 00:51:26.240
 what it knows now, which is being queer is illegal in this country.

00:51:26.240 --> 00:51:31.080
 So if you're on a plane, the flight attendant has to tell the people on

00:51:31.080 --> 00:51:34.560
 the plane that they're going into a dangerous situation.

00:51:34.560 --> 00:51:37.880
 There's this latent sense of fear and

00:51:37.880 --> 00:51:41.560
 danger built into this that is really shocking to me.

00:51:41.560 --> 00:51:46.120
 I would love for it to be like the show Schitt's Creek,

00:51:46.120 --> 00:51:49.560
 where there's queer people, but they're just living their lives and

00:51:49.560 --> 00:51:52.480
 it's not the main thread of their storyline, right?

00:51:52.480 --> 00:52:00.100
 Yeah?

00:52:00.100 --> 00:52:09.720
 >> I wonder how much of this is the data set and how much is the actual AI

00:52:09.720 --> 00:52:15.040
 kind of algorithms not really understanding the context of what it's looking at.

00:52:15.040 --> 00:52:20.320
 Because I often feel you can say someone's sexual orientation and

00:52:20.320 --> 00:52:24.560
 it has no real relevance other than defining the gender of the people involved.

00:52:24.560 --> 00:52:27.440
 And then other times it's actually incredibly relevant and

00:52:27.440 --> 00:52:29.480
 the context really matters.

00:52:29.480 --> 00:52:34.160
 And then the other thing I was wondering about was

00:52:34.160 --> 00:52:40.640
 sort of how you weight the relevance of data.

00:52:40.640 --> 00:52:46.080
 So for example, a pivotal book, would you weight that more?

00:52:46.080 --> 00:52:51.400
 Would you, as values change, how does the data set evolve over a number of years as

00:52:51.400 --> 00:52:55.600
 society values change and people's views change?

00:52:55.600 --> 00:52:57.120
 >> Yeah, good question, thank you.

00:52:57.120 --> 00:53:00.520
 So it's both and the data set and

00:53:00.520 --> 00:53:07.760
 the model are deeply entangled because the way the algorithm is built

00:53:07.760 --> 00:53:13.600
 is really pulling from the frequency of each word appearing in the data set.

00:53:13.600 --> 00:53:17.880
 So if Shakespeare appears in there more than anybody else,

00:53:17.880 --> 00:53:21.520
 it's gonna quote unquote think like Shakespeare, right?

00:53:21.520 --> 00:53:28.320
 So as the question before about how a queer model might be architected to do

00:53:28.320 --> 00:53:32.320
 something differently, that might be another opportunity to intervene and

00:53:32.320 --> 00:53:33.760
 build it a different way.

00:53:33.760 --> 00:53:38.000
 But right now, the two are very much tied together, so

00:53:38.000 --> 00:53:42.480
 it absolutely is the data set and the model.

00:53:42.480 --> 00:53:48.720
 But they are connected too deeply to kind of detach.

00:53:48.720 --> 00:53:53.720
 Thank you, yeah.

00:53:53.720 --> 00:53:55.080
 >> Thanks.

00:53:55.080 --> 00:54:00.800
 We have five minutes left according to the time schedule, but I feel like-

00:54:00.800 --> 00:54:02.080
 >> Yeah, I'm happy- >> It's getting long.

00:54:02.080 --> 00:54:06.920
 >> Yeah, now I'm happy to talk with anybody else after or play with the tools.

00:54:06.920 --> 00:54:09.280
 And thank you so much for all your questions.

00:54:09.280 --> 00:54:19.280
 >> [APPLAUSE]

00:54:20.160 --> 00:54:25.240
 [MUSIC]