Jen Looper: [00:00:00] Hey everyone, I hope everyone’s head is ready to explode with the, you know, the amazing things we’re going to be learning together today. Uh, but all, all kidding aside, I am going to look that up because sometimes I have very strange dreams and you know, maybe, maybe that’s part of that syndrome, but yes, I’m very happy to be here.

My name’s Jen Looper, and, um, I am. I am the former academic team lead at Microsoft. Um, I have 25 years of experience in the industry as a software engineer and as a developer advocate. Wound up at Microsoft, had a great time, spent some time at AWS. Also had a great time as head of academic advocacy. I am now funemployed, so if anyone has an awesome job for me, please let me know.

I would love to hear about it. Um, I am very interested. [00:01:00] together the disparate domains of humanistic studies and technology. My PhD is in, uh, literature, but it’s actually in medieval French literature. So my particular specialty is 13th century prose romance. Uh, one of those growth industries. To be honest, the text that I studied, uh, from the 13th century was an absolute bestseller banger back in the 13th century.

People went gaga for it. They had the, the manuscripts were immense and it was kind of like, um, a soap opera back in the day. Anyway, that is not what we’re talking about today. Today I would like to start with a clickbaity title, which is The Last saree, and I’m hoping that folks kind of have an idea of what a saree is.

Uh, I have a great love for India. to India a couple of times last year, managing huge student communities over there. And we really had a lovely time. And I learned a lot. And one of the things I really love about India are their textiles. All roads lead to India when you start [00:02:00] talking about textiles, and I’m a textile junkie.

I have an antique lace collection. I’ve done machine learning studies of that lace collection. Um, I worked at the Museum of Fine Arts in Boston, uh, learning about. curation of textiles. This is something that I’m really passionate about. And in honor of this, which is probably going to be the last time I give this talk, which was originally designed as a TED talk to be given in India.

Um, I w I was in full saree at that moment, but I didn’t want to do that today. Um, our climate does not quite align with, um, with Indian sarees here in Boston. But I wanted to wear this dupatta, which is the shawl that are created to go with your, um, with your outfits in India. I might take it off because it’s quite warm, but this, um, this has a story behind it, and we’ll go into that in a minute.

But um, I was drawn to this topic by a clickbait, another clickbait title from this online website called Shilpa, uh, which is a fashion website. It is asking whether the eerie era of [00:03:00] saree draping was silently coming to an end. And this is a question that people are asking. It’s kind of difficult to wear these garments.

It’s about six yards of material. Drape, drape, drape, pin, pin, pleat, pleat, toss it over your shoulder and hope you don’t have a wardrobe malfunction. Uh, it takes often two or three people to get you properly into a saree. It’s not an easy thing, but it is beautiful when done properly. So I started to become curious about this idea of an erosion of the cultural heritage that we are inheriting.

I’m not, of course. Indian, but I have a great love for all types of different cultural heritages. So I thought, okay, let’s do a little talk about this. Um, today’s agenda, I want to have a little thought experiment on how to move. From cultural appropriation, which is what we don’t want, to appreciation, and then from comprehension to conservation, using different types of technologies.

I have a couple of demos that I’m going to do, using some, um, some [00:04:00] Various various types of technologies that we can use for what we would call the glam sector. That’s G. L. A. M. galleries, libraries, archives and museums. So whenever I talk about the glam sector, that’s what I’m talking about. They are having a moment where they’re doing a lot of digital transformation right now, and a lot of that has to do with digitizing their collections and their figuring out how to group them together.

This terracotta image on my left is, um, the first image of a saree like, um, draped garment. It’s from 200 or 100, um, years before the common era, so BC. Uh, it’s from Bengal and it shows this kind of draping that was already happening way, way, way back in the day. So we’re talking about a tradition that goes way, way back.

So, What’s at stake? Who cares about sarees, right? I care about sarees. A lot of people care about these things. And a lot of people are trying to evolve them and trying to modernize [00:05:00] them and doing interesting things with them. You’ll have to see what you think about this. But the lady on my left, in the spectacular blue silk saree, is kind of the classic image of a lady in a saree.

She’s beautifully posed. Her blue color reflect, is reflected in the doorway behind her. The pinks kind of are reflected. This is probably done in Jaipur, the pink city. Um, this, the stripes and the top all align. She has her jewelry on fleek. She looks amazing. This is a classic, beautiful picture. And this is how one should look in a saree.

And then we have Zendaya, to my right, who caused a bit of a scandal in this outfit because some people really liked it. It’s a designer outfit. It’s attempting to do something different with saree. It’s much more revealing. The saree in its traditional aspect is actually pretty conservative. It wraps up a lot and is not designed so much to reveal as to conceal.

But here we have Sandhya, who is in [00:06:00] a very revealing outfit, uh, with some kind of top and some kind of, uh, sewn drape. It’s, it’s In my opinion, she got roasted with a little bit of justice, I think, for cultural appropriation because there’s this idea that you’re wearing something that you don’t fully understand.

Well, how can we use technology, perhaps, to help us fully understand what we’re looking at? Some people are doing really cool things. And this is a lady, Natasha Thassan. I believe she’s from Sri Lanka. She can show you how to drape a saree in one minute flat on TikTok. Wrap, wrap, wrap. Pleat, pleat, pleat.

Don’t even bother with the pins. Just tuck it in and go for it. And she does an amazing job. She’s even showing how to wear it without the top. Um, if you wrap it a certain way, you can get away with that too. Um, and I think it’s fascinating that for this one minute drape. video done on TikTok, which is a very powerful algorithm, probably the most powerful social algorithm in the world.

She got 7. 4 million views on this last time I took the screenshot. That’s amazing, right? And that shows that [00:07:00] there’s a desire to think about how we can, you know, not be appropriating things culturally and just slapping things on and appropriating beautiful images, but actually learning how to do this properly through today’s AI driven culture.

So let’s see how we can look at textiles. Which I’m a junkie of. I love textiles. Um, as a catalyst. And you can, uh, you can take what I’m going to be talking about and apply it to anything you like. Right? Anything that you think is collectible. If you like, um, different types of My, my grandfather collected antique models of tractors because he was a farmer.

And wouldn’t it be cool to take his whole huge collection and run it through those systems that I’m going to show you and show that, you know, their provenance and where they came from. So, this is all applicable to anything you care about. For me, I care about textiles, so let’s talk about it. So, if you look down, I’m looking down at my legs right now, and I can see I’m wearing jeans, uh, in this era of, of Post COVID presentations, we have to sometimes do pants [00:08:00] checks.

Well, check your pants because maybe, maybe you’re wearing jeans. Jeans are fascinating, right? They are the quintessential American piece of clothing, right? Well, in point of fact, there’s a lot to be said about jeans as an international garment, an international textile. And you can use jeans as a kind of a motif to take you all around the world and to show all of the cultural heritage that went into the creation of this iconic American piece.

In the middle, by the way, we’ve got James Dean, who is too cool for school, but he’s basically copying, uh, these workers, who are, um, loggers, and they’re wearing their jeans rolled up, cuffed, like he’s doing, so all of these guys are wearing these tough working clothes, but he made it too cool. This is the international story of blue jeans, and I’m going to briefly take you through this very interesting history.

It’s interesting to me anyway. A lot of you might recognize some of the logos on this slide. Um, we’ve got Levi Strauss who, who basically patented this idea that jeans are very, very, [00:09:00] very strong. That you can’t rip them apart even with a four mule team or two mule team, right? Um, but they started dying garments blue way, way, way, way, way back.

Indigo is one of, has been used for ages as, as a way to dye garments, and we can trace it to being cultivated in Peru about 6, 000 years ago. Um, there’s a PBS documentary called Riveted that you can get all of this information out of, by the way. It’s very worthwhile. This blue dye came together to be associated to a kind of pants in the 17th century actually and guess what?

The original blue pants are from, guess where? India. They’re called dungarees, dungarees. It’s a type of indigo dyed Indian cotton worn by sailors. Now whenever we find garments worn by sailors it’s gonna traverse the world pretty quickly, right? So Already in the 17th century, we find a fabric called Serge de Nîmes, that’s where the word denim comes from.

Serge de Nîmes from Nîmes, France. It’s a port city [00:10:00] in France, which was copying a fabric from Genoa, and the French called it Bleu de Jeanne. Bleu de Jeanne, that’s where the word jeans comes from, right? So we have denim jeans, and they’re inspired by dungarees. They’re these practical blue pants worn by sailors.

Um, they could also be brown sometimes, um, and there’s some work pants in the 1850s that were these kind of stiff brown canvas pants for minors. Um, but you can also trace the very sad history of a culture when, especially via fabrics, because we have indigo and cotton. As cash crops, um, that were, um, being cultivated in, um, uh, plantations worked by enslaved people and milled in the U.

S. North via, um, folks, if you go to Lowell, Massachusetts, you could find some of the fabric mills that are being, um, still preserved. Uh, and then eventually all of this. Cultural heritage makes its way to the mining camps in California during the gold rush. They needed strong pants. They were running around.

They needed to make sure [00:11:00] they wouldn’t tear when they’re bending over and panning for gold. So, uh, Levi Strauss came up with rivets. And if you look at the rivets on your jeans that are holding the pockets together, even these kind of Not very great jeans that I’m wearing now have rivets. That’s what was patented by Levi Strauss.

So even if you don’t care about fabrics at all, you can see that there’s a cultural heritage that’s being passed around the globe, and fabrics are kind of an incredible catalyst for passing, passing around these traditions. Now wouldn’t it be cool If we could document the transition of these, of these cultural, um, aspects, uh, that people want to preserve.

And it, and this is the, the task of the glam sector, right? This is what they’re trying to do. They’re trying to figure out how they can take, um, just get their collections digitized, first of all. Get good pictures taken of them, um, and then put those images into some kind of storage buckets and get them referenced in a database.

Now preferably in that database you want to have a lot [00:12:00] of um, metadata. Where it came from, how old it is, what color it is originally, any, any damage, uh, who, Who, who donated it to the museum, where it’s from, this kind of stuff. So all of that provenance, uh, that’s all required to be kept in, in a database somewhere.

But there are also, and that’s very basic stuff, right? You digitize your stuff, dump it into a bucket, dump the images into a bucket, get them referenced in a database, and preferably create some kind of API that people can then reference to take a look at your collection. That’s, that’s, that’s really where the public can benefit.

But some museums are going farther, right? They’re using, More complicated technologies such as OCR, such as AR and VR, such as AI and machine learning to create, um, really engaging, uh, study, uh, opportunities for their collections. The Metropolitan Museum in New York City has their Collections API. You can take a look at some of their, some of the, um, items in their, in their collections.

But they are basically doing the, the simplest thing, right? They’re [00:13:00] digitizing their collections and they’re finding ways to connect to the metadata. to each other so that you can cross reference a, um, a Chinese vase, which would be in the porcelain collection, to, you know, a, um, uh, a painting on silk or on paper, which would be, you know, in a different collection.

So you can, you know, find the, the ways that these things intersect and connect. So you need to have a well designed database and a well digitized collection and a nice API that people can use. So, This is what’s called DAMS, digital asset management. There’s a whole industry around it. I find it really interesting because I love to play with these APIs.

You can create awesome demos using the collections APIs. Here’s an example of the use of OCR with a museum collection. So you can, um, this is a example from, um, some kind of plant collection. I think it might be the heart, one of the Harvard museums, but they’re using OCR to just down, um, document the labels.

They’re figuring out, you know, there’s a lot of typewritten data there and they’re using [00:14:00] OCR to get some kind of transcription of the, of the items to tell, you know, where the items are from. So that’s, you know, your basic use of OCR. We’re also using OCR to, to, um, transcribe handwritten, uh, labels as well.

So that’s another really good, um, use of this sort of thing for, um, presidential libraries, actually. But. Moving beyond even OCR, are there really other options that we can do with even more interesting and kind of edgy technology? Well, in point of fact, yes, there are. So I created a little web app that’s based off of a custom trained model.

And I did this using Amazon S3 and Amazon Recognition. You can do this also with, you know, Any kind of Google technology clarify has great image recognition, but I needed something where I could take my images, do a quick custom training to get custom labels, and then, uh, create a machine learning model that I could call using an API.

So what I had to do. Surprisingly simple [00:15:00] and a lot of fun. It was a fun afternoon. I went through all of the different styles of saree that people wear. So in India, every region and even places within the regions have their own styles of silk weaving. They have their own styles of draping as well, and their own motifs and their own colors.

So you can group them by the I grouped them by the type of fabric, such as batik, which is a wax painting from a place like Bengal, uh, versus, uh, something like a Banarasi, which is a style, that’s a style of saree. So I have, you know, the, the location and also the, uh, the style. So I just dumped these into a nice S3 bucket, uh, I gave them appropriate names, and then I went ahead and trained, uh, a machine learning model.

I only had, I think, a hundred and There’s 27 objects. I think there were only 130 or so images that I scraped off the internet, but I just pointed, um, Amazon recognition to go ahead and make a custom model [00:16:00] to be trained. And it gave, you know, it’s not the best, it’s not, it’s not the most accurate model in the world, but for a first pass, I found it relatively adequate to correct what I said.

I had 358 images and I was asking for 39 custom labels to be created and it went ahead and did that. So here is, uh, let me see if I can play this. Okay, so here is a, um, an image that I have, a website that I’ve created. So this is obviously a Gujarat style saree. It’s detecting the location and it’s detecting that it is a bandhani style.

So, um, That’s, those are like little tie dyes, a type of style. Now this is what I’m wearing now. This is a patola style. So it’s gonna say it’s from Gujarat and it’s patola. This is a video of the web app that I created. Um, going further, we can take a look at this. I believe she’s a Banarasi style. Yep, from Ujjain Pradesh.

So that’s the location. Um, I should probably make two columns, you know, location and [00:17:00] style. Um, This one is probably the most straightforward. This golden white, if you see a golden white saree, you know immediately it’s from Kerala and that’s a Kasavu style saree. So that’s a very classic look. And here’s another gorgeous Banarasi style.

And, and in this web app, you can do detected labels. So it’s a, excuse me, Paitani style from Maharashtra. Sorry about that. I’m saying the machine knows more than I do at this point, um, but you can also choose an image from your own device, upload it, let’s do that, and then it’ll add that to the image. Now this is a Gujarat drape, you have this in front, so that’s going to be another Patola style saree from Gujarat.

So with a few images, you know, a couple hundred images and a custom trained model, you can get to the point where you can start detecting some very esoteric stuff, right? You could detect my grandfather’s, uh, Um, different kinds of, uh, toy tractors, or you can go the [00:18:00] route of fabrics and, uh, and detect the styles.

So that’s, that’s, that’s nothing more than image recognition, right? That’s just straight up image recognition with some custom labels. It’s not, it’s basic machine learning. I wouldn’t call it, you know, um, revolutionary from an AI perspective. So what happens when we try to get some insights about our cultural heritage?

objects, culturally important objects, through the use of AI. So AI is funny, right? AI is interesting. At one, in one, in one point of view, it’s a bit disconnected, right? It’s a bit disconnected from its source. We don’t know where these models have been trained a lot of times. Until everything gets open sourced, we’re not going to be given that insight.

We’re not quite sure. So images kind of lose their reference point. You know, we don’t know where that stuff comes from. Um, attribution is erased. A lot of authors like myself have had our stuff, um, scraped up and trained on. Uh, and it’s certainly, uh, unclear where the [00:19:00] attribution was. I think some Platforms are getting better about putting the source, but, you know, sourcing is often very muddled, especially with image recognition.

Results can be superficial. Um, if you look at this older woman’s hands, I’m quite sure she has at least six fingers. It’s always the hands, right? Um, but, uh, you know, it’s, it’s Given that I gave a very basic prompt asking for some, some images from Gujarat, this is what it gave me. It’s trying to give you something like a, um, um, a Paitani type of about the little tie dye style.

Um, it’s doing its best. However, in an age of AI, These results are very democratized, right? Everyone can be a creator. Everyone can get access to ChatGPT or other types of platforms like that. Anyone with internet access can use. I had a student in India ask if they could use these tools in their native Gujarati language, and I was like, you should try it.

Try it and, and, [00:20:00] and you can help the AI get better because you’re participating. You’re, you know, you’re adding your, your linguistic competency into these models eventually. So this is something that’s really important because with more participation, we’ll get better AI results. So here’s some ladies with a classic.

These are actually real people, not AI ladies. And they have their patola sarees. Uh, and I just want to put a little caveat slide in here that, you know, in the case of the artisan textile, we’re going to ask. generative AI to help us out to appreciate cultural heritage. But I want to just make sure that we understand that we’re not going to be able to improve, um, the something that has a thousand years of history, right?

And we ought to understand also that, that these have deep symbolism. Patola, this dupatta cost me 300. Um, a, a regular wedding saree with a patola. Um, a motif like that would cost over a thousand. So people are going and buying these as an heirloom, as something that has deep meaning to them. So, um, it’s [00:21:00] something that we need to understand that, uh, a lot of connoisseurs will understand that there’s a deep symbolism.

Mine has little fish, um, and I think that’s really cool. So it’s kind of, you know, just something because I like, I like these motifs. But they have meaning. Um, So how would you go about using generative AI to give you a bit more appreciation for your cultural heritage and help people be educated a little bit more on the cultural heritage?

Well, you can try tools such as PartyRock. PartyRock. aws. It looks like this. Um, I generated a little app. It’s, what it does is you give it a prompt and it’ll generate an app that will help you. So I said, I want, uh, a tool that will help me generate a meaningful saree design based on my culture. where I, the, the type of saree I want and the region that it’s from and the motifs like fish and I don’t know, flowers and other things.

Um, so I made this app. You can try it yourself. It’s at this QR code. It’s at silkweave, [00:22:00] s12. de. com slash silk dash weave. And you can try this for yourself. So what you just do is you ask it to create. a saree. So I asked it to create a kantha work saree with big floral embroidery or a madras stylized saree with vines.

It created this. It’s not dreadful, right? It’s okay. For a madras saree, I would have expected a checkerboard pattern. They’re famous for kind of a plaid design. But eventually I got it to generate this, right? So I said I want a saree with clouds, because at that point I was doing a lot of cloud computing. I want clouds.

I want them to be stylized and embroidered. I wanted an outline and it never did get that quite right. But um, I wanted to have some kind of design on the palu. The palu is what you toss over your shoulder. With, um, a kind of a different, uh, gold and purple color motif and a nice weaving style. That’s what it generated, and it was inspiring enough that I actually had it created for me.

Uh, this [00:23:00] is the famous cloud saree that was created for me in India, and I wore this on several occasions. Uh, it’s a, it’s a, it’s a keepsake for me, and it’s all hand done with the, the clouds. So this is the world’s first AI created saree. I’ll have a picture of me wearing this at the end. But what if you want to go a bit deeper and ask for some really interesting motifs?

If you’re spending a thousand dollars on a wedding saree, you probably want it to have exactly the design, exactly the motifs, and exactly the meaning that you’re looking for. So I’d say, because I’m a connoisseur, I want Um, it took me a long time to find my husband, so I want a hunting scene in the Palu.

Um, I’m very, I’m very, um, curious about, um, having good luck embroidered and woven into what’s wrapping me up, so I’m going to have mangoes. So you can say, you know, there’s a lot of symbolism that we can dip into. How could we ask a generative AI system to create a meaningful Uh, saree. Well, in point [00:24:00] of fact, it doesn’t work very well, actually.

So I asked for a Banarasi saree design symbolizing good luck, fertility, and prosperity, and it created this heinous mess with a weird elephants, AI elephants, uh, or it created this kind of generic looking pink. It’s pretty, but it doesn’t mean anything, right? There’s a, it’s, it’s trying, it’s just not very successful.

And that was with a couple of different, uh, generators that I tried. So, there are ways to fix that to make your results better, and the way we do it in Amazon is to use knowledge bases in Amazon Bedrock. So what you’re going to do is you’re going to connect an LLM that you like to use to a vetted source.

Uh, and all you have to do is basically point it to a vetted source. I pointed it back to, um, Shilpa, which is um, one of the, one of the websites that, that I trusted to give me good, uh, meaningful results. And then once I had that, that source connected, the LLM was much more, um, it gave me much better results.

So [00:25:00] I asked it again what motifs would be included in a saree design that would symbol, simply good luck and wealth. And it gave me You know, you can add paisley, you can add a butto, which is kind of like a flower, um, arrangement. Uh, and you can use, um, crowns, you know, you can, you can use all sorts of things to symbolize the harvest.

So, the experiment worked really nicely because it allowed Sonnet to get me better results because I connected it to this vetted source. It gave me quite a lot of detail here. And, uh, it, it gave me References, which I really appreciate it. So you can click on, you know, one, two, or three in here, and you can get this source that it’s getting these ideas from, which is great.

I’m a, I’m a teacher. I want to see, I want to see your work. I want to see the sources. So here it’s suggesting that I, um, use a mango motif, a floral jowl, some paisley, um, bird motifs, which was new to me. I hadn’t thought about that. That’s cool. So once you have that, you want, you have a better result, a better LLM, um, result and a better [00:26:00] prompt.

Then you can feed this prompt into your image generator and get a better result. So here is something that is, um, a little bit more appropriate for maybe, uh, Asaree symbolizing something that you want it to symbolize. So you’re ready for your beautiful party, uh, so, with something that, that you can have as a conversation piece, as something that’s truly meaningful and draws on a thousand years of cultural heritage.

So I kind of like that. I think that’s really cool. So in conclusion, I would just like everyone to sort of think about how we can zoom in. This is a moment in, um, in culture in general, in the West, at least that we’re, we’re, we’re so bombarded by new stuff, new tech, new, everything, everything’s new and fancy.

Sometimes it’s good to step back, look at what’s coming before us, look at how we stand on the shoulders of giants, and also go deep, really go deep, dive deep, and see how we can generate not just stuff, but meaningful stuff, right, so that we [00:27:00] can really be better informed about the heritage that we’re inheriting, and so that we can conserve these traditions.

So here, Midjourney had a pretty decent result here because, um, They were really asking for specific types of saree, even one lady wearing a cotton saree, and it actually does look like a cotton saree. Um, so, the better we Train up these models and get our assets digitized and and create comprehension around what we’re looking at.

The better the AI is going to be and the better it’s going to serve us because let’s not forget that this is just a tool that we should be able to use to make our lives better. Right? So this is the money shot. This is me in the AI saree. Um, the hotel, the lady in the hotel was like, don’t you have any red lipstick?

I was like underdressed, but I did my best and it was a blast. I had an amazing time. If you want to stay in touch and talk to me about your cultural, uh, [00:28:00] experiments that you’re doing and how you’re helping to keep your heritage alive, I would love to hear about it. This will give you access to a website where you can contact me.

And I really appreciate all of your time. Thank you.

Sean C Davis: Thanks, Jen. That was amazing. And I just love the perspective that you brought that, you know, there’s a lot of negative sentiment around AI, and I thought it was great that you’re, you’re turning it around a bit and, and thinking about how we can use it for good for, for culture and preservation. And I have, I have a handful of questions that I want to ask you, but I’m going to keep an eye on the chat and the Q and a section.

So for you folks out in the audience, the questions coming and I will feed them to Jen. But let’s, uh, yeah, let’s, let’s dig in. So one thing I’m wondering is, you know, you talked a lot here about using AI and using you just in general, using technology to help us preserve culture over time. And I’m wondering, um, kind of like the opposite perspective of that.

Is there a risk that if we [00:29:00] don’t act in the right way today, while we’re making a lot of advancements that we’ll lose some of this culture, some of this history eventually.

Jen Looper: So the real flip side that’s happening now is actually a little bit connected to climate change. Because some of the materials that are used to create artisan textiles, um, some of them are, um, under threat.

The types of silkworms that are growing that you have to, you know, feed them certain specific things with certain climate. So, um, that is the flip side and, um, this is another reason why I like to pay a little bit close attention to the amount of energy and the amount of, um, Um, Resources that we use up when working with AI, you know, because this is also, this is also something that is, uh, that is threatening us a little bit, to be perfectly honest.

Sean C Davis: Yes, absolutely. And we hear of that improving, but it’s, it’s like, it’s got, it’s got to move faster. And that’s one type of cost. And we’ve got a, a [00:30:00] question from Ian in the audience who says, can you give us an estimate of cost to train, uh, the images that you used?

Jen Looper: Yes, yes, um, I was surprised, actually. Uh, I let this model run for about a half an hour to train, and it cost me a solid 50 bucks.

Um, it ain’t cheap. Uh, I wasn’t happy about that, and now I’m not at Amazon, and I can say that. And, uh, So, uh, I would say watch out, watch out with, uh, with training on custom models. I decided to take this app and flip it to use, um, some Google technologies. I want to try that, but I’ve always had good results also with, um, clarify.

That’s why I raised it. I think we should, um, I think we should look at cost effectiveness. There’s another thing we should look at. So Sasha Luciani. who is in, um, Hugging Face has un, uh, revealed this energy score that she’s allowing to [00:31:00] be, um, applied to certain, um, machine learning models and AI models. I want to keep a good eye on the energy score because that’s also a cost.

So there’s all kinds of hidden costs, but yes, I, uh, I took a look at that bill and I shut it off really quickly.

Sean C Davis: Well, that’s really interesting. Does, so does she need, she needs access to Information that we don’t have, like, I’m guessing that the, the company is providing these agents are they’re hanging on.

There’s no other way for us to know that. Right?

Jen Looper: Yeah. And she was, she had a little spat recently with, uh, with folks like at open AI, because they’re, they’re not open about the, the energy that’s being used, the water that’s being used, you know,

Sean C Davis: sometimes, you

Jen Looper: know,

Sean C Davis: um, and so back to the, um, the, the dollar cost is it.

Okay. With, um, with any of the ones that you’ve tinkered with so far, can you put. You put like a cap on it, or you just have to kind of write that in programmatically.

Jen Looper: For the training, I [00:32:00] think you’re going to have to expect a certain amount of cost. I haven’t done a comparison yet. I want to try, um, I want to try Gemini.

There used to be this amazing tool at Microsoft called Lobe AI, and everything was free. It was beautiful. And I think they killed off that product probably because it was too expensive. But, um, Maybe there are other ways to do this. Maybe we can do it locally. It would take a while and it would take a lot of, um, you know, local resources, but maybe that’s another option.

So you’re not training in the cloud, you’re training locally. I used to try to do this with things like TensorFlow, um, uh, On my local. I just have a little MacBook Air here. So it’s it’s crying out in pain. But, um, maybe there are other options that we could try. I would love to have discussions about this because working in the glam sector cost is a very big deal.

And the last thing you want to do is expect a museum who should be putting their resources towards conservation and towards, you know, preservation. Blowing it on cloud computing costs. I don’t think that’s a great idea. Honestly, there’s other ways we can do this.

Sean C Davis: Yeah. And, [00:33:00] um, something I saw in the chat earlier from Brian was talking about smaller language models that if you’re, if you’re trying to achieve.

One specific thing or, you know, set, set of things and your conditions can be more limited that, um, you might, you might not need to, you might not need to introduce all of that bloat and all of the information in the world to the model. And I’m wondering if you’ve, um, if you’ve tried both approaches and have found one more or less effective than the other, you know, models trained on like some contextualized information versus ones that are just.

open and trying to get as much information as they can.

Jen Looper: I have not, but I’m sure that that would give us a lot of good results in terms of speed and cost. It’s something I think we ought to be exploring and I think we ought to be benchmarking. And maybe that’s the next step for this, you know, for this, this type of activity, see the different providers, see the [00:34:00] different ways of doing this and do the same thing, you know, on each with the same data on each and benchmark it.

I think that’s a great idea. I’d love to do that.

Sean C Davis: Oh, yeah, that’s super interesting. Yes, absolutely. Um, and. So when the in the demo that you showed when you you would click the button with the photos of the sorries, and then it would always it seemed like it was always listing a certain number of answers, I guess, and and then attributing that confidence score to them.

And, um, if you I think what I was trying to get at or working through in my head while I was watching that is. Is there then like a, a model that you can put on top of that or some sort of other logic that you can use to say, oh, this is correct or it’s not correct? Like, can you make a, can you make it into a yes or no?

Or is it always going to be some sort of, um, sliding scale of confidence?

Jen Looper: I think that it would be, at that [00:35:00] moment, that’s where the connoisseurship comes in, and maybe that’s where the human acknowledgement comes in. What I would consider doing is what, um, there’s a project called In Codice Ratio, and they were taking medieval manuscripts from the Vatican and crowdsourcing students who were clicking on each character, each letter, and saying, this is an A, this is a B, and then other folks would verify, verify, verify.

So it’s the old school. It’s not a very crowdsourced way of doing ML, but it would be a really nice thing to create maybe one textile model based on the collection at the MFA or something like that. And then people could take that and train on top of that. So we really need some base, we need some base culture models, we need a base saree model, you know, and then you can verify.

Sean C Davis: You need the human intervention to help make sure that these things are actually giving us correct information, right?

Jen Looper: Yes, the one thing you’re going to be [00:36:00] sure of is the carol the one because it’s always white and gold. That’s the one I always get right.

Sean C Davis: Yes, exactly. Exactly. Okay, question coming in from Brian.

Brian asks, Was there a specific reason you chose Claude? If that’s he’s, if he’s remembering correctly that it was Claude. He says, I know bedrock can swap models and supports a bunch of supports a lot of different things.

Jen Looper: Yeah, I think if I remember properly, I tried that prompt. With, um, with a couple of different ones.

And Claude gave me the best results, the most verbose and the best results. Um, I think that one just performed the best for me as I remember. But again, we could be worthwhile taking a little benchmark moment to test that again.

Sean C Davis: Yes. And all right. One last question for me is, so, you know, you talked about you, you have a lot of information and education, educational material out there, and it’s hard to know what’s getting scraped and how it’s getting repurposed [00:37:00] and, and used.

With or without your permission, um, and that we have a huge problem right now in terms of accurate citation of the sources being used when these agents give us some answer. I’m wondering if you can talk a little bit about, um, specific considerations that you think that these startups and agents, uh, need to do or take into account, maybe to provide more accurate, um, sourcing, or maybe to build confidence in the human on the other end.

Jen Looper: Yeah, um, I recently did a little test because I was curious who had scraped my textbook. And I tested ChatGPT to say what’s the content of the last paragraph of the first chapter of Computer Science for Kids, and indeed, it had safety mechanisms built in, and I appreciate that. It said, since it’s copyright material, and I was surprised because I know Wiley, which is who I publish with, I’m pretty sure they gave They gave all their stuff away.

But, um, they said, we don’t know, it’s copyright. [00:38:00] So go pound sand and go to the library. Um, on the other hand, I also recently went to, um, perplexities. So there’s a couple research, uh, um, platforms being produced right now. There’s open AI research, I think. I think there’s, they have their own research one, 200 bucks a month, which I’m not paying.

And then there’s the free one from Perplexity. Tested the one with Perplexity, very interesting because the citations are all over the place. And, um, it was giving some reasonable citations. Um, Where I get stuck is that the scholarly journals are still quite locked down, and they’re not being trained on.

Elsevier, Springer, that stuff’s not being trained up on. Um, so the citations you’re getting is only stuff that’s open on the internet now, so still, scholarship is locked down. And I don’t know if that’s a good thing or a bad thing. Um, I think, I think scholarship should be open. as long as it gets off copyright, perhaps, or after a certain amount of time or something like [00:39:00] that.

It’s a really difficult question. I’m trying to figure out, you know, where’s the, where’s the, where’s the barrier, you know? I think as long as your stuff is, is not yet in the public domain, maybe we shouldn’t be scriping on it. But in the meantime, at law, as long as we can have some kind of citations that are built in or some guardrails, when we’ve got, you know, J.

D. Vance saying that we should just innovate and forget safety, I’m 100 percent not about all that, you know, we need to make sure that we have safety built in so that we can protect our authors and we can keep writing. Um, so I’m, I’m threading a line here trying to figure out, you know, maybe we should wait until things are in public domain to scrape.

Maybe we should not be scraping copyright material, but at the very least, when you’re making up stuff for God’s sake, cite it. That’s what I would say.

Sean C Davis: Yes, maybe there’s something here where you can play into, you know, what’s, what’s the future of the library look like? Because even because you can get copyrighted material at the library for no cost, you just have to do the work to actually get that.

So that would be. [00:40:00] Yeah, how that plays into it.

Jen Looper: Yeah, I think, um, where was I recently? Internet Archive has a little interesting little library that you can go and log in and you can check out books, um, interestingly. So there’s that. Um, and there’s a lot of stuff recently that did come into public domain.

I created a machine learning model off of all of Emily Dickinson’s poetry. I had to take it offline. I didn’t know that was still not in public domain. So, um, Uh, but I think just keep an eye on all of the, um, bits and pieces that are coming into public domain and you can get, you can get your hands around them.

Uh, just be aware of, of these issues and copyright sensibilities, I think.

Sean C Davis: And I’m going to, I’m going to, uh, surface something that Linda commented, just to see if you have a comment. She says that, um, we need, we need good bases, always strong foundations. And could they do their own locked source version?

So the info is accessible, but the copyright is protected.

Jen Looper: Like [00:41:00] a copyright AI model, that sort of thing. Um,

Sean C Davis: it’s, it sounds similar. Yeah. It sounds, sounds like something like that. Yeah.

Jen Looper: Perhaps, um, I don’t know. I guess it would then depend on. Access, right? It’s again, the library motif. Maybe you can check it out for a certain amount of time, but you can’t do certain things with them.

You can’t photocopy things out of the library and redistribute them for sale. So maybe there’s a licensing thing and this is a legal, a legal situation. Maybe there’s a licensing, um, subtlety that we need to dive into here. Speaking as someone who’s not a lawyer,

Sean C Davis: yes, fair. Well, thank you so much. Jen. I really appreciate your perspective.

And it was a fantastic presentation.

Jen Looper: Thank you. Appreciate it. Cheers.

Bye. Bye.