The Power and Responsibility of Generative AI with Daniela Braga | S3 E27 (at Web Summit)

Daniela: 0:00

Instead of starting from the coolest, shiniest tool, let's look at my probably most boring problem, and not the coolest, shiniest tool, because the most boring problem is probably causing 30% of inefficiency and span.

Joeri: 0:15

Welcome everyone on the Web3 CMO Stories podcast. My name is Joeri Billast and I'm your podcast host. Today is the third interview at the Web Summit and my guest is Daniela Braga. Daniela is the founder and CEO at Defined. AI and, as a founder and CEO, she is a world economic forum AI technology pioneer and recipient of many entrepreneurship awards. Daniela was a member of the 12th- person task force that advised the US president on AI strategy. Hi, Daniela, so great to meet you. I'm delighted, Joeri, nice to meet you. So, guys, I have a really interesting question. It's an open question about Genitive AI, because so much things are going on. What is for you? How do you look at it? What is the status in the various sectors of Genitive AI?

Daniela: 1:11

First of all, genitive AI is not new. AI has been around for 50 years and Generative AI is a way to create existing or unseen content, from seen content as the whole principle of text- to- speech recognition, speech recognition, but the text- to- speech, which is the content generation through voice, the concept of image generation or image identification. All of these things have been around for 20- plus years, except that now, with ChatGPT in everyone's pocket, it looks like it's new and it's not. It's been around. So that's the first myth AI has been around. Genitive AI has been around. That said, we are seeing now interfaces between natural language and voice, to code, to music, to image, to video, to any content you want to create through these interfaces of voice or text and this is being to a point now that is very exciting. It's unseen so far to music, to sounds, to everything.

Joeri: 2:24

Exactly, I'm also using it in my daily life. But AI has been around so on, but now it is accessible for the audience now with ChatGPT and so on. But of course with new technologies., here are always some ethical considerations you can look at. I know for you. I think it's really important and it's important. Can you talk about for your business and maybe in general, the ethical aspects of AI?

Daniela: 2:52

All AI starts in data, and data needs to be consented, paid for, representative, and biased and transparent. That's what our company is all about. We are the largest marketplace of training data for AI, ethically sourced. Our data comes from partner data and real world data that is all anonymized, consentive, live for and shared with the partners the revenue part and legally made available, commercially available for our customers. We have simulated data that goes through our own platform and that also deals with people.

Daniela: 3:27

We collect data from all over the world people's voices, images, biometric information but under consent and anonymization, it's untraceable to the contributor and everybody understands what they're signing for.

Daniela: 3:43

In the case that a voice, for example, or the likeness of a person is to be used for branding and we have to buy the commercial license of perpetual usage of your voice, for example, but not exclusive, necessarily that has a different price. Everything is clear and transparent. And then, of course, there's another type of data, which is synthetic data, which is really generative, ai based data that all of the three sources of data allow customers to create unbiased models and more diverse models, which, because you never have, for example, african American dialect in enough in real world data that you can create models to address those types of dialects or people who come from like minority groups in America, and don't understand fortune is based speakers or they'll. So those are. That's all the three combinations of sources and our principles and data with values that define what we do. But, of course, ethical AI goes beyond data, goes by, goes around the application, the design, the testing process and the accuracy and reliability of the model.

Joeri: 5:03

So thank you for sharing that, Daniela arm. Of course we are on Web CMO Stories Podcast podcast. So I have also a question about marketing and generative AI. There are a lot of strategies possible. Can you talk a bit about that?

Daniela: 5:21

I recommend to keep an eye on us for podcasts. That keeps us updated in five minutes daily, which is the Podcast from Up to You. In five minutes he gives you the G7 of the tools and of the day Currently for marketing you can use. You can create with Adobe Firefly a whole plot point from text to image, from text to and changing the background, pixelation or images inside the actual image. You can add voice in different languages. Even you can actually have a narrator Think it's the tallest victory and you create the whole video without hiring a narrator in any way, not in any language, but in a good set of languages already and a sequencing as like a story design, a story board, and it designs it for you.

Daniela: 6:24

You can create with the 11 labs is a very good, interesting way to create text to speech, except that it doesn't ask the person for consent. That is a big red flag. You cannot create a deep fake. That that's why the executive board in the United States will enforce a watermark on AI generated content, precisely to not deceive people. So it's a problem of deception. It's all good if the content is created artificially. You just cannot accept, and so on.

Daniela: 6:59

The deep fake part of the 11 labs is very interesting technology, because I'm from the time when you had to collect 50 hours of a person in studio to create some generative content in text to speech. Today, with these foundational models that contain so many languages and voices, you create them. You just need that voice, almost voice adaptation of a minute or two to generate a decent text to speech model. Except that Microsoft, for example, if you are building a voice on Microsoft technology, you have to record a consent form record really On seven, the seven labs, on 11 labs. You cannot do that and you cannot, you don't have the option and that is not gonna be. That is part of the principles of ethically. Especially when a tool becomes viral, there is responsibility associated. If it's just a test, it's all good, but there is a heaviness, a responsibility when a tool becomes viral or it's not so much, in my opinion, around revenue generated. It's about user space.

Joeri: 8:09

Yeah.

Daniela: 8:10

It can change very quickly in this world.

Joeri: 8:13

Absolutely, it can really go quickly and there are a lot of tools out there. And then there is ChatGPT and I see that the ChatGPT is also evolving all the time. And I see that going back to ChatGPT, because With trillions in parameters right now. The tool also need to have an added value over using ChatGPT. There are so many tools.

Daniela: 8:39

Gpt is one. Actually, falcon seems to be ahead of all of them in the leaderboard, and we have Lama.

Joeri: 8:48

And.

Daniela: 8:48

Lama is bringing Lama and Matty is bringing Lama 3 very soon, so Lama being a more efficient model from a parameter building standpoint.

Joeri: 9:00

So, yes, it's not easy to if you are not inside the space you pitch. Tools should I use, because everyone is talking about a model tool and with other prongs and stuff. But yeah, how do you keep yourself clarity on everything that happens? Do you read a lot? Do you listen to podcasts?

Daniela: 9:20

We, first of all, we are enabling the builders of these technologies, so we, by default, have to be up to date to know what to recommend them with the best data to improve their models. So that's one.

Joeri: 9:35

Yeah.

Daniela: 9:36

We? I don't. In principle, we like to try everything, but we have an ethical responsibility to distinguish, for example, the models who script web versus the model that acquired it legitimately and we know who they are. So until we know that, in the case of OpenAI, they have been building a model very fast, they probably never expected to see this usage base. Now they seem to be moving towards a reevaluation of the data they had there, but that wasn't the case before. Yeah, we never worked with OpenAI and we did try. And there is now it seems to be a little bit more opening to look at what data do they have inside. Yeah, we, what we keep up. And then there's you gotta be on top of what's it Journal Tech News one of my favorites, and the app too. Guys, if you find me in its briefings, I like it too. It's really to the point and I guess your podcast too.

Joeri: 10:49

I'm trying to do that. Yeah, I always say to people listening to my podcast when I meet someone, it's like in a restaurant you get a menu, so for you, listen to these and these podcast. I say, yeah, oh, you want to learn about AI? Listen to the podcast episode with Daniela. You want to learn about NFTs? You want to learn about Metaverse. So it is a little bit relevant. I'm learning by listening to podcasts and also by interviewing people like you, Daniela. And Daniela, as everything continues to evolve with AI, what advice would you give to businesses that

Daniela: 11:24

There are two types of businesses. Ai is here to stay, first of all, and there are the big enterprises and there are the smaller enterprises. The big enterprises need to make a decision whether they're going to buy or build, because if you're a big enterprise and you're going to rely a lot on these models, they're going to be expensive, they're going to get expensive. It's expensive to build as well, but it depends on your market. It depends on your customer base. If you go, the world will evolve towards aggregation of offering. Not today. We have probably 300 large language models, but that's not going to be the day tomorrow. No, the evolution will be into almost cannibalization of ones for others, or M&A transactions, mergers and acquisitions and others will just die Towards a world like that. It will be a concentration of options.

Daniela: 12:31

The first question is are you going to build or are you going to buy? If you're a smaller company, you have to pick your point in productivity. What is my main concern? What do I need to? What do I spend more money in the people's side of the house and where can I make the best impact of productivity and then reverse engineer to Instead of starting from the coolest, shiniest tool? Let's look at my probably most boring problem, and not the coolest, shiniest tool, because the most boring problem is probably causing 30% of inefficiency and spend. The best problems to solve with M&A are the boring problems, are not the cool generative problems, all the ones that are more sparkly now, but are not necessarily that what business people should go for first.

Joeri: 13:22

Actually, that's also. I'm doing some AI workshops also for clients to help them, and I always ask them let me know what is the process, or things that are boring, that take time, that we have a look at that. How can we solve this with AI Instead of the other way around? Here is this fancy new tool that you should use Exactly Now when we are recording this podcast here at the web submit chat. Gpt came out not so long ago. With their individual how do you call me? People can make their own GPTs and even if you call it, call it yeah, what are you thinking about? Because this opens a world for more and more possibilities. People can create now GPTs without knowing any coding. Do you see? Are you excited about that?

Daniela: 14:13

It's the dream of the centralized web, but for the GPTs I do see I still don't see that happening, honestly, because even I see wrappers around open source this is a code, you will have an open AI API, whether it's Whisper or Dali or whatever it is, and you will be able to call that API and build it in your application.

Daniela: 14:42

That is not necessarily because for you to like a Lama model and even a GPT, if you are sophisticated enough, you can start building your own expansion of your model and Lama allows, for example, to build on-prem on your own cloud, on your private cloud, which GPT doesn't, because open AI doesn't allow that. Some of these new ones, people don't understand that every time they are interacting with the model, even if they think it's their own, they are actually improving open AI's model and in Lama, if you actually if you can deploy it on your own instance and build from and actually augment it in your private instance without going back to meta. Those are the small details that people need to take account for, and there is no such thing as you build your own GPT. No, you build, but not everybody is sophisticated enough and it's not your GPT, it's on Lama or other options.

Joeri: 15:42

Exactly, but Align is in quality detail. For me, it's a real good insight from this podcast episode if people are now listening to this.

Daniela: 15:50

They all sound, look the same, but they are very different. The way they are made, the way they are deployed, everything is different in all these models.

Joeri: 16:00

Yeah, the time is running, Daniel, because we are at the web-smith. I know that they will kick us out of the booth here very soon. I know you have other things to do also, but if people want to know more because you already teased them a bit with everything that's possible, where would you like to send them?

Daniela: 16:19

See if people want to know more about us. Yeah, what you are doing. Yes, to our DefineAI website, main website and the resources we have, the blog section and the white paper section, and we just launched our manifesto of ethical AI, stating our ground proof of how we operate the data aspect, because all AI starts in data and without the right data and bias data and transparent data, there is no fair AI and responsible AI.

Joeri: 16:47

Actually, that is my background, also with data, but that's for another time. Guys, as there will be a blog article, there will be show notes. You will find all the links that Daniela mentioned over there. So, daniela, thanks again for being on my show.

Daniela: 17:02

Thank you, Joeri, and have a great time in Lisbon again.

Joeri: 17:04

Thank you, and if you are listening to this podcast episode, you think, wow, really cool insights. You can share this with people around you, that other marketers, other entrepreneurs, other people that are interested in AI. If you are not yet following the podcast, this is a good moment to do this and, of course, I would like to see you back next time.