Welcome to Speaking of AI, a podcast where LXT’s Phil Hall discusses current trends in AI with leading experts. In our first episode, Phil chats with Diego Bartolome, CEO and founder of sintetic.ai and an expert in generative AI.
Phil:
Hi, and welcome. Our guest today is Diego Bartolome, founder of sintetic, and a leading expert in the hottest of hot topics, arguably the most widely discussed technology of the past 10 years, generative AI.
Diego has a PhD in electrical engineering, and remarkably, he carried out his doctoral research in parallel with studies in business administration. After that, Diego founded and was CEO of tauyou, a machine translation company specializing in NLP tasks, which was subsequently acquired in 2016 by TransPerfect.
At TransPerfect, Diego led the AI team developing chatbots, document classification, and active learning tools. And he expanded TransPerfect’s AI-based offerings into a diverse range of verticals.
After leaving TransPerfect, Diego worked on cognitive services at Microsoft, focusing on AI for customer support scenarios. In his current role, Diego is CEO and founder at sintetic.ai, an organization focused on generative AI, but also developing a search engine tailored to specific verticals.
Diego, it’s a real pleasure to have you take part in the podcast. And given the excitement around generative AI currently, this conversation is very timely. Welcome.
Diego:
Thank you. Thanks for having me.
Phil:
I’m going to dive right in here. When it comes to generative AI such as DALL·E or ChatGPT, do you think their inherent creativity leads consumers to believe that these applications possess a higher degree of intelligence as opposed to highly functional but perhaps less creative applications, such as autonomous driving or vehicles, which may have been at the forefront of AI consciousness for many people just a few years ago?
Diego:
Probably those applications and self-driving cars and what you mentioned were not mainstream. What happened with ChatGPT in a matter of days was incredible in terms of number of users and the applications that it opens up.
Suddenly everyone started talking about it. It was in the news, at least here, well, everywhere in the world, I think. And all the users were trying it. The interesting fact that we are seeing that people are realizing that if you want these technologies to work for your use case, you need to work on them as well.
It’s very impressive because you start interacting with ChatGPT and all technologies and you get responses that might or might not be true. And then if you want to apply them for business scenarios, then really you need to dig deeper. So they are extremely creative, but they are not that easy to achieve what you want to achieve with them.
And with image, it happens more or less the same. So you can prompt DALL·E or others to get an image. And what you might get is not really what you want. I mean, it’s good to create art, I would say, if you want, because it can be extremely creative. But on the other side, if you want to turn that into reality, it takes a lot of work.
You really need to be careful that the responses are accurate, that the images are accurate, and then, that the speech is even accurate. It’s something that we will see more and more. So, we’ll see a lot of companies working in the space, but trying to find those applications that make sense.
No, so that will be interesting. The thing is with generative AI, is that basically anyone can use it for literally work. And that explodes the number of applications.
Phil:
Yeah, sure is. Our CEO told us only last week, “could you all go and explore whether there are any possibilities of leveraging ChatGPT on your daily tasks?” And I’m quite impressed by some of the mundane tasks that it can do quite quickly and easily. I’m inherently lazy, so it’s quite welcome.
Diego:
We are all lazy and we all like to automate repetitive tasks and what the opportunity I think, for this technologies is as well is that it could allow us to do the more creative stuff or the tasks that require thinking or empathy; those tasks where people are still better today, I think will be reinforced by generative AI in general.
But everyone is looking into how to leverage ChatGPT and other tech. Many years ago when I created tauyou, my first company, it was mostly at the beginning machine translation, but companies we work with, they needed to automate it to automate processes.
So that comes always first. You can look into AI technologies, but you should have the data. You should have IT systems that support your daily work. You could already automate a lot of things with UIs, with automated processes, automated technologies, and software in general, if you want. That comes first. You cannot go into generative AI without having all of that.
So there is a journey in that digital transformation for companies. And that’s important because sometimes people think, “Okay, let’s jump into this and then, okay, but our IT, our software should be better.” So that’s the first step.
Phil:
I said in the last question that this is right in the forefront of the global consciousness, if you like. So, people are thinking about it and people who may have had close to zero awareness of AI a few weeks ago, a few months ago, are now talking about it.
You are not one of those people. You’ve been working in this field for 20 years. So let me ask you, what are some of the changes that you’ve observed in the last 20 years and what do you think are the implications of these observations for the next phase of evolution?
Diego:
What I’ve seen – probably the most important thing – is the speed of development and change. In the past – I would say years – there seems to be something new every day – new applications, new technology.
I think that tipping point were the transformers in 2017 – that paper about attention it opened up so many applications in generative AI. Now, we have text with ChatGPT, but we have DALL·E, a stable diffusion models for image.
We have speech as well, a speed generation. There are a lot of companies creating emotional voices even. We have music; there are people that are creating music even with Google, but they’re not releasing their systems. So to me the most important change has been the pace of development and the number of applications that it enables.
Of course, it’s a combination, I think of two things. One is that technology and the core. The architecture of those deep neural networks, that’s one. Also the availability of GPUs that maybe in 2006 or ‘7, it wasn’t like that. Those to me are the two most important things when I was working there at tauyou for machine translation, how it evolved from 2006 or 2007 to 2016.
I mean, it was quite slow. You could do things differently. There were companies doing really interesting technologies, but when neural machine translation started with deep neural networks at the beginning, the systems were not that great. But suddenly it started to improve at an impressive pace.
And to me that’s the main thing. The change that we need to make in this five years is to make those applications real in business scenarios first, but consumer as well and trying to see where these technologies can bring a return on investment to companies and to people. That’s the biggest challenge in terms of equality, let’s say. I would say, I think for instance, ChatGPT, $20 a month, that’s a good price for many people in many areas in the world, but that might be extremely expensive in other areas. So that’s something to think about as a whole society in the world, how do we want to tackle all this.
Phil:
I think that’s a very important consideration. We don’t want this to be the exclusive possession of the richer nations around the globe.
Diego:
Yeah. Because otherwise it can even generate more inequality in the world. But we’ll get there. I see many interesting companies everywhere. Everywhere in the world you see companies today.
And that’s also very good. There are a lot of people working on low resource languages. There are initiatives in every language to get similar systems to ChatGPT. Even if they’re expensive, to an extent, they are becoming cheaper and cheaper.
To me, that’s something that is really exciting as well, because it’s not only English or Spanish or Portuguese, German, the major world languages; it’s basically any language, anywhere in the world, you can create similar technologies, and that’s quite exciting.
Phil:
Okay. Let me dig down on that a little bit. In an earlier response, you talked about It being likely that there will be a lot of companies working in this space. Is it likely that they will be developing their own generative AI technology? Or will this be something more like another wave of dominance like we’ve seen in the past going back to the early 2000s where Netscape dominated for a period and more recently in 2012 it seemed like Siri was synonymous with speech recognition speech interfaces.
Do you think it’s likely that there will be a proliferation of companies utilizing one or two products, which are developed by very large companies, or do you think there will be actually a lot of development of generative AI by smaller companies?
Diego:
Well, currently the companies that are developing the infrastructure, the models are large. Now, you, you have OpenAI. You have other companies that have received quite a lot of funding such as Cohere.
But the difference in what we are seeing versus what happened there is the open-source movement, if you want to call it. There are a lot of initiatives that help with that. The perfect example is Hugging Face where, in an initiative with other companies, they created this BLOOM large language model. So, for those, of course, infrastructure is critical and it’s expensive. You need quite a lot of investment to make that happen.
But the good thing is that it’s not only one. There are many and that’s difference versus what happened in the past. I think we’ll see what the market ends up doing, because it might be that some of those companies get acquired because they got like a 100 million or 300 million in funding that will require some exit for the investors in the future.
And that will concentrate the market. To me, that would be dangerous. I think a number of companies building those foundation models is critical for innovation. And then also what we will see is a lot of companies developing smaller models. But you mentioned that at the beginning, the first question, specifically solving a task, which is a way of doing things as well.
So, those large language models generalized pretty well, and they can do a variety of tasks, but maybe if you have data and your exact application, the goal that you are trying to achieve, your model could be significantly smaller and therefore cheaper, faster. I was always telling my teams, and I’m telling my teams now that, “You, we don’t need to develop what is cooler at some point in time. We need to develop what is best for the task.”
And that may be a smaller model that is tailored for a particular goal. That will come a lot as well, because now everybody’s trying to do everything with all those large language models, but it might not be necessary in many situations.
Phil:
So your vertical-specific search technology, is that an example of that?
Diego:
That’s an example. Besides the generative part that we are creating content for web, for e-commerce, for customer support. There is another area if you think about generative technologies; on the internet, everything will be generated by machine with feel at sintetic that really if you want the best quality, you need a human touch.
So maybe the generative part gets you the first draft, but you need to tailor it to your audience and do it better and improve it. So the people are needed. But in any case, even today, if you search on Google it’s very hard to find the relevant article.
You find the articles that are better positioned, that have invested more in SEM or SEO but maybe you are not able to find what you want, which was not the case like 20 years ago. So what we are doing with the search engines – with your data or with data you are interested in being able to find the data only in a subset of documents, webpages, documents, research papers, whatever, so that you are able to find, in your vertical the information you’re looking for.
So, we use embeddings for that and then we use Q&A based on an open-source model. We try to fix the information overload that we’ll have with these kind of technology solutions.
Phil:
Very cool. So we’ve talked a little bit about the evolution of AI, and I think you expressed that there was some surprise at the acceleration of development since… What was it? 2015, 2016?
Are there other elements in this AI evolution that have taken you by surprise that, that where maybe the direction of the technology didn’t go where you expected it to?
Diego:
Well, that paper about attention, that was quite the tipping point. What surprised me a little bit is the way researchers and industry and companies have developed applications based on that.
Now, in the end, everything can be coded as a sequence and therefore you can apply this type of technology for basically anything, for anything that you want. And to me, that’s probably the biggest change now, maybe not thinking only about text, but anything can be called as a sequence.
That’s probably the most interesting of that.
Phil:
Great. So looking at the news over the recent weeks, one of the hot themes has been the integration of generative AI into search. Microsoft announced that it’s going to integrate ChatGPT into Bing.
Google quickly responded by introducing Bard and saying that they were working on a similar direction. Some writers have raised a concern that generative AI at its current stage of evolution has qualities that make this risky. And, you know, to be honest, I share some of this concern. We’re already living in an environment where there is a lot of fake news.
So, deliberate misinformation, accidental misinformation, misguided misinformation, whatever way you want to look at it. There’s a lot of misinformation out there. The risk with generative AI seems to be that it’s very good at producing things that look convincing, that sound convincing. The quality of the written expression is very good. It’s probably better than a lot of people could write themselves, but you’ve got questionable content wrapped up in a beautiful package. Are we rushing into this too early? Is this something that’s perhaps irresponsible to turn this into a kind of technology arms race?
Diego:
It, it feels that this is what is happening indeed now with Google, Microsoft. I read something about Amazon yesterday as well, and Meta. There will be a war among the big players for sure. And those, they try to reach everyone in the in the world now. So that, that’s a difference.
To me, it’s a little bit dangerous. We are living in that world where you don’t know what’s true or not, or a search, we were saying before you’ve tried to find something. And then you get tons of pages that are basically the same. It feels like copy and paste – all of them.
So that’s tricky finding the right information and the truth, let’s say, if that exists. That’s really relevant. I think probably it’s part of the marketing efforts in those companies. So who gets there first and how they are impacted by these technologies. But to an extent I don’t know if they could do it, but, launching these type of solutions at a smaller scale, or in only in some cases, could be probably more beneficial.
So, making sure that whatever development is done reaches the market at a quality level that it’s what people need. Because otherwise the only chance that we have is making sure people learn how to use them. And whenever there is a risk, they know what the risks are. And that’s tricky because not everyone does it, or not everyone reads a piece of news, and then they think about it and say, okay, might this be true or not? Or is this real? Or this is fake?
No. And, since we are not doing that as a society, there needs to be some safeguards. An example of that is with my daughters. I have two daughters. They were playing with sintetic, with the alternative part because, “Oh wow, this could be cool for school”.
And then they were trying it and I was telling them all the time, “You need to make sure that what it tells you is accurate or not.” That’s very important because if you just use what it outputs, it might be wrong. So maybe it’s an opportunity for us to develop that critical thinking again, if you want to call it, because I feel sometimes that we have lost it along the way.
Phil:
Yeah. That’s an encouraging way to close on that thought. I like the idea that a product like this is going to stimulate critical thinking. I think that’s a great positive.
Diego:
I hope, because in the end, this will happen, unless politicians ban it. I read some news this week as well about that in Europe that they’re trying to prevent certain systems in some situations, which could be dangerous, I think as well, because then you are not encouraging development.
I think they need to be open and they need to be there, but then we as a society need to change. The challenge that we have, I think, is that this technology has evolved so fast and we, as, as people, as you said, we are lazy. It’s hard for us to change. So then education is changing very, very slowly.
How to balance that, that’s probably the trickiest situation. But I hope it encourages that and then it reinforces the human skills. We need to be able to speak well, we need to be able to reason, well, we need to be able to be more critical, maybe philosophy gets back into society now asking, “Is this really true?” Or these type of things will be extremely valuable in the future. I think, and I hope.
Phil:
Can I ask a question that, uh, I have an inherent interest in? So as you probably know, LXT is a provider of AI training data, AI test data. And one of the questions that I get asked very frequently is, “Well, isn’t it going to get to the point where that’s just not needed anymore? That there’s no need for a human in the loop?”
Now, we’ve already talked about, if you’re using generative AI, you don’t want to just push a button and send, you want to review it, you want to check it, you want to treat it as a first draft and take it from there, so that’s a kind of human in the loop but, after the product’s developed.
But in the development cycle, the testing and training cycle, how long do you think it’ll be that human in the loop is still a relevant part of that equation?
Diego:
To me it’s slightly similar as well, so I think the human in the loop is still important. It feels a little bit as well that this type of solution can help a lot in that training data. I remember creating data for chatbots and that was extremely expensive and time-consuming, and really hard.
And today, many people I’ve seen, they are doing that using these types of models and reaching the same accuracy or similar accuracy with that. But to me, the human validation, I think it’s important. So it might come before or after but, any system, will need to be maintained, and will need to people… We will need to take a look at what’s happening and then as engineers think about the next step to actually improve it, and that requires human thought.
Of course, it’s the same as within AI technology; you said these type of tools of generative tools can create content that is better probably than the average person writing. So, here is more or less the same If the training data was not extremely valuable, “if you want”, then probably generative, data created by generative models will be enough.
If you are thinking about medical applications where accuracy is extremely important or legal, then maybe that data creation is needed, that people do it right. But, but of course it changes the game as well, you know, and I think the example of that also how data annotation companies they started to create, and I feel that many are now using generative as part of their offering.
That’s probably the way to go as well. We need to innovate and we need to play with what feels like a threat. It might be an opportunity.
Phil:
Yeah. Very healthy attitude to that as well. So I’m going close with one last question here.
If you were interviewing yourself, what is the question you would really hope to be asked, and what is the answer to that question?
Diego:
Yeah, I was thinking about that before, and then as we were talking. I think one of the questions that we touched on a little bit is, “What are the impacts in society as a whole besides the technology itself?
But we touched on that a little bit. And that critical thinking, that way of evolving as humanity in a good direction, I think that would be my hope; that these types of technologies or solutions have in our world.
So to me, the question regarding that and the part we touched on, I think that was really interesting and, and to me that was the question I would probably touch on as well.
Phil:
That’s great. Okay, Diego. Thank you so much for joining us today. It’s been tremendously informative. I’ve really enjoyed our conversation and I look forward to perhaps welcoming you back again at some point in the future so we can check in on where things have actually gone. In the meantime, it would be great to have you as a guest again.
Diego:
Cool. Yeah, we can do that, Phil, or we can create uh, uh, a digital twin and then maybe the digital Diego and the digital Phil can speak.
Phil:
I’m very keen. Yeah, that sounds great.
Diego:
Okay.
Phil:
Okay, Diego Bartolome, thank you very, very much for your time. I really appreciate it. Thank you.
Diego:
Yeah.
Phil:
Bye-bye.
Diego:
Thanks to you. Bye.
