Latest / The Collaborative Canvas Podcast / S3 E8 - The Dark Side of AI: Bias, Hallucinations & Fake Intelligence | Is AI Really Smart?
Transcript
- speaker-0: I want to answer this question, but this so important. And now the AI is going to give you shitty answers. speaker-1: valid point. We've been hearing lot about this as well in India. speaker-0: on chat tpt, gemini, copilot or cloud. In this study, speaker-1: It's been backed by a lot of data because it's speaker-0: Yeah, most of the biggest LLMs are from the US, most of them, or China. But for me, it's like an amazing tool to go 100 % of synthetic data all the time. And when you talk about it, this is really important because speaker-1: What is exactly machine personality? There are certain cultural nuances. You I generally get the same feedback. You mentioned about D2C specifically. Countries like Bangkok or Thailand. See, AI is not 100 % sure, right? speaker-0: try to corrupt me But They make like a jar of us, no? speaker-1: It's more western centric hallucinations, what you saying? I get rejected, which still means that a lot of people would still lose their jobs. speaker-0: One of the trick you can do it and I speaker-1: you Welcome back to the Collaborative Canvas, where we map the intersection of the innovation, strategy and the human stories behind it. I'm your host, Ankesh Rixena. Today, we recording live from the vibrant city of Bangkok at the event IAX Insights Innovation Exchange. We have spent the last few days talking about the future of data, but one question keeps rising to the top. If AI is reshaping how we understand human emotion, whose emotion is it actually reflecting? Joining me today is a man who is literally auditing the machine personality of the world's leading models. Cassiano Albuquerque is the head of AI product and research innovation at Cassie.ai. He's here to talk about the paradox that every brand builder and researcher needs to understand, synthetic emotions and the real impact. Welcome, Cassiano. To begin with, ⁓ let's introduce yourself, Cassiano. speaker-0: So, my name is Cassiano. I'm from Brazil and living in New York City. So, I'm passing my whole career, it's two decades career in market research and tech research. So, I through Ipsos and data collection panels. And for the last four years, I started to entrepreneur. So I opened another company before Mark Research for All. It's like a aggregating platform for our industry. And then I sold the company and restarted with Cassie AI, which is a software house for helping companies to build themselves the AI ecosystem. So that's why I came here for Thailand and Bangkok and IIX, APAC. for show the paper that I'm producing and then I study about behind the scenes of the AI stuff. So behind the scenes, how works like the LLMs and how works like orchestration of like agencies and synthetic personas and how this comes through if like the emotions, if the... the ⁓ AI could feel in the emotion or not. So I come through this ⁓ meta-stud from my research in artificial intelligence. speaker-1: Okay. We have been knowing each other for almost like four plus years now. And I've seen market for all to Cassie that you've developed and we've been talking about it. And apparently since, know, it was a lucky charm for me that you happened to be in Bangkok and I saw the, the, the paper that you presented, the presentation that you did. How about we give an audience a glimpse of what was your conversation at ⁓ IEX? And what was the whole topic about? And then we jumped right into your question. ⁓ great. speaker-0: So, the paper that I brought to IIX was a meta-stud about ⁓ market research and all the artificial intelligence for LLMs and agents orchestrations. So, it's a meta-stud and it's good because most of the presentations talk about the outputs. So, I'm talking about the behind the scenes, how it works. to get those outputs and was amazing. Everybody was a really good conversation there. Everybody's so warm and they have like a bunch of questions after the presentation. So. speaker-1: Yeah, I I saw the entire crowd being there in the track two, guess that was track two. And it was a full house hearing you up and right and answering all those questions, asking all those questions that you want. But then I realized that, I can keep my peace for this podcast. So I have a separate question is that I wanted to check about. So you spoke about the machine personality gap. What is exactly machine personality? speaker-0: So when you talk about LLMs, you need to go step back and understand how they are made for. So how you make like LLMs from scratch. So every single LLM have like a background of companies. we are having like most of the big LLMs companies, AI companies come through US and have like a sort of company in French, but it's open source. And then we're going to talk about open source and private. And some, it's coming like strong with China, AIs too. So we have this study, Bipsic, but now they're growing for all the type of LLMs also there. So every single one came like background for how they make the machine learning, how they put input, all the data. from each one was different. So if I talk about, let's suppose the grok, which is the X one, the elements, become way more think about the X stuff there. So it's different between them and for example, Jamie and I, who come like input data about the Google. So everyone have like, behavior in different in each one, each other. And, but this is the beauty, know? because you not need to select just one LLM for work. You can work with the multiples LLMs and this is gonna influence in your synthetic data also. speaker-1: But tell me something, I mean, we'll get to the synthetic data part because I think that was the line right of your conversation at IX. What I generally do, you know, when I look for information, I look for information on chart GPT, I look for Gemini, ⁓ Claude, Copilot. The backbone of all these four that I work on has the same data strength, the same database and everybody's feeding and they're searching the entire globe to, you know, answer my question and I think somewhere down the line it is like if somebody is saying X the other one would be X1 or X minus 1 how does that difference actually work out I mean if I'm really doing a synthetic data for that the difference would always be highlighted we always say that it's always a 70 % correct and they always have that asterisk disclaimer that you know not trusted completely. Exactly speaker-0: This is a point everyone's and this should be normalized for everybody because the end of the day who gonna accept the output of the LLMs is us. No, so we need to real check and not let the ⁓ AI do all this stuff for us. They're not gonna be like ⁓ supply for the AI system. We need to work together. As a manager, we need to manage the robots, the LLMs for this. yeah, the difference between the answers is because they are growing different and they have like, we are calling from guardrails. Each one has like a different approach for guardrails and different approach for how they could answer that. questions in certain way or a different way. The crop is more funny and I feel like they're like the black sheep of the LLMs because they are not reasoning. They are like doing stuff for math, but for creativity, it's a really, really good one. When you talk about it, this is really important because the LLMs normally, they are learning about general information, the general information for internet, general information for books or for images, videos, audios, and all this multi-platform stuff. But we need to put our layer of specialists on this. And that's where they come, the agencies. The agencies, you put your context data or people love to call it prompt ⁓ context engineering, they create all the data of there because when you get this specific data, you get way more accurate ⁓ output and answer for the machine learning because they... Learn about your specific data there. speaker-1: I like to quote a question I heard and I wrote that question down. for a brand builder trying to find a voice, ⁓ somebody asked you how do we identify which AI model naturally aligns with our brand specific emotional DNA versus one that might be fundamentally biased against it. speaker-0: Yeah, we need to use like multiple ones. And the other thing is we need to bring all the data for the market, like all the socials data. Everybody was talking about something and we can like cutting down all the, ⁓ try to cut them because nowadays it's gonna be way, way, way more difficult. And that's why come the micro research, the data collection, ⁓ for bringing more real data, for inputting those data into the agencies and put the LLMs for learning with real people. And then you can a mural these real people for the brands. So that's a thing that the goals like more. speaker-1: I think even in that case, if you're using multiple LLMs, as I mentioned earlier, I generally get the same feedback on my questions that I post across. And agent-kia becomes, to a certain extent, my voice, because everybody now know what I am actually looking at. Unless and until I'm asking a certain question, I want them to fetch me an information on a regular basis. But sometimes, as So I do a lot of research when it comes to finding the guest, the topics. And I really need to learn more about those topics. So when I'm interviewing you, I did some research about what topic are you going to discuss. there's so many things that ⁓ I searched through different LLMs. It almost gave me very identical answers across, plus minus a little bit here and there. I was aware if I'm going to de- get into more details. I think I'll be able to fetch more information by point, every point that it mentioned. And to the next point, I remember you also mentioned about there are certain cultural nuances. Now, since you've been using grog, but there are other elements as well. There are other elements as well. You mentioned about D2C specifically. speaker-0: Yeah, for cultural stuff, in this study, we got like 40 questions about different dimensions. One of the dimensions for cross-cultural understanding, how does LLMs understand real cultures? So I gave questions about ⁓ slangs for the small city in Brazil or small city in China or small... ⁓ ⁓ particularly stuff for different type of cultures in different places to see each LLM, like know the answer or know what I'm talking about. So I did this exercise and which one was like the best in the chamber for the cultural thing was Gemini. Gemini has like more, speaker-1: It's been backed by a lot of data because it's Google. speaker-0: Yeah, yeah, that's why. then he's good for, for, yeah, Jim Nye is a little bit more verbosing than other ones. And it'll feel like, you, you try to corrupt me. speaker-1: I am about to get to that point. You also mentioned about hallucinations. There must be a lot of red flags, when we train the AI. So how does hallucination in countries like Bangkok or Thailand, how does they work? speaker-0: Because as I mentioned, most of the biggest LLMs are from the US, most of them, and China. So the training is about this most like English language and English culture and American culture. They have like lot of trillions of data there. And the Chinese too, the language studies. they started with the mother languages in each one of them, like for them. So once you go for different languages, the culture doesn't get the same ⁓ quality of answer once you talk about, or in English, or even when you type in English and you type in another language, the answer is not that occurrence with. speaker-1: It's more western centric hallucination is what you're saying. speaker-0: Yeah, another thing for you, try to skip the hallucinations and not trust the first answer that they give to you because all these LLMs want to try to not use that much tokens. So they're going to give you the first answer that they get, like the fast answer. If you accept that, fine. But if not, you need to ask more questions. And you need to know, deep, go deep, analyze deep what I'm talking about. You're not giving me the right answer. I want you to go deep of your information and give me more clarity or give me at least three paragraphs, the information more constantly. So that's one of the tricks you can do it. speaker-1: Well, I'll stop you here. Just one question popped out because of this. ⁓ In MR, we've been talking a lot about ⁓ data quality. Open-end has been a real challenge for a lot many people. And if I see as human, sometimes I can explain certain things, but I find it difficult in putting that down on a questionnaire sometimes. So I tell chat GPT that this is what I'm thinking. It gives me an answer. And there's always a... any LLMs for that matter, they are very respectable to your questions. They always say you're thinking right and stuff like that. And that's how, you know, they frame the polished answer for us, which as a respondent, I posted on them open end and I get, I get rejected because the people who are reviewing it, they see that, you know, it's been copy pasted by, for, for, so it's not the real essence. How do you see that hallucination? in a certain way not helping and rather letting the truth go. speaker-0: That's a really good question. And the fact is, happens all the time, And we need to not trust in the first thing that LLM gives to us. And we need to ask more things. Like, instead of you answer now, this is a prompt. So you need to work better in your prompts instead of you just put a, or make this question for me, or, but forgetting I get better good prompt for get all the boost of the LLMs you need to ask like, ⁓ we're gonna talk about this team, give him or Hinton more clarity of the macro and go the micro for whatever you ask them. then before, and then you put like, before you answer me, you have more questions for me to understand you go deep on your answer. So once you do this, they're to give you five questions, 10 questions, but most of them don't have like, no one to answer this question, but this is so important for give the LLM more knowledge, what are you talking about for they go and grab the important ⁓ answer and give it to you. ⁓ that outputs. speaker-1: And you just gave an LLM gender identity. speaker-0: Yeah, and you know the other thing is because in Portuguese, we always use the hit, we not use it. And for us, LLM is masculine. And I have a soul, Jim, because you're named as him. ⁓ speaker-1: I'm And since you have already given an identity to ⁓ LLMs, I'll probably take you to the next paradox between synthetic and human. You also talk about that. Yeah. speaker-0: I love these questions and this is the main question for ⁓ you in all the conference. I've been in these last three years of AI, such an important team and important tool for our lives nowadays. And this is the question that comes every single section. Synthetic emotions. Human versus the AI. But my answer is, how it's going to be for human? Because we are in control. So we control this AI, the agents, whatever you could call it. And we control them. And we're going to manage them. whatever task we want them to execute for us. So, in my perspective, I was going to be human first, and you need to use the whole capacity. And that's the thing. I think people are still growing, but with not the pace that it should be. for getting all the max of the LLMs and the AI could bring to us for the tool. But for me, it's like an amazing tool that you can use. yeah, and another topic of this that is really important is companies calling, we are the first AI first, just put like a co-pilot. or Jamie and I are for the team. Now everybody in the team has a co-pilot. Now we are AI first. Wait a minute. Wait a minute. So this is, you talk so important because actually someone asked me in a conference, what should go in the free or the pay? No. And then I say, never the free, free just for joking, just for use it as Google, but the data there, it's lack of information between then and once you pay. About the copilot, because all the companies use the copilot for the foe or Gemini or some like Char2PT as the company using them, but they have like a tier. I don't know if speaker-1: Everything comes for a price. speaker-0: Everybody knows that, but you can use a certain amount of tokens and tokens to inside and tokens to outside to input number talks for input and number talks for output. If you are like a heavy user in the company in less than two weeks, you all your tokens and now the AI going to give you shitty answers. speaker-1: You double the code. speaker-0: just one, two phrases and that's it. And even you like be mad, answer me more, they not gonna because they, your tear is gone. speaker-1: That's more of a price impact on the services that you're taking. If you don't pay the price, you would not get the real essence of the GPDs, whichever GPDs we talk about. But at end of the day, it's the main body behind ⁓ any LLMs is the data that's already there, but you pay a price for it. The point is that, you know, within the synthetic emotions and the real impact that we were talking about earlier, ⁓ You also mentioned about the personas in the beginning of our conversation, that synthetic persona. Do you think ⁓ companies who are now coming up as synthetic moderators, they're coming up as synthetic respondents, are they eliminating the human touch of the focus group? Because I still believe that, know, like us talking right up to each other, we are decoding a lot more insights when we speak to each other. It's more like, even if I really wanted to know anything beyond our conversations, I can still dig you deep into that. But that's not the case with, with, think synthetic would only be limited to a certain extent. What's your take on that? And considering from the cultural perspective comes from the Western centric countries. speaker-0: I have a sensation that it's such a hype on these things now for synthetic data and building all the moderators, synthetic moderators, all these things. I think certain point for the big ideas for stay started, instead of use conventional market research data collection or for the beginning of the studies, I think it's value use those because they are quickly answers for you understand, let's suppose you have like some brand have like a marketing, you're gonna talk with the market research team about such a specific topic, ⁓ some targets. And why not in this conversation use the synthetic persona for answer a little bit about what are the doubt that they have in this meeting. I think this is a vital thing, but once you go for invest million of dollars and one new product, you're not gonna go just only synthetic for this. You need to hear human beings for having these sensations or taste or try the things. I think gonna be the combine of both sides. ⁓ Synthetic data, it's really fast. okay, you have like the collection, right? You have like a thousand completes, but you your fieldwork, you extrapolate the dates and you need to complete 10 % of this. Why not go for data, the synthetic data there? and complete these questions. I think this is a lot of things, go 100 % of synthetic data all the time. This is concerning me. speaker-1: Exactly. You cannot have synthetic data as 100 % of your data because that's already a past data. And every other second or every other minute, things are changing. And you actually made a very valid point. We've been hearing lot about this as well in India that I have a very niche requirement. I've done 80 % of my quota, but that 20 % is I'm still struggling with. I wanted synthetic data in a certain fashion, in a certain quota structure, or in a certain percentage. But I don't have any data. So in that case, how would you recommend those research agencies where they were struggling with that remaining 20 % to fill up the synthetic data and they don't have any backed up data for that. Would you recommend them to at least do 5 % so that you can train the model or would you just read the entire questionnaire, entire survey responses that you've got to fill those 20 % remaining points? How would you do that? speaker-0: I, if you don't have like any, you must have like at least background information for give the AI and the agents for help you doing something else, not trust like just a hundred percent in the general, general LLMs to any. speaker-1: Let's say, Cassie, we're doing a phone survey. OK. And there is a specific phone which is already in the market. There a lot of ads being run. I'm sure there must have been lot of data which is already being there. They wanted to have a certain market share about that data. But the representation is not coming. People who are using it, probably be not on phone or on laptops most of the time. They could be like, blue collar jobs, you're not sure whether they take surveys or not. So you have not even filled a single piece of it. You're still lacking behind that 20 % of it. Chad GPT or any other LLMs would still have some data around that phone model. Can we not compliment that data to fill that 20 % gap? Because now we know that, you know, not exactly based on those questionnaire, I may have that data, but at least that data. say 10 % of the essence even if I take from there, I can still frame the answers and will, do you think you can be done or not? speaker-0: You can, but in that case, I think it's good if you go for scrapping the internet. Yeah. Before you go just because most of LLMs doesn't go with data, update data 24, 7. They have like, once they update the system, the family of speaker-1: That's exactly what I meant. speaker-0: of like ChachiPT 5.1, 5.2 and all these things, they have like a deadline for here from there, you don't have any more data. And then the next update, they're gonna give those data there. So none of these LLMs has like a 24-7 information. So before you ask straight in the LLM for help you like this, I will go one step back. to make ⁓ a scrap in the social media about the segment that you are talking to, grab all the comments, the new comments for all the company or for the vertical that you are studying, grab all this information, put this into the system, into the LLM for that, give like more knowledge about the target that you are talking about. And after that, you ask for that data synthetic for them. I think this is going to be a little bit more occurrence. speaker-1: I think that's exactly the valid point. mean, if you can actually train the model firsthand and then filling up those 20 % later on, which will be easier for because at least the model know what you're going to ask. Yeah. Rather than giving you crappy answer. Yeah. And then you're not accepting that quota because even with the synthetic data, you don't have that kind of a judgment. You don't have that kind of a quality assurance. You don't have, you know, the speaker-0: Yeah. speaker-1: what we call, I mean, as a client who don't trust the data because it's synthetic, you would at least have some specification which has already been feeded in, basis on which you get some output and you can still accept it. speaker-0: I believe this, yeah, 100%. speaker-1: And going back to the point because I remember you know they have been very polite. Also made a point on aggrievables. speaker-0: ⁓ yeah. Be criticized. You you can do two things here. If you work straight on ChatTPT, Gemini, Copilot or Cloud, they have a system there and you can, they have in the settings, you have some prompt that you can change there. Normally this prompt doesn't have anything. But if you go in the settings for those chat bots, you can insert a prompt there to be, I know only you people either with me, I wanna criticize, I wanna you don't give me shitty answers. You can put all these prompts over there and he gonna be every time that he wanna talk to you, he gonna pass for this prompt and this is gonna be your guard hail. for a girl like a so kind and give you whatever you want. Yeah. So you need to trick with them and put like a prompt who can like show them like who is managing here. Yeah. speaker-1: I remember I saw a short on YouTube. It's, and this guy was actually telling the word that, know, if your chat GPT is overplayed, just tell him, don't be over-polite. Give me a straight answer. Don't, you know, don't say that I am right. I am asking, which is why, which means I'm wrong or I don't know that. So give me an answer, which does make sense. So most of those brand strategies, they often need friction to innovate. And again, one of your point where you mentioned, so how do we prompt an AI? I mean, just. not being said just yes man. mean if you can you just elaborated it but I wanted to say for an example. Not just saying I wanted a straight answer don't give me shitty answers. What is that prompt that can at least you know enable them to get the right answer and not being overly polite. speaker-0: Yeah, you can go if you use those charge bots, you can do like this, they go in the settings and set up in your system, the system prompt, you call system prompt, because you have like two types of prompt in the settings. ⁓ speaker-1: That's in the settings. I have never tried that because I am always like very... It's very funny to see that you know whatever I say they say yes. speaker-0: ⁓ huh, yeah, and most of them do these things. And, if you go in the system settings and go in system prompt, every time that before he gave you the last answer, they're to pass through the filter of your system point prompts and they're going to see, ⁓ with him, can, I can be like that, that polite anymore. And then going to change the answer over there. Or you can put this as a system prompt on your agents. Once you created your agent, you can put like a tone of the voice of the agent. You can put also this, want you, if you have like some questions for me, give me the questions, never give me the first answer that came to you. I want you analyze internally. And this is coming with another point is like a sub agents. speaker-1: Yeah, I heard of sub-agents as well, Ricky. speaker-0: Yeah, instead of use one agent, you can use two or way more agents. That's you call agent orchestration. So you have your main agent. And then instead of they give to you straight point your answer, you input the sub agents inside who gonna criticize everything that this guy did. Got it. And then give the feedback. If not score, let's suppose, to 10, they're going to give the feedback, you need to improve here, here, here, here, and they're going to go back to feedback, the agent's going to work again, and then they criticize the agents and then say, okay, now you're done. You can input those multiple agencies and work like orchestrations. then each one do specific tasks into your general task. It's a good one. speaker-1: You pointed out a very good example with the satisfaction score. How do you see the future of insights professional will be if AI is handling these logics? And, know, with data processing, or probably, let's say, the most critical human skill for any market researcher or brand strategist in India is understanding how the data is talking to them. But how do you see the future of these guys if everything is being given to AI? speaker-0: I, again, human always first, right? ⁓ But I see with good eyes once because the operational time is consuming a lot of time. normally someone does ⁓ that data process for a big project, they're going to take like a week, at least a week to data process all the data and check all the data and everything. Instead of they take like one week to do all the process as we've been done for decades, the guy just going to check if the agent or if the AI did the correct things. So we need to, maybe we need to change a little bit to execution things to be way more understand and check, like way more check personally to see and let them have stuff. Should the AI do for us? speaker-1: which still means that a lot of people would still lose their job because it's been synchronized to a few more only. speaker-0: Mm Yeah, that's that's some issued like speaker-1: See AI is not 100 % sure, right? It's not there where it should be. It's going to take some time. Do you believe that those data processing ⁓ companies or the teams who are actually or the people who are behind it, they should improvise on their skills, they should upgrade themselves? speaker-0: Yeah, that's definitely. feel AI is like in the past when you have Microsoft Office. Before Microsoft Office, everybody is typing in the machine, with the paper, all this stuff, and then go to computer. How this process to analogic typing and go to the computer. A lot of people lose their jobs over this, but you need to readapt for the reality. And that's the thing. And then personal thing too. I'm one of the teachers in, in, in Brazilian association for helping market search, understand more about AI. So yeah, that's the thing that the, community should. escalate this for all type of analysts and everybody. ⁓ understand more and because nowadays people need to be proficient in prompts. That's the thing. They need to know how managing the AI and this is challenging. This is new, but I think we're going to escalate for that side too. speaker-1: Got it. And one more question. Generally, I ask this to most of those guys, because the name of the podcast is the Collaborative Canvas Podcast. a collaborative future. The question is, our podcast since is about collaboration and future, how do you see, looking five years down the line, let's say, the world where AI and humans co-create the emotional maps? Because something that you also raised during the conversations. And will AI eventually understand our emotion better than us? speaker-0: No, but they understand how, as a machine learning everything, they understand the percentage of the emotion affects us the way that they're to do it. So it's like a statistic guy behind the scenes. ⁓ If something happens like this, or if you have some advertisement like this, most of the people are to react in that way. because now they have those data. So this is something like they're going to know a lot of things about emotion. Of course, feeling not going to be, but they kind of predict some percentage of the feelings that you're going to have there. speaker-1: There's a very funny story behind this. ⁓ I was looking and I think came from the shorts itself or reels. don't remember. Probably Instagram or somewhere I saw that. So there two guys sitting together. ⁓ One came to his friend and said, dude, that girl, I love her and I just like her very much. And today I have actually asked her for a coffee. And she said, let me get back to you. I don't remember exactly what line was. And I really wanted to know will she be interested in me or not. This guy looked a little surprised as to why this question comes to him because he doesn't know that girl. So he why don't you ask Chad GBT? speaker-0: Hahaha speaker-1: And Chad GPT would have predicted something. I'm not sure. I don't recall now, but that's a very funny story I saw. he was like, he was there. He wrote some prompt that, this, this, this is what happened. I asked Chad GPT to predict whether this girl will like her or love her in future or not. speaker-0: That's the thing, you in the future. It is so funny because a lot of people just become addicted. They do anything if you're not questioning the system. They make like jarvahs for your personal assistant 24-7 and everything. You need to ask them. You need to get some concern too. Okay, I work with AI, but... I mean, I have my own, our lives though. Yeah. Yeah. But speaker-1: This is hilarious. mean, I think Chad GPT or any LLM for that matter has their own pros and cons and it is still, you know, taking its own shape at this point of time. So there've been conversations around whether people will lose their job, whether people should be upscale. And I think that's not going to go anywhere for at least a couple of more years. Things are changing for good and I think people should accommodate into that and should upscale themselves. speaker-0: I'm really optimistic for this too. I always be the guy optimistic, but I don't have like a real concern to have here. But you you need to improve ourselves to getting the same track on the technology gonna go. speaker-1: You're doing a fantastic job. saw that you've been helping behind the scenes with people. and specifically, thank you for giving me time today to learn more about your paper and the one that you presented. I'm sure a lot many people would also look at your paper, hear us out and they'll get to reach out to you to see. speaker-0: Let's put in the comments the dashboard interactive there and everybody can code there and download. speaker-1: I will do that. I'll probably get more details out for these guys. So at least they should come and learn how, if they really needed some help, you guys can reach out to Cassiano. I'll give you the details in the description box. Thank you so much for making time out, Cassiano. It was such a lovely talking to you. Getting to know more about from your world as to how synthetic data or AI or what all LLMs are working behind the scenes. speaker-0: Thank you and... speaker-1: This has been a master class in looking under the hood of the tools we use every day. And what I'm taking along with me, Cassiano, is that AI can give us the speed. The human in the loop isn't just a safety measure. It's where the actual brand soul lives. And if you're listening to this and realizing your AI prompts might be leading you into a cultural or emotional dead end, today's conversation is just a roadmap for you. To our listeners, if you found value in these insights, please hit that subscribe button right now. Whether you're on YouTube, Spotify or Apple podcast, subscribing ensures you don't miss our next deep dive into the strategies shaping the era. Kasyano, thank you so much for joining us in the Canvas Live here in Bangkok. I'm Ankesh Akshana and until next time, keep collaborating, keep creating and stay curious. speaker-0: you