98 | Making Data-Driven Decisions With AI, A Deep Dive into Data Collection and Analysis with Tianyu Xu

0:04Hello and welcome to Leveraging AI, the podcast that shares practical ethical ways to leverage to improve efficiency, grow your business, and advance your career. This is Isar Maitis, your host, and we're going to talk about a unique topic today. The reality is almost everything we do at work, whether it's marketing, finance, sales, strategy, all of these things require to do research. And the reality is that most of us doesn't do a lot of research, or at least not enough, in order to be more successful than we are. And if you do proper research, you will have a lot more relevant data in order to make better decisions in whatever it is in your business that you're trying to do.

0:44So why don't we do a lot of research? Because it's a lot of work. And most of us have 10 other things that we have to do other than the research. So we either skip that part completely or we just not invest enough in it. But the reality is that with AI capabilities today, you can do multiple aspects of the research from data collection, whether it is primary research or secondary research, and then the data analysis side, whether it's qualitative or quantitative data analysis, all these things that used to take a lot of time, you can do significantly faster and then benefit from it in whatever it is you're doing at work.

1:23And so this is a very important aspect of almost any job in business that most of us do not have the skill to do. But our guest today, Tianyu Xu is an expert exactly on that. He's actually an expert on a lot of other things. And he has an AI consulting agency that's called TYAI. He lives in Singapore, but supports people all over, hence how we met through LinkedIn and started talking and speaking about different stuff in AI. When I learned he's a research expert, and he is a research expert because he did that for several different years in larger companies, including Twitter. So that's real background is doing market research across different aspects of professional business.

2:08And now he's taking AI and applying it on top of that, which I find extremely attractive. And so this is what we're going to dive into today. It's going to be fascinating. We're going to do actual use cases, show you the exact tools, the prompts, everything you need to know in order to do better research in your company. And so I'm really excited to welcome Tian Yu to the show. Tianyu, welcome to Leveraging AI. In the next few years, AI technology will change our world dramatically. Whether you are a business executive trying to catapult your business forward or just somebody who refuses to be left behind and want to advance your career, this is the show for you.

2:50I'm your host, Isar Maitis, a serial entrepreneur and an AI enthusiast. You'll hear invaluable practical tips from innovative business leaders, AI practitioners, and some of the brightest AI minds in our world today on how you can leverage AI in ethical ways to advance your career and grow your business.

3:13Thank you so much, Isaf, and it's a pleasure to meet everyone here. Awesome. Let's dive in. Listen, I'm personally really excited about this because I'm a data freak. I try to run all my businesses based on data and knowing how to do better research faster is a goldmine. So the stage is yours. Yeah, definitely. I think when I first used ChatGPT in December, the first use case, the very first use case I do was to use it for research. It was quite disappointing at that time because that's the first time I learned concepts of hallucinations because when you learn when you ask chargpt to come up with industry insights statistics you will get that in minutes in seconds and then at that time there's no chargpt did not could not connect to the internet so it comes it just fabricated all the data all the insights so i come to the conclusion that okay maybe you can't use it for research purpose and then yeah Then my view of Chachi BG has also been changing along the way over the past 18 months.

4:24And I think fast forward to today, I think the general AI tools are more than ready to handle any type of research task. Because number one, all the tools are, because number one, people are more aware of these issues. So when we design the prompts, when we write the prompts, we know what kind of prompts to craft so that the LIM, so that the AI tools will be less likely to come up with hallucinations so that people are more skilled in handling the AI tools. The second reason is that the large language models themselves do not produce knowledge, but they can connect to a knowledge base. So that's why there is a RIG approach.

5:10Then there's also most of the LRMs nowadays can connect to the internet. So you can pull data from the internet, like how you do the search on complexity. Then it gives more credibility to the data, to the insights you get from generative AI tools. So I think at the moment, this is actually the best moment for everyone to dive in this aspect, the research tasks. No matter whether you are a research professional or you are just a professional who needs some daily research, now is the best time to dive in. I agree. I will highlight two things for people who are new to the show. So if you haven't been listening to the show for a while, there's two things that Tianyu said that you may not be familiar with.

5:54One is RAG or RAG, which is Retrieval Augmented Generation, which is a mouthful. But what it means is you can give the large language model or an AI model the data. So you can give it access to documents, to folders, to whatever data you want, and then it knows how to answer based on that data. So this is the holy grail because then it learns only from your documentation and then it doesn't make stuff up or gets random information off the internet that may not be what you're looking for. So this is number one. And the other thing he said is perplexity. Perplexity is if Google and Chachapiti had a baby, that's what you would get.

6:30So it's a large language model that is connected to the internet and can be used to give you real answers versus search results. So what Google tried to do in the last two weeks, and by the way, now they turned it down almost back to zero because it wasn't great. But Perplexity got it right. It's an amazing tool. I use it every day, probably more than I use Google at this point. And it's a tool that can help you do research. And I'm sure we're going to talk about this later in this episode. Yeah, definitely. And I would also like to elaborate a bit more on the possibilities and the risks as well, because for the possibilities, we know all the tools are ready.

7:08And so to be specific, the tools are more than ready for you to convert anything from raw data to visualization to insights. Before, you may need advanced Excel skills or Python, or you need SQL to query the database. But now, as long as you have the raw data, you can even use ChatGPT to analyze the raw data and get the insights within a few minutes. It's the same case for perplexity. So basically, what perplexity does is that it understands your query, and then it separates your query into multiple queries, multiple Google searches. Imagine you can do 100 Google searches in one go, and then the AI will.

7:58So depending on what AI model you are connecting complexity with, basically it summarizes all the 100 search results into an article, into a report ready for you. So that's the best thing about using Jai for market research at the moment. Yeah, and I'll say one more thing. when it gives you that summary, and it's a beautiful description, how you describe perplexity, but when it gives you that summary, that article about the 100 search results that it read, it also gives you citations of which article with actual links. This is where I find this information. This is where I found that information.

8:37So you can go and check two things. A, that it's respectable sources and not Jane's blog, which, okay, maybe Jane is awesome, but I don't no. And so you can see that it comes from reputable sources, but also you can actually click on the link and see the actual source and see A, more information and B, that it makes sense. And so you can dramatically reduce the chances that the data you're using is crap or hallucinations, just because you can go back to the source and check it. Yeah, absolutely. And I remember a meme I saw last year. So they were saying, so all the memes says that the job of micro-research was, so you spend 80 % of the time doing primary research, analyzing data, and then 20 % of the time doing the report writing and the final checks.

9:27And then after using ChartGPT, you spend 20 % of the time doing the research, collecting data, writing report, and 80 % of your time doing the verification, fact check. So that has been the challenge last year. Only until recently, because of perplexity and because of the improved GPT-4-0 on chat GPT, we will spend less time on fact-check. It is still needed, of course, but we will spend less time on these kind of tasks. Awesome. So let's dive into the practical aspects of this. What kind of use cases are possible, which tools, and how you do it? i can give you one example let's say we if we want to analyze the some raw data from scratch yeah and our goal so there are a lot of use cases but maybe we i will just focus on one use case which is writing writing a report and then analyzing the the property market in singapore okay so i already have the raw data downloaded because in in singapore a lot of the a lot of the data a lot of the public data are readily available and they are the quality is one of the best because you can just download data from the website and then plug into chat gpt and then visualize everything and analyze the data immediately you don't need to clean the data because it's already well structured and cleaned so maybe so today i would like to show you how to how do we convert the hundreds of thousands of real estate transactions into charts and then into insights and then into a report in a few minutes.

11:13And as Tianyu is pulling the data on the screen, for those of you listening, we're going to describe everything that we're looking on the screen. If you want to also see the data, you can also go to the Multiply YouTube channel. So that's Multiply instead of Y with an AI in the end. And so if you do that, you will be able to also see this, but you can stay with us if you're driving or walking your dog or doing the dishes and you can't watch YouTube right now, then stay with us. We're going to describe everything that's on the screen. I will also generalize one thing, which is this is the example that Tianyu has is obviously his data that is public that he can share, but you have such data from multiple sources in your business.

11:55This could be your financial reports. This could be your sales data. This could be your customer service data. We all have tons of this data that we usually try to do in Excel on our own, but sometimes it's very hard because it's a huge amount of data. One thing that I will say before we get started, and if it's something you want to say, I apologize if I'm stealing this from you, but the data, as humans, we usually try the data to be easy for us to read, meaning we add gaps in the data. Like between every month to the other month, we do a space of one column. Between every year and other year, we do two column spaces.

12:32Between rows of departments, we do two spaces of rows. And it completely confuses the AI when it's trying to analyze the data. And it's trying to fix it. And sometimes it will ruin the data as it's trying to fix it. So before you upload the file to chat GPT in this case, but it's the same exact thing if you do it with Gemini, remove all the human stuff. like no spaces, no gaps, no fancy stuff, just the raw data with the right headers and the right columns in place, that's going to make it a lot easier. And the other thing that I've learned, and actually that's more of a question to you, but I can tell you what I found is when I load CSV files versus XLS files, I get better results.

13:13So this is coming from me. I'm not 100 % sure. I'm sure you've done this more than I did, but I'm wondering if you've got the same stuff. So on the point about CSV files, Generally speaking, CSV files are smaller. And then if you analyze data with Python, it's more efficient analyzing CSV instead of Excel. I think because of Python, not because of ChatGPT. Because of Python, it's easier to analyze CSV file. And again, those of you who don't know what Python is, is a programming language that existed for a very long time that was used mostly, not only, but mostly for data analysis by data scientists.

13:50and now both ChatGPT and Gemini write Python code when you give it these kind of tasks. And that's why it's critical. And that's why we can do all these things because it's not ChatGPT trying to figure out. It's literally writing a piece of software to help it do it. Yeah, absolutely. And remember when ChatGPT first came out, there were hallucinations, there were problems with math. And then later on, it becomes smarter. So instead of giving you an answer about math question, ChatGPT will just pull out. We'll just use Python to do the calculation and then return you the accurate result. So in that sense, we are not using ChatGPT to analyze the CSV file.

14:30We are using ChatGPT to instruct Python to analyze the CSV file and then return the insights analyzed by ChatGPT again and then deliver the insights to us. So that's the whole process. Okay, so let's down in. So we have ChatGPT 4.0 open. What do we do next? for all chat gpt4 all is the default i'm not sure about your experience in my experience it's it's faster but it's not as smart as gpt4 it's getting it's quite lazy these days and it just it's easier to it's easier for you to get to get errors in python than gpt4 but since but i think because because of the the timing today uh we will just do it on gpt4 all and then i will break down the long prompt into smaller requests okay in in many cases you can use a super long prompt on chat gpt for example i just uploaded this file this is the real estate this is the resale flat prices of the public housing in singapore from 2017 to 2024 and every single transaction is there so there are about 200 000 transactions a very detailed transaction it's a big file yeah public data is impossible to manipulate on Excel, but we have it here.

15:49Normally, what I would do is to write a long prompt and ask ChatGPT to clean the data and identify the structure, describe the data, and then create the charts based on what I want, and then summarize, create an executive summary. So that is one prompt. But that prompt, the long prompt may also have problems because especially for GPT-4-0, it tends to jump to the conclusion very quickly. So it will always end up with errors or hallucinations. And that's why today I will break down the long prompt into steps. Each step, I will only do one thing. Yeah. I will add something to that. In general, best practices is to break this down into steps instead of giving it really long prompts for two reasons.

16:36One, as you mentioned, it tends to be more accurate on everything that you do, by the way, not just on the analysis, but the other thing, it provides you more control. Because you can see in step one, if it's doing what you intended it to do, and if it really understood the task before it continues to step two and step three and so on. And if it did not do step one correctly, you can go back and correct it and then only then continue to step two. So you would get better results and you would get more control over the process to have exactly what you wanted. And so it's recommended in general, and even by the companies themselves, if you go to the best practices from Cloud or from OpenAI or from Google, they're all saying the same thing.

17:17It's a really complex process, break this down into a step-by-step set of instructions. Yeah, absolutely. So right now, I'll just ask ChatGPT to describe the file and give me data sample let's see what's inside this csv file and what it does is to jump off with analyzing and then you can see the python how it's writing the python code so it's pulling so it's pulling the csv file which is uploaded to the sandbox in the inside chat gbt so this the file in the sandbox will be deleted after three hours so there's so it's not restored there forever and And then we can see that. I'll say two things about the last sentence you said, sorry to stop you, is the fact that these files being deleted is good and bad.

18:07It's good because they don't save your data, which is awesome. So you can upload data that is more sensitive than other data that you may put into ChatGPT. It's bad because if you want to come back to the conversation, it doesn't have the data anymore. And then it asks you to re-upload the data and then sometimes it works and sometimes it doesn't. and so it's if you want to do one of these analysis just stay with it and keep using it versus yeah i'll come back after in the afternoon after i'm done with my meetings which happens to me a lot in chats but in these kind of chats again it will ask you i don't have the data anymore please re-upload the file and then sometimes it just doesn't get it right i don't know why but that's just the nature of it i can totally relate to that yeah the biggest problem is that the context window is not long enough so it doesn't capture the entire conversation yeah that's why you feel conversation is longer than maybe 10 messages and then it only let's assume that it only captures the last the latest 10 messages so then it forgets what's the file name and what library is used and what was done before that and then you can't you just you just you just end up with the same loop with the same error again again yeah that's yeah so i agree with you just do if you want to do it to get the job done, just get it done within one chart.

19:24Yeah. Yeah. Yeah. So right now we have the, yeah, we have a preview. We have a preview of the CSV file, which contains the month, town, and flat type, two room, three room, block number, street name, story range, 10 to 12, first one to three, and then floor area in square meters, flat model, least common state, remaining lease, and the retail resale price. So basically this tells you every single transaction of the resale flat in Singapore from the past seven years. And then China GPT also returns the first five rows of the data set in the form of a plain table here. And then we can start the analysis.

20:13For example, I want to see an overall trend. show the overall trend of the housing of the resale price in the whole period

20:30in the in the line chart for example

20:37and these kind of requests are very simple so it's so you will just create the line chart that shows the average sale price over the past seven years. Over 200 ,000 rows of data. Again, you cannot do that in Excel. It's just not going to work. Well, that's impossible. Yeah. One thing I like about the latest update about ChartGPT is the interactive chart. So if you notice on this chart, there's a button to expand the chart. So if we expand the chart, we just landed in the page. I don't know whether they have a name for it. We just landed in the page of the chart. So on the left-hand side, we can see the chart in the full screen.

21:22And on the right-hand side, we can continue the conversation. This chart is adjustable. We can change the color of the line, or we can even change the type of the chart. Yeah. Yeah. So those, again, those of you who are not watching, we now have a screen where instead of the regular screen of ChatGPT, when you expand the charts, you get about three quarters of the screen is the chart. And then just the sidebar on the right, that is the continuation of the chat, where you can continue manipulating, cutting, slicing, whatever you want to do with the data in that. And that's extremely helpful, but you can also click on different sections in the chart itself because it's fully interactive and change stuff in the chart.

22:12And that's magic. This is probably the easiest user interface to change charts that I know of. It's even easier to do it in Excel or Google Sheets. Yeah. I forgot to mention, as soon as you upload the CSV file to ChatGPT, the fire itself becomes an interactive table. So you can click to expand the table from the fire name, and then you will be able to see the entire table. Can you imagine how many 200 ,000 rows of data all in the same table? Yeah. And I will say something cool about this. When Tianu says the chart, the table is interactive, it's interactive, meaning you can click and select several lines of data, just like you do in Excel and say, I want to create a chart from this and it will.

23:01And so like you're doing in Excel, only you don't have to know what the functions are and how to create tables. You just ask what you want to ask, but you don't have to use all the table. You can usually, you can really be selective just like you are in Google Sheets or Excel. You can select columns. You can select specific cells. You can ask it to create whatever you wanted to ask and you can do it. And it's amazing. Yes. So for them, I just selected two two columns. One is the least command state. The other is resale price and asking Chachapiti to show a histogram of resale prices. So it essentially only takes, just took the two columns and then created the histogram.

23:43And so one more thing that I want to say, because it's really, you and I are thinking geek about this for a week now, but one of the amazing things is we know as humans that are not data scientists or research experts, we know three kinds of graphs, right? A line graph, a bar chart, and a pie chart. That's probably 90 % of what we do. The reality is there's 20 different kinds of charts that we don't know. And what I do sometimes now, because this tool is so amazing, I'm like, okay, this is the data that I have. What would be the best graphic representation for to show this? And they're like, oh, scatter plot trend, blah, blah, blah.

24:20I don't even know how to call it and it will generate it for him. Oh my God, this is amazing. I didn't even know, I didn't even know this thing existed. But if you don't know what is the best way to visualize your data, the data, and you know what you're trying to visualize, you can literally ask, and it can give you one or two or three ideas. And then you can literally ask it to make it and it will. So the Systagram, again, you're not looking at it if you're just listening to the podcast, is an amazing graphical representation for what we're looking at, which I wouldn't have thought about? Yeah.

24:49Let's follow what you said. We, assuming that we don't know what we want, we can just ask ChatGPT, show me, maybe act as a data expert, data analyst.

25:03Identify top three insights about this, from this data set.

25:12and generate three best charts that represent the insights be creative okay yeah we can if we don't know what to look for we can just ask chat gpt yeah to give us a top insight top line insights from this data set and generate the charts based on the insights yeah it's much easier in this way i tried the same thing on gemini but gemini is not as smart as charge gpt because you have to tell gemini exactly what you want otherwise you'll get nothing all right so we have something very interesting here so first we have the average resale price by flat type, you can see that obviously the number of rooms correlates with the resale price.

26:11And then we have the retail resale price trends in different towns. In the line chart, you can see some towns, some of the towns are definitely going through more higher growth than the others. I want to pause you just for one second on this particular chart. So what, again, you need to remember this is a machine. So in this particular case of comparing different towns, there's 50 of them on the chart. So think about 50 lines on the same chart. You can barely see anything, but the cool thing now you can pick that charts and tell it, show it to me for just the top, the three cities with the fastest growth.

26:48And then it will basically pick out of the whole list the three cities with the fastest growth and it will do a completely different chart by just asking for what you want so you don't need to be a data scientist you don't need to go through now how do you even decide which one has the fastest growth right because you have to go through each and every one of them and look year by year and so on you just ask the question and it will do it and Tianyu as I was saying it did it and we're going to get the updated chart yes okay I think there's a there's there's still something that we forget to do at the beginning which is the you still need to instruct chart gpt or what what metric do you use for the measurement for example we are looking at so chart gpt is taking average resale price of the entire apartment as the measurement but in most cases in various states you talk we talk about price per square meter or price per square feet yeah so i i can ask chart gpt if you think that okay in your market price per square feet is the right measurement then we can ask chartgpd to to calculate to recalculate price per square feet and then create new charts based on this for example i want to show let's show let's say show me the show me show me top five towns by the average price per square feet

28:18in the line chart. And here, I want to show you that it's going to make a mistake.

28:27So you've tried this before. I tried this. Or you have a crystal ball, one of the two. Both are going to be cool. Yeah, I want to show you the mistake because sometimes if you do not be specific in your prompt, it may just assume the metrics. For example, one of the columns is floor area in square meter. And then we have another one, which is price per square meter. Okay. Oh, okay. Sorry. I think this time I guess it's right. Last time when I tried the same prompt, it forgets to convert the square meter to square feet. But now, according to the Python code, it's converting the square meter to square feet by times.