Speaker 1: I tested three data AI tools so you can find the best one for your research and I wanted to know the limitations of each. So if we look at the sort of data that I put in, I put in this data initially and this is public healthcare data. You can see it's got a very sort of like specific layout. There's no metadata. It's all just very simply sort of like laid out which is perfect for AI tools. I put that in and I wanted to know what information each tool would give me based on this data set. So I put it in and this is what I came up with. So I asked Julius first of all, this is public healthcare data, provide some insights into what this data shows, include graphs or other visualizations that you think will help. Then it went away and it started looking at all of the information. You can see it's got all of the Python code that it used. It's got the simple language summary of what that Python code does. Then it started spitting out the visualizations. Here's the Python code and then this is what I'm interested in. It's got distribution of hospital codes. It's got distribution of admission types. It's got distribution of severity of illness, distribution of stay lengths and that was it. But you can see that it did its best at sort of like just coming up with those first little insights into that data. Then I did exactly the same thing with Visley. Visley's got a very similar layout. You put your prompt here. You put the files that you want explored here and this is what it came up with. So here, this is public healthcare data, exactly the same prompt and then I put it in. You can see it's got the code here and this is what it put out. It said distribution of hospital types. It's got distribution of hospital regions. So it's chosen slightly different information but importantly, it's starting that process and then one thing that you'll see is that it's given me also a summary of the public healthcare data analysis and then I went on and asked it more information. I did exactly the same with ChatGPT, the new 4.0 and this is what happened. Blah, blah, blah, blah, blah, scroll to the top. So exactly the same prompt here and it first of all gave me the information. So it said this dataset contains public healthcare data. Then it sort of looked at the different columns and then it sort of had an analysis plan and this is what I really like. It said this is what I'm going to do and then it started that visualization process. So here you can see that it spat out a load of different graphs and I can click on these and get them blown up and see what they're like. So it's got distribution of hospital types, distribution of hospital regions, all of that stuff. So it did a really good job initially and the one thing I like about this is this, the interactive graphs. So if you pop this out, you can see that you can start interacting with these graphs on ChatGPT, not just sort of like the images but here you can scroll over and get the actual information. That's one thing I really like about ChatGPT and I don't think any of the other ones had it. You can download the information but these aren't interactive in the same way that ChatGPT is. Then I decided that I wanted to ask a little bit deeper. So whenever I'm looking at data, I'm looking always at this raw data and I'm like, okay, what two things do I want to kind of mash together and get the new insights for? So I asked it, could you provide a breakdown of the distribution of hospital stays by duration and I wanted to know how easily each one of these tools was able to pick up on that nuance. And I'm pleased to say that all of them did a pretty good job. So here is the distribution of hospital stays. So you can see we've got stay duration and the count at the side here. So you can see most people, according to this graph, was 21 to 30 days and that's the histogram that it's kind of like put it in, which is great. And then it says this bar chart provides a clear overview, blah, blah, blah. And that was it. So then I did exactly the same thing here and you can see that it's chosen different buckets to put the data in. So here it looks a little bit different. I actually prefer Visly because you can actually interact with this graph. So I think the graph looks a little bit nicer. It's got a few of these different kind of tools on the side here. And you, yeah, like I said, you can sort of zoom in and out. You can interact with it. You can see once you hover over it, you get this, I don't know why I'm pointing at the screen. That doesn't help you at all. I did exactly the same thing here. And once again, you get that interactive graph, something that Julius doesn't have at the moment, but I'm sure that after this video, they'll put it in. You could see that it sort of like chose different buckets. I don't really like this one, 21 to 30, it doesn't do the lower days or lower stays like this one. You know, this starts from zero to 10. So I think this is a win for Visly over the other two at the moment. But this wasn't the end of the data. I actually have got from my PhD some of this data. It's a text file and it's the IV curves. So you can determine the efficiency of solar cells. And so you can see that this is really unstructured data. We've got metadata. We've got this metadata. We've got the performance parameters that are actually calculated by the software that kicked this out. And you've got the raw data down here. And I wanted to know how well each of these tools manages to deal with this kind of unstructured raw data that would sort of like be spat out by a machine. Like that. That's how they spit it out. They're like, now you deal with it. So I went over to Julius and I put it in. So this was the calculating efficiency. And I said, this is an IV curve from an OPV. That's an organic photovoltaic device. That's what my PhD was in. Can you plot it and calculate the efficiency? So the one thing I liked about this is that initially it started doing this thing. And one thing I love about Julius AI is it self-corrects. It's like, this didn't work. Don't worry. I've got this. And it just sort of like re-evaluates. So there it didn't do too well, but it does re-correct. And you can see that it did actually manage to get the right data even though it was buried under a load of metadata that I really like. So this is all its reasoning. And I love that it was actually looking for the right stuff. And then here we are. This is what I wanted. This only took a matter of seconds. And it says, here is the IV curve of the device. And this is exactly what I'm interested in seeing. You can see it's a nice IV curve. The scale is appropriate for the values that I gave it. And it then says, let's calculate the efficiency of the OPV device, which I wanted. Let's go to Visley and see what happened with that. So here I was asking it for exactly the same thing, but Visley kind of struggled. So this is an IV curve from an OPV device. Can you plot it and calculate the efficiency? And then there was all these issues like show output error. There was an error there. It was looking for different files, which it couldn't find. And so it just kind of like gave up, right? It just said like, it looks like there's an issue with the data file. We need to inspect. And I said, continue the analysis, because I wanted to see if it could push through that. And it did. It did. I was so amazed that just this simple prompt, I was able to then just say, you know, have a look through. And it says, let's inspect the few lines to understand its structure. And it appears the file contains metadata, great, and then the data, blah, blah, still causing issues. So it looked deeper. Then it said, ah, okay, it seems like the data still contains metadata and performance parameters, which are not part of the IV curve. We need to skip these sections. So it's starting to reason with itself, and this is the power of AI. You can just let it do its thing, and eventually it kind of self-corrects. These AI tools are getting more and more powerful. I absolutely love it. So it finally finds the information that it needs, which is this bit, the IV curves. And then it kind of stopped there. And then I said, okay, can you plot the IV curve? And no, that's not the curve I want. That is a straight line. I don't even know what it pulled to get that stuff. But there we are. There's the straight line. And I don't understand why it is like that, but no, that is not right. So I stopped there with that. I was like, you know what, Visley, you clearly can do this. Let's see what ChatGPT could do. So this is what I was interested in. It says, this is an IV curve. Can you plot it and calculate the efficiency? Now, importantly, one thing that hasn't happened to this point is plotting and calculating at the same time. That is about to change. Spoiler alert. So, first of all, it says the provided file contains IV curves. It didn't have to go through all of that kind of like backwards and forwards and looking at the data and extracting the metadata. It just went straight for the IV curve because it kind of has a better reasoning of what's in a file once it's looked at it. That IV curve is perfect. That's just what I want. I could copy and paste that into a presentation. Brilliant. Now, this is the thing. Now let's calculate the efficiency. It went on to calculate the efficiency all on its own, which I really liked. So it said, this is the efficiency. So from the metadata, so it knows it's got metadata and it's kind of like holding that in its memory, which is brilliant. It says the efficiency is already given as that. However, I can confirm this by recalculating. Yes, JGPT. That's what I want. I want you to reconfirm. I don't just want you to use the metadata even though it's in the metadata. I want you to double check my stuff for me and it knows the efficiency for the formula. That was the wrong way around. It knows the formula for the efficiency. That's how shocked I am that it worked. So it gives me the formula, which is great. It shows me what's in it and then it says, let's calculate it, recalculate it. It's approximately 3.149, which closely matches the 3.1488% confirms, blah, blah, blah. Yes, it did everything I wanted. A massive tick, tick, bing, bo-ding, bring, for ChatGPT because it did exactly what I wanted in the two steps and this is using ChatGPT 4. This isn't even using the other newer version. So incredible. I was very impressed. Julius AI, on the other hand, it says, next, let's calculate the efficiency of the solar cell. It tried its best and you can see that it pulled out the performance parameters and put it into its own format, but then it said, let me know how I should continue. And I said, well, calculate the efficiency of the OPV device and it says the efficiency of the OPV device has already been calculated as 3.15. So it used the metadata and it didn't double check it, which isn't a problem if the raw data coming in and the metadata actually works and it actually makes sense. But I do like that ChatGPT had that extra little check. Lovely. Now, I wanted to break these things. I wanted to see how far they could go. So it works well with text. You can put in your stuff in there and the sorts of insights you could get just by exploring the data using AI. I think Julius AI and ChatGPT are my favorite at the moment, but I wanted to see if I could break these things. This is an image of silver nanowires, but that's not silver, that's an H, nanowires and single walled carbon nanotubes. Can you analyze the image for me? I actually put in the original TIFF file. It didn't like that, so I put then in a JPEG and it liked that much better. And so you can see it uploaded it. And the one thing I liked is it did try its best to give me information about that file. So it says here, silver nanowires, appearance, they're thicker, the distribution distributed in a random mesh-like pattern, orientation, single walled carbon nanotubes. So it does pick out the different things. It knows that this is single walled carbon nanotubes and this is a silver nanowire. So I'm very impressed that it was able to do that and notable features, intersections, surface morphology. These are things I definitely talked about in my papers and my presentations when I was looking at this. And then it said, if you need any further information, let me know. And I said, what is the average diameter of the silver nanowires in this image? And it says, this is what I need to do. It didn't do it, but it did give me edge detection, which I could have used. So when I did this, like initially for my PhD and postdoc, I would go in and measure the bar, the size bar, what do you call it, the scale bar, that's it. And then I'd measure the silver nanowires and that's how I'd do it. But here it's actually done it a little bit better, I think, and it's used edge detection. So it allows me then to have a look at this and say, you know what, these are the edges that I'm going to measure across. And it gives it something incredibly kind of more robust than me just saying, well, I'll just pick some carbon nanotubes and silver nanowires to measure. So I do like that. I can use it, but it didn't give me an answer, which I wasn't expecting it to anyway. Next thing is, I asked Visley, Visley, the same thing, and this is what happened. It did edge detection and then it gave up. But here it was strange. I don't understand what it did for me. Maybe someone can help me here. But I said, what is the average diameter of the silver nanowires in this image? And it said 32.3 or 32.4. It says 32.4 units. So I think maybe it's using the number of pixels, but I'm not quite sure. It hasn't used the scale bar, but nonetheless, it has done this for me, which I really like because I could have used that in my PhD and said, I used a fancy tool to edge detect the silver nanowires. You can look at the roughness here as well. You could compare that to silver nanowires without carbon nanotubes on it, that sort of stuff. You can do all sorts of extraction of data for papers. And then chatGPT, I asked exactly the same thing of chatGPT, and it gave me some analysis on structure and morphology, interactions, and then I also said, what is the average diameter? And it said, well, you'd have to use something else, which is absolutely fine. I get it. I was trying to find your limits, and I'm now asking you to do something you can't do, which is absolutely fine. But this is 500 nanometers, so you'd easily be able to work out the diameters using anything but AI. Here it says you can use Fiji or ImageJ. I actually used ImageJ during my PhD, so it is providing actual valuable information. So there we are. Those are the three tools. For my data, I would be sticking using JuliusAI and chatGPT. I think those together are a great tool set for all of the things you want to do with your data. So I would be using JuliusAI, and then anything I couldn't do with that, I'd be heading over to chatGPT. The data analysis has never been so easy. If you like this video, check out this one where I go into detail about JuliusAI and how to use it. I think you'll love it.