6 Ways ChatGPT Code Interpreter Is Already Being Used

0:00On this episode of the AI Breakdown, we're covering the new ChatGPT plugin, CodeInterprepreter, which has the AI community on fire. Before that, however, headlines from Meta, Microsoft, and Google DeepMind, whose CEO says we could see AGI within the next few years. This AI breakdown was originally released on YouTube.

0:24Welcome back to the AI breakdown brief, all the AI headlines you need in five minutes or less. Today, we start with Microsoft Bing getting a huge number of upgrades around its chatbot. And I got to tell you, I'm still not over caring about what Microsoft is doing, but here we are. In short, Microsoft has announced a set of Bing chat plugins. Now, this follows from ChatGPT doing something similar. In fact, later on the main breakdown episode, we're going to talk a lot about one ChatGPT plugin called Code Interpreter. But anyways, Bing has followed suit and is offering all sorts of new ways for partners such as OpenTable and Kayak and things like that to integrate with Bing Chat so that you can use that same chat GPT style interface or chatbot interface to go out and do the things that you would normally do.

1:11Other updates include multimodal answers, so answers that include visuals, charts, graphs, and other non-text formats. On top of that, they're also trying to add in some UI upgrades, such as chat staying contemporary as you click around the web. Lastly, there is no longer a waitlist for Bing Chat. That means anyone who wants it can get it. Now, a lot of folks have pointed out what Pete from the Neuron here does. ChatGPT and Bing Chat keep colliding, and it's unclear how this will end. Why pay for ChatGPT Plus when Bing Chat also uses GPT-4? If you're paying for Plus, why use Bing Chat when ChatGPT will have browsing soon?

1:45And now both have plugins. Is it a battle or just an eventual convergence? We'll have to wait and see. Speaking of ChatGPT, Meta is warning about ChatGPT imposters, and this feels like it was absolutely inevitable. Their latest security report says that there's a lot of malware that's now masquerading as ChatGPT or Bing Chat. So if you are seeing an advertisement or just a promotion of some new chat interface that uses these terms be extremely cautious. Next, we turn to some updates for AI developers. A few days ago, McKay Wrigley wrote, Replit is a sleeping AI giant, unbelievably excited to see how their new custom code models perform.

2:24They're already experimenting with stuff like auto app generation with AI agents. It was already the IDE of the future, but it might be the AI company of the future too. That pronouncement got a little bit more clout today as Replit announced that they developed their own new LLM. This is meant to supercharge the development process and it's being released under a commercial-ready open-source license. Now, on top of that, a company called Modular just released a programming language called Mojo. It's being described as Python-like, and it's specifically designed for AI applications. Mark Tenenholz here says, What programming language will run AGI?

2:57Well, it might be Mojo. He says it's truly paralyzable, fully compatible with all Python packages, and has speed boosts from types but not required. I'm seeing a ton of chatter about Mojo, even though it has just come out. Speaking of AGI, a new interview with Google DeepMind CEO Demis Hassabis is just out, and he says that we could see some form of AGI within the next few years. He says the progress in the last few years has been pretty incredible. I don't see any reason why that progress is going to slow down. I think it may even accelerate. So I think we could be just a few years, maybe within a decade away.

3:31Now, Demis does say that we should be a little wary, saying I would advocate developing these types of AGI technologies in a cautious manner using the scientific method. where you try and do very careful controlled experiments to understand what the underlying system does. The challenge is whether that's of course possible in a world where every company is competing to get these new technologies out faster than their peers. That was one of the concerns expressed by Jeffrey Hinton when he left Google and gave his first set of interviews a couple days ago. Finally, speaking of concerns, there is a meeting today at the White House between Vice President Kamala Harris and the CEOs of companies like OpenAI, Google, Microsoft, and Anthropic.

4:07So we will see what comes out of that and whether it sets a path in terms of how the US might think about regulating or at least engaging with AI going forward. That's it for today's brief, but stay tuned for today's AI breakdown. This is the most excited I've seen the AI community since AutoGPT first showed up on the scene, and it's because of ChatGPT's new plugin, Code Interpreter. All right, so the main thing I want to do in this video is show you six ways that I'm seeing people use Code Interpreter in just the first few days of it being available. That includes data analysis and visualization, academic research, support for development projects, visualizations that are really interesting and fun, basic video editing even, and an approach to data analysis that leads to business strategy.

4:52Before we get there, I just want to do a brief intro on what ChatGPT plugins are and what Code Interpreter is specifically. So ChatGPT plugins are effectively the way that third parties can use ChatGPT and build their functionality into it. So you see here a set of the plugins that they launched with, Instacart, order from your favorite local grocery stores, OpenTable, provide restaurant recommendations with a direct link to book, right? So you can do these things from within the ChatGPT experience. But ChatGPT itself, or OpenAI, created a couple of plugins itself, and one of them was called Code Interpreter.

5:30Effectively, what Code Interpreter allows you to do is to upload some set of data and then have ChatGPT spit out analysis or visualization. And that visualization part is really big. Obviously, that's a really new capacity. So let's move now to the six use cases that I've seen, starting again with data visualization. This one has gotten a ton of attention. John Backus uploaded a CSV of San Francisco crime data and asked ChatGPT to visualize trends. So you can see number of incidents per month over time. You can see crime hotspots in San Francisco. You can see what it suggested they should do with analysis of the data set.

6:11So John says, give me 10 ideas of trends, visualizations, and analyses I could do. So you're not even required to know what analysis or visualization you want. What ChatGPT came up with was day of the week analysis, hourly crime trends, seasonal crime trends, police district analysis, crime resolution rates, top crime categories, crime category trends, crime clustering, comparing crime types by location, correlation analysis, right? Like really interesting stuff. You see here that it did some of that visualization, that number of incidents by day of the week, hourly crime trends, et cetera, et cetera.

6:47So this is sort of a pure play version of this data analysis and visualization that I think is at the heart of Code Interpreter and why people are so excited. Now move over to Ethan Mollick, who is a professor at Wharton, and he says academia is in for a wild ride. What he did just as an experiment was upload census data and a data dictionary into GPT with Code Interpreter. He asked it then, I would like you to generate some interesting draft hypotheses about industries and metropolitan areas and then to test them with the data. Make assumptions if you need to, put it in a paper. So you can't really read it probably from where you are, but it says title regional dynamics of industry characteristics, a comprehensive examination of payroll, employment, and establishments across metropolitan and micropolitan areas.

7:36So effectively, this is taking census data and actually not just analyzing it, but thinking about what is interesting about the data, what comparisons might be worthy of further elucidation or study, and then turning that into a paper which is supported with visualizations. You see here, linear regression, total number of employees versus total annual payroll across metropolitan areas, and another correlation between total number of employees and total annual payroll across MSAs. Now, Ethan says, it's obviously not a top journal paper or anything close, but this took me less than 10 minutes, and I did not do any work to find an interesting dataset or guide the AI in any way.

8:16If nothing else, academic publishing is about to be overwhelmed. Next up, we have coding support. Pietro here says, chat GPT code interpreter is incredible. Here it extracts colors from an image to create a palette. You can see here that Pietro uploads an image that has a set of colors in it. And code interpreter looks at that, goes through a number of steps and eventually spit out a palette that he can now use in the development work that he's doing. A fourth category I'm calling visualizations, but really interesting. And this is the show that this isn't just academic. Ethan again here says this was kind of delightful.

8:54I uploaded a CSV file of every lighthouse location in the US. ChatGPT code interpreter create a GIF of a map of the lighthouse locations where the map is very dark, but each lighthouse twinkles. A couple of seconds later, he got exactly that GIF. You can see the twinkling coming all around the lighthouse borders. Riley Goodside did something similar with an example that I think many of you will recognize. He writes, make a 512 by 512 GIF with falling green matrix letters. Assume no fonts. 30 frames, five frames per second. No talk, just go. Code interpreter says finished working and has exactly the thing that he asked for, the classic matrix falling letters.

9:32Get ready for a bunch of weird, quirky and nostalgic stuff like this. Riley did another experiment though as well where he did basic video editing in chat GPT converting an uploaded gif to a longer mp4 with slow zoom. Riley says I'll upload a gif and you give me a five second mp4 with a dramatic slow zoom in. No talk just go. Riley loves that no talk just go. Code interpreter says sure upload the file and I'll do it. This is the gif that Riley uploaded and this is the dramatic zoom in that came out on the other end, turning a GIF into an MP4. Now, the final way that someone's using this that I wanted to discuss is actually kind of a synthesis of a number of different parts.

10:16This isn't just data analysis. It isn't just data visualization. It does that, but then it goes to another level by actually layering in suggested business strategy on top of it. So what David Boyle here has uploaded is a set of information. It's an untitled spreadsheet about music market revenues in various countries. What you're seeing on your screen right now is code interpreter figuring out what data is actually in this chart, the country, the population, the total recorded music market revenues, the total physical and digital revenues, the year on year change, et cetera, right? This is just code interpreter figuring out what's in the chart.

10:53Now David says he'd like to cluster the data. He's actually driving this process a little bit. He says, I want to find clusters that are attractive. Here you see it working. And then it starts to come up with these clusters. Cluster zero consists of countries with relatively low total recorded music market revenues. Cluster one is countries with moderate total recorded market revenues. Cluster two consists of the United States, which is the highest total recorded music market revenue. However, ChatGPT also says that the way that it bound things was arbitrary and he could further refine the clustering.

11:24David Boyle says, I want a higher resolution view, perhaps five clusters. And then he says, can you bring each to life to an executive trying to decide where to focus, give each cluster a catchy name, etc. Sure enough, Code Interpreter comes back, the Rising Stars, cluster zero, the Untapped Potential, cluster one, the Budding Performers, cluster two, the Music Superpower, cluster three, and the Consistent Contenders, cluster four. And this is where David starts to translate it into strategy. He says, for each cluster, please write a strategy for a company looking to grow its business in that segment.

11:59Sure enough, a customized strategy for each of those sectors comes up. So what I loved about this example is that it really shows how this isn't just some cool academic analysis, but can be a direct line between data analysis and visualization into strategy, business strategy that sits on top of that. Anyway, I think over the next few weeks, we're probably going to see a ton of very incredible explosion of uses of code interpreter. And I am so excited to see what people do with this and how it changes how we think about data analysis and visualization in general. If you're enjoying the AI breakdown, please subscribe to the channel.

12:34And until next time, peace.

12:50You