EP 43: AI-Powered Lottery Analysis on Powerball Patterns

Transcript for this Episode:

INTRODUCTION VOICEOVER: This is Tech UNMUTED. The podcast of modern collaboration – where we tell the stories of how collaboration tools enable businesses to be more efficient and connected. With your hosts, George Schoenstein and Santi Cuellar. Welcome to Tech UNMUTED.

GEORGE: Welcome to the latest episode of Tech UNMUTED. Today, we're going to take another look at lottery numbers. We've done this previously, got an interesting outcome when we did it. We're going to take a look again. We looked at mega millions the first time. We're going to look at the Powerball numbers this time. Before I dive in on this, please note the data discussed in this podcast is based on historical lottery results and is for entertainment purposes only.

Past lottery trends do not predict future outcomes. Every lottery draw is a game of chance with each number having an equal likelihood of being drawn. We will dig down into that. Again, that's regardless of past data. Do not base your lottery decisions solely on this information. Always play responsibly. Understand the odds of significant wins are very low. We accept no liability for choices made based on our content. If gambling is a concern, please seek help.

SANTI: Now, that that's out of the way. [laughs] It's important to understand. Listen, the lottery podcast we did, it's probably one of the most popular ones we did. It's important, folks, to understand that this is just us trying to test artificial intelligence and see what the outcome is. Don't take this and say, "Well, I'm just going to throw all my money at the next Powerball and win because I heard these guys on podcasts and AI figured it out. No, it's entertainment, but it's more importantly, it is watching how artificial intelligence is able to analyze a large set of data and give us very interesting outputs. That's really what this is about.

GEORGE: Agreed. It's a random set of data. We will come back and do another episode and look at data that is not random. I'm exploring a couple of different options at the moment that relate to census data and other data. I just need to get it structured in a way to use it. Two notes on this. This data came from data.gov. It happens to be New York State that loads this data up into data.gov.

That's where these numbers came from. There was a change in the way they do Powerball. I'm not a big lottery person, but they changed the number of red and white balls in the sequencing of the numbers late in 2015. I deliberately only pulled 2016 forward data. With that, I'm going to launch my screen here and walk through fairly quickly a couple of prompts to both understand the data.

I've loaded an Excel file. Again, this file only contains two columns. It has a date of the draw. It also has the winning numbers. The way that it's set up in the Excel table is it has winning numbers plus the Powerball at the end. There is a way to break that down and analyze the data. In the prompts that I use, I'm going to give those specific prompts so that it understands what the data is.

The reason we're doing this is we had downloaded this data before in the previous one, and there were anomalies in the data and other things in it. It took a little bit for ChatGPT to figure it out. I'm going to take a look right now, and I'm going to say-- Look at the attached file. I'm just going to pop these prompts in here really quickly. Tell me what patterns, anything interesting in the data. The first five numbers are the winning numbers, and the last number is the Powerball. It's analyzing, and it's going to come back with some details on this, see what it comes back with.

SANTI: Again, it's important to point out that you're using ChatGPT-4. This is a paid subscription to ChatGPT-4, correct?

GEORGE: Correct, it is.

SANTI: Which, if you did this in Microsoft Copilot, it would use ChatGPT-4 by default.

GEORGE: Correct.

SANTI: I'd be curious to know what Copilot comes out with. This is ChatGPT-4 that's being used, okay?

GEORGE: There's a set of analyses going on here. I can't necessarily tell you what all this is, but it starts to give a breakdown. This is the initial results. The initial result is, what are the frequency of the main numbers? Then what are the frequencies of the Powerball?

SANTI: Very similar to what we saw the first time we did this test with the Mega Millions, right?

GEORGE: Correct.

SANTI: We got this graph, yes.

GEORGE: Now, you should note, this data set's relatively small. It's two draws a week or something. I'm not sure exactly what it is over the course of 2016 through a week or two ago is what the data set was. It's not a straight line. If this thing ran for a hundred years, you would assume statistically all these would be basically at the same place. You wouldn't be able to distinguish between one number or another, but you do see a couple numbers that are peaked in this, that have higher numbers than the other ones. It did frequency of main numbers, Powerball, we saw that above. Let me put another prompt in here.

SANTI: It's nice because it gives you the frequency in a graph format, then it gives you a short narrative for each one. It's almost like explaining the pattern that you're seeing in the graph, which is really good.

GEORGE: Correct. You could have done this by looking at it, but I'm going to say, what are the top five regular numbers, not the Powerball, that appear most often? What pairs are used most often? The only reason I put that in is the last time we did this, without prompting, it identified pairs, and there were a bunch of pairs that had occurred more often. Again, keep in mind, random data. At some future point, we're going to look at like census data or something else that should not be random. There should be clear trending in it.

SANTI: Yes, that would be good. I agree. There was an error analyzing.

GEORGE: An error analyzing. It might catch it itself. It caught it itself. It's going to go back, and it's going to now look and analyze it in a different way.

SANTI: This is interesting. Let's read what it said. The folks who are listening. It said that there was an error analyzing, and it gave us an explanation as to what the error is. Here's what it says. Can you scroll back up so I can read that? Because I think it's important to see that. It seems I missed importing a necessary function for generating combinations. Let me correct that and proceed with the analysis again. The AI read your prompt, started analyzing, realized that it didn't have the right, I guess, function. It's telling you that it's going to reanalyze it using additional resources. I've never seen that before.

GEORGE: What typically happened in the past when you got a failure like this, it would fail, you could try to regenerate it, and it typically would continue to fail.

SANTI: This is new.

GEORGE: Correct. It did the five most frequent numbers, which are down here. It did the pairs, which it also charted. Hasn't really given us any other significant insights.

SANTI: It gave you a graph, right? A simple graph, a visual that correlates with the narrative it just gave-- Listen, this is all fascinating. I'm actually more fascinated by the fact that it made an error, caught itself making an error, auto-corrected, told you what it did, and then reanalyzed it. I've never seen neither ChatGPT or Copilot do that. This behavior's new to me. I've never seen this. You're right. Usually, it just errors out and moves on. Here, it's just corrective action.

GEORGE: I'm going to start the next prompt, and I'll let it go, which is based on trends, what are the most likely numbers to occur? This is sort of asking what are the likely numbers, and we'll see whether it answers it or not. On that piece, there's a couple of things to keep in mind when you're entering data, and you're going to use ChatGPT to analyze it. This has two columns. I mentioned it earlier. There's no comma separation in the numbers in the table. They literally go 15 space, 11 space, 36 space, something. It's able to go in and analyze that. If you were to try to do this with pivot tables or something else in a spreadsheet, I think you could probably get to the outcome, but you would have to break those fields out.

SANTI: Right. It's more work.

GEORGE: Then somehow do some calculations. Would be very challenging.

SANTI: This is another productivity hack, but listen, something happened while you were speaking. Scroll back up. Let's look at your prompt that you just entered.

GEORGE: The prompt was based on past trends, what are the most likely numbers to occur in the future.

SANTI: Look what it did. It sounds very similar to your notification at the beginning of this podcast, your disclaimer. It gave you a disclaimer. It said predicting future lottery numbers based on past trends, that’s a common desire, but it's important to understand that lotteries like Powerball are designed to be random. It just goes on and on to talk about, "Hey, I'm going to do this but know that it's a game of chance."

GEORGE: It got a little annoyed with me almost here. It's like, I already identified the top five numbers.

SANTI: Yes. It's interesting. It has an attitude. [laughs]

GEORGE: Yes.

SANTI: I’ve never seen this before. I'm seeing a different behavior from ChatGPT-4. Just today alone, one out of correcting something, two snapping back at you saying, I already did this. Why are you asking this information again? All right. This is interesting. Listen, I'm more fascinated by the response than anything else.

GEORGE: Here's one by the way. The next one is what numbers would you pick next? This is just for fun. Humor me. Let's see if I can actually get it to give an answer. It approached it randomly. I have done this same prompt before. That's not the way it did it in the previous ones. It did not do a random pick. It picked some of them based off frequency and then it, and then it picked the other ones based off random.

SANTI: It says, disclaimer for a bit of fun, here's a randomly generated set of numbers. There you go. Very interesting.

GEORGE: I'm going to just do two more prompts here. The next one, let's just verify what we think we've been talking about. Given all the data analyzed and understanding Powerball numbers are random, do you think the numbers look like a random distribution? Because eyeballing the original chart up top, it doesn't look like that. There were a couple that were higher and a couple that were lower.

SANTI: That's a great prompt.

GEORGE: The size of the data set is big but not really big.

SANTI: That's a great prompt. Let's see here. I'm curious. It's a lengthy narrative.

GEORGE: It's a lengthy narrative. When I tested this previously, it actually gave it a deliberate answer and said that the distribution was random. Today, it chose not to do that.

SANTI: It gave us two summaries at the top. What were the two summaries? It summarized into two things. It said variation in frequency, and so it gives you a description of that, and then pairs of numbers, and it gives you a description of that. Okay. See, I thought that it would have taken the numbers that it generated for you and answer that question based on the numbers it generated. That's what I was hoping it would do. I was hoping it would follow the conversation, right?

GEORGE: Yes, agreed.

SANTI: That's what I was hoping. Yes.

GEORGE: I'm going to do one final prompt. I'm going to say, can you create a high-resolution widescreen image to be used for a podcast thumbnail on YouTube that represents this chat session, and that uses bright colors with a hint of technology?

SANTI: Interesting.

GEORGE: This may, in fact, be the image we use. We'll have to take a look and see if we like it.

SANTI: That would be fun, actually.

GEORGE: It's creating the image. You've seen us do other podcasts on image creation, mostly on mid-journey. This is not a super complex prompt. Got another error. We're going to see what happens here.

SANTI: Scroll down. See what it says.

GEORGE: It's fixing itself.

SANTI: It's fixing itself. See, it's the second time that it has an error and auto-corrects itself. I've never seen this before. By the way, this is now DALL·E, right? DALL·E is the OpenAI image creator, whereas ChatGPT is the generating text-- Oh, look. Oh, what a great image. Look at this image. This is great.

GEORGE: Then it gives you a read-back on what it did. It did some different things. It said it made it vibrant and eye-catching design and modified some other things in the original request. It's pretty interesting.

SANTI: That's great. I think we've generated the thumbnail for this podcast. Listen, I'll tell you what I really took away from this podcast today, was the behavior of Chat GPT-4 has changed. I can't quite put my finger on what exactly changed, but there's things I saw today that I've never seen before. The auto-correcting definitely won, but then it snapped back at you. It had a little bit of a condescending talk-back attitude, like saying, "I already did this. Why are you having me do this again?" I haven't seen that before. That's interesting because it's behavior almost. It's really weird.

GEORGE: It is apologetic sometimes as well if you say, "I asked you for this, and you didn't give it to me." It will respond back typically like, "I'm sorry. I missed that. Here's the data analyzed." Most often when that happens, and this goes back to how you use prompt engineering with any LLM, you often get caught in a loop. There is something you're trying to correct, and you are unable to correct it. If I went back on that image we just showed, it had what looked like a timestamp in the upper left-hand corner.

If I tried to take that timestamp off, it's unlikely I could get it to be removed. It would start an iteration of things that even exaggerated it more, and it had a little error in it. The first three numbers look correct. The last one was not really a number. I don't know what it was, right?

SANTI: Right.

GEORGE: There's what you need to understand about prompt engineering. In that case, you usually have to just start over, do a much more complex prompt upfront. That full prompt will likely get you the outcome that you're after.

SANTI: It's a good thing you brought that up, because we do have a podcast where we talk about some best practices for structuring a prompt. If you're listening to this podcast, and we've done a lot of podcasts around prompts lately, but specifically, we're going to talk about how to structure, just some best practice, how to structure a prompt so you can get the best results. Look for that one. George, this was real fun. Every time we do the lottery thing, people get excited because it's just one of those things where, if you play the lottery, you're always trying to find that next set of numbers. Hey, you can have ChatGPT generate your next random set of numbers. Just don't come back and say, "Hey, I lost all my money because of you guys." Because we warned you, it's a game of chance.

GEORGE: It is random.

SANTI: It is random.

GEORGE: Again, we will take a look in a future episode at more structured data that is some census or survey-based data. The difficulty that I've had with it is, it did very well on two columns of data in this analysis. When you put tables or multiple columns in place, and it needs to cross-analyze, it has a much harder time.

SANTI: Interesting.

GEORGE: What I'm trying to do is structure it purely as columns in a spreadsheet, columns in rows. I think it'll be better at analyzing that. One of the data sets I downloaded was government data, it was housing data in certain markets. It was table after table after table. It could not figure it out. It failed. It went back similar to what we saw today and tried two or three different things, and then it gave up.

SANTI: Interesting. I love when we do this because we're putting AI to the test, and folks don't get a chance to see this. Everybody talks about how great of a productivity tool AI is, which it is. It's also good to see how it reaches these conclusions, and not only that, how we can push it and maybe get failures. Because if we can understand where some of these failures are coming from, that helps increase productivity because we know how to get around it.

I love when we do this. Anyway, I think we’ve got to bring this podcast to an end. Folks, this is a good opportunity to subscribe right now to Tech UNMUTED on your favorite podcast platform and even on YouTube. Until next time, remember this, stay curious, stay connected.

GEORGE: Visit fusionconnect.com/techunmuted for show notes and more episodes. Thanks for listening.

CLOSING VOICEOVER: Visit www.fusionconnect.com/techunmuted for show notes and more episodes. Thanks for listening.

Episode Credits:

Produced by: Fusion Connect