In the Mouth of AI Maddeness

I’m not sure how I’m gonna get my hands entirely around this post, but damn it all I’m going to try. It’s been a while since I came at the blog with a long one, but if I don’t start documenting some of the AI Maddeness work I’ll be too far in to find my way out 🙂

Strange thing is despite the post count this month on the blog, I’m writing a lot. But I’m writing into a text box postmarked to oblivion. I’ve become fascinated with what it might spit out. It’s often strangely wrong and oddly glitchy, but the generic beats it hits and the sense of over-the-topness it deems a personality where there’s none is mesmerizing. It’s all very confounding, I’ll be the first to admit, but I’ve come to the machine on a mission. I’m employing it to take on the personas of stock sports writers, fans, and a strange assortment of characters that populate an imagined universe. It’s become particularly fun, at least for me, but it might be worth taking a step back to explain what the hell I’m talking about.

Looking for an excuse to explore AI, I decided to play an entire season of the 25 year old video game Madden 2001 and use AI tools to create a universe around a single season of video game football. I imagined it would be very rudimentary to start (which it definitely is) and build over 17 to 20 weeks (20 assumes playoffs and more) to a more fine-tuned content production machine standing in for localized sports media.

What’s more, thanks to Tom Woodward, we’ve automated the bots that populate the discussion in that universe using an API call to ChatGPT that’s tied to unique character profiles that use those parameters to comment on blog posts (there are no social media bots yet, but I imagine that’s in the future). The average bot leaves 1 to 2 comments a post, but after a few posts I’m realizing we need to add more commentator bot profiles so that we can have fewer repeaters. Tom blogged how he made it happen, and I appreciate as always his practical approach to making all the cool things associated with this project happen—funny how great things materialize when you leave Non-Programmistan.

AI Commenting Chaos Incarnate

So, the thrust of the project is playing and recording a game of Madden 2001 each week throughout the season as the NY Jets. This is the real reason why I’m doing this, to be clear. I’m looking for an excuse to play this video game and somehow have it matter. One of the ways I’ve justified this amazing time sink is to simultaneously play with AI given Madden 2001 was that game in the early 2000s that I first saw AI called out as part of the settings interface. It is also the first game I encountered wherein the complex sets of rules and impressive attention to detail helped bring the NFL to life as a videogame.

These factors made it feel like a new breed of sports game and it was really appealing to get high and play it for endless hours when it came out. According to legend, John Madden would not put his name on anything that didn’t capture the authenticity of an 11-on-11 game. Tiburon Entertainment met that benchmark during the late 80s and early 90s, and by 1998 they’d been bought and brought into the EA empire. As a result an iconic game franchise was born.

So, as you can imagine, playing this game on the Playstation in 2000 was an experience. As mentioned above, one of the things that struck me was how the menu integrates the ability to customize the AI for the game.

In fact, when you enter this menu, you get a sense of how rudimentary those settings are with the Coaching Strategy for the CPU (the machine!) moving between normal, aggressive, or conservative. The Defense Strategy has three options too: normal, blitz, or zone. The slider moves between pass or run.

You can move from coaching to Defense and Offense were you can switch between the CPU and Human (one of the first times I recall this dichotomy in a mass entertainment product):

So you can fine tune the game’s “AI” to create the type of gaming experience you want for not only the CPU, but also the human, which in many ways already assumes a majority of the experience is not dependent on that puny human with the controller in their hands.†

As my son tells me, it’s not the same AI as ChatGPT, and he goes on to quote some smart Youtuber he’s watched that explains neural networks, LLMs and the like, all very good sense to be sure. That said, it was no practical help for me given I’m trying to pass off this rather frail connection with anything called AI as further justification for the project. Dare I say work? You’ll be glad to know I didn’t let his appeal to reason and finer distinctions stop the AI Maddeness show, I just decided build an AI-based media empire around it as further justification.

I spent much of this spring and summer playing Madden 2001 with the idea I’d ramp-up to an entire season this month, just with teams from 25 years ago. It’s kinda wild writing about players from 25 years ago, it’s as if they were still playing and the simple act of re-visiting this season via Madden somehow collapses time.* I also wanted to play the games at roughly the same times as the real 2025 season, so the first week of the 2025 NFL season would also be the first week of AI Maddeness. By doing this I’m hoping to approximate the real time of a season, the build-up, the sense of growing intensity, and the real physical and mental toll it takes on the players of a real team, albeit virtually. What’s more, a huge part of that equation is the media storm surrounding those 17-21 weeks of play. So even if I play the game for less than an hour Sunday or Monday night, I can spent the rest of the week trying to automate the ensuing hype-building media machine.

The media world I’m creating is somewhat insular in that it’s AI first and foremost. It’s not trying to blog like I blog, it’s trying to comprehend and reproduce a reflection of sports reporting from a scraped web. What the hell does that look like? I mean, that fact is itself somewhat alluring, if not downright seductive—as it was in my sorry case. So after months of training on the game and trying to figure out what this AI Media world might look like, I spent much of the summer vacationing, which was awesome. But, it brought me to the beginning of the season, the first week of September, with a half-finished plan.

But who cares, it’s a playful way to learn about AI somewhat outside the polarizing admonitions from either camp about what AI is and isn’t. That said, I know sports are political, I’m playing the season just before 9/11, which provides a crucial inflection point that marks a historical moment were surrendering constitutional rights for the sake of national security was increasingly normalized—in many ways a direct precursor to  the sorry state of the union in which we currently find ourselves. So much of sports media has become increasingly empire adjacent, the media closing ranks against Colin Kaepernick, a process led by the NFL, is a good recent example of this. In fact, the. rise of bot and meme culture just further exacerbates these divisions, which makes the political world of sports a media-driven hall of mirrors. It’s hard to situation yourself when you have questionable facts, hot-takes, and editorials coming at you from all angles.

Post-9/11 was also the era when the military flyover of NFL stadiums took on new meanings, as did the omnipresence of the flag. These things were not just sideshows, but became central to a consolidated display of national pride that seemed rooted in the deeper symbolic comparisons constantly made in the sport. The reference to gladiators, dynasties, and endless “wars”  campaigns made football the embodiment of a more militaristic form of nationalism. So it makes sense that much of sports media both mimics and helps shape the norms that define this kind of chest-puffing patriotism. What might this AI  technology add to an already well-oiled assault on reason that social media has achieved so well in recent years? There’s surely a deeper mouth of Madness beyond AI, but this technology is “trending”so it makes sense to try and understand as much, or little, as I can.

Anyway, I started playing the season early this month and currently heading into week 3 of the season, with the Jets (my team) going 2-0 in a couple of tough games. I recorded the game play and put both games on Youtube. I was originally going to provide commentary on top of John Madden and Pat Summerall, but having tested this in the spring I found I was too busy playing to have anything intelligent to say. I’m older now, so I have to think beyond my guns.

The games come out as a tight 48 minute affair, and so far the recordings have been solid. For week 2 I tried to have the video automatically stream at a pre-determined time, but somehow the time zones got messed up, so I was never streaming. No worries there because I was also recording locally, so I could upload and premiere the game on Youtube after the fact. I’m not sure there’s any need for it to be a live stream, but I still want to figure that out. Apart from that, you’ll notice if you watch the videos that during halftime and the end of the game I go through all the stats provided somewhat meticulously. This is so I can capture the halftime and end of game summaries, the scoring summaries, as well as the various players stats. I collect these via screenshots and put them in a folder that I will then take to ChatGPT.

After the game I have a whole series of data that I can provide to ChatGPT to come up with an article about that week’s game. Given the stats, it does a good job of conforming to the details of the game. So, at the point of asking ChatGPT to write the article, it has more than enough details to fill in around. Turns out I spend a good amount of time fashioning the request for an article I want. I was a quasi-sports writer in high school with my friend, who actually did all the real work. I liked seeing how quickly he boiled down the game to its basics: names, plays, scores. I have no desire to write like this, and in many ways with outlets like the NY Post, Daily News, and Newsday all continuing to struggle in the world of online media, the sports journalism industry has simultaneously exploded and declined.

The best way to get the data out of Madden 2001 is to feed screenshots of the game stats, scoring summaries, and player stats into ChatGPT, which then organizes the image text into quite accurate tables of data that can be exported via CSV. That’s still pretty much magic. Then I’m able to import the data manually into Google Sheets, but it’s clear once I have the sheets set with consistent data I want to collect that this piece can be automated weekly. Also, with the data from the screenshots in readable format, I can ask ChatGPT to write an article. from various voices, as well as create graphics (a struggle), social media posts, and even scripts for a podcast or video highlight reel.

Of all the media ChatGPT creates, I’ve been struggling most with the graphics. I spent a lot of time trying to correct strange repetitive inconsistencies in the image results. This rarely happens with the text, to be honest, but it’s non-stop with the graphics and posters. Despite the struggles, I also find the images oddly impressive. So I fed Chat GPT this screenshot of the final score:

I asked it to create something similar, more specifically I said:

Use the following image to create a graphic in the Madden 2001 style with the icon titled “AI Sports Maddeness 2001.” The graphic should have the score Jets 21 Patriots 16 and have their helmets facing each other on the background of the Jets football stadium

It came up with the following on the first try:

I was impressed. I loved the grainy, almost baseball card like aesthetic. Maddeness was spelt “Maddness” so I tried to have that cleaned up, but that’s usually where things get further and further from what you want. Because every new ask leads to an entirely new image with often increasingly bizarre mistakes and omissions, I returned to the chat with the following ask:

This good, but “Maddness” should be “Maddeness” and can we make the graphic’s orientation landscape. Also, you have to put Jets under the 21 and Patriots under the 16

And I got the following image, which was damn near perfect, save the fact that there was no score for the Patriots and AI Sports reverts back to EA Sports:

And so it goes with graphics, it’s almost better to start fresh with the graphic you want, as I do for the Gangus Green fan art stuff, than feed a  graphic that is meant to take the data off a specific screenshot—at least that’s my experience. Nonetheless, after multiple attempts that led me further and further from what I wanted, I settled on the above image and just photoshopped in the 16 after Patriots.

As for the text articles, they usually come out spot on, but after a couple of weeks they’re already feeling formulaic. I’ve been posting posting on AI Maddeness, AI Sports Zone, and Gangus Green, and if you read them you can already get a sense of uniformity across the postings. I wonder how folks mix that up more? Is the idea you can get it to sound general enough with a unique voice? I mean re-caps of sports events usually sound route, but this is somehow different, it’s like routine with a fake persona. Would love to get any advice here.

I wrote week two’s game recap, spending time outside the machine. When copying it into ChatGPT it proceeded to strip out a fair amount of details to be tighter and more organized. Mine might have been better in my heart, but it also might have been long-winded and boring–I’ve been known. But if what I’m trying to do is give each journalist/media outlet a voice, I’m struggling with how to make a voice (or voices) from the machine that sound more compelling than a bad novel?

In fact, the comments on the posts are a lot of fun at first, Tom made-up some awesome stock football fans that show and respond to the post, but they quickly become identifiable as bots. One solution is just more commentators, or better yet a more nuanced prompt?  I don’t know, my head hurts.

Beyond the sites AI Maddeness and AI Sports Zone, I created a sports fan that I basically told ChatGPT was a huge Jets fan, they think Curtis Martin is god, and continually breaks QB Vinny Testaverde’s chops. The fun thing to see is where this leads, but soon enough it becomes repetitive, how do we make our fake fans more convincing? A.I. Jock, the author for AI Zone and AI Maddeness, is something of a mesh between organizational marketing and sports writing. They kind of get away with their pieces, even if they are repetitive, because that is almost what is expected, whereas a fan’s personality is actually harder to nail.

The cool piece is I can define these character’s personas right in WordPress, which will control their writing style. Again, almost like a series of characters in a strange fake blogroll. I’ve been laughing hysterically at the comments from the NFL conspiracy theorist and the overly pedantic fan. I am going to have to do some more character studies if this project is going to have anything resembling a soul. 🙂

The other piece is my sites are all just stock WordPress themes at the moment. I’ve never programmed a theme, but I’m going to use CoPilot to guide me through the process on top of the base theme Understrap. That’s this weekend’s project, and I hope to have some early 2000s web pages up sometime soon, maybe even in time for the Jets showdown with division rival the Buffalo Bills. They’re home, and they have momentum on their side!

So far the most frustrating thing has been creating supporting graphics and other charts and tables based on the data collected. I’ve spent way too much time on these, and I  may need to diversify to other AI image solutions like Dall E or Midjourney, but that’s how they suck you in. Soon I’ll have a new tool for every media element, it’s already the case with ElevenLabs for AI Audio voiceover. I had fun with the NFl Films style voiceover  while figuring things out this spring, but I’ve not gotten around to that level of media entertainment just yet.

The easiest images to create thus far have been the fan posters for Gangus Green’s blog. They usually consist of one set of descriptions and done. For example, after ChatGPT takes my game description and various stats and turns it into a weekly recap by AI Jock, it can then be re-written from a rabid fan’s perspective, which is the Gangus Green blog. After that’s done, I simply ask for some fan art, like so:

This works, now I am wondering if you have any ideas for some fan art from Gangus Green based on this post

From which I got the following:

I had a field day with these.

For no. 1 it created the following image, which is interesting cause I don’t know what “DOWN” means in this context: “Martin Keeps Moving the Chains Down?” I mean Down Field Makes sense, but just down loses the narrative a bit. But still, an awesome image for my purposes.

Also, no. 2  is fun because it takes a direct quote from the article, and seems to love this idea of Testaverde throwing wiffle balls throughout the game given he put up some rough numbers. The image is awesome, but I also love the “Houstcon” —so close and so far all at once. There is also this strange sense that ChatGPT “knows” when it has good lines, which is bizarre to me

Another awesome one is no. 3, this image depicts the defensive backs as ball hawks given they had a solid second quarter keeping Bledsoe under wraps. The weird thing is the supposed jersey numbers are inconsistent, at times wrong, and often repeated. Not to mention the hawks are downright scary. You almost had it, ChattyG!

Finally, the bonus idea ChatGPT threw at me was a trading card for Curtis Martin’s epic game. I took it up on the offer and this one came out quite well. One shot and done:

Worth noting that ChatGPT’s top 3 image recommendations were also mine. Spooky.

When it comes to more complex, data-driven graphics with summaries, ChatGPT gets regularly confused, repeating the same lines, make the graphic text run off the edges, jumbling table formatting, and so on. Below are a few examples of some of the weirder ones.

The other piece I’m trying to have feed into the AI Maddeness and AI Sports Zone sites are updated reports like player stats, divisional standings, the schedule, and more. I think these will be easy once I have the workflow for moving OCR’d into ChatGPT that then are pulled directly into an updating Google Sheet that’s hooked into a specific page in WordPress. I’m sure there’s more I can do with the stats, and as I learned from my early tests and experiments during the sping and summer, the cumulative stats are important for play-off predictions, wild-card berths, and more. I found getting them post facto from week 17 backwards all at once a pretty daunting task. I did it, but the OCR data was not nearly as consistent so that much of it was manual work.

This brings me to a final point. I’m certain this hasn’t saved me any time—on the contrary. The time I spend manually creating GIFs, getting the stats in order, and trying to get the graphics to comply is a total black hole. But the comments on the site and the quick creation of blog posts and social media posts does give the illusion of time savings. That said, the lead up work is a lot. It’s like prepping to paint: the wall-prep, taping, tarps, and cleaning takes much longer than the 30 minutes of painting, Prepping is everything and it makes me wonder how folks out there manage and organize the various chats they use to take care of repetitive tasks like getting text off the image and putting them in CSV tables. At this point it’s almost automatic with the Madden 2001 images no matter what chat I’m in. It seems that the system in general has been training on what I’m asking it across chats, which is interesting.

Anyway, that’s where I’m at with AI Maddeness thus far, I would write more, but I have a game against Buffalo this Sunday that I’ll have to both market and practice for 🙂

_____________________________________________

*I imagine Curtis Martin or Vinny Testaverde searching their names and finding a post about stats for a game they never played.

†For Madden this is definitely the case given you can only control 1 out of 11 players on your team at any given time—so “AI” is controlling the other 21 players on the field