Originally published as an Advisory Hour Substack post on 2026-08-26. This first-party site copy preserves the text, localizes public supporting media, and connects the record to its project trail.
Original post
Mario Walker is an experimental mashup of Mario Bros and Star Wars. The build made use of new techniques around character sprite generation, animation, and world building. The bigger idea is that in an AI powered world it’s worthwhile to just try combining two competing ideas to see what you end up with.
Mario Walker pulled lessons from V4 of IslandLab and also a trick involving Chroma Key for generating sprites. Techniques now largely obsolete with the advance of transparent images off of the gpt-image-2 API. Still, I was able to put the Chroma approach to work for both Mario Walker and Founderman2, which I’ll cover tomorrow.
I’m shipping Mario Walker with over 60 stages. These are not great stages, but directionally you can see the scale and width of my game projects is expanding. I’ve gone from doing one or two stages, to procedural worlds, and now I’m getting a sense on how to handle 60+ predefined (not procedural) stages driven by AI. I wrote the other day of an observation about the profound problem of time with the vast sea of video and music that humming GPUs in my basement happily create while I’m asleep. This is magnified in the face of game worlds this vast.
While you can pick any retro stage you like (and these are not all authentic recreations-only the first stage I gave a little more attention to on that count), the star wars specific stages demand you complete them in order. There’s a good chance you won’t be able to complete every stage: I am in the early innings of overhauling my quality assessment testing. More on that in a moment. Here’s a screenshot of the stage selection screen to give you an idea of how much content is shipping in this fan-made game.
I shipped with this Mario Walker a debugging tool so you could see how I taught an AI model how to see sprites. It’s very effective for an AI model that has tooling. If you look closely, you’ll notice the broad outlines of the robot lab from island-v4. I reasoned that if it worked well for robotlab, then surely it’d work well for sprites-and it did.
The character lab has every single sprite implemented in a debug pane. AI allows us to move beyond simple unit testing or integration. Why not character specific testing? My sad regret is not earmarking enough time for this Mario Walker experiment to implement every animation sequence available. Should you browse the art sets, you’ll see that I have far more animations defined than were implemented in the game. These animation tools even include the retro Mario sprites, which to the best of my knowledge no one has ever built anything quite like this for Mario.
You can also look ahead at the sprites in the character lab to see the boss of the retro stages-including a few defects that I didn’t quite get to fixing. The boss is a weird mix of three boss-like characters: a seasoning of Warhammer 40k, a bit of Star Wars, and a Captain profile from a strategy game I can’t quite recall the name of.
I wrote before on Mario Walker and the power of combining different ideas that seem unrelated. The exercise stretches you creatively and it opens up new concepts that may not even have occurred to you before. If you’d like to learn more about the origin of this project or early recordings of gameplay, you can see that here.
EricGamelab: What if... Mario Walker?Explore mashups when it comes to what’s possible with AI to bring forward entirely new types of games. We are at the very beginning of an artistic renaissance when it comes to games reminiscent of the flash era. For example, what if Luke Skywalker were in Mario? How might that work? I think it’d work something like this clip-with obvious rooms for impro…Read more4 days ago · 2 likes · EricLearnings
What I learned building Mario Walker is that once an agent gets to work on a long problem there isn’t much reason to be on the computer. I can’t even claim I checked in from the phone. There wasn’t much point. I setup the build on a task that I intuited would take a couple days to complete (my longest running single code task stands right around the four day mark) and just… sorry, but the princess is another castle my way to the gym, and then the bar in that order.
It’s a fascinating and sad thing for me. I remember many cold winters in Iowa when night exceeds day there isn’t much to do. So, during the weekend and evenings when darkness was intercepted only by the glow of the monitor illuminating my fingers poking at the keys for whatever new and exciting project I had in mind. Looking over those projects is laughable compared to the vast size and scale of what I’ve built with AI over just the last four months, and as I write this I realize I’ve swapped subscriptions with Unity3D and other asset providers for local AI devices as well as an AI subscription. The entire industry has shifted fast.
On the best of days, it really is like being a kid again. The power and promise of a computer that can do anything I want. I’m not bound by anything more than the power of ideas and words, which are two things that I’ve always orbited in one way or another. If you subscribe to me her on substack, then you’ve seen me being intentional about learning specific words or ways of writing. Phrases are real magic in the world-the right prompt could cure a disease, solve a thorny weed problem in the yard, or even help you replace a ceiling fan.
The bigger risk is having clear exits. AI takes an acceptance criteria and completes it faster (in practice) than you have time to flip your phone on to check the socials. How far do you complexify your project or take a project or experiment? When I use AI it only ever wants to add and expand. My personal website is a great example of this challenge because I’m nearing a personal milestone of 1,000 content items on the site. When is a personal site not just a site but something else?
Like you, I’ve seen many apps cursed by complexity and AI. Does my grocery savings app really need to have so many features and buttons when the only reason I signed up for it was to get a $0.05 savings discount on gas? I do wonder when all the features being stuffed into apps is going to end up causing a kind of cognitive collapse when it’s realized that humans can’t possibly comprehend apps with so much complexity. If you ever bounce around AWS portal, GCP portal and Azure portal on the same day, then you know that feeling of “maybe it’s time for a cancer screening” with the UIs.
I had a lot of fun building Mario Walker just the same.
I know, right? I know the game didn’t need an AT-AT walker blowing up goombas in the background. I iterated on several variations before landing on the current setup. It works and it’s just a joy to watch. If you’re wondering what got me even thinking about this weird combination of characters, well, I’ll blame YouTube for this. There are many wonderful independent animators on YouTube that do weird and interesting twists with classic games. If it weren’t for several late night binge sessions, the idea likely wouldn’t have crossed my mind.
Where to Play
Try it out on the gamelab (on a desktop):
https://www.ericrhea.com/gamelab/mario-walker.html
Keys:
S key runs
WAD or Arrows move
Shift key lightsaber
Spacebar jump
Mobile support was given next to no time. If it’s present, you should buy a lotto ticket.
Connected work
- Mario Walker: See the public overview, screenshots, play flow, and Character Lab.
- Play Mario Walker: Put the cartridge on Channel 3 and choose from all sixty-two stages.
- Mario Walker Character Lab: Inspect the animation controls for the hero, enemies, bosses, and Retro roster.
- What If... Mario Walker?: Return to the earlier build note before the experiment reached its full stage count.