Using Google’s AI, I created some hilariously bad Nintendo knockoffs.

This week, I got my hands on a new generative AI tool that allowed me to whip up some amusingly bad knockoffs of 3D Nintendo worlds. Check out my take on something reminiscent of Super Mario 64.

I wasn’t a fan of Metroid Prime 4: Beyond, but it still outshines my attempt at a Metroid Prime experience. And then there’s my version of The Legend of Zelda: Breath of the Wild, complete with a paraglider and, for a brief moment, a second Link.

All of this was made possible by Project Genie, an experimental research prototype that I had the chance to explore this week, even if I might not be using it quite as intended. Google DeepMind has been hard at work developing AI “world” models that can create virtual interactive spaces based on text or image prompts. They unveiled the impressive Genie 3 model last year, but it was only available as a limited research preview. Now, Project Genie is rolling out to Google AI Ultra subscribers in the US, giving more people a chance to experience what Genie 3 can do.

The timing of this release is partly to gather insights on how users engage with it. According to a product manager at Google DeepMind, the goal is to discover new use cases that they hadn’t anticipated. The company is particularly excited about Genie’s potential for visualizing scenes in filmmaking or interactive educational content. For instance, you could snap a photo of your kid’s favorite toy and use it to inspire a Genie-generated world. There’s even potential for helping robots navigate real-world environments. However, it’s important to note that Project Genie isn’t yet a fully polished product meant for everyday use.

With Project Genie, you can choose from a variety of worlds crafted by Google or create your own by defining prompts for the environments and characters you want. After a brief wait, Genie generates a thumbnail, and then you can dive into the world. Each generated world can be explored for 60 seconds, with a resolution of about 720p and a frame rate around 24fps. While exploring, you typically control your character using the WASD keys, jump with the space bar, and adjust the camera with the arrow keys.

One of the worlds, called “Rollerball,” features a blue orb rolling through a snowy landscape, leaving a trail of paint behind. As a “game,” Project Genie didn’t quite hit the mark. There were no objectives or goals, no sound, and frustrating input lag that was even worse than what I sometimes experience with cloud gaming. (This might be due to the poor Wi-Fi in my office.)

During the 60-second experience, Genie occasionally forgot to display the paint streak where I had rolled, and sometimes the orb would stop leaving a trail altogether. This inconsistency made me question Genie’s reliability.

Another world, “Backyard Racetrack,” was a bit more engaging since there was an actual track to follow. My racing lines were terrible—thanks to the input lag—but I enjoyed trying to navigate the turns. However, near the end, part of the track unexpectedly turned into grass, breaking the immersion, and the wheel rims looked pretty rough.

I had a blast pushing the boundaries of Project Genie to create 3D, AI-generated games featuring familiar characters, like my Super Mario, Metroid Prime, and The Legend of Zelda-themed worlds. While they were amusing, these worlds lacked scores or objectives, leaving you with little to do but walk or jump around. Even if there were tasks to complete, the input lag made the experience nearly unplayable. (Again, this could be a Wi-Fi issue, but I still faced lag even when closer to my router.)

I ran into some limitations with what I could create. For instance, Project Genie wouldn’t generate a world based on my Kingdom Hearts prompt, which included Disney characters in a steampunk setting. When I removed specific character names and described them instead, Genie generated a thumbnail that looked strikingly similar to Sora, Donald, Goofy, Jack Skellington, and Cloud. However, when I tried to create the actual experience, Genie blocked me.

When I inquired about why I could generate worlds with Nintendo characters, a product manager explained that Project Genie is designed to follow user prompts. They are closely monitoring feedback as part of this experimental phase. The Genie 3 model was primarily trained on publicly available data, which likely explains why Link used his paraglider in my test. Just before publishing this, Project Genie stopped allowing me to create worlds based on Super Mario 64 due to concerns from third-party content providers.

If Google tightens restrictions on generating interactive worlds based on established gaming franchises, I can’t imagine Nintendo will be thrilled with what I created. As it stands, Project Genie isn’t particularly impressive. The input lag and 60-second limit hinder the interactive experience. At times, I couldn’t even control my character, only the camera. After the oddities with the paint trails and the track turning to grass, I felt uncertain about the consistency of the worlds.

Project Genie is an improvement over some AI-generated worlds I tried last year, but it still falls short compared to a well-crafted video game or interactive experience. The potential for blending different media types is exciting, but there’s a long way to go before it reaches that point.

Maybe my expectations are too high, considering Project Genie is still an experimental prototype. Perhaps my perspective will change as the technology evolves. However, I doubt many will want to spend extended time exploring these AI-generated worlds anytime soon. For now, it seems we don’t need to worry about the genie being out of the bottle just yet.