The Roblox Corporation recently tried to show off its experimental generative AI tech, one the company hopes will let players fabricate entire game worlds soon, but the best the tech could come up with for now is Clair Obscur: Expedition 33... if you squint really hard and pretend that game really sucked, actually.
Roblox lives and thrives off of user-made games, so perhaps it's unsurprising that the Roblox Corporation is also chasing the generative AI dragon. In a series of recent tweets, the company details what it calls 4D Generation, a feature launched yesterday allowing bloxers to generate "interactive 3D objects like cars, plans, and more" via prompts.
The company then moved onto something it calls 'real-time dreaming,' a more experimental idea being worked on in its research lab, and this is where the nightmare fuel comes in. Roblox Corp says users will have the "ability to generate fully playable video worlds prompted from any text or image," and then convert it to "Roblox native as a way for many people to play simultaneously." In a dystopian knife twist, the company calls the idea "Dream Theater – where one user is dreaming, while others watch and prompt them."
Want to see what real-time dreaming is capable of right now? Well, it's capable of generating a game of the year winner. No, I mean literally – Roblox put out a video showing off the tech, and apparently a prompt as vague as 'woman in a glowing cave' excretes what to me looks obviously like Clair Obscur: Expedition 33's Maelle running around a blurry version of the acclaimed RPG's Flying Waters area. Check it out below.
In our research lab, we are building “real-time dreaming” - the ability to generate fully playable video worlds prompted from any text or image.Our real-time, action conditioned world model (currently running internally at 16fps at 832x480p) is trained on a combination of data,… pic.twitter.com/R5CKJkFuh2February 5, 2026
"Why did it make Expedition 33?" one viral tweet reads. "Expedition 33 looks so bad in this, how did it win game of the year?" another says.
Roblox co-founder and CEO David Baszucki explains that the company "built this model with internal Roblox data, as well as open source video data." That internal data includes the 13 billion hours Roblox players spend in-game every month, which is now being "used to train world models, as well as AI-driven NPCs."
