Roblox Showcases Experimental ‘Dream Theater’ AI Tech With Disorienting Expedition 33 Demo

• 
  
  Copy link
  
• 
  Facebook
• 
  
  X
  
• 
  Whatsapp
• 
  Pinterest
• 
  Flipboard
• 
  Email

To demonstrate the current capabilities of “real-time dreaming,” Roblox released a demo video. Despite using a prompt as generic as “woman in a glowing cave,” the output looks suspiciously like Maelle from the acclaimed RPG Clair Obscur: Expedition 33 running through a hazy version of the “Flying Waters” stage. You can view the comparison 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

“How did it end up making Expedition 33?” one viral post questioned. “Expedition 33 looks terrible here; how did it manage to win Game of the Year?” another mocked, highlighting the visual artifacts typical of generative video.

Roblox co-founder and CEO David Baszucki clarified that the model was trained using “internal Roblox data” as well as “open-source video content.” This includes the staggering 13 billion hours players log every month, which is now being harvested to refine these world-building models and power next-gen AI NPCs.

In other community news, a former Square Enix executive recently criticized the viral Roblox hit ‘Grow a Garden’ as ‘somewhat predatory,’ questioning the ethics of paid mechanics for growth and theft within the game.