How AI Will Transform Roblox Games into Photorealistic Worlds | CEO David Baszucki
Summary
Roblox’s AI thesis extends a 20-year plan to build a photorealistic, physics-backed “Holodeck” where 10,000 people can modify a shared world in real time. Baszucki sees a continuum from human “coexperience” to solitary “real-time dreaming,” with NPCs at the latter end. If 4D simulation succeeds, “video is the downsampling” and Zoom becomes a “legacy analog mode.”
Baszucki centers the hardest technical problem on synchronizing the state and history of 10,000 participants. Roblox is testing whether that state belongs in video latents, native 3D or a new hybrid format. Its potential advantage is 13 billion hours of monthly interaction data that can be reconstructed from any camera angle rather than stored as fixed raster video.
Roblox’s NPC roadmap progresses from general game-playing agents to opt-in virtual doppelgangers and then agentic substitutes. Baszucki imagines an NPC learning a user’s gestures, the way they look at things and their speech, then briefly playing with that user’s child while the parent works: “Could my virtual doppelganger fit in for 15 minutes?” The discussion also imagines creators prompting characters such as “a happy Ben Franklin.”
Baszucki envisions a hybrid AI architecture, not one model running the entire world. A high-efficiency multiplayer engine could synchronize state while dedicated systems handle NPCs, 3D upsampling and local 2D photorealism. Roblox is “doubling down on a hybrid, multi-AI tech stack that supports multiplayer natively,” even as direct video-generation models offer a possible parallel path toward personalized dream worlds.
Cheaper generation will raise output quality without necessarily making game development less competitive. Baszucki’s counter to the asset-cost thesis is that “the expectation of quality from consumers goes up at exactly the same velocity.” Five-person teams may produce “astounding” work, but users will normalize that standard; agents could meanwhile spend 24 hours testing builds with 20 simulated players across phones and other clients.
On AI relationships, Baszucki answers “zero right now.” The transcript inconsistently attributes the accompanying “I don’t think we’re close to crossing the human-AI barrier” line, but Baszucki is more receptive to an always-present coach or therapist and says mental-health advice and coaching are already major use cases. Sarah calls that arguably valuable.
Management’s intermediate target is 10% of global gaming content, framed as roughly 300 million DAUs, about 20 billion in an unspecified metric and about 3x Roblox’s current position. Baszucki pairs that measurable waypoint with the much larger Holodeck vision: “Knowing both 3x and 10x is very helpful.” The conversation uses Microsoft Excel—still recognizable after 40 years—as a frame for the possibility that a correctly built Roblox could likewise be “kicking around in 40 years.”
Deep dive
1. AI turns Roblox’s old Holodeck specification into an engineering program
Baszucki’s starting point is a nearly 20-year-old business-plan deck proposing “a new category of human coexperience.” Its physics- and reality-backed Holodeck would let users build and drive a car, blow out birthday candles, or chop down trees and turn them into a house.
AI expands that vision along a spectrum: at one end, high-fidelity communication among real people; at the other, “real-time dreaming,” where everyone else is an NPC. Baszucki sees short-form video’s dwell-time feedback as a glimpse of the latter and Vanilla Sky as its extreme.
In business communication, 4D simulation could become “a superset of video”: colleagues could walk through a virtual office, while spatial attenuation could make a thousand-person gathering sound more natural. Even synchronized singing might require extrapolating each voice forward to offset roughly 30 milliseconds of latency.
The hosts’ realism pushback is important: Roblox’s viral successes have not depended on photorealistic physics. Baszucki agrees, citing Dress to Impress and Grow a Garden; his answer is to improve the underlying simulation and cloud primitives, then let the free creator market—which Roblox cannot control on content generation—decide how to use them.
2. Synchronized state and vector history are the prospective moat
World models look promising, but Baszucki calls the “ultimate synchronized state for 10,000 people” the deeper question: how should a platform retain where everyone was, what they did and their memory over the last five hours? The answer might be video latents, native 3D or an undiscovered hybrid.
Roblox’s ambition is to store “the history of everything on Roblox” as vector data, covering 13 billion hours each month. Unlike raster video, an event could be replayed from new camera angles—useful for investigating a safety incident or creating a video of a meaningful family moment—subject to privacy-compliant, thoughtful and judicious access.
The same archive creates training material richer than passive video: embodied movement, WASD-style actions and interaction with editable 3D environments. Baszucki contrasts that supply with people asking where they might obtain 200 million hours of video-plus-action data.
His NPC ladder has three rungs: an agent that becomes competent at any Roblox game; an opt-in “virtual doppelganger” trained on a person’s speech, gestures and the way they look at things; and a simple interface for dispatching that doppelganger to act. Allaire adds that the Roblox model would be privacy-compliant and that the data would never be shared.
3. Creation shifts toward agents, procedural assets and modular AI
The hosts compare the NPC transition with self-driving systems moving from hand-mapped heuristics to end-to-end learning. Baszucki accepts the analogy: replacing decision trees with “a purely more general modeling solution” is the compelling computer-science opportunity, though he offers no date for when it will outperform carefully authored NPCs.
The hosts raise directly generated video worlds with no physics engine underneath. Baszucki says an interactive system could eventually accept prompts such as “Put me in a western world,” then let a user turn left, turn right and add horses. He nevertheless envisions several specialized layers: synchronization, NPC inference, raw 3D structure, cloud 3D upsampling and device-side 2D enhancement.
For creators, Baszucki does not believe “AI slop” displaces human judgment soon: “The human touch has a long runway on it.” Instead, five people may produce astonishing quality, while agents could spend 24 hours modifying and testing a build and sending 20 NPCs through it on various client emulators.
Asset generation follows the same cloud logic. A sound, mesh or texture could exist at several levels of detail—or ultimately as an AI prompt, generated procedurally when needed. His specimen is a primitive world whose creator requests a free, on-demand transformation: make it “medieval and more realistic.”
4. Embodiment makes companions stronger without erasing the human boundary
Roblox can give an NPC more than a summary prompt: because its actions and conversations can persist, it could access “everything they’ve ever said or done” in a compact format. The hosts argue that embodiment plus cloud memory should make these characters substantially more engaging than today’s text companions.
Baszucki’s Black Mirror thought experiment sends 1,000 virtual doppelgangers through simulated lives with another person’s copies; if 98% report a good outcome, the humans schedule a date. He explicitly distances Roblox from dating, but uses the example to show how difficult new social products are to forecast.
Asked whether children will have a serious AI relationship during their lifetimes, Baszucki answers “zero right now.” The transcript is inconsistent about whether the accompanying claim that humans are not close to crossing the human-AI barrier belongs to him or Sarah. Baszucki is more confident in less invasive, over-the-shoulder coaching, and says mental-health guidance and coaching are already major use cases; Sarah calls that arguably valuable.
5. Roblox pairs a 40-year destination with weekly operating loops
Creator businesses are becoming larger and more competitive: top teams reach the tens of millions, with some making “$30 million-plus a year,” while people as far down as creator No. 1,000 can make a living. Cloud-based live operations now mean daily or weekly updates rather than static releases.
Discovery is becoming more transparent, partly because that transparency keeps pressure on Roblox: someone could game discovery. Baszucki says distribution is consequently richer and less spiky: the top 20 experiences are now all really good and vying for No. 1, unlike three or four years earlier.
His management quadrant combines “take the long view” with “get stuff done.” Teams iterate weekly toward a stable specification, while the 10%-of-gaming waypoint translates that distant vision into roughly 300 million DAUs, “about 20 billion” in an unspecified metric, and approximately 3x the present scale.
Hiring follows the same systems logic. Roblox acquired Imbellus to build 3D problem-solving assessments and now largely ignores university prestige, finding strong candidates at community colleges and smaller Midwestern engineering schools. Baszucki closes with “the power of time and the power of compounding”: AI changes daily, but products such as Excel can retain their basic utility for 40 years, while the conversation frames a correctly built Roblox as potentially “kicking around in 40 years.”