Roblox's 20-Year Bet: Building the Holodeck with Native AI Models

Baszucki says Roblox still has a business plan deck from nearly 20 years ago that the team pulls out from time to time and is surprised by how accurate it is. It sized social networks, YouTube and toys, and it pitched a new category he calls human co-experience, which others label the metaverse or the Holodeck. The idea was a high-fidelity simulation backed by physics where people could build a car, go to a birthday party or chop down trees to build a house.

In that framing, AI is first of all a way to superpower the creation of the Holodeck. It could make the world photorealistic sooner, put 10,000 people in a space instead of a hundred, and simulate acoustics that sound right with a crowd that size. Baszucki notes that the vision has existed in science fiction for decades, with Snow Crash among the examples he cites.

He contrasts this with a second vision, which he calls real-time dreaming: one person alone, dreaming up whatever they want. Short-form video already shows traces of it, since it reacts to dwell time and favorites at 2 a.m., and the extreme case is the film Vanilla Sky, where everyone around the dreamer is an NPC. Roblox sits near the communication end, but Baszucki expects many products in between, including odd categories nobody has imagined yet, such as a voice prompt that becomes a French café where you practice talking to an NPC.

Asked whether the same space will serve consumers and businesses, Baszucki traces a long history of people wanting more fidelity in communication, from cave paintings and smoke signals to the telegraph, the phone and COVID-era video. Many young people, he says, used Roblox as an early test case of co-experiencing during that period. As multiplayer 4D simulation becomes more photorealistic, he expects video to become a downsampled mode of it, a "legacy analog mode" you could switch to if you wanted a Zoom look.

He uses "4D" because it covers not only shape but function, and one thing it adds over video is acoustics. In a thousand-person company meeting on Zoom you get a wall of squares and a hard audio-mixing problem, whereas in a 3D space you hear people with attenuation based on distance and can walk closer to someone to talk. He suggests this may be a better human interface in its own right.

Baszucki also points to a latency problem that no medium has solved: when a group sings happy birthday, everyone hears everyone else tens of milliseconds late. His guess is that systems will eventually extrapolate each person's voice forward in time and remix it so the group sounds in sync. He sees this as one example of the interesting conferencing solutions that could emerge on the business side.

Roblox grew out of a physics engine, but Baszucki says its viral hits such as Dress to Impress and Grow a Garden succeeded for less physical reasons. Grow a Garden leaned on the cloud persistence Roblox provides so that plants keep growing when players are away, and Dress to Impress did new things with contests and layered clothing. The company keeps pushing on physics, including full acoustic simulation, but it cannot control the free market of what developers choose to build.

On world models, he says Roblox is watching closely but its main focus is a 10,000-player experience that can gracefully fall back to 100, 10 or one. He frames it as a deep infrastructure problem: how to synchronize the state and memory of 10,000 people over the last five hours, including where they were in 3D space and what they did. World models store memory in video latents and show early promise, but he says it is still open whether the ultimate format will be video latents, native 3D, or some hybrid 3D video latent space.

Roblox's own plan is to store the history of everything on the platform as vector data instead of raster video. That would allow playback of anything that has happened across about 13 billion hours a month. Baszucki mentions Arthur C. Clarke's *The Light of Other Days*, where total playback erases privacy, and says Roblox would use such a capability thoughtfully and in a privacy-compliant way. Examples include reviewing a safety incident from five camera angles or reshooting a favorite family moment as a video.

Baszucki says there is a growing push to find hybrid video-plus-keyboard human interaction data, with people asking where to find 200 million hours of it. Roblox's data can be reproduced from any camera angle and interacts with the 3D space, which makes it unusually powerful. The company is starting to train NPCs on this data in a privacy-compliant way, and recent demos show them navigating and playing games.

He lays out three levels of ambition. First, everyone gets an NPC that can get pretty good at any Roblox game. Second, if a user opts in, a model could learn their gestures, gaze and way of speaking to create a personal virtual doppelgänger. Third, there would be a simple interface for agentic doppelgängers, such as sending your virtual self to play with your child for 15 minutes while you work. He adds that no ship date exists and that the data will never be shared outside Roblox.

For creators, Baszucki imagines prompting an NPC with a figure like Benjamin Franklin, perhaps a happy one, for an American history experience. The interviewer compares the shift to self-driving, where hand-built heuristics gave way to end-to-end deep learning, and Baszucki agrees that a general modeling solution would be a fascinating alternative to decision trees. He notes this would require cloud-connected platforms to run inference rather than local ones.

On the future of the game designer, he is optimistic that new classes of work will appear and that human touch has a long runway against "AI slop." He expects teams of five to produce astounding quality, but says audiences will then come to expect it, so quality and iteration both rise. He imagines creators sending agents away for 24 hours to tune and test a game, with NPCs playing on various client emulators, including a phone.

Asked about research that generates games directly as video with no physics engine underneath, Baszucki says the fidelity is beautiful but memory remains the issue. He expects the ultimate architectures to be multi-part rather than all-in-one, with a hyper-efficient thousand-person synchronization engine, photorealism done partly client-side, and pipelines that go from a raw frame to 3D upsampling to local 2D upsampling. He says Roblox is doubling down on a hybrid stack that supports multiplayer natively.

He sees world models being useful even in their early state as a creative front end: you narrate a video as you walk around, saying put me in a western world, go left, add some horses, and that video becomes a multiplayer experience. The hosts raise serial micro dramas as a format well suited to the technology, since AI could mass-test plotlines without live actors. The conversation then drifts into the Fermi paradox, with the joke that aliens may be stuck in virtual worlds, against Baszucki's more optimistic Holodeck picture of people hanging out with each other.

On AI companions, Baszucki notes that Roblox is not doing dating, and uses a Black Mirror episode in which a thousand virtual doppelgängers live out relationships before a real date as a thought experiment. He says such uses are hard to predict, and Roblox's focus is raw, high-performance cloud infrastructure with AI as a service. For NPC memory he imagines each character having a lean record of everything it has said or done, which it could retrain on. Asked how many people think their children will have a serious AI relationship, he answers zero for now and says humans and AI are not close to crossing that barrier.

Why hasn't cheap asset generation transformed gaming? Baszucki's answer is that consumer expectations of quality rise at exactly the same velocity as creation gets easier. He says the technology of the industry has to move toward the cloud and vertically integrated asset management rather than shipping on a DVD or in a giant download. Roblox has launched dynamic level of detail for textures, with meshes on the way.

He describes assets in the future existing in several forms: traditional textures and meshes, or an AI prompt generated procedurally on demand. As with video streaming, low-end devices need low detail and high-end devices need 4K, with AI generation on top. One demo covers 3D upsampling, where a creator with a primitive experience adds a prompt such as "make it look medieval and more realistic" and the assets are upsampled in the cloud for free.

On AI coding, Baszucki notes that Roblox Studio is one of the biggest development environments in the world. Users already connect tools like Claude Code to Studio, and Studio also has its own native assistant and code engine. He expects Studio to follow industry best practices for code while adding environmental generation unique to it, moving from a prompt in text, image, video or a world-model walkthrough to a primitive 3D skeleton and then a fully functional game.

On the creator economy, the top creators now make well into the tens of millions of dollars a year. Baszucki sees a healthy long tail, because earnings for the creator ranked 1,000 are growing faster than those of the top creator. The environment has become more competitive, with much more live ops and daily or weekly updates, and Roblox now publishes how its discovery algorithm works, which he says puts useful pressure on the company.

Asked how a leader known for long-term conviction squares that with the speed of AI, Baszucki says Roblox pairs two of its values: take the long view and get stuff done. Inside the company that means breaking a six-month ship plan into weekly iterations toward a clear target. Teams working on AI, facial age estimation and safety operate on weekly cycles, and AI itself moves day to day.

He argues the Roblox spec, which is 10,000 people, real-time multiplayer modification of the environment, NPCs and photorealism, has been stable for 20 years, and one of the hosts calls that the main takeaway for software CEOs worried about durability. Baszucki says the company has sharpened it: a goal of a billion daily users was too hard to operationalize, so Roblox adopted a stepping stone of 10% of global gaming content, which is roughly 300 million daily users and about 3x today. Knowing both 3x and 10x matters, he says: without the 10x it's hard to sleep at night, and without the 3x it's hard to plan.

On hiring, Roblox wanted values-aligned, creative problem solvers but found it hard to separate signal from noise. About five years ago it bought Embellis, a company building 3D assessment tooling, and tuned it for its college-grad and intern program. Baszucki says tens of thousands of applicants go through the same assessment, which the company has tested for fairness, and its public tests are built in Roblox, such as programming a factory or a robot in a turtle-graphics-like language.

One edgy finding, he says, is that the university someone attended has little correlation with the assessment results, so Roblox effectively ignores it. Community colleges and small Midwestern engineering schools produce strong candidates. Asked about a long-term conviction, Baszucki calls himself more of a believer in the power of time and compounding than in amazing inventions, pointing to Excel, which he used in the 1980s and which is not that different 40 years later. He hopes that if Roblox is built right, it could still be around in 40 years.