No Priors

How AI Will Transform Roblox Games into Photorealistic Worlds | CEO David Baszucki

Elad Gil and Dave Baszucki on roblox’s AI roadmap: Holodeck realism, NPCs, and creator leverage.

Elad GilhostDave BaszuckiguestSarah GuohostElad GilhostDave BaszuckiguestSarah Guohost
Feb 5, 202643m
20-year “Holodeck” vision and human co-experience4D simulation for communication beyond ZoomPhysics simulation vs. photorealism priorities10,000-player state synchronization and memoryVector-based historical replay of Roblox eventsNPC training beyond LLMs; embodied agents and doppelgängersCreator tooling: AI coding, environment generation, cloud live-ops

In this episode of No Priors, featuring Elad Gil and Dave Baszucki, How AI Will Transform Roblox Games into Photorealistic Worlds | CEO David Baszucki explores roblox’s AI roadmap: Holodeck realism, NPCs, and creator leverage Baszucki frames Roblox as a 20-year mission to build a high-fidelity, physics-backed “human co-experience” platform—often likened to the Holodeck/metaverse—where AI primarily serves to accelerate realism, scale, and creation.

At a glance

WHAT IT’S REALLY ABOUT

Roblox’s AI roadmap: Holodeck realism, NPCs, and creator leverage

  1. Baszucki frames Roblox as a 20-year mission to build a high-fidelity, physics-backed “human co-experience” platform—often likened to the Holodeck/metaverse—where AI primarily serves to accelerate realism, scale, and creation.
  2. He contrasts that multiplayer future with an opposite extreme: “real-time dreaming,” where AI generates personalized worlds populated mostly by NPCs, and suggests many hybrid products will emerge between these poles.
  3. A major technical emphasis is synchronization and memory for 10,000-player simulations, plus storing Roblox’s history as replayable vector data (not raster video) for safety, user replay, and training next-gen NPCs.
  4. He anticipates NPCs evolving beyond LLM chat into embodied agents trained on Roblox interaction data, potentially enabling opt-in personal “virtual doppelgängers,” while creators gain leverage via AI-assisted building, testing, and live-ops iteration.

IDEAS WORTH REMEMBERING

7 ideas

Roblox’s north star is stable: physics-backed co-experience at massive scale.

Baszucki describes a long-standing spec—10,000 people, real-time modification, photorealism, NPCs—arguing AI is mainly an accelerator toward that destination rather than a reason to change direction.

“4D” is about function and interaction, not just 3D visuals.

He uses 4D to mean simulation with behavior (physics, acoustics, persistence) that enables actions like walking around an office, natural proximity audio, or manipulating objects—capabilities video calls can’t match.

The hard problem is multiplayer synchronization and shared memory, not just graphics.

Baszucki highlights the challenge of efficiently synchronizing the state/history of 10,000 participants—where they were, what they did—suggesting future systems may blend native 3D representations with video/latent approaches.

Roblox wants to store “everything that happened” as vector history.

He proposes recording Roblox events as vector data to enable playback from any camera angle, support safety investigations, and let users remix meaningful moments—while emphasizing privacy-compliant, judicious use.

Next-gen NPCs will be trained from embodied gameplay data, not only text.

Roblox is training NPCs that can navigate and play games using platform interaction data; the roadmap escalates from generally competent NPCs to opt-in personalized models capturing gestures/behavior, and eventually agentic “virtual doppelgängers.”},{

Asset generation won’t “cheapify” gaming because expectations rise in parallel.

Baszucki argues that as generative tools reduce asset costs, consumer quality expectations increase at the same speed; the bigger shift is cloud-connected, multi-LOD assets and on-demand procedural/AI generation.

The future creator workflow is cloud-native: agents that build, test, and iterate.

He envisions Roblox Studio combining standard AI coding assistants with environment/world generation, plus cloud-based agents that can run automated playtests (NPCs across devices), tune experiences, and support rapid weekly live-ops updates.

WORDS WORTH SAVING

5 quotes

We do have that business plan slide… it imagined… the Holodeck… the ultimate high-fidelity simulation where people can come together and do stuff.

Dave Baszucki

As the technology for multiplayer 4D simulation gets better and more photorealistic, it’s almost gonna be like video is the downsampling.

Dave Baszucki

Part of the vision for us is to ultimately store the history of everything on Roblox… not raster like video, but… vector… the ability to play back anything that’s ever happened on Roblox.

Dave Baszucki

The data we have, which is 13 billion hours a month, is… very powerful data… around how to create great NPCs that are more than just an LLM.

Dave Baszucki

No matter how cheap it is to create assets, the expectation of quality from consumers goes up at exactly the same velocity.

Dave Baszucki

QUESTIONS ANSWERED IN THIS EPISODE

5 questions

When you say “4D simulation,” what minimum capabilities must be simulated (physics, acoustics, persistence, etc.) for it to feel like a true superset of video?

Baszucki frames Roblox as a 20-year mission to build a high-fidelity, physics-backed “human co-experience” platform—often likened to the Holodeck/metaverse—where AI primarily serves to accelerate realism, scale, and creation.

On the 10,000-player goal: what’s the biggest bottleneck today—networking, authoritative simulation, client prediction, or state compression/synchronization?

He contrasts that multiplayer future with an opposite extreme: “real-time dreaming,” where AI generates personalized worlds populated mostly by NPCs, and suggests many hybrid products will emerge between these poles.

How would Roblox implement “vector history” in a privacy-compliant way—who can access replays, under what consent model, and how would retention work?

A major technical emphasis is synchronization and memory for 10,000-player simulations, plus storing Roblox’s history as replayable vector data (not raster video) for safety, user replay, and training next-gen NPCs.

For NPCs “beyond LLMs,” what training objective matters most: goal completion in games, social believability, safety alignment, or long-horizon memory consistency?

He anticipates NPCs evolving beyond LLM chat into embodied agents trained on Roblox interaction data, potentially enabling opt-in personal “virtual doppelgängers,” while creators gain leverage via AI-assisted building, testing, and live-ops iteration.

What product experiences become possible only when NPCs have embodiment and persistent memory (vs. today’s text companions)?

EVERY SPOKEN WORD

Install uListen for AI-powered chat & search across the full episode — Get Full Transcript

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.

Add to Chrome