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

2026-02-05 - 44 min - source - Read full transcript

Sarah Guo (host)Elad Gil (host)David Baszucki

Key insights

Roblox's AI push is framed as accelerating a 20-year-old vision, not replacing it.

Baszucki says the company still has its original business-plan slide from almost 20 years ago describing a physics-simulated 'hollow deck' where users can build, interact, and socialize at high fidelity. He argues AI's role is to superpower that existing goal - photorealism, more concurrent users, realistic acoustics - rather than redirect the company toward a new product category.

roblox-long-term-vision

Future immersive products sit between two poles: shared communication platforms and solo 'real-time dreaming.'

One extreme is a high-fidelity multiplayer communication platform (Roblox's stated direction); the other is a single user experiencing an AI-generated world that reacts to them alone, which Baszucki compares to doom-scrolling short-form video and to the premise of the film Vanilla Sky. He expects most future products to land somewhere in between.

world-models-vs-physics-simulation

Roblox is training NPCs on 13 billion hours a month of vector-format interaction data, not raw video.

Because Roblox stores playback as reproducible 3D vector data rather than raster video, it can replay any past session from any camera angle. Baszucki argues this gives Roblox a uniquely large, structured dataset for training NPCs that go beyond being simple LLM wrappers, addressing a data-scarcity problem other teams face when trying to train on human keyboard-and-video interaction.

ai-npcs-and-virtual-doppelgangers

Baszucki lays out a three-stage roadmap for NPCs, ending in agentic 'virtual doppelgangers.'

Stage one is NPCs competent enough to play any Roblox game. Stage two is opt-in, privacy-compliant NPCs trained on an individual's own gestures, speech, and behavior. Stage three is giving those doppelgangers a simple agentic interface so a user could send their virtual self to act on their behalf, e.g. filling in for 15 minutes of play with a child.

ai-npcs-and-virtual-doppelgangers

He expects a multi-stage AI pipeline for world rendering rather than one end-to-end model.

Baszucki predicts platforms will combine a hyper-efficient synchronization engine for thousands of concurrent users, server-side and client-side photorealism stages, 3D upsampling, and local 2D upsampling, with world-model research plugging into different stages rather than replacing the whole stack with a single generative model.

world-models-vs-physics-simulation

Cheaper AI-generated assets haven't changed game-industry economics because consumer quality expectations rise at the same rate.

Baszucki argues that no matter how cheap asset creation gets, users' quality bar rises in lockstep, so the real shift is structural: games need to become cloud-connected and vertically integrated so assets can stream at variable LOD and eventually be generated on demand from a prompt, rather than shipped as fixed files.

ai-native-content-creation

Roblox Studio is building both a native AI coding assistant and integration with external coding agents.

Users are already gluing third-party coding-agent tools ('cloud code') into Studio to control it directly. Roblox is simultaneously building a native assistant and code engine, and expects a parallel 'environmental generation' track where a world is built by prompting and iterating from a rough 3D skeleton to a fully functional game.

ai-native-content-creation

Roblox reframed its north-star metric from 'a billion DAUs' to '10% of global gaming content.'

Baszucki says the older billion-DAU target was hard to operationalize across product teams. The new target (~300 million DAUs, about 3x current scale, with a higher internal bar for the US market) can be decomposed market-by-market, which he says makes it more useful day to day while a bigger 10x number keeps long-term motivation intact.

roblox-long-term-vision

Roblox's hiring assessment shows little correlation between candidate quality and university prestige.

After acquiring Embellis, a company that built 3D problem-solving tests (including for SAT-adjacent assessment work), Roblox has run tens of thousands of intern and new-grad candidates through the tests. Baszucki says the results deliberately ignore where a candidate went to school, and that this has surfaced strong talent from community colleges and smaller engineering schools, not just top-tier universities.

talent-assessment-and-hiring

Baszucki is skeptical that AI-human relationships will become common soon, but expects near-term value in AI coaching.

Asked what share of children will have at least one serious relationship with an AI in their lifetime, he answers 'zero right now,' saying society isn't close to crossing that barrier. He is more confident about a nearer-term use case: an always-on, low-friction AI coach or therapist 'in your earbud,' which he compares to mental-health and coaching use cases already emerging on text-based AI platforms.

ai-npcs-and-virtual-doppelgangers

Roblox pairs a stable multi-year vision with weekly iteration cycles across teams.

Baszucki describes an internal principle of 'take the long view, get stuff done': teams like AI, safety, and facial age estimation ship on a weekly cadence even while the company holds a 6-to-12-month and 20-year vision. He argues this fast reaction time is what keeps a long-term vision credible while the surrounding technology changes daily.

roblox-long-term-vision

Transparent discovery algorithms are a deliberate competitive strategy, not just a trust feature.

Roblox publishes what factors drive visibility in its discovery algorithm. Baszucki argues this transparency keeps constant pressure on Roblox to make discovery genuinely good, since creators can see exactly what to improve, and credits this shift with a less 'spiky' top-20 experience list than three or four years ago.

ai-native-content-creation

Books referenced

The Light of Other Days - Arthur C. Clarke - Baszucki cites this sci-fi novel's premise of infinite, fully transparent playback as a cautionary comparison for Roblox storing 13 billion hours a month of user history.
Snow Crash - Neal Stephenson - Named alongside 'the holodeck' as an earlier sci-fi vision of the shared, high-fidelity virtual space Roblox has pursued for 20 years.

Media referenced

Ready Player One - movie - A host compares Roblox's original business-plan vision of a physics-backed, high-fidelity shared world to this film.
Vanilla Sky - movie - Used as the reference point for the 'real-time dreaming' extreme, where a solo user lives inside an AI-generated world where everyone else is an NPC.
Black Mirror (dating episode) - show - Referenced when discussing virtual doppelgangers - an episode where a thousand simulated versions of two people 'date' each other before a real match is suggested.
Grand Theft Auto - other - Cited as the current high bar for hand-built game NPC behavior, which Baszucki contrasts with a future general-model approach to NPCs.

Companies

Roblox - Subject of the episode; 150M+ daily users, ~13 billion hours of engagement per month.
Embellis - 3D problem-solving assessment company Roblox acquired about five years ago; its tooling now screens tens of thousands of intern/new-grad candidates.
OpenAI / ChatGPT - Referenced as an existing example of AI coaching and mental-health-adjacent use cases that Baszucki expects to extend into embodied 3D NPCs.

Techniques and frameworks

Dynamic LOD (level of detail) for assets - Roblox is rolling out dynamic LOD for textures (meshes next) so every asset can stream at a quality matched to the device, moving toward on-demand AI-generated assets instead of fixed downloads.
Embellis 3D assessment testing - A university-agnostic, puzzle-based hiring test (e.g., programming a virtual factory or robot) Roblox uses to screen new grads and interns for problem-solving ability.
'Long view, get stuff done' - Roblox's stated operating principle: hold a multi-year product vision while running weekly iteration cycles inside teams like AI, safety, and age estimation.
3x / 10x goal framework - Baszucki says a CEO needs to know both the 3x version of a target (to plan operationally) and the 10x version (to stay motivated long-term); Roblox's own reframe is '10% of global gaming content,' about 3x current scale.

Summary

David Baszucki, founder and CEO of Roblox, joins Sarah Guo and Elad Gil to walk through how AI intersects with a company vision he says has barely changed in 20 years: a physics-backed, high-fidelity shared world he calls "the holodeck." His framing throughout is that AI is an accelerant for that existing plan, not a pivot. He repeatedly returns to an old Roblox business-plan slide from nearly two decades ago and argues the goal was always photorealism, massive concurrency, and realistic physics and acoustics for up to 10,000 people at once - AI just makes those things newly tractable.

Much of the conversation is about NPCs and what Baszucki calls "virtual doppelgangers." Roblox's advantage, in his telling, is data: 13 billion hours a month of user interaction stored as reproducible 3D vector data rather than raw video, which can be replayed from any angle. He lays out a staged roadmap - NPCs that can competently play any Roblox game, then opt-in NPCs trained on a specific user's own behavior, then agentic doppelgangers a user could send out on their behalf. He's careful to note this is explicitly not a dating product, while acknowledging the obvious adjacent territory (he brings up a Black Mirror episode about simulated dating unprompted). Asked directly whether children will grow up having serious relationships with AI, he says "zero right now," positioning himself as more skeptical than most people he talks to, while still expecting near-term value in AI as an always-on coach or companion.

On the technical architecture question - whether world models will replace physics engines outright - Baszucki predicts a multi-stage pipeline rather than one model doing everything: a synchronization engine for large numbers of concurrent users, upsampling stages (3D then local 2D), and world-model components plugged in at various points, possibly split between server and client. He's skeptical of a single end-to-end approach dominating anytime soon, drawing a partial analogy to the shift from hand-tuned heuristics to end-to-end deep learning in self-driving, while noting Roblox is currently "doubled down on a hybrid multi-tech stack."

On content creation economics, he pushes back on the premise that cheaper AI-generated assets have changed the gaming industry: consumer quality expectations rise exactly as fast as creation costs fall. What's actually changing is infrastructure - Roblox is rolling out dynamic level-of-detail streaming for assets and moving toward on-demand, prompt-generated assets, both of which require the industry to go fully cloud-connected rather than shipping fixed downloads. In Roblox Studio, he describes both a native AI coding assistant and support for external coding agents plugging directly into Studio, alongside a parallel "environmental generation" track for prompting a rough world into being and iterating it into a finished game.

Strategically, Baszucki describes reframing Roblox's north-star metric from "a billion DAUs" to "10% of global gaming content" (roughly 300 million DAUs, about 3x current scale), which he says is easier to decompose and operationalize by market. He credits an internal principle of pairing a stable long-term vision with weekly iteration cycles for keeping that vision credible as AI capabilities shift daily. The conversation closes on hiring: Roblox acquired an assessment company (Embellis) that built 3D problem-solving tests, and Baszucki says running tens of thousands of candidates through them - deliberately ignoring university pedigree - has surfaced strong talent from community colleges and smaller schools, not just traditional feeder universities.

Notable Quotes

"We feel if we build Roblox right, it might be the kind of thing that's kicking around in 40 years." - David Baszucki

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

"What proportion of you think your children will have at least one serious relationship with an AI throughout their lifetime? I think zero right now. I don't think we're close to crossing the human AI barrier." - David Baszucki

"Take the long view, get stuff done... inside the company, rapid iteration." - David Baszucki