Inspiration
AI progress is usually presented through benchmark tables, model launches, and infrastructure announcements. We wanted to turn that abstract competition into something immediate: a motorsport showdown where compute is the finish line.
That became THE AI RACE, an unofficial fictional orbital sprint between twelve frontier AI labs. Instead of another dashboard, players get a focused burst of speed, risk, drafting, overtakes, and spectacle on the way to the unclaimed HELIOS orbital compute array.
Our goal was to build a browser game that launches instantly, rewards immediate replays, and makes the AI theme part of the mechanics—not merely a visual skin.
What it does
You pilot OPENAI against eleven AI-controlled rivals around a 2,725-meter anti-gravity ribbon. The course climbs, banks up to 30 degrees, and passes through six named sectors before reaching HELIOS.
Drafting behind another ship charges a wake lock. Releasing and re-engaging the inference burst converts that lock into a slingshot attack. Contact costs speed, shield energy, and boost, while eight off-line data cores reward aggressive route choices.
Three difficulty levels alter rival speed, grip assistance, recovery, damage, and racecraft. Four mission contracts challenge players to win, collect every core, complete a contact-free run, or deploy two slingshots.
Completed contract runs create MODEL N-1, a deterministic holographic personal-best rival with a live time delta and six-sector comparisons.
An active race-control narrator calls named leaders, overtakes, sectors, attacks, damage, and the exact winning margin. The winner triggers a 7.6-second HELIOS takeover sequence with rack-light cascades, an ownership banner, an orbital energy beam, claim rings, and dedicated cinematic cameras.
The game supports keyboard, gamepad, portrait mobile, and landscape touch controls. An optional Sam Altman selection is an explicitly unofficial text-only tribute role. It uses no likeness, endorsement claim, or identity-linked voice.
How we built it
THE AI RACE is a self-contained static browser game built with JavaScript, HTML, CSS, and a vendored Three.js module. It requires no account, API key, installation, or cloud request during gameplay.
The racing simulation runs at a deterministic fixed 120 Hz. Ships operate in track-space coordinates on a true three-dimensional manifold. Curvature produces lateral force, banking counters it, and slope changes acceleration, making the course geometry mechanically meaningful.
Each rival has its own curvature preview, preferred racing line, passing offsets, risk, burst timing, and wake-lock behavior. Simulation randomness is isolated from visual randomness so matching difficulty-and-contract runs remain comparable for MODEL N-1.
The renderer uses WebGL raster rendering with adaptive Balanced, High, and Ultra profiles; instancing; custom shaders; PBR materials; bounded dynamic lighting; radiance shafts; HDR bloom; ACES tone mapping; and adaptive resolution. Original aerospace material atlases and environment plates were authored in Blender and rendered locally on an NVIDIA RTX A6000.
Audio combines the Web Audio API, AudioWorklet-driven engines, procedural music and effects, spatial rival audio, and priority-based race control. Original narration lines were synthesized with Kokoro-82M's generic stock voice. A predecoded 49-clause audio sprite constructs exact names, actions, ranks, and finish margins without requiring a cloud request during the race.
Codex served as an engineering and QA collaborator throughout the project: decomposing features, editing the implementation, creating deterministic test seams, building Playwright regression suites, analyzing screenshots, and iterating on mobile, graphics, narration, and performance. Human direction established the concept, priorities, feel, creative judgment, and acceptance decisions.
Challenges
The first challenge was making the anti-gravity course physical rather than decorative. Steering, banking, gravity, lateral grip, recovery, collisions, and rival AI all had to agree on the same moving coordinate frame.
The second was deterministic ghost racing. Scenery and AI initially shared randomness, causing supposedly identical races to diverge. Separating the race seed from visual randomness made reliable personal-best comparisons possible.
Narration presented a different timing problem. Full prerecorded sentences could not describe every possible overtake or finish margin, while runtime browser speech was inconsistent. The compositional audio sprite solved coverage and startup latency, although smoother phrase-level prosody remains an area we are improving.
The final challenge was preserving the same visual identity across desktop GPUs and mobile devices. Capability-based profiles and adaptive scaling reduce fill rate, shadows, particles, and resolution while retaining the ships, materials, lighting language, and readable HUD.
Accomplishments we're proud of
- A complete twelve-ship 3D racing game delivered as a static website.
- Mechanically meaningful banking, elevation, drafting, collisions, shields, boost, and slingshots.
- Distinct rival strategies and three difficulty levels without scripted finishing orders.
- Deterministic personal-best ghost racing with six-sector telemetry.
- Exact named race-control calls and finish margins.
- Four mission contracts with separate progression records.
- An original orbital-compute visual identity grounded in aerospace and data-center references.
- A cinematic game-state-driven HELIOS ownership sequence.
- Automated desktop, mobile, narration, gameplay, graphics, and genuine RTX A6000 verification.
What we learned
Spectacle works best when it reflects game state. A light show is attractive; a station changing ownership to the actual winner while race control announces the precise margin feels earned.
Determinism also became a product feature rather than only a testing convenience. The same architecture that made automated QA reliable enabled MODEL N-1.
On mobile, preserving the game's identity is more important than preserving every expensive effect. Deliberately reducing fill rate and secondary effects produces a better result than replacing the presentation with a generic low-quality mode.
What's next
Our immediate priority is smoother, more conversational race-control delivery using longer phrase-level programs and an original or properly licensed narrator voice.
We also want another physical-device QA pass, expanded accessibility and onboarding, additional contracts, deeper post-race telemetry, more courses, and optional leaderboards that preserve the instant-play static build.
Log in or sign up for Devpost to join the conversation.