Agora-1 by Odyssey: GoldenEye N64 in Multiplayer AI Simulation

The artificial intelligence lab Odyssey recently unveiled Agora-1, a groundbreaking world model that allows up to four players to simultaneously immerse themselves in an AI-generated environment inspired by the classic game GoldenEye for the Nintendo 64. This innovation marks a significant advancement in the field of multi-agent game simulations.

Agora-1 operates by separating simulation and rendering. One model is tasked with continuously calculating the shared game state, while a second model generates an individual perspective for each player in real-time. This approach enables each participant to experience a unique journey while interacting within the same virtual world.

Odyssey also introduced an interactive audio-video model named Starchild-1. This model is designed to generate synchronized visuals and sounds while responding to textual inputs in real-time. The company sees potential applications for these technologies in training AI agents and collaborative robotics, where multiple robots could work together in simulated environments.

Agora-1 is one of the first multi-agent world simulations, a domain where world models have previously been primarily limited to a single active player. With Agora-1, up to four players can engage in a shared deathmatch simulation, where each interacts with the same world generated in real-time. The model acts as a learned game engine, allowing for the creation of new levels without losing the mechanics of the source game.

Unlike video generators such as Sora from OpenAI or Veo 3 from Google, which produce fixed clips, Agora-1 continuously simulates a game state and renders a distinct perspective for each player. Odyssey divides this process into two parts: simulation and rendering. One model learns from the internal state of the game how the world changes based on players' actions, while a second model, based on diffusion, learns to transform that shared state into visuals.

In the demonstration of Agora-1, up to four players can compete in a simple level. Previous multi-agent approaches like Multiverse or Solaris primarily faced challenges when players lost sight of each other, according to Odyssey. Agora-1 aims to provide consistent views of the same world from multiple angles, thus overcoming these limitations.

Starchild-1 Adds Audio and Text Interaction

Odyssey also introduced a sister model called Starchild-1. The startup describes it as an interactive audio-video world model that generates synchronized visuals and sounds while responding to ongoing textual inputs. It operates on modern hardware at 24 frames per second. Unlike Agora-1, Starchild-1 focuses on a single user but adds voice interactions and ambient audio. There is currently no public demo, just video samples and a technical document.

Google's Genie 3 is currently the most well-known competitor among world models. It produces visually more impressive environments but only supports a single user. Veo 3 and Sora 2, on the other hand, generate short, standalone videos without control during playback.

Agora-1 is playable as an early research preview on Odyssey's website. The team, led by CEO Oliver Cameron and CTO Jeff Hawke, is also considering applications in collaborative robotics where multiple robots need to reason together about actions and spaces, as well as in training AI agents in fully simulated environments.