GAIA-2 (Wayve)

Key Claims

• Latent diffusion world model — generates multi-camera driving video from structured conditioning
• Controllable generation — conditioned on ego-vehicle dynamics, agent configurations, environmental factors (weather, time of day), road semantics
• Multi-camera spatiotemporal consistency — crucial for AV use since perception stacks rely on synchronized multi-camera input
• Three-geography training — UK, US, Germany — demonstrates cross-region generalization
• Addresses core AV-specific challenges that general-purpose generative models miss: multi-agent interactions, fine-grained control, multi-camera consistency

Why This Matters

GAIA-2 is the strongest commercial validation of generative world models from the AV industry. Where Genie 3 is a general-purpose demo with research-preview access, GAIA-2 is a production tool at an AV company used for sim-to-real training, counterfactual scenario evaluation, and rare-event data augmentation.

Key commercial signal: Wayve launched GAIA-3 in 2026 explicitly positioning it as advancing "from simulation to evaluation" — suggesting generative world models are moving from data-augmentation tools to first-class components of the AV evaluation pipeline.

Positioning vs. Other Camps

• Direct counter-example to LeCun's critique: GAIA-2 is pixel-space generation, not JEPA, yet produces operational value at a deployed AV company
• Architectural opposite of V-JEPA 2: GAIA-2 generates high-res multi-camera video; V-JEPA 2 predicts abstract representations then collapses to control via V-JEPA 2-AC
• Reinforces the "split into two products" thesis (Thesis 4): generative dominates simulation/data augmentation; JEPA targets control

Notes

Lineage: GAIA-1 (arXiv 2309.17080, Sep 2023) → GAIA-2 (Mar 2025) → GAIA-3 (2026, Wayve press release, not yet an arXiv paper). Future compile pass should ingest GAIA-3 technical details when the paper drops.

Source: GAIA-2 by Wayve