The next frontier of AI is the physical world—spanning embodied AI, industrial simulation, and scientific prediction. Because the physical world changes at millisecond timescales, models must refresh predictions of environments and actions at ultra-high frequencies.
Why do traditional inference paradigms fall short, and how do Action-Conditioned World Models (ACWM) & World Action Models (WAM) differ?
Token-by-token autoregression (LLMs): Strictly serial—each token must wait for the previous token to pass through all Transformer blocks, locking output frequency to single-token latency regardless of added compute.
Autoregressive denoising (ACWM / WAM): Generates continuous frames in chunks from pure noise. Once a chunk completes a single denoising step, the next chunk can immediately read its noisy KV cache and execute its own corresponding step—allowing adjacent chunks to overlap in parallel and scale freely with compute.