According to @KyeGomezB, Fast Byte Latent Transformer uses diffusion at byte level for parallel decoding, slashing inference passes and latency.