Chronological feed of everything captured from Luke Zettlemoyer.
Tuna-2 challenges the prevailing assumption that pretrained vision encoders (e.g., VAEs, CLIP-style representation encoders) are necessary for competitive multimodal understanding and generation. By replacing modular encoder designs with simple patch embedding layers operating directly on pixels, Tuna-2 achieves state-of-the-art performance on multimodal benchmarks while enabling fully end-to-end optimization. Although encoder-based variants converge faster in early pretraining, the encoder-free architecture scales more effectively, particularly on fine-grained visual perception tasks. This suggests that unified pixel-space learning is a viable and scalable alternative to latent-space approaches for both generation and perception.