Joint embedding architectures that do not attempt to reconstruct perform much better than reconstruction-based methods (autoencoders, VAEs, masked autoencoders, diffusion models) for learning image representations that transfer well to downstream tasks.