GenAI generates GenAI – A Diffusion Model That Generates Neural Networks (github.com)

Researchers have introduced NNGen, a groundbreaking diffusion model that generates generative AI models instead of images. Using a text-to-image architecture, NNGen takes textual descriptions to produce working neural networks with specific functionalities, effectively mapping English instructions to neural network weights. For example, when prompted with "a tiny MLP drawing agent that traces a five-pointed star," NNGen outputs a new neural network capable of executing that task. This innovation elevates generative modeling by creating neural networks directly from abstract textual representations, showcasing a novel application of existing diffusion techniques. The significance of NNGen lies in its potential to streamline the development of specialized neural networks without the need for extensive human-labeled datasets. Instead, it self-generates a corpus by training numerous controllers in simulated environments, employing a compact text-conditioned UNet to denoise weight canvases. The model is versatile, capable of performing various tasks, such as drawing and game behaviors, with a focus on measuring compositional generalization. This approach not only demonstrates the feasibility of generating AI models from user-defined tasks but also highlights the evolving landscape of generative AI, where even the architecture of intelligent systems can be produced on-demand through refined algorithms and computational creativity.