Show HN: Self-growing neural networks via a custom Rust-to-LLVM compiler (github.com)

NOMA, an experimental systems programming language tailored for machine learning, has been announced, showcasing its innovative approach to neural network architecture. NOMA treats both model training and topology changes as intrinsic language features, allowing for dynamic growth of neural networks during training. Unlike traditional frameworks that require rebuilding graphs whenever the model structure is altered, NOMA makes learnable parameters explicit memory buffers that support real-time resizing. This design modification preserves the optimizer state upon expansion, enabling continuous training without disruption. The significance of NOMA lies in its potential to enhance workflow efficiency within the AI/ML community. By implementing reverse-mode automatic differentiation as a compiler pass, NOMA ensures that gradients are derived at compile time, resulting in robust performance improvements. Its compiled programs generate small, standalone binaries, contrasting sharply with the overhead typically seen in existing frameworks like PyTorch or JAX. In benchmark tests, NOMA demonstrated competitive execution times and superior accuracy in retaining optimizer state, thereby reinforcing the utility of language-level semantics for differentiable programming and dynamic architecture evolution. This project invites further exploration into compiler-based automatic differentiation and adaptive neural network structures, marking a pivotal advancement in the landscape of machine learning frameworks.