Salt: Systems Programming, Mathematically Verified (salt-lang.dev)

Salt, a new systems programming language, has been launched with a focus on mathematically verified safety, offering significant benefits for high-performance computing. Designed to enhance tasks like optimizing LLM inference and managing bare-metal unikernels, Salt leverages Z3, a theorem prover, to assure that function contracts hold at build time without adding runtime overhead. Notably, Salt eliminates the need for garbage collection and complex borrow checking by using arenas for memory allocation, which are freed in bulk, ensuring that no references outlive their contexts. The implications for the AI and machine learning community are profound, as Salt demonstrates its capabilities through various implementations, including an inference engine for Llama 2 and a neural network for training on the MNIST dataset. Salt achieves performance that rivals traditional languages like C and Rust, meeting or exceeding their benchmarks across multiple scenarios. The language also incorporates an intuitive syntax with easy-to-use features like pipe operators, making it accessible while maintaining rigorous safety through compile-time checks. By addressing foundational challenges and promoting performance and safety, Salt is poised to impact the development of resource-intensive AI applications significantly.