Sutra: A Programming Language That Compiles to PyTorch Neural Networks

Researchers have created a new programming language called Sutra. It's a typed and purely functional language that compiles into a PyTorch neural network during its forward pass. The compiler integrates the whole program—including its primitives and control flow—into a single tensor-op graph built on a fixed embedding substrate. Operations like rotation binding, unbind, and polynomial Kleene logic are converted into tensor operations, with Kleene connectives shown as Lagrange-interpolated polynomials that accurately reflect the {-1, 0, +1} truth grid. To validate its effectiveness, two tests were conducted: one involved running the same program on four frozen embeddings across two modalities, achieving 100% accuracy with width k=8, and the other confirmed PyTorch autograd processed the computations correctly.