Show HN: Charl – ML language with native tensors and autograd (charlbase.org)

Charl is a new programming language designed specifically for ML/ML engineering where tensors, automatic differentiation, neural layers and optimizers are first‑class language features rather than external libraries. The announcement shows native tensor types with built‑in operations (add, mul, matmul, reshape, transpose), automatic memory management, a computation‑graph autograd (tensor_backward), and direct support for common layers (Linear, Conv2D, Pooling, BatchNorm, LayerNorm, Dropout) and activations (ReLU, Sigmoid, Tanh, Softmax, GELU). Losses (MSE, cross‑entropy) and optimizers (SGD, Adam, RMSProp with momentum/adaptive rates) are integrated; there’s also a WGPU backend for GPU acceleration and cross‑platform device support (Vulkan, Metal, DirectX) with explicit tensor transfer between CPU and GPU. Example snippets show a complete training loop (forward, loss, tensor_backward, adam_step) implemented idiomatically in language syntax. For the AI/ML community this matters because bringing tensors and autograd into the language core can reduce friction, improve type safety, and open new optimization opportunities (compiler‑level graph/agent optimizations, better memory/aliasing analysis). Native support also streamlines experiments and deployment across platforms. Tradeoffs include ecosystem maturity, library bindings, and benchmarking vs established frameworks (PyTorch, TensorFlow, JAX). If Charl delivers robust performance and tooling, it could simplify model development pipelines and enable more compact, safer ML codebases — but adoption will hinge on interoperability, community libraries, and real‑world benchmarks.