Building an ML Language from Scratch: Introducing Charl

I spent the last year building Charl - a programming language designed specifically for machine learning. Not a library on top of Python, but a language where tensors and autograd are native features.

Why?

PyTorch and TensorFlow are excellent, but they're libraries bolted onto general-purpose languages. I wanted to explore: what's possible when ML is built into the language itself?

What Works Today

• Native tensor operations
• Automatic differentiation (dynamic graphs)
• Neural network training (validated on MNIST)
• 22x faster than PyTorch CPU
• GPU support via wgpu
• Type-safe (static typing with inference)

Example: Training a Neural Network

// Network: 2 -> 4 -> 1
let w1 = tensor_randn([2, 4])
let w2 = tensor_randn([4, 1])

while epoch < 1000 {
    // Forward
    let h = nn_relu(nn_linear(x, w1, b1))
    let pred = nn_sigmoid(nn_linear(h, w2, b2))
    let loss = loss_mse(pred, target)

    // Backward (automatic)
    tensor_backward(loss)

    // Update
    w1 = optim_sgd_step(w1, tensor_grad(w1), 0.5)
    w2 = optim_sgd_step(w2, tensor_grad(w2), 0.5)
}

Result: XOR converges from 0.259 → 0.006 loss (99% accuracy).

Current Limitations

• Alpha quality - expect bugs
• Small ecosystem
• Missing features (modules, generics)
• Not production-ready

This is a research experiment, not a PyTorch replacement.

Try It

• Website: https://charlbase.org
• GitHub: https://github.com/charlcoding-stack/charlcode
• Docs: Full API reference + 20+ examples

Looking For

• Testers (try it, break it, report issues)
• Feedback (what's confusing? what's missing?)
• Ideas (what would make this useful?)

Questions?

1. Is the syntax intuitive for ML work?
2. What's the first thing you tried that didn't work?
3. What ML use case would make you actually use this?

Genuinely curious about feedback. This is an experiment - let's see where it goes.

machinelearning #programming #opensource