Lush: my favorite small programming language

A blog post all about Lush, a small language that the authors present as a “Lisp Universal SHell” whose main features are its ability to integrate simply with underlying libraries with the interactivity of Lisp. And it’s a full Lisp, with a compiler, a macro system, an object system, and (importantly for what comes next) high-performance arrays and tensors.

Lush’s main claim to interest is that its authors include Yann LeCun, Yoshua Bengio, and the other researchers who went on to have a massive influence on the resurgences of interest in deep learning neural networks. Indeed, the first version of the Torch package, which later became PyTorch, was written in Lush. Lush’s arrays make use of a sub-language modelled on APL that leverage Lush’s ease with foreign functions to call into underlying C kernels for tensor operations.

This is a great example of how Lisp’s claim to fame as a language for doing AI has continued into the “new” AI era of deep learning models. It also shows that, if you believe that Lisp’s dynamism and flexibility are an advantage to you, there are no barriers to continuing to benefit from that even if you need to access a wider ecosystem of tools. It’d be interesting to understand the performance trade-offs in this: what is the performance of a Lush algorithm that makes extensive use of its array language, and how can this performance be maximised using higher-level language constructs of the sort Lisp encourages?