Mojo Outperforms Rust?

There was an article floating around saying that the new "Mojo" programming language outperforms Rust.

Stating that something outperforms Rust is nothing new. It's often used as clickbait because it works.

Conversations about performance typically center on programming languages' frontends rather than underlying systems, and issues usually lie in flawed implementations.

Rust is an LLVM (and/or Cranelift?) backed language. You can achieve any performance you want, though you might have to write code with the performance in mind, write inline assembly or hints for the compiler. Saying that Rust is slower than C/C++, Zig, Mojo, or whatever is meaningless, because all of them can be compiled to the same machine code.

See further discussion here: https://reddit.com/r/rust/comments/1al8cuc/modular_community_spotlight_outperforming_rust/.

Mojo is meant to be a Python-like language, or a faster Python. The obvious target market is AI.

But, does anybody actually care about performance in AI? Why would Rust or anything other than Python be considered for AI anyway?

In ML/AI, a significant portion of compute is performed on the GPU. The written code is usually highly experimental, ugly and spaghettified. Python is great for this kind of work, and with tools like Google Colab, you can move blazingly fast. Why would you need an efficient systems programming language? While doing AI, you need to write lots of ugly code quickly. In my opinion, the idea of Python will stay, and Mojo can make it more efficient.

There's also a difference between ML in a Jupyter notebook, serving learned models, online/incremental learning, learning on streams of data, recommendation systems, big data, processing video, etc. Some parts could need extra performance, and Rust can help.

Here are some interesting Rust projects to look at:

• Qdrant (https://qdrant.tech), a high-performant vector database. It is written in Rust. It powers X's "See similar posts" feature, afaik.
• Arroyo (https://arroyo.dev) and Vector by DataDog (http://vector.dev) for stream processing.
• Burn (https://github.com/tracel-ai/burn) a deep learning framework.
• tch-rs (https://github.com/LaurentMazare/tch-rs) Rust bindings for PyTorch.
• rust-bert (https://github.com/guillaume-be/rust-bert) for running NLP transformers models for summarization, sentiment analysis, embeddings, etc.
• Candle (https://github.com/huggingface/candle) a minimalist ML framework.