Mojo 🔥 — a new programming language for AI developers (first look)

Key Takeaways

hi everyone I'm Patrick and welcome back on my channel so yesterday Chris lattner the co-founder of llvm and the Swift programming language announced a new programming language called mocho which Builds on top of python and is designed specifically for AI applications and in their keynote they claim that it's up to 35 000 times faster than python so let's take a quick first look at it so Mojo combines the usability of python with the performance of C or C plus unlocking unparalleled programmability of AI hardware and extensibility of AI models so this is one of the main motivations we all know one of the main drawbacks of python is its speed and poor low level performance so what ends up happening is often we use python as a glue and have the actual implementations in C plus plus or c for example this is what happens with numpy or tensorflow or pytorch or for AI applications we even go a level deeper and write code in Cuda which is optimized for your tensor operations on this on the GPU so Mojo tries to unify all of this so here it claims write everything in one language no C plus plus or Cuda required and at the same time it should be as simple and familiar as python so let's have a look at some of its features so first of all Mojo is designed as a superset of pythons so you can also write python code and in fact the hello world looks the very same as in Python so a lot of the language features and the functions and the syntax are the same but then it has additional features on top to extend this so I will go over this in a few moments but what I also want to mention is that right now they mentioned this in their docs how compatible is Mojo with python really and at the moment it is still very early and missing many features but they want to work towards a full compatibility in the future so we have to wait how this will go so let's look at some of the features so here we have the softmax function in Python and we can rewrite this in numpy which looks similar for example we also have this Define a function and then we return something but here we have this struct which I will show you in a moment again so let's go over some of the features here first of all they say it has Progressive types so we can leverage types for better performance and error checking then it has zero cost abstractions so we can take control of storage by inline allocating values into structures so this is what I just showed you here we can define a struct which looks similar to a class but it has some differences so first of all for example you can also write your init function but then you might notice these typins here and with this we can take more control of the storage then we can have ownership and borrow Checkers so we can take advantage of memory safety without the rough edges and in fact they claim to be as safe as rust then you can have portable parametric algorithms so you can leverage compile time meta programming to write Hardware agnostic algorithms and reduce boilerplate and then you get language integrated Auto tuning so you can automatically find the best values for for your parameters and take advantage of the target hardware and then some more features they list here and I also want to point out the documentations that you can check out so for example they have the section about basic systems programming extensions so for example one of the extensions are the let and wear declarations that you can use on top of your normal python code of course so here you can use this for immutable and mutable types then you get these struct type that I already showed you which looks similar to a class but has some differences then you get strong type checking then you can overload functions so this is similar to C plus plus or Java in Swift so here you can write a function that has the same syntax as two init functions for example but it has different arguments and then here also you have this function definition so this is similar to a def keyword in Python and you can also write def function but then again here with this you get some more functionality and for this I recommend to check out the documentation then you get a copy in it and a move in its special method and then you also get this section about parameterization and yeah so again if you want to go into more depth then check out the documentation I will put the link in the description so how can you try Mojo and at the moment it is not yet publicly available so you have to sign up here and wait until you get access and once you have access you can try try it out via the so-called Mojo playground which is just a two-parter notebook with also the Mojo compiler in the background so so then here you can write for example also python code and then also cells with Mojo code and then try this out but again it is not yet publicly available so you cannot simply install this on your own machine so you have to use it via their playground and also it is not yet open source so they have a repository and here they write we plan to open source Mojo progressively over time but it's changing very quickly now and we believe that a small tight-knit group of Engineers with a shared Vision can move faster than a community effort so we also have to wait until this will go open source so yeah this is just a very short overview of Mojo at first glance it looks and sounds very exciting but as you can see here it is still very early so we have to wait and see so yeah let me know in the comments what you think about Mojo if you also think it could change the way how we write AI applications in the future and if you want to try it out yourself and also if you want to see more tutorials in the future on my channel let me know and then I hope to see you in the next video bye