We are building a next-generation AI developer platform for the world. Read our latest post on how Mojo 🔥 - A systems programming language presented at LLVM 2023
Rust isn’t good for fast iteration (Python with GPU-accelarated is still the winner here). Mojo could indeed turn out to be the next big thing for AI programming. But there are a few issues:
The #1 thing IMO, Modular still hasn’t released the source for the compiler. Nobody uses a programming language without an open-source compiler! (Except MATLAB/Mathematica but that has its own controversies and issues. Seriously, what’s the next highest-ranking language on Stack Overflow rankings with only closed-source compilers/interpreters? Is there even any?)
Even if Mojo is successful, the VC funding combined with the compiler being closed-source mean that the developers may try to extract money from the users. e.g., by never releasing the source and charging people for using Mojo or for extra features and optimizations. Because VCs don’t just give money away. There’s a reason nearly every popular language is open-source, and it’s not just the goodness of people’s hearts…not just individuals, but big companies don’t want to pay to use a language, nor risk having to pay when their supplier changes the terms.
Modular/Mojo has lots of bold claims (“programming language for all of AI”, “the usability of Python with the performance of C”) without much to show for them. Yes, Mojo got 35,000 speedup from Python on a small benchmark; but writing a prototype language which achieves massive speedups without compromising readability on small benchmark, is considerably easier than doing so for a large, real-world ML program. It’s not that I expect Mojo to already be used in real-world programs, I don’t, I get that it’s in the early-stage. It’s that Modular really shouldn’t be calling it the “programming language for all of AI” when so far, none of today’s AI is written in Mojo.
What makes the Mojo developers think they can make something that others can’t? The field of programming languages is dominated by very smart, very talented people. So many, it’s hard to imagine Mojo developing anything which others won’t be able to replicate in a few months of dedicated coding. And if Mojo doesn’t open-source their compiler, those others will do so, both with funding from competing companies and universities and in their spare time.
With all that being said, I have decent hopes for Mojo. I can imagine that the bold claims and toy benchmarks are just a way to get VC funding, and the actual developers have realistic goals and expectations. And I do predict that Mojo will succeed in its own niche, as a language augmenting Python to add high-performance computing; and since it’s just augmenting Python, it probably will do this within only one or two years, using Python’s decades of development and tooling to be something which can be used in production. What I don’t expect, is for it to be truly groundbreaking and above the competition in a way that the hype sometimes paints it as.
And if they do succeed, they need to open-source the compiler, or they will have some competitor which I will be supporting.
Why so many people are negative to Mojo? As a ML Researcher, I really don’t like rust for fast development iteration.
Rust isn’t good for fast iteration (Python with GPU-accelarated is still the winner here). Mojo could indeed turn out to be the next big thing for AI programming. But there are a few issues:
The #1 thing IMO, Modular still hasn’t released the source for the compiler. Nobody uses a programming language without an open-source compiler! (Except MATLAB/Mathematica but that has its own controversies and issues. Seriously, what’s the next highest-ranking language on Stack Overflow rankings with only closed-source compilers/interpreters? Is there even any?)
Modular/Mojo has lots of bold claims (“programming language for all of AI”, “the usability of Python with the performance of C”) without much to show for them. Yes, Mojo got 35,000 speedup from Python on a small benchmark; but writing a prototype language which achieves massive speedups without compromising readability on small benchmark, is considerably easier than doing so for a large, real-world ML program. It’s not that I expect Mojo to already be used in real-world programs, I don’t, I get that it’s in the early-stage. It’s that Modular really shouldn’t be calling it the “programming language for all of AI” when so far, none of today’s AI is written in Mojo.
With all that being said, I have decent hopes for Mojo. I can imagine that the bold claims and toy benchmarks are just a way to get VC funding, and the actual developers have realistic goals and expectations. And I do predict that Mojo will succeed in its own niche, as a language augmenting Python to add high-performance computing; and since it’s just augmenting Python, it probably will do this within only one or two years, using Python’s decades of development and tooling to be something which can be used in production. What I don’t expect, is for it to be truly groundbreaking and above the competition in a way that the hype sometimes paints it as.
And if they do succeed, they need to open-source the compiler, or they will have some competitor which I will be supporting.