Given that NVidia now decided to get serious with Python JIT DSLs in CUDA as announced at GTC 2025, I wonder how much mindshare Mojo will managed win across researchers.
There is also Julia, as the black swan many outside Python community have moved into, with much more mature tooling, and first tier Windows support, for those researchers that for whatever reason have Windows issued work laptops.
Mojo as programming language seems interesting as language nerd, but I think the judge is still out there if this is going to be another Swift, or Swift for Tensorflow, in regards to market adoption, given the existing contenders.
Mojo (and Modular's whole stack) is pretty much completely focused at people who are interested in inference, not training nor research so much at this moment.
So going after people who need to build low latency high-throughput inference systems.
Also as someone else pointed out, they also target all kinds of hardware, not just NVidia.
What about Candle, made by Huggingface? Seems to at least allow the basics and has lots of examples, all of them run on both CPU and GPU. Haven't dived deeper into it, but played around with it a bit and found it good enough for embedding purposes at least.
I think the big value add of Mojo is that you are no longer writing GPU code that only runs on one particular GPU architecture.
In the same way that LLVM allows CPU code to target more than one CPU architecture, MLIR/Mojo allows GPU code to target multiple vendor's GPUs.
There is some effort required to write the backend for a new GPU architecture, and Lattner has discussed it taking about two months for them to bring up H100 support.
Currently looks more like CPUs and eventually AMD, from what I have been following up on their YouTube sessions, and whole blog post series about freedom from NVidia and such.
There's pretty broad support for server-grade and consumer GPUs. It's a bit buried, but one of the most reliable lists of supported GPUs is in the Mojo info documentation.https://docs.modular.com/mojo/stdlib/gpu/host/info/
Already GPU code, kernels, and complete models can run on datacenter AMD GPUs using the same code, the same programming model, and same language constructs.
not sure, modular is focusing mainly on enterprise applications. but if you look at the current PRs you can see people hacking support for standalone consumer-grade Nvidia and AMD gpus because it is easy, you just add the missing or different intrinsics for the architecture in the lowest level (in pure mojo code) and wire it up in a few places and voila you already program and run code on this GPU. iGPU/Apple GPUs are still not supported yet but it would interesting to see their integration
Julia has GPU compilers for Nvidia, AMD, Intel, and Apple, and we have KernelAbstractions.jl for writing a kernel that is portable between all of them (plus the CPU!)
Just as LLVM doesn't automatically have a backend or every new CPU architecture, Mojo/MLIR doesn't automatically have a backend for every new CPU/GPU/TPU.
However, writing an LLVM backend for RISC-V sure did add support for a whole lot of different programming languages and the software you have access to through them in one fell swoop.
The same is true here.
Instead of rewiting all your GPU code every time you need to target a new GPU/TPU architecture, you just need a new backend.
Chris Lattner (the tech lead behind Mojo, LLVM, Clang, Swift and MLIR) appeared on a podcast a bit over a week ago and discussed the state of Mojo and where it is going.
He also discussed open sourcing Mojo and where the company expects to make its money.
Funny how the already weak case for not working on Julia instead of creating a new language is becoming even more flimsy :
FAQ:
> Why not make Julia better?
> We think Julia is a great language and it has a wonderful community, but Mojo is completely different. While Julia and Mojo might share some goals and look similar as an easy-to-use and high-performance alternative to Python, we’re taking a completely different approach to building Mojo. Notably, Mojo is Python-first and doesn't require existing Python developers to learn a new syntax.
>We oversold Mojo as a Python superset too early and realized that we should focus on what Mojo can do for people TODAY, not what it will grow into. As such, we currently explain Mojo as a language that's great for making stuff go fast on CPUs and GPUs.
> Julia: Julia is another great language with an open and active community. They are currently investing in machine learning techniques, and even have good interoperability with Python APIs.
The share of projects that start out “we’re going to do X, and conpletely support using the syntax and semantics of existing language Y, but add more on top” that end up “we’re going to do X, in a language vaguely reminiscent of Y” is very hard to distinguish from 100%.
I think the real reason is that Chris Lattner doesn’t want to work on Julia. He’s likely very (and justifiably so) opinionated about these topics, and probably wants a certain degree of creative authority over his projects.
Although from the way some Modular videos are done, it seems nowadays he is more in an overseer role and setting up the direction, while others are the ones actually pushing Mojo and Max tooling forward.
Maybe I just misunderstand it from the presentation format.
He works on Mojo a lot, mostly on weekends. In the past few months, he has worked on strings, collection literals, dependent types, reference captures,
comprehensions, and many other nice language features.
Additionally Julia works on Windows, lots of its issues have been ironed out throught the last decade, and many folks are invested into it, instead of a single company product.
Yes many of the mainstream languages started as single company product, but lets put it this way, would anyone be writing one of such languages today, had those not been languages gatekeeped to access a specific platform?
So outside accessing Max and its value preposition as product enabler for XYZ, who would be rushing to write Mojo code, instead of something else.
I think the modular proposition is to solve the AI infra problem using a modern systems programming language but with a python syntax (to keep the mental overhead low) it is mainly trying to replace C++/CUDA. I am not sure Julia is suitable for such endeavour. Julia is not marketed as a systems programming language and I find it hard to believe that it can be one as dyamically typed langauge with GC and JIT.
Could Julia replace python in the high-level dynamic code space used in research and training? maybe. but I find it really hard to believe that it can replace CUDA/ROCm/C++ .. etc.
Lisp Machines from Xerox PARC, Texas Instruments and Genera proved their point, even if the market wasn't up to their price point versus UNIX graphical workstations.
Dylan was going to be Newton's system programming language, and while the language group lost the the C++ team (Apple had two competing teams for the Newton OS), it was still NewtonScript for everything userspace, and it was getting a JIT by the time the project was canceled.
Objective-C is dynamically typed beyond the common subset with C, and was used even to write NeXTSTEP drivers.
I don't know how much of a chance Julia has against CUDA/ROCm/C++, especially now that everyone on the GPU space has decided to give feature parity to Python on their hardware, via day one bindings to the compute libraries and JIT DSLs, so that makes Mojo even less of a chance than Julia has.
Julia has an established ecosystem, and presence on the scientific community with ties to MIT.
Python is the champion, and most folks writing CUDA/ROCm/C++ are already using it.
So who would be reaching out to Mojo, instead of Python JIT DSLs/bindings or Julia, when having Fortran, C, C++ allergy?
Creating DSL/Bindings for Julia or python to underlaying platform like CUDA is not really replacing it, it is yet adding another layer on top of the existing platforms or creating a prototyping envirnoment for research. The question is not can Julia interface with CUDA, it is can Julia replace C++/CUDA/ROCm in an end-to-end scenarios (preferably with portability across GPUs vendors)? if not then there is no comparison between the stated goals of Julia and Mojo. they are different langauges targeting completly different use cases.
Julia does not just have bindings to CUDA. Native Julia code can compile to build .ptx kernels https://cuda.juliagpu.org/stable/development/kernel/. This same code can also generate kernels for AMD GPUs, Intel GPUs, and Metal.
We for example built software that generates kernels on-demand that embed user functions for all 4 of these systems and showed it's much faster than just CUDA bindings for array functions for certain nonlinear systems (https://www.sciencedirect.com/science/article/abs/pii/S00457...)
Julia is much more like a static language than you might realize. In fact, within a fixed world-age, as far as julia's JIT is concerned, it is a static language. Our JIT also isn't like other JITs, we sometimes call it a "Just Ahead Of Time" compiler because it is built much more like a traditional compiler than a tracing JIT.
We have quite fantastic GPU compilation stuff too, and julia functions can be compiled to Nvidia, AMD, Intel, and Apple GPUs through their respective GPU compiler packages, and one can use KernelAbstractions.jl to write code that is GPU vendor agnostic and works on all of them.
We're also getting an (experimental) fully ahead-of-time compiler built into the language with v1.12 that spits out an executable or dylib.
It's the Latent Space podcast, published June 13. It's a pretty decent podcast for keeping on top of AI stuff. Swyx, one of the co-hosts, is active here on HN.
From Chris Lattner on Modular discord few days ago:
Yep, that's right. Int behaving like a machine integer is very important for systems performance. Leaving the "int" namespace untouched allows us to have a object-based bigint in the future for compatibility with python.
As others have mentioned above, it is still a goal to be compatible with python in time, just not a short term priority
I guess they abandoned the python superset idea? I followed them for a bit when they first publicly launched and they said "don't worry, we'll be a real python superset soon" and the biggest omission was no support for classes. A few years later, it looks to be missing the same set of python features but added a lot of their own custom language features.
It was highly aspirational goal, and practically speaking it's better right now to take inspiration from Python and have stronger integration hooks into the language (full disclosure, I work at Modular). We've specifically stopped using the "superset of Python" language to be more accurate about what the language is meant for right now.
Being a Python superset and being fast are fundamentally in tension. It would be possible, maybe, to have a Python superset where you get the highest performance as long as you avoided the dynamic features of Python. However, I suspect it would make the user base grumpy to have a hidden performance cliff that suddenly showed up when you used dynamic features (or depended on code that did the same).
The dynamic features of Python are no different from the dynamic features of Smalltalk, Self, Common Lisp, but people have been educated to expect otherwise due to the adoption failure of dynamic compilers in Python community.
It's not a specific bitwidth in a sense that it maps to whatever largest integer type is natively supported by the target architecture (i.e. basically 32-bit or 64-bit).
The Python superset concept was always a gimmick. The goal was always to juxtapose Python with a language that superficially looks like Python in order for you to completely migrate from Python to Mojo. It is just providing a smooth ramp for you to do so in the same Apple migrated folks from Objective-C to Swift.
Python is a hacky language that was never designed other than laying it's eggs against the grain of what we seem to mostly agree is godo — e.g. functional programming and composition.
Big tech spends a lot of money to avoid python in critical infra.
I'm glad the blazing-fast tech is finally making it to python. Those guys were pretty busy with Rust for like the past 10 years. I'm glad they are finally starting to making a blazing-fast python stuff.
I am really rooting for Mojo. I love what the language is trying to do, and making it easier to run SOTA AI workloads on hardware that isn't Nvidia + CUDA will open up all kinds of possibilities.
I'm just nervous how much VC funding they've raised and what kind of impacts that could have on their business model as they mature.
If they can manage to make good on their plans to open-source it, I'll breathe a tentative sigh of relief. I'm also rooting for them, but until they're open-source, I'm not willing to invest my own time into their ecosystem.
They already released their code under the Apache 2.0 license. Not everything in their stack is open source but the core things appear to be open source.
One of the big differences here is that the Mojo/Python interop package is doing a lot of the heavy lifting of compiling and loading the code. The goal here is to have a language that's close to Python that gives you hardware acceleration as you need it, without having to configure and build C++ code for every platform.
We've run a few events this year (disclosing again that I work for Modular, to put my comments into context), and we've had some great feedback from people who have never done GPU programming about how easy it was to get started using Mojo.
C++ is a powerful and mature language, but it also has a steep learning curve. There's a lot of space for making GPU (and other high performance computing) easier, and platforms/languages like Triton and Julia are also exploring the space alongside Mojo. There's a huge opportunity to make GPU programming easier, and a bit part of that opportunity is in abstracting away as much of the device-specific coding as you can.
I was a HPC C++ programmer for a long time, and I always found recompiling for new devices to be one of the most painful things about working in C++ (for example, the often necessary nightmare of cmake). Mojo offers a lot of affordances that improve on the programming experience at the language, compiler, and runtime levels.
I am not that intrigued that Python that can call some pre-compiled functions, this is already possible with any language that produces a dynamic library.
The space that I am interested in is execution time compiled programs. A usecase of this is to generate a perfect hash data structure. Say you have a config file that lists out the keywords that you want to find, and then dynamically generate the perfect hash data structure compiled as if those keywords are compile time values (because they are).
Or, if the number of keywords is too small, fallback to a linear search method. All done in compile time without the cost of dynamic dispatch.
Of course, I am talking about numba. But I think it is cursed by the fact that the host language is Python. Imagine if Python is stronger typed, it would open up a whole new scale of optimization.
I would rather image CPython being like Common Lisp, Scheme, Raket, Smalltalk, Self compilation model.
Sadly the contenders on the corner get largely ignored, so we need to contend with special cased JIT DSLs, or writing native extensions, as in many cases CPython is only implementation that is available.
Jitting like you mentioned is supported by the MAX graph API: https://docs.modular.com/max/tutorials/build-custom-ops. It could have a nicer syntax though to be more like Numba, I think you have an interesting idea there.
> I am not that intrigued that Python that can call some pre-compiled functions, this is already possible with any language that produces a dynamic library.
> The space that I am interested in is execution time compiled programs. A usecase of this is to generate a perfect hash data structure. Say you have a config file that lists out the keywords that you want to find, and then dynamically generate the perfect hash data structure compiled as if those keywords are compile time values (because they are).
I'm not sure I understand you correctly, but these two seem connected. If I were to do what you want to do here in Python I'd create a zig build-lib and use it with ctypes.
Can Zig recompile itself if I change a config in production? I am talking about this
```
python program.py --config <change this>
```
It is basically a recompilation of the whole program at every execution taking into account the config/machine combination.
So if the config contains no keyword for lookup, then the program should be able to be compiled into a noop. Or if the config contains keyword that permits a simple perfect hash algorithm, then it should recompile itself to use that mechanism.
I dont think any of the typical systems programming allows this.
The thing I focus on when writing compiled extensions for Python isn't the speed of the extension, but rather the overhead of the call and the overhead of moving objects from Python -> compiled and compiled -> Python.
Is there a zero-copy interface for larger objects? How do object lifetimes work in that case? Especially if this is to be used for ML, you need to haul over huge matrices. And the GIL stuff is also a thing.
I've never been thats sold on Mojo, I think I'm unfairly biased away from it because I find new languages interesting, and its big sell is changing as little as possible from an existing language.
That said, importing into Python this easily is a pretty big deal. I can see a lot of teams who just want to get unblocked by some performance thing, finding this insanely helpful!
> its big sell is changing as little as possible from an existing language.
This is not really true. Even though Mojo is adopting Python's syntax, it is a drastically different language under the hood. Mojo is innovating in many directions (eg: mlir integration, ownership model, comptime, etc). The creators didn't feel the need to innovate on syntax in addition to all that.
You're right- I probably should have said something like "part of its sell" or "one of its selling points" or something.
I didn't mean to undermine the ambitious goals the project has. I still wish it was a little bolder on syntax though, Python is a large and complex language as is, so a superset of Python is inherently going to be a very complicated language.
The creators got burned on Swift for TensorFlow, their first MLIR project. One of the problems with that first venture under Google was that the language was not Python.
> as I'm definitely in the market for a simple compiled language that can offer Python some really fast functions
So, Nim? https://github.com/yglukhov/nimpy
The real point of Mojo is not the language, it's the deep roots into MLIR which is an attempt to do what LLVM did for compilers, and do it on GPUs / ML hardware. Chris Lattner is leading the project and he created LLVM and MLIR.
FTA (emphasis added): “Chris Lattner mentioned that Python can actually CALL Mojo code now”
So, the message is that it is possible to create nice Python bindings from Mojo code, but only if your Mojo code makes the effort to create an interface that uses PythonObject.
Useful, but I don’t see how that’s different from C code coding the same, as bindings go.
Both make it easier to gradually move Python code over to a compiled language.
Mojo presumably will have the advantage that porting from Python to Mojo is much closer to a copy paste job than porting Python to C is.
For a language that announced itself (and raised a lot of money on the premise of) claiming to be "a Python superset", this does not sound like a huge achievement.
In all fairness, their website now reads: "Mojo is a pythonic language for blazing-fast CPU+GPU execution without CUDA. Optionally use it with MAX for insanely fast AI inference."
So I suppose now is just a compiled language with superficially similar syntax and completely different semantics to Python?
I think it was pretty clear immediately that running python code was a far away goal. There was a lot more talk about lifetimes and ownership semantics than details about Python interop. Mojo is more like: Can we take the learnings of Swift and Rust and solve the usability and compile time issues, while building on MLIR to target arbitrary architectures efficiently (and call it a Python superset to raise VC money).
That said, the upside is huge. If they can get to a point where Python programmers that need to add speed learn Mojo, because it feels more familiar and interops more easily, rather than C/CPP that would be huge. And it's a much lower bar than superset of python.
It marketed itself explicitly as a "Python superset", which could allow Python programmers to avoid learning a second language and write performant code.
I'd argue that I am not sure what kind of Python programmer is capable of learning things like comptime, borrow checking, generics but would struggle with different looking syntax. So to me this seemed like a deliberate misrepresentation of the actual challenges to generate hype and marketing.
Which fair enough, I suppose this is how things work. But it should be _fair_ to point out the obvious too.
Absolutely. The public sales pitch did not match the reality. This is what I meant with the "Claim to be Ṕython to get VC money" point.
To first order, today every programmer starts out as a Python programmer. Python is _the_ teaching language now. The jump from Python to C/Cpp is pretty drastic, I don't think that it's absurd that learning Mojo concepts step by step coming from Python is simpler than learning C. Not syntactically but conceptually.
Maybe young generations have some issue learning polyglot programming, I guess.
While I agree using Mojo is much preferable to writing C or C++ native extensions, back on my day people learned to program in K&R C or C++ ARM in high school, kids around 12 years old, hardly something pretty drastic.
Many famous Speccy and C64 titles, written in Assembly, were written by bedroom coders between the ages of 14 and 16 years old, getting some pocket money writing them on the UK scene.
Get hold of Retro Gamer magazine for some of their stories.
I've tried learning C a couple times and given up because the curve is too steep to be worth the climb. It's not even the language itself, it's the inherited weight of half a century's worth of cruft. I can't spend weeks fighting with compiler nonsense, header files and #include. Screw it, I'll just use Go instead.
I'm learning Rust and Zig in the hope that I'll never have to write a line of C in my career.
Geez, what a comment. C is much much more simpler than Rust. You’re not supposed to be spending weeks fighting includes or compiler errors, that means you’re have some very basic misconceptions about the language.
Just read K&R “The C programming language” book.
It’s fairly small and it’s a very good introduction to C.
C syntactically is straight forward, but conceptually may be harder than Rust. You’re exposed to the bare computer (memory management, etc) far more than with a GC language or even Rust arguably, at least for simple programs.
Towards deployment is even harder. You can very easily end up writing exploitable, unsafe code in C.
If I were a Python programmer with little knowledge about how a computer works, I’d much prefer Go or Rust (in that order) to C.
This is true, but when you get something wrong related to the memory model in C, it just says "segfault". Whereas in Rust it will give you a whole explanation for what went wrong and helpful suggestions on how to fix it. Or at the very least it will tell you where the problem is. This is the difference between "simple" and "easy".
That applies only if you take "memory model" to mean modeling the effects of concurrent accesses in multithreaded programs.
But the term could also be used more generally to include stuff like pointer provenance, Rust's "stacked borrows" etc.
In that case, Rust is more complicated than C-as-specified. But C-in-reality is much more complicated, e.g. see https://www.open-std.org/jtc1/sc22/wg14/www/docs/n2263.htm
The model you're referring to, a Memory Ordering Model, is literally the same model as Rust's. The "exception" is an ordering nobody knows how to implement which Rust just doesn't pretend to offer - a distinction which makes no difference.
I do sympathize with the parent: The language itself might not be that difficult but you also have to factor in the entire ecosystem. What's the modern way to a build a GUI application in C? What's the recommended way to build a CLI, short of writing your own arg parser? How do you handle Unicode? How do you manage dependencies, short of vendoring them? Etc.
Errors too. When, inevitably, you make mistakes the C might just compile despite being nonsense, or you might get incomprehensible diagnostics. Rust went out of its way to deliver great results here.
I am not arguing about how good or easy it is to use C in production, I’m merely stating that parent complaints about weeks of insolvable errors and issues with includes screams that he needs to read some good resource like book, because he is definitely misunderstanding something important.
THe thing is, if one is an expert it is incredibly difficult to understand the beginner perspective. Here is one attempt:
C is simpler than Rust, but C is also _much_ simpler than Python. If I solve a problem in Python I have a good standard library of data types, and I use concepts like classes, iterators, generators, closures, etc... constantly. So if I move to Rust, I have access to the similar high-level tools, I just have to learn a few additional concepts for ressource management.
In comaprison, C looks a lot more alien from that perspective. Even starting with including library code from elsewhere.
I think one of the "Python superset" promises was that any particular dev wouldn't need to learn all of that at once. There could exist a ramp between Python and "fast python" that is more gradual than the old ways of dropping into C, and more seamless than importing and learning the various numpy/numba/polars libraries.
FWIW generics are already a thing in pure Python as soon as you add type annotations, which is fast becoming the default (perhaps not the least because LLMs also seem to prefer it).
I suppose if you accept the innocent-looking "#"+"#"=="##" then your example kind of algebraically follows. Next it's time to define what exp("#") is :)
* does different things depending on the types of the operands, which is Python's strong typing at work, not Perlesque weak typing. Repeating a string is a useful thing to be able to do, and this is a natural choice of syntax for it. The same thing works for lists: [1]*3 == [1, 1, 1].
It does unfortunately mean that sometimes `*` will work (and produce an incorrect result) rather than immediately failing loudly with a clear error message in the context in which it's actually intended to be numerical.
More broadly this is the same argument as whether overloading `+` for strings is a bad idea or not, and the associated points, e.g. the fact that this makes it non-commutative - the same all applies to `*` as well, and to lists as much as strings. At least Python is consistent here.
Although there is one particular aspect that is IMO just bad design: the way `x += y` and `x = y` work. To remind, for lists these are not equivalent to `x = x + y` and `x = x y` - instead of creating a new list, they mutate the existing one in place, so all the references observe the change. This is very surprising and inconsistent with the same operators for numbers, or indeed for strings and tuples.
The real unique selling point of Mojo is "CPU+GPU execution without CUDA", specifically, you write code that looks like code without worrying about distinctions like kernels and device functions and different ways of writing code that runs on GPU vs. code that runs on CPU, and mojo compiles it to those things.
While much has changed since then, the architecture is effectively the same. Julia's native CUDA support simply boils down to compiling via the LLVM .ptx backend (Julia always generates LLVM IR, and the CUDA infrastructure "simply" retargets LLVM to .ptx, generates the binary, and then wraps that binary into a function which Julia calls), so it's really just a matter of the performance difference between the code generated by the LLVM .ptx backend vs the NVCC compiler.
> For a language that announced itself (and raised a lot of money on the premise of)
claiming to be "a Python superset", this does not sound like a huge achievement
I feel like that depends quite a lot on what exactly is in the non-subset part of the language. Being able to use a library from the superset in the subset requires being able to translate the features into something that can run in the subset, so if the superset is doing a lot of interesting things at runtime, that isn't necessarily going to be trivial.
(I have no idea exactly what features Mojo provides beyond what's already in Python, so maybe it's not much of an achievement in this case, but my point is that this has less to do with just being a superset but about what exactly the extra stuff is, so I'm not sure I buy the argument that the marketing you mention of enough to conclude that this isn't much of an achievement.)
I've written this somewhere else before, Modular did not raise $130m to build a programming language, nobody does that.
They raised that much money to revolutionize AI infrastructure, of which a language is just a subset. You should definitely check some of the things they've put together, they're amazing
Yes. They are revolutionizing AI infrastructure but I guess a lot of world is just babbling about AI, but not every developer needs to worry about AI.
And so his improvements in mojo and now calling mojo code from python just make a lot more net positive to the community than being, some other Ai infrastructure company.
So I do wish a lot of good luck to mojo. I have heard that mojo isn't open source but it has plans to do so. I'd like to try it once if its as fast / even a little slower than rust and comparable to understanding as python.
Agreed, Modular is walking a very fine line, and they're doing so by trading on the reputation of Chris Lattner.
On the one had, as the other poster noted, no one raises $100M+ for a programming language; programming languages have no ROI that would justify that kind of money. So to get it, they had to tell VCs a story about how they're going to revolutionize AI. It can't just be "python superset with MLIR". That's not a $100M story.
On the other hand, they need to appeal to the dev community. For devs, they want open source, they want integration with their tools, they don't want to be locked into a IP-encumbered ecosystem that tries to lock them in.
That's where the tension is. To raise money you need to pretend you're the next Oracle, but to get dev buy-in you have to promise you're not the next Oracle.
So the line they've decided to walk is "We will be closed for now while figure out the tech. Then later once we have money coming in to make the VCs happy, we can try to make good on our promise to be open."
That last part is the thing people are having trouble believing. Because the story always goes: "While we had the best intentions to be open and free, that ultimately came secondary to our investors' goal of making money. Because our continued existence depends on more money, we have decided to abandon our goal of being open and free."
And that's what makes these VC-funded language plays so fraught for devs. Spend the time to learn this thing which may never even live up to its promises? Most people won't, and I think the Darklang group found that out pretty decisively.
I don't think investors look at what makes a net positive to the community when making large investments like in Modular. I was calling out the part of the post that said Modular raised a lot of Money to develop Mojo, that isn't entirely true as just creating a language isn't enough reason to invest $130m into a company, no matter how much net-positivity the language would bring.
It was never going to have Python semantics and be fast. Python isn't slow because of a lack of effort or money, it's slow because of all the things happening in the interpreter.
> Further, we decided that the right long-term goal for Mojo is to adopt the syntax of Python (that is, to make Mojo compatible with existing Python programs) and to embrace the CPython implementation for long-tail ecosystem support
The Python function is implemented in C and uses a faster algorithm [1], and this particular factorial is so small they put it in a lookup table [2]. It is a strange and very unequal choice for a demo.
My guess is that the slight overhead of interacting with mojo led to this speed discrepancy, and if a higher factorial (that was within the overflow limits etc) was run, this overhead would become negligible (as seen by the second example). Also similar to jax code being slower than numpy code for small operations, but being much faster for larger ones on cpus etc.
My impression is that mojo is not python, there are similarities, but under the hood mojo is much more similar to c++/rust.
As part of this it has a stronger type/lifecycle/memory model than python.
Maybe you could write, some level of transpiler, but so much of the optimizations rely on things that python does not expose (types), and there are things that python can do that are not supported.
I’m someone who should be really excited about this, but I fundamentally don’t believe that a programming language can succeed behind a paywall or industry gatekeeper.
I’m always disappointed when I hear anything about mojo. I can’t even fully use it, to say nothing of examine it.
We all need money, and like to have our incubators, but the LLVM guy thinks like Jonathan Blow with jai?
I don’t see the benefit of joining an exclusive club to learn exclusively-useful things. That sounds more like a religion or MLM than anything RMS ever said :p
> the LLVM guy thinks like Jonathan Blow with jai?
I would not compare Chris Lattner with Jonathan Blow. Lattner is a person with a reputation for delivering programming languages and infrastructure; whereas for Blow, it seems like an ego issue. He's built a nice little cult of personality around his unreleased language, and releasing it will kill a lot of that magic (running a language project with actual users that has to make good on promises is much different than running a language project with followers and acolytes that can promise anything and never deliver on it).
Lattner has a record of actually delivering dev products people can download and use. Mojo is closed source to make raising money easier, but at least you can actually use it. Jai isn't even available for people to use, and after a decade of Blow dangling it in front of people, it's not clear it'll ever be available, because I'm not sure he wants it to be available.
> Functions taking more than 3 arguments. Currently PyTypeBuilder.add_function() and related function bindings only support Mojo functions that take up to 3 PythonObject arguments: fn(PythonObject, PythonObject, PythonObject).
Lol wut. For the life of me I cannot fathom what design decision in their cconv/ABI leads to this.
There was a similar pattern in the Guava library years ago, where ImmutableList.of(…) would only support up to 20 arguments because there were 20 different instances of the method for each possible argument count.
> We have committed to open-sourcing Mojo in 2026. Mojo is still young, so we will continue to incubate it within Modular until more of its internal architecture is fleshed out.
> That rules it out of any production deployment until 2026 so.
Has that stopped everyone before? Java, C#/.NET, Swift and probably more started out as closed-source languages/platforms, yet seemed to have been deployed to production environments before their eventual open-sourcing.
I don't think Java (when it was owned by Sun) nor .NET (even currently) run the risk of a VC "our incredible journey" event causing "barrel bending" nor the backing company running out of money. In the first flavor, they'd want their pound of flesh and the "our new compiler pricing is ..." would be no good. In the latter, even if they actually opened the platform on the way out, it still would require finding a steward who could carry the platform forward, which is a :-( place to be if you have critical code running upon it
I guess the summary is that neither Java [at the time] nor .NET were profit centers for their owners, nor their only reason for existing
They certainly were, because no one other than a few hardlines were writing Java or C# and VB code in bare bones editors and compiling from command line, as only the bare bones SDKs were free beer, and for the desktop.
IDEs, implementations for embedded and phones, were all paid products, IDEs by developers or their employers, the others by OEMs.
My point is the early days, JCafe, Visual Age, Visual Studio, Forte, before free beer IDEs for them became common.
Java side with Eclipse/Netbeans, .NET side with the Visual Studio Express editions.
Yes, and there is a reason for that: Both are deeply integrated in Microsofts ecosystem, and whether one likes that or not, that ecosystem is the dominant platform for desktop computing, especially in commercial settings.
Yeah, I also get confused with references. I was annoyed from the start, when "Mojo" was announced as a Python family language. Mojolicious uses the "Mojo" namespace and is referred to as that quite often. I know Perl is not as popular as it used to be, but Mojolicious is probably the most popular framework of a language that is roughly in the same "space" as Python, so that naming choice was very ignorant IMHO.
"1001 Ways to Write CUDA Kernels in Python"
https://www.youtube.com/watch?v=_XW6Yu6VBQE
"The CUDA Python Developer’s Toolbox"
https://www.nvidia.com/en-us/on-demand/session/gtc25-S72448/
"Accelerated Python: The Community and Ecosystem"
https://www.youtube.com/watch?v=6IcvKPfNXUw
"Tensor Core Programming in Python with CUTLASS 4.0"
https://www.linkedin.com/posts/nvidia-ai_python-cutlass-acti...
There is also Julia, as the black swan many outside Python community have moved into, with much more mature tooling, and first tier Windows support, for those researchers that for whatever reason have Windows issued work laptops.
https://info.juliahub.com/industries/case-studies
Mojo as programming language seems interesting as language nerd, but I think the judge is still out there if this is going to be another Swift, or Swift for Tensorflow, in regards to market adoption, given the existing contenders.