Cue responses from all the people a week ago moaning about it not being open source on day one, despite there being a precedent from Chris Lattner that it would all go down exactly like this.
Mojo is now open source
101–108 of 108 posts
Re: Mojo is now open source
#102Love this intentionality: ” Our open source approach has been deliberate: we’ve found that small and tight-knit design teams (not committees) are the best for finding the “soul” of a language, but that feedback from a broader community is essential to escape an echo chamber. As such, we first open-sourced the Mojo standard library, then released hundreds of thousands of lines of kernel code written in Mojo, tools, an…
I wish they never abandoned the goal of being a true Python superset. That would have been something.
It is still a goal. Mojo's semantics are still driven by Pythons, and can only deviate where safety or performance is an issue. Eventually it'll get the dynamic parts of Python
Re: Mojo is now open source
#103Earlier quoted context omitted.
It was in their roadmap: https://forum.modular.com/t/mojo-as-a-python-superset/2490
I think it was there for marketing purposes and that there was little actual intention to work on it. From the roadmap: >Mojo may or may not evolve into a full superset of Python, and it's okay if it doesn't.
That's a bold claim, I would be surprised if that were why they made the decision, there's other Python-like languages that don't need to pretend to be a superset to be successful.
Re: Mojo is now open source
#104Earlier quoted context omitted.
I think people underestimate the fact that Nvidia's GPUs are really good. Modular/Qualcomm's stack will provide a lot more optionality, but from what I saw in the Modcon conference on Tuesday, most of the target audience is towards other hardware vendors. At the moment, after designing a new hardware, you'd have to also build the entire software stack already existing on Cuda, ROCm, etc - this can double or triple yo…
For AI? GPUs actually kinda suck. NVIDIA’s GPUs are best, and their software stack is great, but they aren’t really adapted to AI (training or inference).
I wouldn't say they suck, but they're general-purpose. Anyway, you've gotten into the core of what makes Modular's stack useful. GPUs are great, but it has become obvious that they leave performance on the table, so dedicated AI accelerators get built; however, these don't have software infrastructure and kernels that have been tuned and optimised for years, like CUDA. The bet here is that you can build great hardware, and with a little bit of effort you can unlock all the necessary software for your hardware.
Re: Mojo is now open source
#105Earlier quoted context omitted.
For AI? GPUs actually kinda suck. NVIDIA’s GPUs are best, and their software stack is great, but they aren’t really adapted to AI (training or inference).
I wouldn't say they suck, but they're general-purpose. Anyway, you've gotten into the core of what makes Modular's stack useful. GPUs are great, but it has become obvious that they leave performance on the table, so dedicated AI accelerators get built; however, these don't have software infrastructure and kernels that have been tuned and optimised for years, like CUDA. The bet here is that you can build great hardwar…
Even general purpose matrix multipliers built for large linear systems instead of 3D graphics would get you orders of magnitude improvement. That’s essential what the current generation of AI accelerators are. Dedicated hardware could improve that further. I think that counts as “kinda sucks” in the absolute sense. It’s just far faster and more efficient than any other thing you can buy off the shelf today.
Re: Mojo is now open source
#106Earlier quoted context omitted.
I think it was there for marketing purposes and that there was little actual intention to work on it. From the roadmap: >Mojo may or may not evolve into a full superset of Python, and it's okay if it doesn't.
That's a bold claim, I would be surprised if that were why they made the decision, there's other Python-like languages that don't need to pretend to be a superset to be successful.
Being a superset of Python may not be necessary to be successful, but perhaps non-committally proposing it was necessary for Modular to win the funding they wanted. While lots of people indirectly run GPU kernels with PyTorch, I think the number of people wanting to write their own kernels is considerably smaller and that this niche already has established players (plenty of whom even already build on top of MLIR), making Mojo's most obvious use case difficult to sell.
Re: Mojo is now open source
#107Earlier quoted context omitted.
I was excited to try Julia, but then, as I saw it was an interpreted language, with bulky runtime, slow startup times and bloated library sizes - my initial interest quickly faded. For the same reasons I see languages like Python, Java, C# as inferior.
How is Julia an interpreted language? I sometimes wish there was a robust interpreter for Julia when the compilation latency is not worth the execution speed.
I still classify JIT languages as interpreted, because they require a runtime and it kills the convenience of sharing small binary e.g. in C or Nim.
And Julia docs are also unambigious in stating that code is sometimes interpreted[0]:
> jl_toplevel_eval_flex() then uses some simple heuristics to decide whether to JIT compile the AST or to interpret it directly.
[0] - https://docs.julialang.org/en/v1/devdocs/eval/#Julia-Executi...
Re: Mojo is now open source
#108Earlier quoted context omitted.
Chris Lattner is now Qualcomm's VP of AI Software and Tooling, so he is now responsible for all of Qualcomms AI software. Mojo wouldn't be doing any dying
Chris Lattner is now very very rich. Once his lockup expires he will not need to work for anyone. I'm sure Qualcomm would love him to stick around but suspect that they'll maintain mojo even if he leaves.
he was rich already with all his previous jobs + all the funding the company got