Mind-bending new programming language for GPUs just dropped... 

Time to become a code bender.

Time to add 10 years of experience to my resume/CV

That video was *extremely* well done. Seems like someone has read the docs! Thanks for making it. We're long-term fans of your content here at HOC. If you or any other content creator wants to reach out, we're available to talk about the technology and answer any questions. We know there's a lot of work to do, but we're excited about it. You can expect Bend to become faster with every release. There are also many missing features (64-bit numbers, larger memory limit, etc.) that will be added very soon!

Folds have been a mainstay of the functional programming world for a very long time. I remember teaching students how to implement folds in Python a few years back and it blowing their minds. The basic concept that all loops are simply a subset of a wider class of iterative constructs called folds is transformational - I love the simplicity and inherent understanding of a fold. This new concept of a bend is incredibly interesting, as like the fold it appears to be able to emulate any kind of loop or iterative function, but instead of going from many to one, goes from one to many. I will be playing around with this over the next few weeks I am sure!

Watched this video on 8000 Cores in parallel, added 500h of bend experience to my resume

“1 week of problem, instead 7 days with 7 computers” 🤣🤣🤣

<a href="#" class="seekto" data-time="80">1:20</a> “may even lead to conflicts with demons”😂

We needed this 17 years ago. It is happening 17 years too late. Imagine what we could have done with those core2quads and Phenoms !!

someone tomorrow will start recruiting people with 10 years of experience in this

Saw this on Twitter/X and was like 'no way, this is a scam'. Then I saw their pfp was an anime avatar and ngl kinda trusted them immediately lmao

People are rediscovering functional programming without knowing it.

You should do a video about Lean, it's a language that can understand mathematical logic and be used to prove theorems. It also has a very unique type system.

But GPU guy said we will never have to code again.

Yoooo new programming thing dropped that's not a new js library finally letsgoo

That diagram showing the combinators being untangled is just so awesome.

Props for covering this project

7 days 7 computers joke was hilarious 🤣🤣🤣

Parallel Language: Check. Supports GPU: Check. Built with Rust(has thread safety, fast, beautiful, runs everywhere): What? ALL IN. let's integrate to Python and push to production now.

Yoo!! The guy who wrote that paper you mentioned at <a href="#" class="seekto" data-time="135">2:15</a> was one of my teachers at my first year of my maths bachelor, that's sick

I heard a lot of positive stuff about you and then I met this video. First video of yours i am watching... Totally hooked and I subscribed before I even finished watchng. Thanx you.

<a href="#" class="seekto" data-time="30">0:30</a> So THAT'S why my CUDA code wasn't running close to optimally... it's not my sophomoric understanding of algebra, it was that I wasn't leveraging REGEX!!!

Imagine running this on cloud servers. They'll hate you for maxing out their resources constantly lol

I'm not in tech but I've been following this one after a few people on Haskell Twitter mentioned it, the first of whom exclaimed something along the lines of "this guy is turning GPUs into real modern LISP machines!" which is exciting in and of itself, and then I think Ed Kmett engaged with something indicating that this does reduction of something akin to the lambda calculus really efficiently with GPU acceleration which is REALLY exciting

That quite comfy, that by modifying launching argument, you can specify on the fly, if you would like to use single or multithreaded and either CPU or GPU! Thanks for the video!

takes the order, cleans the toilet and cooks the food, in that order lmao 😂😂

I’m just gonna stick with C, like God intended.

The parallelism promise sounds revolutionary for computing. The way you explained the workings makes it seem less complex. Looking forward to exploring more about this language.

the way you talk about code is so satisfying

The multi-threaded example on Python isn't correct, the code shown in the video around <a href="#" class="seekto" data-time="71">1:11</a> is creating Python threads which are executed under GIL meaning that only one thread is executed at a time under a single system thread (being Python interpreter itself). So while providing multi-threaded languages feeling, the threading module in Python is not suitable for CPU bound tasks and you'd have to mingle your software parts around to achieve parallel execution on multiple threads/cpus.

The first video on all of youtube about this language. Nice.

You're in a league of your own bro, these videos are hella informative and have me rolling lmao

I love that he still puts the “Hi Mom” references in. Gets me every time.

Babe wake up Fireship uploaded

Amazing to see this featuring on fireship crazy stuff from the HOC guys

Holy moly you made this fast!

Love your humor, 1 week in serial or 7 days in parallel with 7 different pcs

4 minutes and one poo later. I now have 10 years experience of Bend. Thanks fireship 🔥

Good content bro!

I just love it Jeff .... This is exactly the kind of content that I am interested in =)

the picture used for the daemon conflict was perfect 🤣

That's actually so sick. Gonna check it out now

<a href="#" class="seekto" data-time="40">0:40</a> I so wanted to see that program written as a bunch of threads each writing one character, with synchronization between them.

Episode 5476 of "coders" afraid of C++ looking for an easy way out.

Funilly enough, I just spent my afternoon reading about Bend, and then stumbled upon this video! It's definitelly very interesting, and it seems promising. Honestly, I wish this had come out 3 weeks ago, maybe it could have saved a project of mine...

APL was made with built-in compatibility with parallel algorithms and reasoning. It could do all of it on a modern interpreter.

"Hi mom!" ❤

Great presentation thanks 😊😊😊

Threading package in python std lib is still GIL bound and just time slices a single thread, you must use multiprocessing to achieve true parallelism in python (unless you have compiled 3.12 using noGIL or have written some native library that uses it outside the GIL)

Brazil mentioned! Let’s gooo!!

This sounds like a nightmare for compiler bugs

Nitpick: multithreading in python are still single threaded because of the GIL. You can spawn threads that are waiting on I/O, but number crunching on the CPU with python's GIL won't ever use more than one vcpu. To leverage more than one cpu in Python requires the multitasking library (spawn additional processes, each single-threaded).

Nice knowledge about parallel programming

<a href="#" class="seekto" data-time="74">1:14</a> But Python still runs a single thread at a time though. This isn't a good example, this will only help when you're limited by waiting for I/O. Compute workloads will be just as fast as without threading. The GIL (Global Interpreter Lock) only allows 1 thread to run at a time. There are ways around this, such as the multiprocessing library but each method has its own caveats.

Modifying your Python code to take advantage of multiple threads just adds a few gotchas. Nothing too crazy except, oh yeah, IT DOESN'T ACTUALLY RUN ANYTHING IN PARALLEL.

Does this also support multiple computers? In our CS department we use a lot MPI, which allows for massive parallelism, and bend would be pretty great as a higher level substitute...

Amazing! I saw this on X 3hrs ago and you already have a funny, efficient, and insightful short with your own DEMOs on it! Do you have a time machine or something?

I DID NOT UNDERSTAND I need a follow-up tutorial on folds and bends

Will this help me build my todo app

Great video ! Just bear in mind that no matter the underlying framworks, not all algos are parallelizable :)

Data partitioning is the hard thing about parrel programming. If blocks have a morass of dependencies between each other then you get mostly sequential execution

We are looking for a code bender with 10 years experience in code bending. I liked my own comment cuz I like it that's why I posted it. GigaChad

<a href="#" class="seekto" data-time="102">1:42</a> “two completely random numbers” Mhmmmmmm

You dropped the mic and I dropped my jaw. Insane stuff.

thank you for blessing simpletons like me to understand this. dont ever stop FireShip.

What a time to be alive!

<a href="#" class="seekto" data-time="138">2:18</a> "high level languages like python and haskell" two very similar languages on the same difficulty level XD

Please don't stop posting content. I am so addicted to this channel already!

This is it! Time to recode my portfolio that I haven't finished yet.

Finally. Procedural Haskell.

<a href="#" class="seekto" data-time="62">1:02</a> "handling one instruction per cycle" *** CISC has left the chat ***

Could whatever techniques used to implement fold/bend be ported over to other programming languages?

Sounds fun when it works. Good luck debugging it.

came after i spent 3 months learning cuda

But now I don't have to write .sType in vulkan 17,000 times. :(

The question is can single threaded computation in bend rival that of single threaded C++. There is supposedly a massive overhead in bend's interaction combinators system.

Ran across this on the Rust sub-Reddit this morning, and now there's already a Fireship video. Dope!

I just started to learn coding a few days ago, admittedly i have no idea on what anything said in this video is, but it sounds good for me!

Time to fold 10 years of bend to my resume

Honestly, as someone who deals with parallelization pretty often (scientific computing), this is kinda huge

There is another language called Parasail, written by one of the maintainers of Ada, and it also automatically parallelizes code, but it is more conventional, and allows loops in the language. For those interested in Parallel computing, well worth the effort as it is very polished at this point.

Damn, It's like everyone apparently knows how to write progrmaming languages and I'm here trying to understand how it all works, despite all the resources available telling me, in complicated ways, how it all works... I'm a simpleton.

We had a tool for that, it was called *compute shaders*

Should have used matrix's spoon bend as thumbnail and add mic hitting the floor sound at the end

Time to learn it!!

"Two completely random numbers". Yeah i believe you :)

0 days without AI

The way Bend works reminds me of recursive functions in functional programming languages like Haskell

Note the shown python code will not do parallel computing at least in current versions with default configuration, because of the GIL

I mean, it sounds great and all, but at least right now, I don't see how this is going to be much better than the solutions we already have for running code on the GPU like compute shaders (OpenGL/OpenCL/CUDA) or a library like tensorflow or pytorch, and the multiple solutions for multiple threads and async/await in all the major programming languages, all of which already abstract away everything quite a lot. Trying to make hardware accelerated computing and general parallelism more accessible is great, but this is the sorta thing where you get into the nitty gritty bits of your algorithm's performance and with that explicit's almost always better than implicit especially with algorithms which require complex human thought and logic to optimize, something much harder to put in a compiler compared to say, the comparatively much more basic optimizations made by the C compiler, and when GPU cores can be easily compared to the CPUs of the yesterdecades: not exactly something you want something running high level code in an application worried about performance to this degree. So you end up left in a weird limbo where most applications are fine with a single thread, most of the few applications left that aren't delve into async/await and stay in the same language, and then the applications that really care about performance like simulations end up going with the more performant options mentioned previously that are already the industry standard anyways, and for good reason too. In my opinion, I love the project and its idea. It sounds great and I want it to succeed, but what I think it'll be great for and largely used for will probably be smaller scripts and data science computations and especially as an educational tool, but using it as something more than that, like to replace the industry standard solutions in an actual large project doesn't feel like it'd catch on nearly as much as just using the standard in the current tech landscape and most likely for years to come. Edit: Yes, this has promise for doing the things industry standard solutions can do and yes it abstracts a lot away which especially helps with things like race conditions (the industry standard solutions I listed also help significantly with that), but what I'm saying is that even then, most of the tasks are ones that are better off done with those industry standard solutions. Also, as for rj7259a's comment I doubt using a GPU for compilation say, for languages like C/C++ is a good idea at least in most cases.

perfect name for a new "programming language"

and here i was using a combination of psutils async and os module to manually parallelize my code in python when i should just have been folding and bending them all along in rust

I was literally talking about using the gpu for multithreading with a friend 2 days ago at work

Reminder: protogens go beep boop

<a href="#" class="seekto" data-time="81">1:21</a> wtf is that picture is that a devil . Chat is this real ?

This totally *bends* my mind

I was so confused when you show the image of a right fold on <a href="#" class="seekto" data-time="185">3:05</a>, which is inherently sequential as you carry an intermediate result from the previous computation. You're looking for the fold with the tree structure.

GPU hackers incoming

So glad to see my experience from a semester of multiprocessor architecture is now obsolete.

I think that would be pretty big for hardware health as well!

Hrm. This bend thing is a neat concept. I feel like the syntax is too indirect though. I can think of a couple of ways to implement it as simply a generic function that takes a starting value and a function that takes the current value and a function to recurse. It's really cool as a built in language feature though as I'm sure it was much easier to implement.