JAX in 100 Seconds

[@rafaelsanchezmartinez9393]

Now I have 10 years of experience in JAX. NICE!

[@ArthurKhazbs]

Well done! Congratulations!

[@user-pe7gf9rv4m]

parallelized

[@hoskdjciianbflso]

Enough years for Jax

[@Cfomodz]

Good because that's the min req for the newest jobs using it

[@LabiaLicker]

*writes in resume

[@justingolden87]

"Imagine you're writing a nuclear warhead in your mom's basement" Oh I don't have to imagine fireship

[@gezenews]

Lol worst day to possibly make that joke.

[@Roiadas]

Relatable

[@JoaoLucas-fo8gv]

FBI would like to know your location

[@GLUBSCHI]

@@gezenews?

[@franciscos.2301]

​@@GLUBSCHI he must know something we don't 😳

[@jonathansmith8070]

Active: Jax enters Evasion, a defensive stance, for up to 2 seconds, causing all basic attacks against him to miss. Jax also takes 25% reduced damage from all champion area of effect abilities. After 1 second, Jax can reactivate to end it immediately.

[@deathpacito8702]

Active: Jax enters Evasion, a defensive stance, for up to 2 seconds, causing all basic attacks against him to miss. Jax also takes 25% reduced damage from all champion area of effect abilities. After 1 second, Jax can reactivate to end it immediately.

Imagine if Jax had a real weapon?

[@mysterioustraveller7165]

ACTIVE: Jax Dash dashes to the target unit's location. If the target is an enemy and they are in range upon arrival, Jax deals them physical damage. Jax can cast any of his abilities during the dash

[@kalpamonx]

Gotta try AP Jax someday, one time I had this one rando Jax who went AP in my team. He was melting everyone!

[@Dom-zy1qy]

Was just scrolling waiting to find a comment like this.

[@4dtoaster819]

"Why did you start using JAX?" Me: "I like the logo"

[@wlockuz4467]

As a software engineer with 10 years of experience, I understood some of those words.

[@MaxMalm]

I felt so stupid watching this

[@yad-thaddag]

@@MaxMalm You're not the only one. 😉

[@poulticegeist]

Yeah, I guess that goes to show that those years of experience were wasted

[@mohamedtrfnx6632]

This is what happens when people become software engineers from the internet lol

[@twentyeightO1]

As a ML student I understood everything. That just means you're not into this stuff.

[@randomeone100]

Time to add JAX in my resume

[@TheRadischen]

10 years experience

[@injSrc]

😂😂

[@MiroslawHorbal]

Good idea. Thank you for the career advice

[@ClaudioRodriguez-ln2ez]

Gg no top

[@dominicanfrankster]

I watched this vid too, I'm going to do the same

[@mritunjaymusale]

Jax sounds like less abstract version of pytorch and more complicated version of numpy

[@Imperial_Squid]

Pretty much yeah

[@kittydaddy2023]

huh

[@Michallote]

However performance is above torch somehow

[@Imperial_Squid]

@@Michallote because JAX basically compiles down to machine code when run, and while a lot of Pytorch iis done at a low level, there's still some overhead that's python running things. But conversely torch allows you to do dynamic ML (edit: actually pytorch also let's you compile but I haven't played with it personally), whereas JAX has to all be perfectly laid out from the start so it depends what you're trying to do. Torch is also just more useful to more people so it's more well supported so even if Jax would be better for your project, there's incentive to follow the herd and use the thing everyone else is using.

[@user-pe7gf9rv4m]

jax borrows a lot of ideas from haskell and functional programming it just has a more familiar syntax (python-style) and calls things differently (scan is a bifunctor for instance)

[@unclecode]

The sample code u shared at the end, so simple & elegant shows what training is all about. Approximation of a function F defined as multiplication of multiple matrixes (its layers), then calculating its error over known data E(F(x), y), computing gradients of that error E', & then updating F's parameters, such E' approach to zero. I will share this code to my students. Thanks again.

[@alcoholrelated4529]

i'm interested in computer science, calculus and statistics, but i only have 300 seconds in total

[@the_everything999]

Then you're in the right place.

[@artyomloukashov636]

Then you need Matlab - short, fast, mathematical!

[@dishankshah5080]

give 100 seconds to each and you're done.

[@potato_zeq3571]

"Imagine if i had a real weapon!"

[@hlab3524]

I WAS LOOKING FOR IT xd

[@sephroth201]

now we need another library call annie

[@ChrisCarlos64]

My thoughts exactly, especially the color scheme looking like his default costume

[@potatosheep]

League x Programming = Legendary

[@mindaugask_]

​@@potatosheep found the scripter

[@rick-ed1sn]

that's a very clever logo

[@LouisChiaki]

Then you should also love Tensorflow.

[@yash1152]

2:43 that sure looks cool and inspiring. but whats clever about it?

[@joseville]

@@yash1152 it incorporates an impossible geometry

[@yash1152]

@@joseville is that related to the branding in any way? > _"it incorporates an impossible geometry"_

[@joseville]

don't know JAX about JAX The impossible geometry in the logo is similar to a Penrose triangle

[@picoplanetdev]

Oh boy another Fireship video!

[@ortana-v2]

truely the first

[@12crenshaw]

At least that one haven't made me anxious about AI taking over the world

[@joseville]

0:25 wow, they incorporated an impossible triangle into their logo! that's really cool!

[@Henrique-up4su]

"Let's start with X: Accelerated (...)" "???"

[@tienatnguyen8899]

That's because X is pronounced "eks", and accelerated is pronounced like "eks-elerated"

[@bruwyvn]

Clearly you haven't watched BEN10

[@birdbeakbeardneck3617]

ai fireship strong with this one

[@onça_pintuda999]

@@bruwyvn For non native speakers XLR8 is just XLR8 in the translated language, usually who watches/ed ben10 at 10yo would not speak english, outside of native speaking countries. It took me years to realise that xlr8 is spelled accelerate

[@kintrix007]

@@onça_pintuda999 Nah it really depends. For example, in the Hungarian dub its name would literally translate back to English as "Lightning Gremlin". They just took the message of it, and gave XLR8 a name a 10 year old could have come up with. Instead of a letter sequence that is otherwise nonsensical. Or, in the Japanese dub, they just called it "Accelerate", as they kept the pronunciation of all the aliens, not the spelling.

[@8ball437]

0:55 the points of inflection are marked wrong on the graph. The yellow points are local extremums, not inflection points.

[@bramvdnheuvel]

Do Elm in 100 seconds! You'll love it! It's an older JavaScript framework that had a lot of thought behind it. It's funny to see more popular ones catch up, but Elm did a lot of the things very well and it deserves some love!

[@VisibilityO2]

I have never heard the best explanation of derivative from calculus like 'if you want to know how big is the mushroom cloud is growing ' .

[@NoahElRhandour]

Active: Jax enters Evasion, a defensive stance, for up to 2 seconds, causing all basic attacks against him to miss. Jax also takes 25% reduced damage from all champion area of effect abilities. After 1 second, Jax can reactivate to end it immediately.

[@MarcusAurelius-kl5yz]

Dang missed opportunity using the Jax from league of legends on the video

[@Gm-ce5kg]

missed opportunity to take a shower

[@msghia5252]

Who?

[@utorque]

Could you do a ROCm video (AMD's CUDA couterpart) ? Love you and your vids !

[@zstrizzel]

great video but still very sad we did not work in a mortal kombat jax reference

[@fernandogurgel]

Was looking for this comment. I always wanted to see a library or tech called Jax, because all I got was AJAX concept from Javascript, which would not entirely fit it, but somehow always reminded me of the MK fighter.

[@4RILDIGITAL]

This video really clarified how this isn't just another accelerated linear algebra library, but a tool developed by Google for futuristic new hardware. The constraints on immutability and the way it compiles to low-level code for accelerated Hardware seems ideal for machine learning. And the automatic differentiation feature is truly next level.

[@FasutonemuMyoji]

Thanx totally not chat gpt or equivalent

[@jjhacker810]

Can I suggest R? It seems to be gaining traction in Google Courses, EdX, Uni, Etc. It'd be good to get a high level overview of the langugage that can be used to summarize it for business, friends, family, etc

[@叵]

*jax from the amazing digital circus* ⁉️

[@thecatguy893]

What i thought

[@SirusStarTV]

Nah, from Mortal Kombat

[@Console.Log01]

please end it all my bro

[@teamllr3137]

I was sure there would be at least 1 comment like this...

[@Fafr]

No it's the classic creator from Geometry Dash, the one behind the legendaray The Nightmare

[@athlone_tch]

- “Mum, could I practice my predictions from automatic differentiation in our garden?” 1:47

[@robertholtz]

Thank you, Fireship. Up until I watched this video, I thought I was smart.

[@joshuaevans4301]

Holy shit automatic differentiation is so freaking cool. Imma need to remember this when I jump back into calculus

[@Jack-ul2rb]

How you edit your videos? . It was an amazing edit And which platform do you use for such arrows and animantions .

[@jackelhuevo2133]

Very cool video! May I suggest you tackle Dyalog APL (or APL in general) in the future?

[@codyandersan]

For those who don't know: Fireship lives on a planet whose gravity is stronger than our Earth, that's why his 100 seconds are our 226..

[@codexcursors]

Thanks, I can now add 10 years experience of JAX to my résumé!

[@MrDejvidkit]

Please do scala next!

[@ferdinandbardamou5508]

My experience: ~1200 to 1500x accelaration of a function with numpy operations on a conventional 3070 GPU, by just adjusting the numpy code to JAX syntax. Just remember that array dimensions must be constant.

[@KennaBlackburn-tm9xn]

100 seconds is a bit of a stretch, but well what can you do. Keep up the great work.

[@netcodedev]

I just opened KZbin and thought it must be about time for a new fireship video. And bang! There it is

[@alphaios7763]

Hoped to get some tips on top lane

[@LetyQ]

1) Always keep track of jungler pathing. 2) Freeze lane when ahead, even if the opponent does not contest and thus does not die, the gold and xp lead from the minion is enough to snowball. 3) Death timers are very low until level 7, don't greed a plate if a takedown is taken, you will lose in tempo and therefore reduce the advantage from the kill. 4) If opponent roams, hard shove then follow up if action not ended. Best case scenario the roams does not connect and opponent will lose wave. Worst case scenario, the roam is successful but as the lane is under the opponent tower, they can neither base nor push for plates, therefor the play is even in tempo even if it is losing in gold/xp.

[@Channel-iu6de]

Man that logo design for Jax is amazing

[@hectors.1644]

A must like video, due to your effort to research this type of content, keep that good work

[@Rahul.Gurung.]

Brilliant back at it again, everywhere I go I see brilliant.

[@kushagrachaturvedy2821]

Love the visuals in this one

[@kellymoses8566]

Automatic differentiation is fascinating

[@YandiBanyu]

Oh, that is actually useful. The abstraction is very much welcomed here. Can't wait to see RandBLAS and RandLAPACK too!

[@boriscrisp518]

This is te best youtube channel on the internet. I just wish youtube wasnt censored so i could expresshow much I enjoy the content

[@foxy_robin]

You videos with Brilliant sponsorship are max quality ser.

[@masicbemester]

ah yes that one character from The Amazing Mortal Kombat Circus

[@NY_Mountain_Man]

I'm an artist who worries constantly about saying the wrong thing on occasion. Meanwhile, I'm watching some guy talk about nuclear warheads. What a fascinating channel. 👏

[@RiversJ]

If it doesn't maintain internal state (with assorted memory overhead) or actually produces new copies of arrays, this is Hugely wasteful on the GPU memory bus. Wide scale SIMD architecture does Not model after OOP or FP concepts perfectly, there is a damn good reason they aren't driven the same way CPUs operate. Both approaches seem faulty and unavoidably slow/wasteful translation layers to CPU styled paradigms. Far more useful would be having access to well designed wrappers on most languages to the underlying APIs which too often require a shit ton of boilerplate and horrible docs along with badly designed interop calls.

[@geoffreybrunet5948]

There is an error for mushroom example with array, it's a 3 elements array but you try to get arr[3], je gradient is [60., 4., 3.]

[@nafrost2787]

This sounds extremely similar to pytorch, tensorflow, keras and the likes. I'm surprised I haven't heard of it by now. Anyone here on the internet have any ideas why is it less known? Maybe it's a latecomer?

[@f3arbhy]

its pretty well used for machine learning in the scientific community. Jax probabiliy has the best api in terms of writing pure functions and transforming them into functions that returns gradient's or hessian. For that reason, one can write code just like writing an equation, which is well appreciated by the scientific community..

[@SuperMaker.M]

nice I just literally started learning JAX after using TF and torch for a long time

[@dovs96]

That's exactly the thing I needed lol, thanks a lot.

[@2puff-nn5sv]

mathematics in 100seconds

[@howird]

JAX actually stands for "Just After eXecution" according to the original white paper

[@nictibbetts]

jaxing off

[@apps9129]

Actually understood a lot more than I thought I would randomly clicking on this video.

[@monkeyofjp]

"Imagine if I had a real weapon"

[@TimmyM]

I don't understand anything you say, but I love watching these videos anyway.

[@raidtheferry]

Hey Fireship, what software do you use to edit/produce your videos? thx

[@EigenCharlie]

Really want to see how impoved versions of KANs wil perform using JAX

[@edbrito-swdev]

I saw the title and thought it was going to be on Jax-RS. I'm glad it wasn't!

[@costitica]

Imagine I tell the interviewer I have 100 seconds of experience in JAX

[@cinderwolf32]

I would love to see Smalltalk in 100 seconds

[@dsjgfhidupgjret]

Finally something I’ve used in this channel :)

[@pasmanitan]

Recruiters are gonna ask for ppl with 15 years of experience with JAX

[@R3dX5118]

Mr fireship, please don't scam us, this ain't no 100 seconds anymore, Thank you ❤

[@XyndraNerd]

You know you're balling if you have your own TPU

[@ItWasEnder]

I said "Im not going to use JAX probably not going to watch". I'm glad I did, it actually looks useful lol.

[@soflass1293]

The most understandable 100 seconds ever

[@Salah-YT]

JAX in 203 Seconds

[@vintagegenious]

All those new languages and libraries are converging closer to functional programming

[@GOUST3D]

I feel like I'm always just out of the grasp of understanding how something like this is useful for like 3d animation data... idk how the array of partial derivatives would be useful but it seems like it would be for some of the stuff I'm thinking about

[@lenionio]

I have 20 years experience in JAX and I must say I love it

[@soundstudio7803]

tfw you realize 100 seconds videos aren't actually 100 seconds long

[@um8078]

Then at your interview they're gonna require you to have 5 years of experience in JAX

[@eulmdev]

common W fireship vid i love the 100 seconds series lol

[@HardbassTV.]

Jax from TaDC?

[@nafrost2787]

Who else is proud of our little Jeff for graduating to a brilliant sponsorship?

[@vicradon]

You explain partial derivatives in a way I've never thought about before. So a partial derivative exists because the function being differentiated has multiple inputs. When it's a single function, there's no need for them. Right?

[@skejeton]

"i invented a new hyper parallel execution language!" new or shading language? "shading language"

[@toxiq5295]

That logo is sickly designed

[@zainkhalid3670]

Just about time I add it in my CV

[@Gstyle85255]

Good to see its not just JS languages in the merry-go-round

[@loserpacman]

The linear algebra field going to the moon with this one

[@hanif72muhammad]

My brain too smooth, what are you talking about?

[@syrus3k]

Wasn't Jax a character in mortal kombat?

[@Andeiih]

My head is exploding like a nuclear mushroom cloud after this one

[@somerandomfinn668]

100 seconds (adjusted for inflation)

[@thedelanyo]

The only in 100 seconds without "Hi mom"

[@pixelsandpointers]

You forgot about ease of parallelization and vectorization of functions which are the true primitives of Jax, alongside autodiff. C:

[@FaiqAli-cc3sl]

You explained it better and faster than my ML professor

[@_rd_kocaman]

Thanks for the video mate now I can add JAX to my resume

[@the_everything999]

Jax seems really cool.

[@amooozk7777]

that was almost 204 seconds😂

[@Nekoeye]

"Imagine"... stop it right there. I'm lost.

[@thatguyoverthere32]

Is this just giga chad MATLAB? Was my first year compute physics course useful after all?

[@Tubeytime]

Thought this was a TADC compilation video, but I'll take this instead

[@tommypospa]

Glad I have 4 years minimum experience with JAX.. oops AJAX.

[@J3dotgg]

Thanks Fireship, this library will definitely accelerate the development of my nuclear warhead!

[@Rebel101]

You know it's a sponsored video, when there is no flat earth.

[@antonchinaev3495]

What happened with "blazingly" fast? I liked it