?? Work at Pickup Points � Flexible schedule

How AI Discovered a Faster Matrix Multiplication Algorithm

Quanta Magazine 1 views

Download

Download Video

Comments

I've no idea what you're talking about
Comment from : @nolikeygsomnipresence270

And yet we keep poisoning ourselves with PFAs and nanoplastics on a daily basis Go figure
Comment from : @bonatoc

Just hyped story as usual for AI stuff
Comment from : @t175har

I did something similar for affine transform matrices in my game engine 15 years ago to combine scale, translate and rotate matrices into minimal algebraic operations It ended up being used in my first app using my own graphics engine on iPhone3
Comment from : @wywarren

LLMs and OpenAI have been stealing the limelight for some time, but DeepMind produced AlphaZero, AlphaFold, new material discovery, Proteo and this algorithm discovery system I really think they are doing the more serious work
Comment from : @Paul-rs4gd

At 5:47 there is a short clip of a long-vanished line printer like the ones I used for my chemistry dissertation at NYU in 1968 The Courant Institute there had a Control Data CDC 6600 (Cray) supercomputer with an extensive library of mathematical functions, written in a mixture of FORTRAN and Ascentf machine language (I could not have finished without the incomplete gamma function subroutine) Documentation for the function MSAB called it “a very fast method of multiplying matrices following the laws of matrix algebra” It also said the subroutine runs “approximately seven times faster than the equivalent FORTRAN code” If Strassen discovered his algorithm a year later, that implies that the improved speed was strictly due to coding the standard algorithm in Ascentf, but seven times faster? I managed to cut the run time in half to solve 3 linear equations in 3 unknowns using Ascentf, but not by one-seventh Maybe I didn’t try hard enough
Comment from : @sjpbrooklyn7699

"is a frontier that is only now being fully explored" 🤣🤣🤣oh my precious sweet summer child
Comment from : @artimes5542

I took issue with the statement in the video that suggested that AI model training tries to find optimum solutions Instead they almost always find localized maxima and minima in the space of possible solutions They are not designed in a way to find optimum solutions any more than the human mind natively is It's important to remember that these systems are intrinsically heuristic in nature We should never take their solutions as perfect without the same level of mathematical rigor via proofs as we expect from regular mathematicians
Comment from : @jamescomstock7299

How many real world problems involve matrices having only element values in the set -1, 0, 1?
Comment from : @physicsguybrian

The only D result on my bac diploma
Comment from : @Stefch2

So, packer isn't maintained anymore?!
Comment from : @moiattube

04:26 htop on Ubuntu
Comment from : @niamotullah99

Finally a new cheat is released for the 5*5 ima use it in ma next test
Comment from : @sillysnowboot

Will this make my graphic card faster???
Comment from : @Christopher4700

AI improving itself
Comment from : @isgoodovsubhuman1972

2:55 is something censored?
Comment from : @superwhizz114

This could potentially do wonders for video codecs
Comment from : @StephenMcGregor1986

They started in Characteristic 2 and extrapolated the algorithm to all numbers Brilliant strategy!
Comment from : @moereese5254

Strassen's algorithm is still far from obsolete This method is only good for limited matrices with 1s and 0s
Comment from : @purringraven3648

I just recently learned strassen's algorithm in algorithm class
Comment from : @udaykiran2427

O(n^3) is still polynomial complexity imagine if you have exponential or factorialbrAbout the spelling in German V is read as F and their W is read as your VbrSt is read as Sht
Comment from : @holyshit922

Сразу видно, что вы профессионалы
Comment from : @HarshitaSingh-l7t

After doing electrical engineering and being traumatized by Matrices, it blows my mind how easily you explained it
Comment from : @daviddlamini4290

And the practical number cruncher doesn't bother with the multiplications at all but instead uses a rainbow table
Comment from : @brashmane2749

Is Strassen's still the fastest when the matrices are not binary?
Comment from : @perezfranco

How often is anyone just using 0 and 1 in a matrix though?
Comment from : @sethdrebick

Блин, реально, первым кругом прокрутил минимум, получит профит, сейчас сделал все 3 круга, в принципе на эти деньги можно ремонт сделать
Comment from : @Agiigamingg

Связка работает отлично, лаб молодец!
Comment from : @EARNTPR

The conclusion is ridiculous The core idea was how to formulate the question No AI will ever be able to do that
Comment from : @kosterix123

Арбитраж криптовалюты подарил мне надежду на лучшую жизнь, я тут действительно прокачиваю свой уровень жизни, могу себе позволить арендовать кв получше - это прокачка!
Comment from : @TestBuildupview

Вообще классный обменник, мне удобно переводить usdt, спасибо Министр:)
Comment from : @casahousehold

could it be reduced to a set of bit shifts?
Comment from : @coupsdestylo

Нихуя не понял, но очень интересно! 😂
Comment from : @havanasyndrome3024

I heard Kasparov formed a chess league of human players who use computers as a team rather than an adversary It seems this video is rediscovering a similar step forward in this field of math
Comment from : @christopherinman6833

I remember that this technique was taught to us when we did our math major in mid 90's
Comment from : @ushi-l7z

Whaooh 🎉
Comment from : @timetableforschoolteam3

why was the paper wrong?
Comment from : @nunoteixeira2129

I wonder what is the role of intuition in solving complex mathematical problems
Comment from : @BenIsraelSeatriz

muito bom mano, entendi foi tudo
Comment from : @LuisEduardo-qp1cq

I dont understand, all you 'smart' people, yet none of you has bothered to implement this in hexadecimal with bit flips i dont get why y'all cant see these methods are still way too slow
Comment from : @VertegrezNox

By brute force
Comment from : @satychary

These AI videos with AI voices are worse than Musk fanboy channels AI can’t discover anything it’s not programmed as an “AI” it’s an LLM which means it stole it and reproduced it Most industries after they fired the idiot that suggested AI are dumping AI
Comment from : @halvarmc671

Who are these seasoned mathematicians that can't master matrix multiplication?😂
Comment from : @dasbogenhaus

Is this the reason of why google ai progress is slower than openai?
Comment from : @jianjunsun7699

Did they tho ?
Comment from : @Sumpydumpert

AI doesn't invent ANYTHING, it just regurgitates the content of some paper some poor Chinese guy published but no one ever read
Comment from : @algonix11

It’s so annoying that humans get butt hurt by the thought of being replaced by AI Such narcissism Meditate to overcome it!
Comment from : @mikezooper

Deep Mind is being used by Google to win at business (which included manipulating politics and other people) Corruption?
Comment from : @mikezooper

Welcome to the new age of mankind
Comment from : @FyaaahS

stupidocracy is coming
Comment from : @gnarusshiaku6644

Disappointed in the lack of Matrix movie references
Comment from : @jimmyjam-vc6rf

9:00 Why prompt engineering is a job now :)
Comment from : @moseschung3220

Your knowledge and approach to binary options trading is truly impressive Thank you for helping us understand this complex area
Comment from : @evgeniyhohlov2637

0 and 1 multiplication has same cost as 0 and 1 addition right?
Comment from : @hariharamadhava

(Title of ) video starts @3:55
Comment from : @themultiverse5447

The restriction of it being values of 0 or 1 seems to kind of negate the benefit of reducing multiplications, as you could just do a bitwise AND operator to multiply a whole row of values at once, which is a trivial operation for computers, unless maybe its floating point values that range from 0 to 1?
Comment from : @jitspoe

Ai가 아니라 우주의 물리적 작용이라고 보는게
Comment from : @양익서-g8j

I think in the very near future, because of this discovery We will be able to measure the acurate Mass of people and objects Heath will be measured by mass Food cost
Comment from : @ontheotherhand463

Yea, a lot of people are going to be obsoletebrBut the repair jobs are always gonna be there
Comment from : @panzerfaust480

Error codes, multiplications, everything is done faster by splicing it One thing I noticed that when I used every day D6 dices for a few moments a day just to sharp my addition in a trivial way my arithmetical tasks went smoother back in the days I did split everything I could to smaller chunks in my mind Universe is a weird fractal :)
Comment from : @tehdii

The fact that the AlphaTensor algorithm appears to be referred to as the "FBHHRBNRSSSHK algorithm" in actual papers is _hilarious_
Comment from : @linkhyrule5800

So, mathematicians works on how to make themselves usless, ok
Comment from : @StockMalthiel

i need the warp dirve and artifical gravity for spaceship can be done this decade? Thnks hehehe
Comment from : @slevinshafel9395

is playing that music in the background really necessary?
Comment from : @mikl2345

Can AI do math ? I mean real math
Comment from : @jediTempleGuard

AI is fucking stupid, anyone who uses it is a traitor to mankind
Comment from : @apbecker

Do anyone know what software is used to create animations like this video?
Comment from : @MuhammadIqbal-zb3dy

How do you know?brMaybe they replace mathmaticians😂
Comment from : @Rorablink

8 minutes 8 minutes to get to the point Thankfully I skipped through the video quickly to skip the off-topic filler
Comment from : @JD99-zj3gq

wut
Comment from : @basedbukkakeman7892

Nobody's talking about the machine elves none of this AI stuff is surprising if you consult with the machine elves
Comment from : @scotthjackson5651

I so respect guys like Fawzi Amazing Imagine what is going to happen within the next 30 years?
Comment from : @qafmbr

Reinforcement training has a major flaw it cant find solutions, that initially demand a sacrifice, but get much better later on It mostly goes for whats better now ( in terms of "score" )
Comment from : @RaskaTheFurry

Mathematicians use proofs to verify mathematical claims Just because AI makes the claim that it has found a shorter algorithm, the claim cannot truly be accepted until a supporting proof is produced
Comment from : @JohnBerry-q1h

unbelievably cool
Comment from : @PrecioustheMovie1

6:30 subtle brain fuck here: the single 3D tensor object includes [the instructions for] "multiplying any two matrices, [along with their result]"
Comment from : @splaytrees

He doesn't explain the new algorithm?
Comment from : @wayneswildworld

6:37@92
Comment from : @raheelrehmt

2:23brQuanta Magazine: Going from eight down to seven multiplications may seem trivialbrme thinking: for 2x2 sure but for something like 8x8 it is significantbr2:29 bshows 8x8/bbrme: lol
Comment from : @unkn0vvnmystery

Geometric progression
Comment from : @Larry-j9b

AI was not used, every content creator thinks they know what AI is 😅😅😅😅
Comment from : @atlantic_love

And this will make our lives better how ?
Comment from : @dedskin1

Very cool stuff I feel bad for the mathematician though He started humans still still be required in the future, which will be proven false once we achieve AGI
Comment from : @harbirsingh7266

So another PR relation for 'AI'google what part have i missed because because deep mind as all so called ai can only answer what has been defined by a human before
Comment from : @tarektahar8632

Please drop the annoying doofy music why do you even need it? You're going through extra effort, time, and expense to make the video worse Bailing out at 7:32
Comment from : @betaneptune

Great video
Comment from : @khalidelgazzar

So the Machines are getting Smarter way Smarter Wile the Humans are getting Dumber way dumber with each new iPhone Now I can see why the Top AI Guys have that spooky look about them like they know something that they are not telling us yet about the future
Comment from : @michaelwells7348

So you can get a paper published in Nature by reducing the number of multiplications in an algorithm from 49 to 47? And without mentioning the number of additions? That is the AI hype machine in full swing
Comment from : @trapkat8213