This is what DeepMind just did to Football with AI...

Dr. Petar Veličković with his science colleagues at DeepMind have just released a new paper called “TacticAI: an AI assistant for football tactics”. The approach identifies key patterns of tactics implemented by rival teams, and figures out effective responses, a huge achievement for modern football.
Petar is a Staff Research Scientist at Google DeepMind, Affiliated Lecturer at the University of Cambridge, and an Associate of Clare Hall, Cambridge. He holds a PhD in Computer Science from the University of Cambridge (Trinity College).
(Note: There was a minor mistake in the intro, Petar was at Mila when he invented graph attention networks, not DM)
deepmind.google/discover/blog...
petar-v.com/
/ petarv_93
Support MLST:
Please support us on Patreon. We are entirely funded from Patreon donations right now. Patreon supports get private discord access, biweekly calls, early-access + exclusive content and lots more.
/ mlst
Donate: www.paypal.com/donate/?hosted...
If you would like to sponsor us, so we can tell your story - reach out on mlstreettalk at gmail

[@schumzy]

Based on the constraints - known starting positions, short time period into the future. this makes sense for the NFL

[@renanmonteirobarbosa8129]

maybe when NFL plays "football"

[@einarjuel]

​@@renanmonteirobarbosa8129maybe when you watch the video

[@purnp5897]

​@@renanmonteirobarbosa8129 😂😂😂

[@Tiriondil]

My thoughts at the beginning of the talk. And I'm really surprised that they didn't contact any NFL team to do that.

[@einarjuel]

@@renanmonteirobarbosa8129 goshdarnit, how dense are you? Corners are set pieces, thats why they are good for this kind of modeling. American football is only set pieces

[@CodexPermutatio]

Things are getting wild.

[@ximono]

I'm actually a little bit underwhelmed. Like he says, it's mostly subtle suggestions, offloading some of the work coaches already do, potentially giving them more time to be creative. Which is cool, but not revolutionary. A bit like having an assistant who only focuses on set pieces and gives you advice on minor adjustments. Which of course could give your team that extra edge and tip the balance in your favour.

[@happylittlemonk]

in 70s a statistician worked out that more goals were scored from long forward passes and gave that info to Graham Taylor who was managing a 1st division (I think) and they went up rapidly because that. So this study was done decades ago. Now AI can do it automatically.

[@MNC2k]

but that data was flawed and led to incorrect long ball tactics that didnt actually work out long term

[@andrew-729]

@@MNC2k If it worked at all even short term I would argue it was successful...

[@kushcapone]

Soccer*

[@piotrek7633]

The video is about football not american football

[@kushcapone]

@piotrek7633 association football Aka soccer is not football. It isn't even American and was played across Europe back to the 12 century and even further back in ancient Greece. Look up mob football.... that sport is from Europe and is the original football, NFL style. Soccer is not as old as football therefore stop calling it football.

[@phpn99]

"Football". It's pretty easy to see that this is the sort of tech that has urgent military applications.

[@tonykaze]

You're saying this like Israel hasn't been publicly using AI to select targets using data collected from online AI bots and AI-piloted drones sorting through data using AI to prioritize them using AI and strike them using more AI-piloted drones and AI-powered bombing targeting systems.

[@olegt3978]

Russians will probably also have a gnn similar to this for optimising their military tactics.

[@Crytoma]

Yep

[@cldlsolshinegrowing4366]

Urgent? How about potentially apocalyptic?

[@Meta-sg7iz]

This is how the dystopian sci fi future begins if we're not careful...

[@alextrebek5237]

@18:10 "similar thing to warfare..." thats the premise of "I Have No Mouth And Must Scream"

[@thePyiott]

Google is just blatantly developing weapons now.

[@sapienspace8814]

American Football and baseball is likely an even more applicable use case for TacticAI in what Petar is describing, in that each play state in American Football is well established because play is more frequently stopped (while in soccer, like basketball, it is almost a continuous, or longer, play state), so baseball, and American Football is kind of like a digital time step at the end of a play in machine learning (just like penalty kick state, and corner kick state is in American Soccer or English Football. I have to blame the Alantic Ocean for contributing to the fuzzification or mix up of the word "football" between the US and England, where in the US we call it "soccer").

[@christopherd.winnan8701]

How many years before an AI coach wins the superbowl and the world series?

[@sapienspace8814]

@@christopherd.winnan8701 Probably as soon as the first coach applies it, though it depends on how many other coaches apply it at the same time.

[@christopherd.winnan8701]

@@sapienspace8814 - Have you heard of other teams already employing AI?

[@sapienspace8814]

@@christopherd.winnan8701 Not yet, but I suspect that if an NFL coach does, it would be a huge advantage, until other coaches adopt AI too.

[@butette]

@@sapienspace8814 Only if the NFL implements rule changes that would allow it. I don't believe you can even use live electronic devices on the sideline or in the booth other than to capture snapshots of plays.

[@simonmalaret5506]

Petar is very articulate, great job!

[@dankplatypus420]

Agreed. Interviewer also did a good job asking provoking questions and not interrupting.

[@MarcCastellsBallesta]

Happy to have discovered this channel!

[@williamikennanwosu]

Petar is a great communicator of technical concepts.

[@artscollab]

Moneyball: The Return

[@SigFigNewton]

Databall

[@kaba1996]

Was just thinking about it recently. Great that someone made it. Yes AI is for that things with it being so great

[@orlanino]

Hari Seldon, psychohistory, hey? Thae old Asimov was such a visionary!

[@treesareafewofmyfavoritethings]

I've been kind of fascinated with whether it would be possible to quantify 'flow' in team sports and how things like errors can be contagious. I think you get a very clear sense of how a team is working together when you watch them play but I've never seen that aspect broken down and analysed. What are the subtle timing errors and signals that propagate out through players and is it possible to recognise them and enhance the cohesive ones and reduce incohesion.

[@danielfazolo8622]

Football is such an awesome sport because no matter how close to a "perfect squad setup" AI generates, the game is played in the pitch, so it shoudn't "break" the game.

[@frank6048]

I find him so relatable for some reason I think even that reddit mod beard is somewhat relatable, not because it look good, but because we've all taken awful decisions regarding our facial hair at some point

[@bachtran957]

But can one apply this research to my FM 24 saves?

[@noahhughes127]

But did it predict "CORNER TAKEN QUICKLY ORIGIIIII"?

[@gammalgris2497]

Predicting the future outcome of player setups (players and positions and possible moves) in certain play situations based on past outcomes. Is that what it's about?

[@GCdevine1]

This kinda reminds me of what F1 & NASCAR has done to integrate AI into their strategy. It will know the optimal patterns and play styles to counter each opponent. The part that I question is if that is what wins then who is really making the decisions and what would be the point of human coaches?

[@vanitytm1012]

Did you watched the video to the end? He explains quite well that the AI doesnt make decisions, that it only can give you suggestions or see patterns and its used to free coaches time to focus on the creative part instead of watching the hours of plays to find those patterns.

[@harimonting01]

Maybe it will make their jobs obsolete. But what can we do about it? You can not stop technology.

[@Kartik_C]

absolutely loving this GDL/NAR/Category Theory arc 🤍

[@niklashenckell8051]

Very interesting application. Software ate the world, AI is gonna swallow it whole. The future will be amazing. One thing I thought interesting is how he described the influence of different players on the success or failure of a set piece. I could imagine taking this on open play and analyzing the players could be huge in scouting.

[@MontyCraig]

Next step is to turn these graphs into hypergraphs.

[@rickybloss8537]

Stephen wolfram was here

[@TruthDragon.]

I literally had this idea yesterday and was wondering how long it would take for Billy Bean/Paul DePodesta 2 to emerge using AI instead of regression models. I guess I did not have long to wait. Billy Bean and Paul DePodesta left their mark on professional sports.

[@cameron1617]

Would be interesting to determine the player/s with the most efficient/probable beat the defender runs based on the best run to make, finding the best player positioning and movement.

[@bernard2735]

Counterpoint: human foibles and unpredictability make life (and games) fun

[@AltoidsYob]

Not all games are played for fun.

[@MNC2k]

depends on the degree to which one is intrinsically or extrinsically motivated. the Premier league demands a lot, and every marginal gain will ultimately benefit them so its a no brainer for the big guys no?

[@piotrek7633]

wait till you get ai bots in your games, and everything you can do irl will be automated already. Then you'll have looots of fun

[@CesarGomez-kp5lm]

Countercounterpoint: wtf does foibles mean?

[@sp123]

The biggest dilemma with technology is that it makes things so efficient they are no longer fun. Analytics created from technology have ruined sports, music, dating, etc by finding the best way to win at the expense of the consumer

[@brycebyte]

Excited to listen to this. I worry that we continue to venture into a pay-to-win model where the best teams across sport are the teams with the most money/resources

[@bobjones7274]

They mentioned it could be applied to other sports. But they will not make this open source in any way, or available to collaborators, am I right?

[@ColinTroyy]

I wonder how they collected the data for training the model.

[@emoneytrain]

System 1 vs 2 is a reference to Kahneman?

[@earleyelisha]

“The game of football has so many latent things, so many unobservables…and sometimes a human might, in the moment, do something different.” Similar curse to self-driving cars and other unsolved problems where models aren’t embodied or have a theory of mind.

[@neighbor9672]

I think these systems that people are developing will overlap over time, become pieces that are all part of a complex puzzle that will ultimately make up truly intelligent systems that will have power and capabilities far exceeding that of any human being. When that moment occurs I don’t know what that means for us, but it will likely happen so long as people still keep developing all these technologies and finding ways to interconnect them all.

[@earleyelisha]

@@neighbor9672 we can hope. It’s sobering though to consider that before we can ever reach human intelligence and beyond, we have to develop systems that have much more basic intelligence. Your average house cat or mice have far greater embodied understanding of the world than any current model.

[@andrewferguson6901]

The technology required to make properly intelligent systems that surpass humans, I expect, is already here. It needs only to be combined in exactly the right way

[@earleyelisha]

@@andrewferguson6901 depends on how one defines “technology”. - a taller ladder wouldn’t have gotten humanity to the moon - rockets aren’t going to get us interstellar travel - more DL/scale isn’t going to bootstrap a system to common sense, true perception, or internal world models.

[@stevemeisternomic]

The thing about self driving cars is that it needs all the cars to be self driven to get the maximum benefit. Human drivers are a problem. They have moods. They have conversations while they should be focused on the road. They drink and drive. They do stupid things at the worst times. They really don't belong behind the wheel.

[@omagro8267]

Of course it was with my club Liverpool. Amazing to hear!!! Cutting edge. Lets gooooooo

[@NotoriousBINGO]

It might explain all our late goals from our subs this season lol

[@pooroldnostradamus]

This is geometric deep learning heritage, I sink so, I sink so.

[@christopherd.winnan8701]

Does it learn from MOTD video or does it need more carefully crafted data? Good for the big international teams. What about the Accrington Stanleys of the world?

[@jumhed994]

Well, all those small teams are already using GPS trackers on their players, which was unthinkable just 10 years ago. These small teams will get this tech soon, but just behind the big boys

[@textbook1263]

Indeed it should help the Accrington Stanley's of the world, AI and such helps with the democratisation of high level sport.

[@afroboi7454]

excited for when this take makes it to video games like FIFA!!

[@bigbogeyface]

Awesome 👌

[@GigaChadRealington]

I’m curious how they monitor hydration, sleep, vitamins, minerals, mental state, traction of grass/turf on given day, etc.

[@breakfreak3181]

Then teams will just change tactics, as has happened for decades. New systems are adopted, teams get used to them, and then newer tactics / systems evolve. If you have a good team and coach, they'll adopt different tactics during the game. Admittedly, few coaches do so though, and prefer to rigidly stick by "their" system no matter what.

[@r0land014]

Yes you could do this but also it brings into position game theory possibilities where teams choose particular corner kick strategies based on the balance of probabilities with a dominant strategy using game theory.

[@schumzy]

Thanks - I have been pushing back on LLM's will fix everything narrative for a while, and this video has shown me my instinct was right. The more specialized a model and better suited the technique, the better for a specific job. I think we will diverge again from this whole "AGI" talk and go back to the specialized models of 2016-2020. Also like the fact it is light weight and easily refreshed.

[@aibutttickler]

yeah, definitely doesnt sound like you have confirmation bias or anything, it's not like LLMs have been rapidly improving since 2023 and show no sign of slowing down lolz, you're so right man

[@rahulkushwaha7896]

by defination AGI should we able to do all task better than individual models

[@mateuszputo5885]

Of course, it's todays copium. Some prominent scientist even belive in this sort of "prediting next token is enough for *AGI" religion.

[@Houshalter]

Until someday we get a model which is specialized at making specialized models. What then?

[@Emerson1]

Nice back yard!

[@sivababu5297]

Why don’t we use this for route optimisation ?

[@johnk7025]

Did they get all Data from existing providers?

[@dretheblack]

Just watch the games?

[@milckshakebeans8356]

No offence, but I'm so happy that this will apply only for set-pieces in the near future

[@lmao4982]

What about footedness? Seems discarding symmetry is a baad idea?

[@regulargold7065]

This is what basketball needs. The quants and the quals get to settle their qualms.

[@MrYehaha]

Even more advanced stats is what nba needs lol

[@lymphomasurvive]

@@MrYehahaPlayers who play defense is what the NBA needs.

[@thePyiott]

I need this in Arma 2

[@umarmars47]

NBA is scripted like WWE

[@kanjinha]

Someone who understands the game. Too sad NBA is not basketball anymore 👌🏼​@@lymphomasurvive

[@kellymoses8566]

The pentagon will hire him.

[@mariterodriguez5349]

Curious!

[@JensGulin]

Interesting that the corner cases that were easy to approach were the actual corner cases...

[@Rubinhogoal]

So more high corner kicks on Virgil, or ...?

[@alliedatheistalliance6776]

I was thinking at some point maybe AI can watch and analyse football matches, and automatically rate players top speeds, agility etc, or maybe even rank the best players. Exciting times!

[@Trizzer89]

This would be my dream job. But I will never be able to get a PhD in machine learning without burning myself out at a second job. Damn, I wish companies would invest in people. Maybe they can offer courses where you can get hired if you pass

[@NickKrige]

is there not a concern that constraining the network will prevent learning about things such as certain aspects being more important to the topic than was originally thought?

[@christopherd.winnan8701]

So who will be the first team to start laying off their coaching and management staff?

[@TomSchaffer]

depends on which team belongs to the most stupid club in the world.

[@Alice8000]

very nice👌🏼👏🏼👍🏼 clap👏🏼👏🏼👏🏼👏🏼

[@haseebpavaratty6249]

Well, that explains the 'Corner taken quickly' vs Barca

[@TomSchaffer]

no it doesn't

[@Trizzer89]

Why dont they just simulate the game where each player has given statistics. Then 2 AIs play each other a billion times until both teams are almost perfect strategy? Any surprises would be handled and tested already

[@toCatchAnAI]

"Tactical" Football. it's the best way to make max profit out of this.

[@jeremyo168]

I developed a similar GNN architecture for a well known professional American football organization.

[@LimabeanStudios]

Is the public missing cool progress happening in the private internal sports tool world?

[@zoomingby]

Here's a cookie.

[@iancurtis123]

​@@LimabeanStudios people have used GNNs for football publicly for a few years now

[@thePyiott]

This is going to be great for Arma 2

[@westwardquest]

Imagine a team of footballers all fitted with vibrating implants that guide where they need to be on the field in real-time and we all watch AIs play against one another. Could call it AFIF.

[@StonedApe420]

Every team member has an earpiece and AI is guiding him " Player number 7, a bit to the left, please"

[@hq1n]

Pep has entered the chat...

[@acdc2468]

How many years away are we from "Lets replace human players with humanoids who can respond better to our AI tactical algorithm providing real time positioning, "velocity" and instructions"? Convince me the game will actually be better with all teams working through tactics generated by machines. We will arrive at a very bland brand of the game where everyone predicts each other and responds accordingly to a dire state - individuality will be punished, collective styles will be punished, unique team compositions will be punished. etc. Look at what Petar assumes coaches actually do - 17:09 - "It will free up coaches to do LESS of staring at situations and more of the actual creative and actionable decision making part" - He doesn't understand that creativity isn't something you arrive at, its the result of the very process that technology is replacing. Ask an artist if their creativity is truly manifested when they prompt stable diffusion to spit out 20 images in a minute OR when they spend 2 hours painting 1 image? The PROCESS is the creative part. The "Creative Decision Making" IS the result of the boring "Staring at situations" part - without that you might as well give a monkey 4 options to choose from and go from there. Also, Who would actually enjoy that sanitized version of the sport? Another case of technologists pillaging industries/fields they have no attachment or stake in (so they don't care what they destroy). They're doing it to art, doing it to sport....go solve real problems that are facing humanity - ecological crisis, affordable housing crisis, global energy crisis etc. instead of continuing to destroy the last remaining bastions of human existence & creativity. AI technologists are the human version of lemurs running us off the cliff. Ask how many technologists studied the humanities alongside their software development, statistics courses etc, next to none. This is why you have technologists who don't understand the social and human impact of the work they are pursuing, more importantly, why they don't care. Is the technology cool? Sure but just because you can do something, doesn't mean you should.

[@3p1cand3rs0n]

very well stated, thank you.

[@ejenkins4711]

And wot about the after math

[@fiscallylogicalsocialhuman]

Who wants to watch or play a sport where the players have to be brainless human drones doing what the AI tells each team. Then it’s AI vs AI with meat sacks as the chess pieces. This is going to get dystopian real quick.

[@fungool]

Isn't that NFL football? Lots of people watch it. The whole point of playing is that players are not perfect robots carrying out instruction exactly. Mistakes are made, improvisations are made, and not every person has the same exact characteristics (leading to a variety of tactics that emphasize their strengths).

[@sp123]

This has already happened to NBA where no team has a unique identiy because every team has to shoot 30 three pointers per game

[@djcardwell]

i doubt it

[@elvoandro7087]

was this not a plot to the anime tzubaza

[@VoidLantadd]

This is a fascinating topic, but it's hard to focus because the greatest mystery is why he decided on that facial hair. I can see where he has stubble on his cheeks and further down his neck, so he could grow a full beard, but he's gone for a ridiculous looking chin strap. He needs to grow it on his cheeks, but most importantly his neckline is WAY too high and it's so bad it's actively distracting.

[@StonedApe420]

How can i get my hands on "Punters Wet Dream". I would make this baby sing.

[@supermodal]

SportsBetAI when?

[@alexijohansen]

This sounds like very complex and hard to collect data?

[@MeatCatCheesyBlaster]

Can't be more complex than learning how to play Go

[@tuckerbugeater]

@@MeatCatCheesyBlaster how many dimensions are in go?

[@Frankbug]

well i guess clubs will save money on managers

[@halim7725]

I've never been into football and although this is very interesting technologically speaking, it makes me even less eager to watch football. It will significantly reduce the flexibility of football teams to play if everything is about data and sticking to a plan predefined by super calculating robots.

[@fungool]

Maybe if you watch video games. I'm not sure that this is going to lead to boring games at all. Why would it not lead to more varied types of tactics instead of every team playing tiki taka, playing out of the back, etc. As Petar said, coaches will/should have more time for creativity if they are not focused on some details that can be handled by a system built on this tech.

[@citizenoftrone6570]

А помните дом Фландерса?

[@aytunch]

It still amazes me that AI has not found any cures to cancer or similar problems. They have all the data they need. Petabytes of MRI and Medical condition documentations. Why spend time on corner kicks and not medical side? I guess it is about funding or to put it better, the lack of funding if they decide to disrupt the already money making sectors.

[@RandomGuyOnYoutube601]

counting on symmetry in football seems pretty stupid. Football players are not chess pieces, they have affinity for one side or the other.

[@jonnygemmel2243]

Ultimately, football crowds will come to realise that they have seen pretty much every game possible play out and then the unthinkable will come with a loss of excitement and interest in football as a game that was once full of endless possibilities

[@fiscallylogicalsocialhuman]

Exactly! It will become pointless to watch, or even play.

[@fungool]

@@fiscallylogicalsocialhuman Oh yes definitely. Since GPT4, reading, writing, painting, music have become pointless because everything has been created. My wife has left me for an AI clone of myself! There's no reason to live anymore, the machines can generate everything!

[@codevev]

This clearly hasnt helped Liverpool lately 😂

[@maxwelltsholofelo7581]

Im a Liverpool supporter but I agree with you on this one.

[@hermancharlesserrano1489]

This is the moment we discover that those unpredictable variances he’s talking about are predictable in themselves, that the model proves predestined behaviour, that the universe has no free agents, no free will…jk Feels like if this tech were implemented, managers would seek to deliberately confound it, so you’d need realtime updating to revise any plan you might derive/seek to implement…then managers and coaches would be second guessing themselves against the output… sounds risky, but you run the models and see what can be taken advantage of, right

[@ernesto8738]

baseball is the perfect sport for it, so discrete

[@nathanielneveryman]

Yeah but when are they going to do football, which seems perfect for AI.

[@edisco3643]

Chelsea spend 200 mil for AI Corner tactics only for Sterling to fail to beat the first man

[@rolestream]

Wow, a video that makes football vaguely interesting. =p

[@bananalord8575]

A Serbian AI researcher uses his power to beat football. Ofcourse!

[@ShadowD2C]

Im versed in CNNs but I only understood 25% of what's being said here...

[@elingrome5853]

surely the main limitation is that the AI cant directly control the game...

[@bennourmedhsin2441]

Who's this footbal club investing in this?

[@bennourmedhsin2441]

Liverpool it seems. Noice now we know why Klopp is on another level.

[@paxdriver]

All we need is a coach, can it do that? Lol #YNWA

[@The_Quaalude]

The NFL is gonna have ai coaches soon 😭

[@nickoslekkas3336]

didn't help vs atalanta :P

[@malakiblunt]

AI dopeing

[@abdullahal-madani]

nice video but you need to rilax with the camera blur

[@abdullahal-madani]

Do not confuse rilax with relax. Rilax is pronounced 'Rie-laks' but does convey the same meaning

[@TheBoogerJames]

That man is fully committed to that neck beard. You can see it trying to grow up to his face, and he said "nope."

[@happylittlemonk]

Soon AI will give you the winner before they even play so there will no need to play the game.

[@fungool]

Yup, people are known for carrying out all instruction perfectly 100% of the time...Don't people watch chess? No need to watch or even play as we should be able to calculate the winner of each game before anyone plays right?

[@RogueElement.]

😂 exactly lol. Some people are stupid bruh. Games are fund due to human error​@@fungool

[@Jr_2132oo]

How to completely suck the life out of human competition..

[@BalkanSpectre]

Genuinely it seems to be a false approach for what football is and it is used as an easy to grab funding case study for development in other fields. An AI used in football should aim to use available data that are hard to be collected by humans (in mass) to highlight tendencies and weaknesses of players to perform certain tasks (that are expected of them -i.e. the will be replicated during the game-). Similarly, trying to aid in figuring why a certain player can't perform their role in your own team.

[@Alice8000]

Very cool! Can you make my wife love me again? 🕯

[@spiffymagicman7284]

I'm invading Shanghai

[@guyincognito1985]

Football? 🏈🏈⚽⚽🏉🏉

[@augustwanger2651]

Keep AI out of football. The only real thing we have left

[@MeatCatCheesyBlaster]

The real money is going to be in predicting the winners.

[@TreeLuvBurdpu]

Note: it's actually just about soccer.