Transcript of Inside the Octagon: How AI Brings UFC’s Fastest Moments into Focus

This is an iHeart podcast.

Guaranteed Human. I'm Malcolm Gladwell, and you're listening to Smart Talks with IBM. Most of what happens in a UFC fight is too fast to see. A fighter drops their shoulder for a split second. Did you catch it? A shift in position that looks insignificant but changes everything. Were you watching? Alon Cohen is the head of R&D for UFC. His job is to help people see those moments, not by slowing things down, but by knowing what to point to after it happens. By the time you realize something mattered, his data systems have already figured out why. It's taken 15 years to get here, first with paper scorecards and a TiVo, now in partnership with IBM. And along the way, Alon's learned something that applies far beyond fighting. The best technology is the kind you don't notice at all. It just helps you see. I thought you were gonna be like tattoos, muscle shirt. I thought you were gonna like, I thought you were gonna represent the brand in that respect. That you yourself would be a kind of, you know, mixed martial arts type. You are not in fact a mixed martial arts type.

I did a little taekwondo in my past. I am, I am not in the way that you put that, not a mixed martial arts type. But I think if you, if you come through the building in Vegas, if you come through the building when we do a show, you're going to see a lot of the mixed martial arts type. Yeah, it even surprised me. I came out of the tech world, the young tech world in the 2000s, and everybody was talking about, you know, we have to have a mix of viewpoints and a mix of, you know, people from all walks of life and all of these other things that evolved into, into something else. But yeah, that didn't happen to me until I went to work for UFC, which is like a weird way to get there.

Where were you in the tech world before? What were you doing?

So when I left school, I went to a startup that today we would call a big data company. And it was there at a moment where everybody was doing relational databases and relational databases from MicroStrategy and from Cognos and from whoever. So we did that. 2001 hits after I graduate, 9/11 and the dot-com bust and all of that kind of stuff. And we all looked at each other and we said, no one, no one of these institutions that needs this kind of information is going to take a risk on a small company right now. They're going to retrench. And I went to law school.

Oh, I see. You start in tech, you briefly have a foot in the 21st century, and then you decide, no, I'm going to go and get a law degree.

I did. People said, you feel like somebody who would benefit from this.

Yeah.

And I wanted to want it. And the first attorney I worked for out of law school, he looked at me and goes, you need to go be in business. Like, you're not an attorney. Yeah.

So you're, you're, you're a failed lawyer. They kick you out. Where, where do you go next?

So that's what— so in 2008, yeah, there were a bunch of lawyers who were doing just fine, but the bottom had fallen out of the legal market. I was helping a friend off the side of my desk. Who had come to me with a banker's box of paper and said, I have been watching The Ultimate Fighter, the reality show where they pick, you know, new entrants to the UFC. And I had been watching. He was a sportswriter before, and he very much— this gentleman's name is Rami Genhour. Rami comes to me and he says, there's no stats for this stuff. I'm used to writing about baseball. So I did a full regression analysis after watching 100 hours of fight time. And now we're at closer to 200 hours of fight time. And I've been scoring these fights., and the UFC's talent have been reading my blog and they want me to do analysis for them and I, I can't be doing it on paper. And so he came to a technologist and said, what do I do? And we talked about double keying the data in Indonesia and building a database and APIs. A year later, we were making small salaries and we had gotten this thing up off the ground and going, and the UFC was starting to use our statistics and we had turned it into a real data product.

And, you know, we were able to score electronically at that point and, And the rest is, at this point, 17 or 18 years of history.

So wait, back up. This is really interesting. Pretend I know nothing about UFC. Okay. So Rami comes to you, and Rami's issue is what? That the way in which these matches are scored is too subjective?

The way that these matches are scored is inscrutable to the average person. Yeah. So his analogy, and it's, it's the one that he experienced is when you write a summary of last night's baseball game, right? What do you do? You say, here's my thesis statement. So-and-so had a great outing and middle relief collapsed or whatever it is. And last night at 7:00 PM at this stadium, the following happened. Here was the final score. And then this middle paragraph is, you know, middle relief had this as a CRA and so on and so forth. And you'd run an analysis and then you talk about, you know, so-and-so has been performing well and he had this outing and that outing and blah, blah, blah, blah, blah. This is of a trend.

You tell, you tell the story of the game in the context of a, a stat— a core of statistical knowledge.

Exactly right. And this is at least how he wrote. And he— you get to that paragraph in an MMA story and it's nonexistent. It's so-and-so is a strong striker, so-and-so is coming off of two losses, so-and-so is on a three-fight win streak. And that is the totality of what's currently available. And Rami's approach is "We can count everything, but everything's not important, and you can't count everything practically," because by the time he gets to the end of this first iteration, we have 67 data points per fighter per round. So it's a ton of stuff.

At this point, you're collecting the data how? Just visually?

He has a piece of paper, and he is on a TiVo. Remember TiVos? We had TiVos until long after they were dead, 'cause we needed them. And he is playing and pausing and rewinding to be able to—

He's doing it the way people in the old days in baseball, remember they would score, They would take their piece of paper and they would score the game as they were watching.

100%.

He's scoring the matches.

His history is, yeah, I go to baseball games with Dad. Dad has a scorecard. And you as a kid are both drawn to it and averse to it because you're like, Dad, you're missing the game. And at some point you find yourself marking down the scorecard. And he brings a lot of that to this.

Yeah. So you say in his first iteration, he had 72 data points. Say that again. 67. 67 data points per—

Fighter.

Fighter.

Per round.

Per round. Yeah. So that's consistently there, or—

So this—

That's an average of—

Well, so at any points on the sheet, there are 67 different things that we are tracking.

Oh, I see.

So there may be more head strikes landed, power head strikes landed, and there may be some other head strikes landed, but there's probably not a tight submission or a kimura or something like that. But these are different kinds of submission types that we would track. And so all of them are available on the sheet so that you can just tick those off and move on, right? So those are the 67 independent boxes on the sheet that we could mark something down on.

Give me examples of things that, non-obvious things that you would learn if you did a formal statistical analysis of a match. Like as a viewer or as a fan, how is this enhancing his experience of what he's watching?

A couple of things are happening inside of an MMA match that are, worthy of note to the average fan. It is scored per round. And one of the problems that you have when you score per round is that people don't think per round, right? They don't get to the end of round 1 and say, round 1 is over. I will make a decision about round 1. Let us now go to— they don't do that. They look at the totality of the match. The shorthand I used for this for a long time is you make the invisible visible to people. Like, what exactly did I see? The other thing that— the last thing that the numbers do for us is they tend to combat recency because you just forget what happened in round 1, or it just doesn't leave as strong an impression. And one of the— and of course, fighters fight to that. What do you do right at the end of the round? You take a dude and you take him down because you were going to finish the fight? No, because you wanted to leave the impression— the impression in the mind of the judges like that dude controlled that round.

And so you were trying to score these big shots at the end of a round. And the numbers allow us to battle some of that for our fans to say, like, how well did he actually do? Did she actually do in this fight? So that's one of the big learnings is you're able to go back in time and not have to watch the video. You just get a spot to be like, oh yeah, there were 2 takedowns in that first round. Yeah.

And so on. So you're allowing people to construct a much more complete and accurate narrative of the, of the match.

I think so.

Yeah.

And I think for new fans, there's a different calculus. 'Cause new fans don't know how to watch the sport. And when you, it's, I joke that when we all watch the Olympics, like, I don't know anything about synchronized swimming and I shouldn't pretend to, but the Olympic broadcast allows me to pretend to. Right? So that, and that's a little bit of what's happening here is that they can look at this and say, oh, I get it.

Yeah.

Right? Like I'm watching all this action. And then we condense that down to this is the number of leg kicks this person landed. And they're like, oh, I did see, okay, now I know to focus on that. This is part of the storyline., and it, it is a signpost. Yeah. And that's a big deal.

So in the beginning, you're just a contractor with UFC.

We are a contractor entity. Yeah. Yeah. For the first 7 years of our existence.

Yeah.

We are independent and are the official data feed of the UFC.

And, and, and at a certain point, UFC says, come and join them.

Ari Emanuel at WME IMG at the time. Yeah. Buys UFC and he says, I wanna sell the data. Where's the data? And they said, well, these two nice kids in the Georgetown Waterfront, they have the data. And he goes, what do you mean they have the data? Like, it's our data. And that started an acquisition conference. We were a fold into that larger acquisition. Yeah.

Going back to the fights, the data, how is it, how is the scoring working at this point?

So the TiVo I brought up earlier.

Yeah.

You're able to watch it with slow-mo and rewind. And we have at the time a tablet or a touchscreen. And we have taken our piece of paper and we have split it in twain, and you're able to mark what you see as you see it, save a round.

You've got score— you've got trained scorers.

Yes. Yeah, they are some of my favorite people. I— we have had trained scorers in our group that go back 13 years with us, folks that we plucked out of the ether, right? You put something online looking for people who are interested in learning this stuff, and then you have to whittle out the people who can learn how to do it and then do it quickly. The first time you score something like this and try to make sense of the of the scoring page, it's going to take you 15 minutes to do a 5-minute round. Might take you more. And that means that at the end of the fight, you're probably 35 minutes late. And so you have to do this for 6 or 7 months on a regular basis, paying you, make sure that, you know, you're fairly compensated to get you to the point where you're able to make decisions rapidly. It just comes as second nature to you. And this is the dirty little secret of UFC, but we joke we've been trying to put ourselves out of that business. For at this point 11 or 12 years, and we aren't close yet.

Put yourself out of the business of having to rely on a human score. A human score.

Yeah. And we said that joshingly, but there's a sense of which— there's a sense in which at some point you understand that compute ought to catch up with you. Right. And we said we should be out ahead of this as sort of the Netflix model, right? If DVDs are going to be a problem, you should figure out what's coming next and you should do that also. And disrupt yourself. And, you know, we wanted— initially we have used it as additive data, so we wanted to do motion tracking. So 11, 12 years ago, we bought a motion tracking software and it couldn't make sense of any of what it was seeing inside the octagon. Like, it really couldn't understand. And this wasn't even striking. This is just like red is here, blue is here, and the ref is here. And, and so we taught a human system how to do it, but we keep going back to the, to the machine well, hoping it will get better. And today, in fact, we are We are closer than ever, but nowhere near close to overtaking us as scores. It's weird. I have been playing, as I said, with my motion tracking computer vision experiments were 12 years ago.

10 years ago, I was hanging cameras to attempt to do skeletal tracking to see what that could give us. And it didn't work. You know, we have been on a 10-year journey to discover what it can do and being fostering of it, right? That's what R&D is supposed to say when something fails. Say, well, is there something of this that I can retain? Does this add some value enough that it really makes sense to just continue to invest in it some? So if it can ever make good on its promises, we'll get to see exponential value.

What's driving you here is what? You want to get your data quicker, or do you want more data? Do you want qualitatively different data? What is it? What's, what's, what's in the back of your head?

What are they here? Man, so many motivations are pushing this for us. One motivation is faster data. So imagine the following, very simple. If I see a significant strike landed in the middle, and so let's say the first person to 100 significant strikes landed in this fight, that is potentially betting data. Somebody could bet who could get to that first. Right now I can only offer that data unofficially because my humans need to wait for my official data to come in. If computer vision could really get that right, then I could stand behind that number and say that is betting ready. And you could resolve that bet at the moment it happens. And so you open up all of these markets for one. Like that's a, that's a big motivator for this kind of stuff. The second is I am, when I'm adding these AI technologies, I'm often looking at them and saying like, what else can I do with this? What else is it telling me about the fight that I'm not already getting? Motion was a big one of them, right? Where are people in the octagon? When I mentioned earlier that one of the criteria for judging a fight is control of the octagon space.

Well, okay, everything I was tracking with my scores before doesn't talk about that, right? If I'm hitting you more and you're, you know, defending poorly or whatever it is, maybe that's control. But what if you're a counter puncher, right? Famous fighters who are counter punchers, I draw you in and then I hit you. Does that mean I'm ceding control to be able to take back control? What does that mean? And so we didn't have any, any numbers on where people are. That's an immediate use of AI that is additive to what I am doing. And all of this is always in the service of storyline, right? What is not being captured in the fight? And what should the commentators be talking about? And how can I support what they're talking about? How can I give them something that makes them think,, oh, actually the thing I wanna be talking about here is this, right? That's ultimately the goal. The goal here is, I mean, let's go back to the first thing you said. You said, I don't look like the kind of person who might be an MMA fan. There is a point at which if I throw up a number on the screen and it's like, this person is 27.24, and people are like, nerd, like, what are you doing?

And there, there is a space for that in other sports, but sometimes we're doing it just to seem technical. I think if you're doing the tech right, you should forget about the tech. You should forget about the fact that there's AI behind it. You should forget about all of this. The thing that hits the screen is so-and-so has never lost a fight if they have landed 4 takedowns. We're on number 3. Now we know what we're watching. Yeah, right now you have a binary, simple thing to look at. So-and-so has a has a takedown accuracy of 94%. So-and-so has a takedown defense of 75%. Let's see how they match up against their classic averages. And now you're getting into some more techy numbers. You're like, this guy is below his average and that guy's above his average. Okay, now we have a way of measuring very quickly who's winning and who's losing across at least this part of the fight. And so the goal is to get to storyline as fast as possible in human understandable terms, right? This is in some ways the great moment of the LLM as well, right? Who cares how many weights and how many parameters and how many whatever?

When I talk to it, it feels like I'm talking to a human. That's the great unlock. The technology suddenly disappears and the experience stays. And that's at our level, at our level of sophistication, that's what we're doing.

Yeah, I wanna go back to make sure I understand. When you talk about control of the octagon, of the octagon, conceptually, what is control of the octagon?

It is the last of the criteria in the judging rules. And it should mean that in a fight, one of the things you're trying to do is assert dominance. The most basic level. It is left vague. Different commissions around the country will tell you slightly different things about what it means. Some of them will boil it down to if I hit you more, I control the fight. But a lot of others of them will say like, it is really there to be that sixth sense feel of was I able to put you where I wanted you to be when I wanted you to be there? And in our statistics, one of the ways you can talk about that is advancing or center control. If I always own the center, you always have the wall at your back, so I control the space. Right. There are ways of finding small analogs, directional pointers to tell you that. But when you talk about what a judge is looking for, they're looking for that ephemeral sense of, you know, who's the big dog.

And that is, if you're strictly kind of scoring the match, you— that's one thing you're not picking up because you're too focused on position, Who's striking whom? Yeah. That kind of larger gestalt. You're talking about a gestalt, right? Sure.

Yeah. Somehow the Nevada State Athletic Commission did not get to criteria 5 and go, Gestalt. That's what we're looking for. We're trying to deal as a matter of course. We want to bring people into our arena, have this experience be devoid of Weltschmerz and focus on the gestalt. It just didn't make it. That's what I'm saying.

So you've got— your motivations are, one, I can enhance the viewer experience in terms of story. Yep. I can facilitate things like betting on the match, and I can aid the judges in getting a sense of who the dominant party is.

One and two, yes. Three, we actually scrub all of our— all of our video feeds into the arena are scrubbed of our statistics. I do not want the judges— their job is to follow the criteria set by their commission. So we're in this weird spot where, like, my sport has— my fighters are independent contractors. It's a regulated sport by the state commission that we're dealing with. And my job is to, is to bring the event together. And as a result, I don't want anyone pointing at us saying, you said, and so I decided, like, look, you understand your criteria, you're a fully trained judge, you need to decide who won this fight. Our statistics will tell the story our statistics tell. And by the way, we can have an occasionally antagonistic relationship where we thought that was a poor decision, right?

Yeah, that makes it interesting. Exactly. Yeah. So, okay, let's go back to the AI. So you start experimenting with this a while ago. Oh, yeah. And you're not entirely happy with the results. Is that fair? Or you understand it's embryonic?

Both. Right? You— why should I start down the project of making this work if I don't believe it's going to work? Right? I try— if I'm going to bother dedicating the time and the effort and potentially a lot of money, I want it to work. Uh, and I was disappointed time and again, and I have found things to bring along. Right? So early skeletal tracking has become the basis for today.

Skeletal tracking is— define the term.

Skeletal tracking is when you hang cameras nowadays, you are very quickly able to discern what is and is not a human inside of your space. And then you're able to, to identify points on their body. Uh, nowadays more points than we used to. It used to be like shoulder, shoulders, head, torso, hips, knees, feet, right? And nowadays you have two additional points on the feet so you can see their orientation, hands, you can see the orientation, and face, you can see where I'm, where I'm looking. Those are basic skeletal models, and you need that to be able to understand from an AI perspective who is striking and was that a miss, was that a hit? We can get into collision detection, which is very difficult because skeletal tracking isn't, it's not the outside of your arm, it's the inside. And then beyond was there a collision, we get into problems of intent. So intent is the hardest thing actually to solve. There's a lot of times where two guys are fighting and one guy's going like this. He's range finding, just putting a hand out there to see is, you know, how far away are you and sort of to bat you away.

And that is not a strike attempt. How do you teach the computer that that's not a strike attempt? Because sometimes it looks like it and sometimes it doesn't. And then same thing with kick to the head. It makes contact with the outside of my arm, which I brought up. In our world, that's a blocked strike. Yeah, but teaching a computer what exactly that means and when and how, like when it's up here, when my arm is up, that's a block. When my arm is down and hits my shoulder, that's not. It's, it's those nuances that proved incredibly difficult for machines to be able to handle for a very, very long time. And the, the acceleration in the last few years, which has allowed us to onboard this technology, has solved so many of these problems at the same time. We no longer need fixed cameras. I now do it off the broadcast cameras, which are on people's shoulders and they, they shake and they rotate and whatever, doesn't matter. It can still figure out what it's looking at. This is the octagon. These are the fighters. Here are their bodies. That's the ref. He doesn't matter.

All of that stuff. It can now figure out, it can begin to understand what real landed and missed is, and it can do so on, on much less perfect data than we needed the first time. That acceleration has been incredible. It, If I didn't know step 1 and 2 and 3 and 4, I'm not sure how ready I would have been for where we are now.

Yeah. So let's dig into where we are now a little bit. So you go— I'm interested in the partnership with IBM and the kind of transition from this period of experimentation to where you are now. Sure. So where— at the time, you guys, you're aware of the development of AI. Yep. And it's getting better and better and better and better and better. And you clearly have an intuition that, oh, this is, this could potentially really open some opportunities for us. Yep. What's the point in time where you're, where you start to think, oh, this could be real?

Our, our relationship with IBM is born of a little bit of suffering on our behalf is the best way to put it. We had a partner as part of the larger business. That really wanted to focus on AI with us. And we said, we would really like to do an insights engine. We think it's really additive. These storylines in real time would change the way that we talk about the sport in broadcast. And for 2 years we went down that road and it crashed and burned. We can go into the reasons, but they didn't understand the sport. Uh, they tried to brute force the problem. And if you try to brute force this problem, it's going to kill you. And because there are just so many permutations out there. And IBM starts a conversation with us and very quickly as we engage, a couple of things emerge and they're not about the tech. They say to us, we think we can do it with the tech. We say, we have to believe you that you can do it with the tech. We don't know. We're not experts in it. They can bring people who understand sports analysis to the table.

So all of a sudden, when my phone call ends, when somebody at IBM has a question, they can talk to somebody. Better yet, within the first couple of weeks, I said, it would really be helpful to have somebody who's an MMA fan because some of your questions are specific to the sport and we could answer them all the time. But I think that the velocity here goes up if you can just say, hey Brad, what's happening here? Can you explain what are the, what is the culture? What's important? What's not important? Right? 'Cause you can come up with a lot of ideas, And fans would be like, I, I don't care. And it gets worse from there. So we, we quickly eliminated that with IBM. They had people on the bench who understood just the, the area and then the specifics of our sport, and they were able to fill in some of the gaps for us. And so we could move at pace. Mm-hmm. And it was all about the people at the beginning. And obviously beyond that, they had to make good on the, our tech can actually do this. We can use the AI.

It's not so much that the AI needs to run the final product, right?— that's not what we're talking about. It's can the AI speed up the creation of the product? Can AI help us expand its reach? Can AI suggest to us the things around the corner that we didn't see? That's where the AI came in. That's really the genius of this. Once it's all hard-coded down, the AI does take sort of a backseat to the code because once you have to go fast, once you need to be efficient, you know, that's the way that you do it is you run a whole bunch of traditional code and then you run it through an interpretation model to be able to give you natural language at the end. And that's super efficient in AI, but it's, it's an understanding of how to deploy both of those things in concert and have the right kind of people.

So walk me through here, IBM, what are they doing? They're, you, they're taking the video feed.

Is that how it starts? They're taking the data feed.

The data feed.

Correct. It's instead of trying to have them extract everything outta the video feed.

Yeah.

We've already done a lot of that stuff and we're doing it at speed. So we have my data feeds.. And then we also have any augmented information that we may be getting from external AI computer vision systems. And that's coming in as the source data to IBM. So now they're getting, they're getting a data feed and they're getting it at speed. Their job is to bust open that data feed into all of the various questions that somebody might ask and think about the interrelationships between data. So sometimes it's as simple as, this person is on pace to set the largest number of takedown attempts we've seen the last 12 months in this division. It's like a really nice talking point, relatively simple, all the way to here are two comparative metrics for these two fighters, one of which is this is his strength and this is his strength, and here's how they're matching up right now and, and everything in between. And to do that at speed.

Yeah. So they're taking two things. They're taking the data stream from your scores. And then also you, so you, you have these vision cameras.

Yeah, it's a computer vision AI stat system that we're using to augment our stat system.

And how, how does that work?

So the way that works is we'd finally found somebody who could do two things for us, give us accurate counts and be able to do so without us attaching a whole bunch of cameras to the octagon. Right? So they can do it directly off of my handhelds and they're able to do things like give us combo percentage. They can do, uh, punch and counter punch because they can assess that. They can talk about style. That's a cross, that's a roundhouse. That's right. Those kinds of things. And all of that is additive. Plus, by the way, they brought back all the motion stuff because now my AI computer vision can actually do motion for real. And so they brought all that back in as well. Into this.

Why can it now do— just because the tech is better?

Tech's a lot better. The ability to just know there's blue tape on the blue fighter's wrists and red tape on the red fighter's wrist. And we've gotten to the point now where skeletal tracking and computer vision is able to identify people by that and also take additional notes. That dude's in green shorts and that dude's in yellow shorts, right? And it'll do those kinds of things to just figure out who's who. And the hardest thing for a computer vision system to do as it does motion is when we become the 8-legged beast and we come apart, it can know that we're purple, right? We're red and blue at the same time. But when we come apart, it has to rapidly understand who's red and who's blue again. That's where a lot of these motion systems had a lot of difficulty. Refs in black, pretty easy to figure out, like somebody who's dressed in black head to toe.

You've got these data streams and presumably there is some overlap between them, but that's fine. You're just feeding as much data as you can from these two different sources into the IBM black box. And the IBM black box is generating stories for us, ideas, observations, insights.

I can give you a little more detail on how it happens. I think that the beauty of unpacking how AI works is there are moments of magic and moments of drudgery, and they have to work perfectly together.

Right?

AI, the data comes in to IBM and it has a storage system called Iceberg and it runs a classic data process called ETL, which is extract, transform, load. We care about our data at the fight level or the round level. IBM is trying to tell us things about our fighters, so it is consolidating all of its information in these big fighter tables that it can then sic the AI on, and then the data moves from the this ETL, Iceberg level, up a level. They actually, I think they do it by metals. So you start at bronze and you move all the way up to gold and you move up a level. So now your data is mezzanined basically for the, for the actual AI system to be able to draw on it and efficiently process this data into insights. And then it moves up into insight delivery where we have an interface that allows us to not just look at insights, but then I have a human factors problem. My human needs to see the insights and he can't see, or she can't see all of them, right? That's way too many. So you have to have a scoring system that says this insight is more relevant than this insight and surface only those so that a human, actually this was true of Watson a long time ago, right?

When it did medical diagnostics, it couldn't give you the answer. It could give you 7 answers and a doctor would look through them and said, got it, got it, got it. That's silly, that's silly. Hadn't thought of that one. And that was the value right here. And we have a similar thing, right? Where we'll get a whole list of them and you get to pick one of them and be like, go. And so that's how the data comes in. And to everybody else, it should be a black box, but I'd prefer it not be a black box to us.

In the course, this might possibly be a naive question, in the course of a fight, how many insights are being generated through AI?

Yeah, that's an excellent question. So in an active, let's say we'll pick, what I would consider to be the Cadillac of fights, right? So it's 5 rounds. It's a, it's a championship fight. So you have 2 fighters with a long history, good records. In that fight, I mean, it could be, it could be 1,000.

1,000?

Because, because every permutation is possible.

Yeah.

Right? So the system is generating just an enormous amount. And the real, the intelligence is in filtering it out so that my humans don't see 1,000. Do my humans see 15?

15.

Yeah, because remember the moment that, in a fight, there's not a lot of chances, right? I'm not gonna spit out all 15 onto the screen. In traditional statistics, I've got 2 stats at a round, maybe we don't have any stats at the first half of the first round. If something amazing is happening, my job is to move outta the way and just let that sit on the screen. And so my insights are there to be the most valuable thing I can put on screen at that moment.

Yeah.

And that's gonna happen in a fight twice, and over the course of the night, 10 times. But when it happens, everybody should have been talking about this fight like this, and now they should be talking about this fight like that.

Give me an example of the kind of insight you get that you probably would never have gotten in the kind of previous universe. Sure, sure.

I think compound insights are the most complicated, right? Where it's, we used to get basic, basic but difficult to answer questions. When was the last time someone won a fight after being down 100 strikes? Like, that's a hard thing to throw into a SQL database at the last, in a moment's notice and get that answer. But you can do that kind of thing with Insights Engine, right? Was like, you're like, oh, it is looking for, it's one of its pipelines, it's looking for those kinds of disconnects between the data and the reality. This is the largest differential, striking differential between two fighters, you know, in a middleweight title fight since— and here's this other iconic fight that everybody remembers, right? It's those kinds of things that you might in the back of your head think to yourself like, huh, this feels like— I don't think I've seen this in a while, but you can't, you can't give voice to it. And in a live fight, there's very little time. So what'll happen is two fighters, let's say it's a striking statistic, two fighters have not had this much of a striking differential in a fight since X point.

Yeah.

It is currently 3 minutes into the round of round 3, let's say. And the commentators are talking, we're watching the action, everything's happening. You get this sense in your head, I need to get in the next 2 minutes, the last 30 seconds are already, we're going out to commercial. Like we're not interrupting that. The fight could end at any moment. So there's time pressure. I've gotta hit a promo at some point in the middle of this thing. So you see how like my time is getting shorter and shorter to be able to deliver this. I've gotta have my producer be able to communicate it to the producer who's gotta communicate it to commentators, who's gotta be able to get it out of a system into a graphic and then sold in to be able to put on screen. All of that has to happen in, call it between 30, I said to you like, I have 3 minutes, so I have 2 minutes left in my round. Probably only have about 40-second window, and I'm gonna lose it starting the moment I think of it until the moment it hits screen. So I've gotta have about a 15-second turnaround.

And so it's not, you don't have to ask a very complex question for you to exceed the 15-second turnaround unless it's at your fingertips. And that's what Insights Engine is for. It's to kind of come in and be like, I know you have some cool ideas, but we're never gonna be able to get 'em into fight time unless something is— it's pushing, not pulling.

You need to have buy-in from the commentators.

Yes, we do. So first of all, being a commentator is a very sophisticated undertaking. Almost anything you see, and it's one of those things that you watch happen and you say, well, that doesn't look that difficult. But a commentator has an IFB, a little earpiece in, and the— I've watched them do this and I can't wrap my head around doing it, which is you must be looking at the camera and presenting information in coherent prose while the producer is talking in your ear. Not— and then when you stop talking, he starts talking. No, no, no. Talking in your ear while you're talking and you've got to be able to process that information and 2 sentences hence make it part of your delivery. And that information processing capability is by and large sophisticated. Now, people are people. Some people are more pro-stat, some people are more, you know, anti-stat. It doesn't really matter. You just have to know who that is. But they have a relationship of trust with their producer. And if the producer says, I want to go here, they can push back. But those two people work together on the regular. Producer is going to sell you the things he thinks you're going to buy.

Yeah, right.

Or at least the things he thinks, this is as far as I can push it and you'll really go there because he wants you to sound authentic. And he certainly doesn't want you to be fighting with him on air. Yeah, right. So that, that very human process is taking place. And by and large, one of the reasons Insights Engine produces human-readable stories is because I would have pushback if somebody said IBM Insights Engine says that so-and-so is 57.246% likely to win the fight. Be like, okay. And that doesn't feel authentic. But so, like, we've noticed that he's advancing and he's a counterpuncher. This is off. And so now hopefully the commentator is going like, wow, the movement in this fight is different than we would expect to see. And let me give you some things to bolster that, that point. Or, and people have done this, we have given them an insight and they're like, oh yeah, you're right. Like, that's what's weird about this fight.

Yeah. It's fascinating because genuine engagement with the insights is actually from the from a viewer's standpoint, the most interesting thing. The idea that there would be— there'd be some tension between what we're seeing and what's actually happening, that is one of the most interesting things in sport, right? Oh, I thought, you know, Steph Curry had a great game. And then I look at the stat line afterwards and it was like— or the opposite is even better. He looked terrible out there. And I look at the stat line, it's like, oh my God, he was like— He shot 70% from the field.

I feel like I do that more often than not, where you're like, that looked terrible. So a couple of things that are true in fighting that are, I think, true elsewhere is if people have a favorite, they tend to watch their guy. And so their guy always, unless he looks terrible, he looks great because you're not watching the other guy, you're just watching your guy. And that stats help bridge that gap a lot. Our commentators in general don't fall into that trap. And so when people fight with the commentators, when fans fight with the commentators, they're watching one guy. They're not watching both guys. I will watch fights and I will say, you know, I'm not watching, I'm not seeing the fight you're seeing. Some of that is because a lot of our commentators are practitioners. They're seeing something I'm not seeing. And then, but it is then their job to explain to me what I'm not seeing. And that's fine. I think that the, the other place that I would want to take this and where insights I think are very useful is our fans watch in a lot of contexts, a lot. And how many times have you walked by a game Right?

The classic is I'm at a sports bar and maybe the audio is on, but maybe the audio is not. Maybe it doesn't matter because it's loud. I want to be able to provide an experience for that person that is meaningful and potentially more meaningful. So yes, my insights are going to interact with my commentators, but if I can't hear the commentators, how nice to be able to put a storyline, not just a number, but a storyline directly on screen. Right? We'll have this with milestones. So-and-so has exceeded their best ever, whatever. And it's just a moment where you're watching, you're like, that's cool. And you'll go back to watching. And that's really, that's what we're going for.

Yeah. Are we fundamentally changing the relationship between the fan and the sport?

Interesting. I, I didn't set out to have Insights Engine fundamentally change the relationship of the fan and the sport. I wanted, I wanted something that fit like a glove, right? If, if you were sitting next to someone who had an encyclopedic understanding of the sport, and they were in and they were fun, right? That was a big part of it. What are the kinds of things they could tell you where you're like, oh, cool. And it's, it's more fun to watch because of that. And not, I think, where a lot of the push was in adding tech to the sport, which is we're going to make it nerdier. Like, that was the thing I don't want to do. I'm interested in leaning in to what makes this great and, and letting you see, right, the classic, the stats make the invisible visible, letting you see those things, but in a narrative arc so that you can contextualize them. You can say something more intelligent about sport, but like it— I wasn't here to make you feel like I changed the sport. I was here to make you feel like you were spoken to more crisply about it.

You were able to speak more crisply about it. And, and the fight just looked a little shinier, a little brighter because of this kind of thing. And like I said, when this tech does its best work, it just disappears.

As hearing you describe it, it strikes me that what you're doing is you're moving people up a level.

You're right. We are bringing it to a place where you feel like you can, you could even have an opinion. Because you understand enough of what's going on. You are 100% right. We're moving everybody right, right up the scale on a daily basis. I want to go back to Insights Engine is not here to feel technical, and that's the genius of it. Its simplicity, its narrative is the thing that allows you to bring people forward because you're not using jargon, because you're not just into the numbers, right? That's, that's really, that's really the key here.

When you look back at your experience with fight data, are there a few categories that are essential to winning?

I think the easiest way to answer that is the entire statistical system that we built is based on only one question. We don't track style, for example. I don't track whether somebody hit you with a cross or a jab. Even more ridiculous, in the official statistics, we don't track whether we hit you with an arm or a leg. And that's because the entire statistical system is built on this question: what moved you closer to winning the fight? Mm-hmm. And it turns out that what moves you closer to winning a fight is where you take damage, right? And how much damage you take. So if I hit you in the head, arm, leg, doesn't really matter. Style doesn't really matter. It's did it land? So the whole statistical system is built on that question of how much of each of these things did you do, these things which we have directly correlated to victory. And with live statistics and the augments, we can do lefts and rights and things like that that people wanna see. But that's the fundamental question. So for the question is, what's the most important thing you could do? Land strikes, land significant strikes to the head.

That's going to end most fights. Can you end a fight with significant strikes to the body that massively outpace your opponent? You can. They're at a second level down. Can— and then there are things that are a little harder to talk about that way. If you have more ground control, that's putting you closer to completing a submission. So that again is, is the kind of thing. And obviously once we get to the number of submission attempts you attempt are going to make it more likely that one of those will be successful. But it is all built from this basic question of all of our stats. Does it get you closer to winning the fight?

Yeah. One last question. I know you thought about this a lot, which is next. What's that like? If I, if I, if I invite you back, we have this conversation 5 years from now. What is the fun wrinkle that you're working on?

We have a couple of things that are in the works. Some of them are very sideways to this, but they're— So one of them, which we're prototyping, is the ability to bring additional voices into the mix seamlessly. So I'll give you a perfect example. If the coaches who are on the side and they're yelling in whatever language they are yelling, they are attempting to influence the fight. And it is— And we do this occasionally. We'll cut to the coaches' cam and we'll listen to them for a moment, but they're doing it constantly. Wouldn't it be interesting to understand what they're trying to do and what it means? So for example, if I'm saying you're doing great, but you're not doing great, am I being encouraging? If I'm saying you are doing great and you are doing great, what does that mean? What is my tone of voice? What's my level of stress? All of these kinds of things so that you could very subtly be able to distill that information and bring it in and say like, by the way, if you wanna know what the coaches think, feel, do. That's interesting. So that's like a really nice one.

That's relatively close to you.

Now I have a sense of the psychological context of the coach and the athlete.

Let's go further away, right? You said 5 years in the future. I think 5 years in the future, I have 2 beliefs. I don't believe people want a lean forward experience as much as everybody would love them to want a lean forward experience. If you could watch a Steph Curry performance and I could let you control the camera, No, I don't want to get a dog camera. I don't believe that people do.

Yeah.

But that doesn't mean that those technologies don't have a place. So people are talking about volumetric capture, Gaussian splatting, all of these kinds of things that allow you to create 3-dimensional space and move through it. I am working on, I have a belief that one of the places I can't put a camera, imagine the two of us were fighting, is right here, right in front of my face and watching your face. And yet the explanation you want to give of, I hit Malcolm in the shoulder because he had dropped his hand and it was open. Show me, make it available to me. Speed up that process fast enough, create the telestration tools fast enough. And I think that in 5 years, those things are going to become rote. And we will expect impossible angles, all of these kinds of things to become part of our world. We are going to have to rethink what it means to watch, to be a fan of something, right? If you only watch a 15-minute cutdown have a NASCAR race. Are you a fan or not a fan? I don't know. Do you watch 15-minute cutdown of every race?

And how do I service that person and how do I engage with that person? If you do have this very— you said earlier that people only watch half a show, for example. If you have a fragmented set of attention, I can either fight it and rage against the dying of the light, or I could lean into it and try to service that person and make them a real fan and bring them into the rest of our ecosystem. And I think that that is going to continue to happen. I think we're going to continue to make an argument for live events and the pull of that. One only needs to go, as I'm sure you've seen, when you're in the room, there's nowhere you'd rather be. And it is so interesting, even though it is so much less good than the angles you see on television, you can't help but look away. We have— we're bought in on that thesis. But, but the moment you step out of that room, how do I give you the greatest experience possible?

Yeah. Yeah. Alon, thank you very much. Thank you. You've made me want to become a UFC fan, which is a statement I never thought I would ever say.

I'll tell you what, come watch one with us. Come watch one with me. Yeah, I bet you walk away and you're like, I'm not sure I'm watching every week, but holy moly, is this amazing.

Smart Talks with IBM is produced by Matt Romano, Amy Gaines McQuaid, Trina Menino, and Jay Carper. We're edited by Lacey Roberts. Engineering by Nina Bird Lawrence, mastering by Sarah Bruguière, music by Grammascope, strategy by Cassidy Meyer and Sophia Durland. Smart Talks with IBM is a production of Pushkin Industries and Ruby Studio at iHeartMedia. To find more Pushkin podcasts, listen on the iHeartRadio app, Apple Podcasts, or wherever you listen to podcasts. I'm Malcolm Gladwell. This is a paid advertisement from IBM. The conversations on this podcast don't necessarily represent IBM's positions, strategies, or opinions.