Daryl Morey on AI and the NBA

"I think it's the biggest tech advance of my lifetime since the actual computer"

One key mission at 🏀 5x5 will be to shed light on perhaps the most important innovation of the past few decades: artificial intelligence (AI).

We want to provide a guide to how AI and machine learning are making an impact — and will potentially make a much larger impact, for better and worse — on the NBA in particular and the basketball world in general.

Part of our series on making AI and the NBA easy to understand, including its potential for both positive and negative outcomes.

Earlier this month, Philadelphia 76ers president of basketball operations Daryl Morey appeared on several panels at the 2024 MIT Sloan Sports Analytics Conference, the massive annual gathering that he and Jessica Gelman co-founded in 2006 — which I’ve been fortunate enough to attend most years.

Though this year my participation was limited to catching the YouTube feed, it was with special interest that I watched the panel discussion that capped Day 1 of the two-day event: “Winning with AI: The Future of AI in Sports.”

The whole discussion was worthwhile, and you can find it here. But Morey’s comments — given his significant role in an NBA front office, leadership in the sports analytics movement, and candid speaking style — were especially noteworthy, not to mention pithy and cogent.

So as a way to introduce the topic of AI and the NBA, we are providing a portion of what Daryl had to say to the gathering in Boston.

Here are some of the topics discussed, along with Daryl’s responses, very lightly edited for readability:

How significant is AI?

I think it's the biggest tech advance of my lifetime since the actual computer, which came out for PCs in like ‘78, ‘79. I think it's bigger than the iPhone pretty easily because it's even more fundamental.

We don't have multiple panels [on AI at the Sloan Conference], like I said earlier, because every panel should just be about AI, because it's that important for the future of, really, everything. 

---

How do the 76ers use AI?

Basically, all it is is prediction, right, whether you get into the large language models — just prediction of the next word, based on a huge heuristic space. 

Getting to the Sixers, though, we use it for productivity — on the back end, just like you were mentioning, [Microsoft] Copilot, things like that — and then we use it to try to improve our predictions. 

So far, in terms of improving predictions, we haven't found a ton there that beats the stuff [we already use], but that will change. That’s definitely going to change.

---

Click here to see the panel discussion on YouTube.

---

The mystery of AI

I've been in computer science for a long time, and this is the first thing we don't understand. Like, there are literally papers [saying], “We don't know how it's making that prediction, and we think it's this, and its trillions of nodes.”

That's mind-boggling; we've actually now created something with 0s and 1s where every step, we made the creation, but we don't actually know how it's producing the results. [This] is completely new, actually, and people don’t talk about it enough.

---

The scary side of AI

So there's a lot of scary things with this, but it's like it sort of is. This is happening. There isn't going to be really [a] way to like do all this safety stuff. There isn't. There's going to be open source, there's going to be multiple countries, multiple actors. 

So there isn't actually going to be any way to control it, so you really just should, you know, lean into it, honestly, for me — to help your business, to help what you’re doing. 

And there could be a very scary thing you hit, but what's the alternative? Not embracing it? It makes no [sense].

---

What data or AI skill sets are important in your hiring decisions?

The question’s important, and I think it's important for the audience — especially folks who are either just starting their career or even later.

It really is about adaptability. It really is about having problem-solving skills, the ability to learn something quickly, the ability to make decisions in a group and work well in a group.

Even as you add machines and AI into your group, they just become another component of your team, helping make decisions, helping produce a result. And so it's those foundational skills way more than, you know, Python or C++ or whatever the du jour coding is at that time.

The roots of AI

Anyone who’s been in statistics and predictive modeling, AI’s been around forever. 

I took a class from Geoffrey Hinton, here at MIT, who’s like one of the godfathers of neural nets, and Michael Jordan — not that one — was his co-professor. And just to show you the transition, Jeffrey Hinton was being made fun of by Michael Jordan, that neural networks would never amount to anything whatsoever. And he had no argument back. So this was in ‘97. Which is a fun story.

---

On the NBA using AI for player health and safety

I know they've done a lot of studies on player health and safety and done a lot to fix our schedule, which was a big, big issue. The word AI can get overused, so I would just call it a very well-structured algorithm to set up the schedule to minimize the number of times players have to play multiple times. 

They also commissioned a study on what things drive injuries and which schedule things drive injuries, and I know they factored that in.

In fact, shout out to Evan Wasch, former lead of the conference, who worked on that with a team at the NBA.

---

What are the limits of AI (vs. human knowledge)?

If by AI we mean like the more advanced machine learning techniques, which is large data sets pointed at problems — which we have overhead camera data, we have pose data, which you mentioned.

At least currently, the general “take the big data, point the recipe of models at it, and see if it can predict which draft picks will be better, which players will perform better in the future” … So far, none of those are beating our internal stuff. That doesn't mean it won't happen. So that's sort of the current state — we haven't found one.

In terms of just taking the pose data, taking the overhead camera data, and saying, “Tell me which players are going to predict better …” That day may come, but it hasn't come, at least for us yet. 

The challenge of machine learning and AI

So, pick-and-roll, it's like the simplest play in the history of basketball. It's literally, a person's driving and the person [guarding them] gets hit by someone else, so that person can no longer just guard the guy. It's the most simple play.

Turns out to get complicated in practice, because one of the best ways to confuse the defense is to go act like you're about to set a screen, don't actually set the screen, and run away quickly — slip it, or there are other terms as well. Or you can just run past each other like you're going to [set a screen] and just keep running. And the defense has to, “Oh, there's a screen coming. I got to … OK, quick … Oh, he didn't set the screen.”

So if you put all that in [to a computer], coaches don't agree on what necessarily is always a pick-and-roll or what wasn't. It's a very fundamental component of our game, though, so tagging it's very important. So really it's almost like you have a dial of, like, this is 95% a pick-and-roll, this is 5% a pick-and-roll. 

But you have to set a threshold and say it is or isn't for most of the analysis coaches want, which is basically, like, “Whenever we run a pick-and-roll between these two people, what happens? Do good things happen or do bad things happen?”

So, it's just an example where AI is helpful. It creates the classification, but if you go back and look, all the ones it tagged as pick-and-roll, it's not going to get them all right — and neither would a set of coaches.

If a human can't do it, it probably means AI is going to have some trouble. Often they're superior [to] humans, though, in those environments.

---

---

Would he tell the Sixers’ owners he wants to reject a trade based on data produced by AI?

It wouldn't go exactly like that, but something close to that does happen, right, because Josh Harris and David Blitzer — our two primary owners, very intelligent guys as you might expect. So, when people are, like, “That person has final say of X,” that's not really how sports works.

You have generally smart owners, you generally have very smart GMs. You make your recommendation, like, “I want to do this trade.” But it's very important you then also walk through the reasons, like, what are your main tentpoles of support for that recommendation? 

And one of them can be a model says, “This guy is better than that guy.” Then you have to follow it up with, “And that model has very good reliability and reproducibility in this environment. We're in a slightly different environment, so we think the variance is a little higher, but we're still quite a bit above the threshold of a trade.”

Yes, we do go into that detail. I assume there are some teams where maybe the owner isn't as involved, but honestly, sports teams are bought by billionaires generally. They love sports, and I just always ask, “If you were a billionaire and you bought a sports team, would you just not be involved at all and just be, like, ‘Ah, they're fine.’” 

You'd want to hear and be in the action, right? And I think it's actually appropriate, because it sharpens the team to be able to explain why they're recommending something.

And so, yes, an AI model could spike a trade. You would have to feel comfortable that model has produced good results in the past.

You're only as good as your past data, really with anything — humans, too. People make like it's a new thing, but humans are just a product of their environment data as well. And so, if you think that you're in the same or similarish-enough environment that the model is producing to what it's saying is the right thing, that ups your confidence in making that decision.