Can this AI really take over coaching football?
The news is filled with stories about AI taking over many jobs in the coming years. The jobs noted to be the most at risk are those that are repetitive or that follow patterns or rules — even jobs in accounting, healthcare, marketing and law are at risk. Others say that AI will work by our side, making our jobs easier by automating the routine, boring tasks. However, something like coaching football, with all its strategy and leadership components should be immune from AI — right?
Well what if we could replace a portion of that job, that for most coaches, is boring, repetitive and a waste of time. What I am referring to is the post-game interview.
Every week coaches (and players) are subjected to pointless, repetitive, inane questioning from the bottom 5% of journalism degree graduates, otherwise known as the sporting press. At best, these questions are along the lines of:
- What do you need to do to win? (I dunno…score more points?)
- What could you have done better in the game? (Not do [bad thing], do more [good thing])
- How do you feel about that bad play/good play?
- How does this win feel? / How does this loss feel? (It sucks to lose, it feels great to win. What else is there?)
- Can you talk about… (Sports reporters love doing this thing where they lead the question with their own editorial and then say “talk about that.”)
At worst, reporters ask complete “clown” questions:
- “Did you guys lose this game, or did the other team win this one?”
- “Can I measure your bicep?”
- (To Kobe Bryant) “Can you guard?”
- Of Thurman Thomas before the Super Bowl “How do you get pumped up for the big game?”
Because the questions are boring, repetitive and cliched the answers then become cliched as well. Like “one game at a time,” “we’re not looking ahead,” “we are going to practice hard and prepare this week,” “the other team was the better team this week” or “we were the better team this week.” or “they made [more/less] mistakes thats why we [won/lost] the game,” etc. etc.
Actually, a lot of people are fed up with sports reporters anyways. Remember Marshawn Lynch? All this doesn’t even touch on the fact that newspapers are already using AI to write sports stories, because…NEWSFLASH, you don’t need to be a genius to tell people who won or lost a game and what the score was. Would you even notice if we just replaced the interview quotes with text generated by an AI? Let’s find out.
To create the AI coach-bot, I started by scraping transcripts from over 1,700 college football press conferences from the last five years. After a little cleaning and formatting, I combined all the questions and answers into one long 20+MB text file that had a “Q: ” and “A: ” format. From there, I fine-tuned a GPT-2 (117M token) model on the transcripts file using a GPU in Google Colab. And viola, our coach-bot is born. (A lot of credit goes to Max Woolf for creating the simple GPT-2 repo which I used for this project as well).
The AI works by taking a prompt (in this case it’s prompted with a ‘real’ question) and then, based on the training data, it continues to generate its own answer to that question.
The following is a real question posed to a player during a press conference and one of the answers (A1 or A2) below is the ‘real’ answer given at the time by the real player and the other answer was generated by AI:
Q: “How did you think that dynamic worked for you two this year in the backfield? Nobody asked you a lot about it. You obviously want as many carries as possible, but how do you think that worked between you two?”
A1: “Between us in the backfield we’ve got a lot of protection jobs that are critical. So it’s like we’ve got a combination of all those things. We worked from that standpoint. We were able to establish a lot of situational ball leverage situations. And it’s been good for us in the run game.”
or
A2: “Well, it helped keep us fresh — or fresher, you know. It was a great bond, and we made some great memories. So I think it was a great time this year.”
Can you tell which is which? A2 is a bit repetitive ( a lot of “great” things happening), and A1 is a bit more nuanced and provides a longer explanation…so? The fake answer is A1.
Heres another real question with an AI generated response following:
Q: “Brett, we had a social media question. The day after the bowl game is the Army All‑American game. Did you participate in that, and do you remember any recollections or any of the players you correspond with still today?”
A: “You know, I think, you know, I just wanted to be very personal; just the fact that it happened I think is a great honor to speak for the Army. In terms of what was happening, that was the first time I was there on that trip, four years ago. I know it was a big milestone for me, and to be a Lord Chancellor, you know, being able to go to a National Championship game four years ago is a great accomplishment, just kind of a reward for the kind of effort that we’ve had to prepare for that game.”
Not perfect…the AI was going along pretty well until it started talking about being “a Lord Chancellor”, but theres nothing cliche about claiming to be Lord Chancellor. Compare that to the real answer given to that same question:
“I played in the Under Armour game.”
Seems like a win for the AI to me.
One of the first iterations of the model I trained was only on the answer portion of the press conferences and I think some of the quotes it generates are worthy of being printed on a Vince Lombardi motivational type poster:
“I think if I get caught, I’m only allowed to get caught if I’ve been caught to win the game. But that’s like a mirror image to me of what I have to do to win the game.”
-AI Coach-Bot
Finally, the best part about this, is that, due to the way the AI was trained, using Q: and A: formatted text, we don’t even have to prompt the AI with a question. We can just prompt it with “Q: ” and it will start asking and answering its own questions. Here is a completely AI generated Q and A exchange:
Q: Coach, what makes a good quarterback?
A: I think the guy that can really help you out is the guy that can run the ball. He’s a little bit undersized. So you can see in his legs that he’s a very elusive player, a sixth man, and he’s really more of a scrambler. And then also he has a quick release ability. It’s a pretty simple thing to play with. So I think it’s all a little bit of the same.
Q: And you’ve got your vision guy, right?
A: Yeah, well‑coached quarterbacks, you have to always have your couple eyes in the first pocket. That’s harder to do. You have to have your first and second and third and fourth-and‑six. So you have to make sure you can get around that guy. And then your goal is to just keep him up. So that’s pretty much a tight game plan for me.
Yeah, it mentions a “sixth man” but who hasn’t mixed up their sports analogies before?
Here is another fully AI generated exchange:
Q: Do you think that this is a great development for you guys and just to go into the season with a young quarterback that hasn’t shown any signs of slowing down?
A: Yeah, I think so. But I think you look at the freshman class that we have in the past and the fact that they’re seniors. You get first‑round draft picks and they’re all whoopin’ boners. You got a guy like Blake Sims who is mature, smart, and obviously very talented, and he’s a very good player, but there are a couple young kids that are playing very well and the combination of them is a key for us.
I have no idea where it got “whoopin’ boners” from but hey, why not? I feel like they’d call people “whoopin’ boners” in Canadian football, eh? I will also point out here that since it was trained on college football interviews, it will pick up on real names sometimes, like Blake Sims. Other times it completely fabricates names like “Jimbo”:
Q: Do you still have a chance to get back with Sam and Jimbo and how would you characterize your relationship with Ryan and Jimbo?
A: Yeah, just talk to them; maybe not that much, just talk to him, talk to him a little bit. I mean I’ve made some friends over the years. But I really like it just because we have the same way of doing things. Like I said, it’s not one thing. It’s two to tango, it’s not one thing
And sometimes it will go off the rails completely:
Q: Could you have liked to have been a junior and possibly have grandpa here?
A: I like to think about it like I am about my family when we make out here. I never really had a grandpa. A lot of people said, I can’t think of any grandpa here. I’m just glad that I came here and stayed. It’s a great place where I can go out and have a great time.
Q: (No microphone.)
A: It definitely helps. I like all the kids. They are always around me.
Remember it was trained on transcripts and sometimes “inaudible” or “no microphone” pops up.
Lastly, I thought I’d try to pull something a bit more literate out of coach bot by prompting it with Shakespeare quotes:
By the pricking of my thumbs, Something wicked this way comes. Open, locks, Whoever knocks! Lock my hips up and start trying to roll a spot block. Seniors can kick them in the head in the first down game! That’s a Thursday game, but we’re young, and we’re a young team.
To be, or not to be: that is the question. And I think the answer is no. I’m not as — I think it’s not really the issue, but I think there’s a great deal of opportunities in the country
Some are born great, some achieve greatness, and some have greatness thrust upon them. Whatever you have to do, that’s in your power. If you go out and change‑‑ if you do your business, you’re going to be fine. If you don’t, then you’ll be in trouble.
So maybe coach-bot isn’t quite ready to take over, but perhaps given more training data, it can go out there and become the whoopin’ boner of a coach it was meant to be.
