How the relentless drive for optimization made baseball impossible to watch

The unlikely similarity between modern baseball and the AI boom.

Here’s a question for major league baseball’s opening weekend: What does Philadelphia Phillies left fielder Kyle Schwarber have in common with artificial intelligence?

At first glance, not much. Schwarber — who has the rough physical dimensions of a human keg of beer — is best known for hitting the baseball really, really far when he hits it. Counting the regular season and postseasons, he hit a National League-leading 52 regular and postseason home runs last year — Schwarbombs, as Philly fans like to call them — including one against San Diego in the National League Championship Series that may have been able to achieve orbital flight.

GPT-4 is capable of doing many things, but mashing a baseball 119.7 mph for a total of 488 feet, as Schwarber did against San Diego’s Yu Darvish, is not currently one of them.

But viewed another way — and I say this as a lifelong Phillies fan — Schwarber is the roly-poly human incarnation of an out-of-control AI. He’s Skynet madly swinging a 31-ounce Louisville Slugger. He’s HAL from 2001 in custom Home Run Derby cleats.

And he must be stopped.

The relentless pursuit of optimization

Allow me to explain.

In 2022, counting the regular season and the postseason, Schwarber had 743 plate appearances. Along with those 52 home runs, he walked 101 times and struck out a truly mind-boggling 218 times. That means that almost exactly 50 percent of the time Schwarber came to the plate, he achieved one of baseball’s “three true outcomes”: a strikeout, a walk, or a home run.

A true outcome means that the ball isn’t put in play and no member of the defense is involved beyond the pitcher throwing the ball and the catcher attempting to catch it. That means no fielder trying to make a Gold Glove play on a batted ball. It means no speedy runner trying to go first to home on a ball hit in the gap.

The three true outcomes are baseball at its most bloodlessly — and boringly — efficient. And over the past 20 years or so, efficiency as represented by the three true outcomes is precisely the direction baseball has been headed.

On the defensive side, that has meant teams loading up on pitchers capable of throwing the ball really, really hard, in search of strikeouts. A ball put in play, after all, could fall for a hit. The average four-seam fastball last year was 93.9 mph, up from 93.1 mph in 2015, the first year such data was collected.

It used to be that only freaks of nature like Nolan Ryan could hit triple digits on the radar gun, yet last year there were 3,356 pitches thrown at 100 mph or higher — nearly twice as many as the year before, and far and away the most in baseball history.

On the offensive side, the emphasis is on the other two outcomes: walks and home runs. I never believed it when my Little League manager used to tell me that “a walk is as good as a hit” — kids like to swing the bat — but that’s now gospel in the majors, especially if that walk is followed up by the next batter hitting it out of the park for a multi-run homer.

With pitchers throwing harder than ever, just making contact is difficult enough, so batters have tried to compensate by focusing on raising the launch angle generated by their swings, increasing the chance that when they do put wood to ball, it has the lift to leave the park. Schwarber’s titanic shot off Darvish, for instance, had a launch angle of 25 degrees, well above the league average, which is higher than it used to be.

The downsides of taking all those Ruthian uppercuts, especially if you’re facing a guy throwing 100-mph gas, is that much of the time you will swing and miss. That means lots of strikeouts and considerably fewer balls put into play in the field. In addition, the recent adoption of advanced defensive positioning has meant that managers are moving fielders before the pitch to where they think a hitter is most likely to put the ball, what’s known as a shift.

No batter in the National League faced the shift more often last season than Schwarber, who would often see three infielders on the right side — where a left-handed power hitter like him pulls the ball — with a second baseman essentially playing shallow right field. Those tactics meant that even when hitters like Schwarber made solid, non-home-run contact, they were less likely to result in a hit. Partially as a result, league-wide batting average fell to .243 last year, the lowest since 1968, while the humble base hit is practically an endangered species.

The result is a game that is highly efficient and highly boring, with lots of strikeouts and way fewer hits, punctuated by the occasional burst of dingers. Just look at last year’s World Series, which featured Schwarber’s Phillies against the Houston Astros. Game 3 saw the Phillies win 7-0, with all seven runs coming on five homers, including one by Schwarber. And then came Game 4, when the Phillies became only the second team in baseball history to be no-hit in the World Series, with an incredible 14 strikeouts.

Yes, as a Phillies fan, it was considerably more fun to watch the glorified home run derby that was Game 3 than seeing them be no-hit the next night. But even in their win, the Phils struck out 11 times. All in all, the Fall Classic, baseball’s crown jewel, mostly involved watching pitchers throw the ball really hard, batters swinging really hard (and generally missing), and everyone else pretty much just standing around.

The upshot is that the most efficient, effective way to win baseball — a strategy basically every team is pursuing — happens to have produced as a side effect the most boring kind of baseball to watch, as evidenced by declining ratings and attendance.

In the relentless pursuit of optimization, baseball may be killing itself.

Unaligned on the diamond

So what does this have to do with AI? It helps to go back two decades to the Moneyball Oakland A’s and their general manager Billy Beane, the Miles Dyson of baseball’s Skynet situation.

The A’s had a problem: They were broke. (As Beane, played by Brad Pitt, put it in the movie Moneyball: “There are rich teams and there are poor teams. Then there’s fifty feet of crap, and then there’s us.”) Unable to compete with rich teams like the Yankees for free agents, Beane had to compensate by using statistical analysis derived from the work of maverick figures like the statistician Bill James to identify players who were undervalued for their actual production.

This meant going against the conventional wisdom, which valued players for things like speed and batting average, and pursuing players who could reliably get on base any way possible, as this scene from the film shows:

The surprising success of the Moneyball A’s helped speed a league-wide revolution in using advanced statistical models in baseball. Out were traditional numbers like stolen bases (the risk of getting thrown out and losing one of the 27 outs a team has in a game was higher than the reward of advancing a base). In were stats like on-base percentage (which measures the most important thing a batter can do — not make an out, whether through a hit, a walk, or getting hit by a pitch).

Rosters were constructed in the front office and teams were managed in the field with an eye toward maximizing the three true outcomes. No one broke any rules. (If anything, baseball became cleaner in the 2000s and 2010s, as performance-enhancing drugs were phased out). The problem was that the most efficient way to win baseball games under the rules as they existed turned out to be highly inefficient for the purpose of entertaining the spectators and TV audiences who make major league baseball major. (Not incidentally, Beane didn’t watch his teams play. He feared that the act of spectating would lead to a “visceral reaction” that might outweigh his Moneyball rationality.)

In the world of artificial intelligence, this is an example of “misalignment.” Through highly detailed rules, which now run to 191 pages, the creators of baseball tried to construct a game that would be entertaining — meaning action, running, excitement. Think Wille Mays making an over-the-shoulder catch in the World Series or Rickey Henderson stealing home. That was their goal.

But the goal of the teams that play baseball is to use the rules to win games. Excitement doesn’t factor into it. Which is more or less how you end up with the 2022 Phillies, a team explicitly built to hit a bunch of home runs while also setting a record for the most strikeouts in a World Series.

We’ve seen this happen repeatedly in artificial intelligence, especially in games. One AI that was trained to play a boat racing game learned that the most efficient way to score the most points wasn’t to win the race — which is what the designers and presumably most humans would aim to do — but instead to drive around repeatedly in a circle, hitting a handful of targets over and over. This was not fun to watch — seeing the boat mindlessly spinning around and around reminded me of watching Game 4 of the World Series — but that didn’t matter to the AI. What mattered was racking up points by any legal means necessary.

To its credit, baseball is trying to fix its alignment problem by changing some of the rules to encourage more hits and more action. So this season, extreme defensive shifts of the sort faced by Schwarber are outlawed — two infielders have to be on either side of second base when the pitch is thrown, and infielders won’t be allowed to start in the outfield.

Baseball has also added a pitch clock of 15 seconds per pitch (20 seconds when a runner is on), which should both speed up the game and tilt some of the advantage back toward the batter. And bases themselves were increased in size from 15 sq. ft. to 18, partially in the hopes of encouraging more stolen bases. (There were just 2,487 stolen bases last year, down from 3,264 three decades ago.)

Will it work? Through spring training, games have been shorter, stolen bases have gone up, and slightly more balls put in play are going for hits. That may not be enough to save baseball — few disciplines, after all, are as relentless in their pursuit of optimization as professional sports, and players like Schwarber will ultimately be judged on their stats and their win-loss records, not how entertaining their playing style is.

But at least baseball is taking proactive steps to nudge their sport in the direction that fans might actually enjoy. On Opening Day, the average length of a game was down, while stolen base attempts were up. As for Schwarber, he went 0-5 in an offense-heavy 11-7 loss to the Texas Rangers, striking out twice for a three-true-outcome rate of 40 percent. Which I guess counts as improvement, albeit not the sort that any Phillies fan is likely to applaud.

As humanity faces down large language models that can throw the equivalent of 100 mph or more, it’s worth being very, very careful about the rules and the goals we program them with — lest we end up in an eternal no-hitter.