Playing Balatro

Is Like Being Elon Musk on Ketamine

I just got back from the Game Developers Conference, where one of the big topics was “Are you playing Balatro?” The game is a rogue-like (because you play it over and over again) deckbuilder (because you modify a deck of cards over the course of each game) that combines a standard 52-card deck and poker hand rankings with a bunch of extra stuff that interacts in various crazy ways. It’s a fundamentally simple game with a ton of charm and polish in the execution, and it’s a lot of fun. But what makes it especially interesting to me is the way it combines skill and luck, specifically, the way it has a whole lot of both.

The relationship between luck and skill is a somewhat under-theorized aspect of games. It’s easy to think about games like Chess, where the outcome is completely determined by the strategic decisions of the players, and pure gambling games, where the outcome is completely random. But what happens when, as in so many games, strategic decision-making and randomness intermingle?

There’s a nice chapter on this topic in Characteristics of Games and Greg Costikyan touches on the subject in Uncertainty in Games, but both of these treatments are fairly basic. One useful idea is the distinction between input randomness (random elements that the player responds to by choosing an action) and output randomness (when randomness influences the result of a player’s action), a concept originally developed by Geoff Engelstein on his podcast, although I first learned about it from Keith Burgun. But, overall, the interaction of luck and skill in games remains somewhat mysterious. Balatro is a good opportunity to experience that mystery directly, and an invitation to consider what it means.

Although it borrows the hand-rankings from Poker as the foundation of its scoring system, Balatro doesn’t really resemble Poker that much. At its heart, it’s what I call a shopping game. Over the course of each run, you are offered different game-modifying Joker cards to acquire. Jokers combine in various ways to create powerful point-scoring engines. Assembling these engines and tweaking your deck to work with them is the core of the gameplay.

But the thing that Balatro does share with Poker is that it is both deeply strategic, with lots of opportunities, and incentive, to calculate the effects of the actions you choose, and high-variance, meaning the outcome of any one individual game session is only loosely correlated to the strategic choices you made.

I love games like this. Why? I’m not exactly sure. Partly its just the high-octane cocktail of two complementary emotional ingredients - the degenerate gambler’s love of intoxicating action mixed with the pretentious intellectual’s love of sober problem-solving. But I also love the fact that games like this are confusing, misleading, hard to analyze. I’ll never forget the feeling I had when I first realized that you could study Poker the way you studied Chess. Sneaky! I thought. A game that is less dumb than it looks. (This is true, to a degree, of all games.)

But the confusion goes both ways. Over the past few weeks I’ve spent a considerable amount of energy arguing with one of my game developer friends about how to interpret Balatro’s luck/skill dynamic — the degree to which our results in the game reflected the quality of the strategic choices we were making.

“It’s a slot machine” I said.

My friend countered by pointing out all the different ways in which your strategic choices could influence your odds of winning, and gestured to the fact that our results were improving over time as evidence that strategy had a significant impact on outcome.

It’s not that he was completely wrong. But the thing I wanted him to understand was the fundamental principle of high-variance games, which is that they drastically limit our ability to interpret the results of any single run as evidence of the quality of the strategy we are using. Take any strategy, X. (A strategy, in this sense, means a complete description of what actions to take under what circumstances. So, for Balatro, not just using a certain Joker or buying a certain kind of card, but which Jokers and cards to buy under which specific conditions.) Run 100 sessions of the game using strategy X and plot the results. You will see a broad distribution of results, clustered around some point but falling off in both directions. Taking a random sample from that distribution gives you next to no information about the shape of the overall curve. It’s not exactly zero, but it’s practically zero.

This is the way in which Balatro and Poker are similar. In Poker, a common mistake for beginning players is to win a hand, or have a winning session, and think “Ah! This new approach I’m using is working well!” This mistake is referred to as being results oriented. It doesn’t mean that Poker isn’t deeply strategic, it just means that that’s not how you evaluate the quality of different strategies in Poker. In high-variance games like Poker and Balatro, you evaluate strategies by theory-crafting — analyzing the game with math and logic, and developing a complex model of the game to understand its important strategic principles. Over time, you can look at the statistical shape of your long-term results as (fuzzy) evidence for how good your strategic principles are, but here’s the important thing: your key insights must come from taking the game apart and understanding how it works, not by scrutinizing the signal of your immediate results.

Why does this matter? Why was I so adamant about insisting on this aspect of Balatro? It’s a fun game, a bit silly, who cares?

Because slot machines are dangerous. It’s fun to play with dangerous things, but you can blow your head off with them. For me, a big part of the fun of playing with slot machines is knowing how they work, and a big part of how they work is by manipulating the cause and effect mechanism in your brain in exactly this way, by hijacking your brain’s reward function and giving you a false impression about the amount of agency you are exercising. Our brains are hyper-tuned to see signals in our environment, to interpret them as expressions of an underlying structure. Slot machines are creatures that have evolved to feed on this behavior by generating a signal that sounds like the message “keep digging” but is really only a complicated kind of noise. You can eat them, and they are delicious, but only if properly prepared. Which means, first and foremost, recognizing the trick and not falling for it.

It’s good to see the Chess in Poker, and realize how that game is smarter than it looks. But it’s also important to see the slot machine in Balatro, and realize the ways in which it is dumber than it looks. Again, this is not to deny that Balatro is, unlike actual slot machines, full of complex, interesting decision-making. If it wasn’t I wouldn’t still be playing it. I still often find myself in situations where I’m not sure what to do, and by carefully thinking through these situations I can feel myself learning, developing a better understanding of the game, and improving my overall strategy. But there’s a sense in which this process is happening despite the game, in a kind of wrestling match with the game’s explicit reward structures, the way it doles out wins and losses and big scores.

You can see this in some of Balatro’s specific structural features:

• The game is organized as a series of scoring challenges. In a good run you will often be scoring far more than the necessary targets, but the game doesn’t pay attention to this excess score, it doesn’t affect the outcome at all. As a result, you often find yourself continuing to tinker with your deck, making it more efficient and improving its scoring power for the fun of it, in a way that the game is basically ignoring.

• Then, every time you “win” a game by beating the final scoring challenge, the game invites you to continue on to “Endless Mode” (really, just a series of additional, rapidly-scaling scoring challenges.) This moment is deeply confusing. Wait - what was I doing? Was I trying to get a win or was I trying to see how high I could score in Endless Mode? In reality, the best strategy for maximizing your chances of winning are different from the best strategy for maximizing your chances of going deep in Endless. The former requires adapting, scrappily, to whatever combo pieces the game offers you, the latter requires taking greater risks to seek out the smaller set of overpowered combo pieces that are capable of producing astronomical scores. In fact, it’s already known how to “beat” endless mode by generating more points than the game’s scoring function can handle, it just requires waiting patiently for the necessary ingredients to show up. By blurring the difference between these contrasting goals, Balatro amplifies the sneaky disconnect between player action and game result.

• Meanwhile, win rate, the one thing that all of the player’s clever strategies and complex problem-solving is optimizing for, is completely ignored. The game doesn’t track it at all.

• Once you get into the game’s higher difficulty levels you quickly find yourself “start scumming” — restarting a run as soon as the first few levels go poorly. Yes, you could try hard to play optimally in order to eke out a few extra levels in these unlucky runs, hoping to encounter a viable combo piece, just like you can try hard to optimize the performance of your deck in a lucky run where you’ve already guaranteed victory half-way through. In both cases the game isn’t paying attention at all.

One of the reasons I love games like this is that, unlike Chess, this is a wicked problem. What’s the best way for an agent to extract information from a noisy environment? We make sense of the world through stories of cause and effect, but, on close examination these stories are far less simple than they appear. In The Book of Why, computer scientist Judea Pearl identifies three levels of causal inference:

1. Association. Discovering, through observation, the connections between phenomena.

2. Intervention. Predicting and testing the effect of our own actions on the phenomena we observe.

3. Imagination. Considering counterfactuals, hypothesizing about what might happen if things were different than they actually are.

In a game like Balatro we encounter an even harder version of this problem: how do you learn in an environment that isn’t just chaotically noisy but that is loaded with adversarial signals that have evolved to mislead you?

The real world is full of examples of people wrestling with versions of this problem. In a recent column, Donkeyspace hero Matt Levine described Elon Musk’s particular version of the struggle:

The Elon Musk theory of corporate governance is that everything that Elon Musk does is good for his investors:

1. Elon Musk has made many hundreds of billions of dollars for investors in his various companies.

2. Those investor returns come, in large part, from Musk’s personal attributes and efforts: Those companies would not be successful without his perfectionism, drive, management style, techno-optimism, etc.; nobody else could have built those companies, and even today they would not perform nearly as well without Musk’s intense personal involvement.

3. Those personal attributes, which are good for investors, are inseparable from the rest of his personality, which is … oh, you know. “I reinvented electric cars and I’m sending people to Mars in a rocket ship,” Musk once said on Saturday Night Live; “did you think I was also going to be a chill, normal dude?” His personality only works as a complete package; you can’t get the shareholder value creation without the bad tweets.

4. Therefore, the bad tweets are also good, because they enable the good stuff; they are the compost that feeds the flowers of shareholder value.

This theory is … I mean, it’s probably not wrong, is it? It is pretty unhelpful as a theory, though. It proves too much. It implies that any conflict of interest between Musk and his shareholders is impossible, because whatever Musk wants is necessarily good for shareholders — even if it is bad for shareholders, it is good, because the bad thing is just the price they pay for Musk’s goodness

Anyway, here is Elon Musk on the Elon Musk theory of corporate governance:

Elon Musk said that his prescribed use of ketamine alleviates periods of low mood and is in the best interest of investors in Tesla Inc. and the other companies he runs.

For Wall Street, “what matters is execution,” Musk said in an interview with former CNN anchor Don Lemon streamed Monday on YouTube. “From an investor standpoint, if there is something I’m taking, I should keep taking it,” he said referring to Tesla’s success.

Musk said he takes the drug as prescribed periodically to treat what he described as “chemical tides” that lead to depression-like symptoms. 

I am sure Elon Musk has good doctors, and I am not going to second-guess the medical advice he is getting, but the corporate governance theory here is “if there’s something I’m taking, I should keep taking it.” That logic applies to taking ketamine on medical advice, or taking cocaine against medical advice, or anything else. Whatever he does is necessarily good.

Each of us, in our own way, is in the same boat: trying, and often failing, to understand the connections between the things we do and the results that follow. Balatro is a laboratory for experimenting with this problem, a laboratory with blinking lights, soft carpets, and free drinks.

The reason gambling is fun is that we are not our reward function. Maybe once we were, maybe once, long ago we were nothing but a little knot of signals that was able to remain stable far out of equilibrium with its environment, a little loop of Bayesian inference that was able to climb the signal gradient towards more and more complex and robust forms of stability. But now that we know that, now that we can look back and see that, we are no longer fully embedded in our reward function, it has become something we can use, something we can play with, instead of simply who we are. With some effort we can partially extract ourselves from the loom of our reward function, look around, and ask: “Where am I?” and “Where do I want to go?”