Nerd Note-vember

Hey hey - this is the first official installment of Nerd Notes...the newsletter for people who want to go waaaaaay too deep down the Pursuit of Play rabbit hole.

Why deck generator doesn’t work

In the Dortmund writeup, I made a throwaway reference to the Pokemon Deck Generator tool I had prototyped, which I ultimately concluded was a failed prototype. I didn’t really get into why it doesn’t work, but it’s something that I spent a lot of time thinking about. I think it’s interesting in considering what are good and bad generative AI use cases at the moment.

So let me first say this: the idea of a Pokemon deck generator is probably feasible - just not with the current generation of publicly available generative AI. What Deepmind has done with AlphaStar is instructive, both in terms of its current capabilities and in terms of the amount of resourcing necessary to make it actually work. With enough computational power, I think it could work. And while that would involve a complex neural network, it would probably not be an LLM.

The reason that an LLM is ill-suited to this use case is interesting though. Arguably, a chess strategy LLM probably would work, and the reason that would work and Pokemon wouldn’t has to do with 2 things: the static vs dynamic nature of the two games, and the degrees of turn-based complexity.

The Pokemon deck generator doesn’t work because the data that the LLM is trained on has become obsolete to the game itself. It is a probability machine that is trying to guess what word it should say next, and there’s a lot in its training data related to Pokemon TCG but all of that data was about past versions of the game that are no longer playable. It recommends sooooo many cards that have long since rotated. What’s funny is that it does this even when I give it a specific database of currently legal cards from which to draw. Maybe I could improve the prompting to make it more reliable, but I went through A LOT of iterations and could never get it to work right. Contrast that with chess, where the rules are static and so the vast amount of training data related to chess all remains relevant. An LLM is still probably not better than a chess simulator like Stockfish, but the two things can complement each other nicely.

And then there’s the complexity. What happens in a single turn of Pokemon has some constraints: you draw 1 card to begin your turn, you can only play 1 supporter card, you can only attach 1 energy card from hand, you can only retreat 1 time, you can only attack 1 time. If you do only those 5 things with clear constraints around them, then a single turn consists of 5 discrete actions. Of course, that’s not all that happens in the course of a single turn. Beyond those things, a player can play Item cards, Stadium cards, and Tool cards. Some of their Pokemon in play will have special abilities they can use. Every one of those actions can affect the game state in so many different ways, and all of that flows from the unique build of a 60 card deck and how it needs to respond to the opponent’s unique 60 card deck. That’s a lot of strategy related to lots of different game dynamics and mechanics. It is beyond the capability of an LLM.

Live and learn!

The elite Pokemon player who no one copies

In the last edition of Pursuit of Play, I went off for a little while about Tord Reklev and his ability to innovate and define the standard approach that other players will take to certain meta builds.

There are innovators, and then there are mad scientists. Innovators get copied, mad scientists continue to toil in isolation. If Tord is the innovator, then Sander Wojcik is the mad scientist. Want proof? OK, fine. Without getting into the finer points of Pokemon, one basic thing to know is that Pokemon need energy to attack or retreat; to get those energy, a player has to attach an energy card to a Pokemon card. Easy enough, right?

So take a look at this deck that Sander Wojcik played to a 12th place finish in the Masters Division at Lille. And just in case you’re not feeling like parsing the nuances of a 60 card deck, the part that matters is here:

His deck didn’t have any energy in it, so he couldn’t ever attack…and he finished in 12th place.

Sander’s nickname, coincidentally, is the King of Control because he’s playing the game in control mode. Normally, a player is trying to win the game by knocking out their opponent’s Pokemon…which requires attacking…which requires energy (except in rare cases where a Pokemon has a free attack). Sander engineers decks where he prevents the opponent from attacking him and ultimately forces them to run out of cards in their deck, triggering the other condition for winning a Pokemon game.

Because I’m drawn to novelty I’ve been playing this deck online the last couple days, and I’ve come away with 2 big insights about it:

• it’s brutally effective. Once you get the basic logic of it, it wins a lot.
• it’s not exactly fun. There is some challenge involved in playing it - you have to be able to identify what deck the opponent is playing and therefore what they’re going to do so that you’re setup to counter it. If you can do that, you’ll probably win. I suspect the real fun and challenge was less in the playing of the deck and more in the building of it. Sander is like a Swiss watchmaker who wears his prize creation in public. I’m like some rando who bought a Patek Philippe and wears it ostentatiously.

If you want, I made a video of me playing this deck - and then I edited out all of the turns that were just "I draw a card and pass to you because I can't attack, then you draw a card and pass to me because I've prevented you from attacking." But this gives you the gist of how it works.

Sander doesn’t show up for every tournament; in fact, he didn’t show up for a single tournament until the 2nd to last one in Stockholm last season because he didn’t think the format was well-suited to the control style. When he does show up, however, he has a finely tuned, completely unique deck in tow that no one has ever seen before - the community refers to his deck type as "A Sander Pile," usually with a sense of reverence. He has crafted it to oppose the most common meta decks, and he has engineered it so he can play just the right cards against just the right matchups. If you watch him play 3 different matchups, you might see him play the deck 3 different ways.

Whereas Tord’s build of a deck becomes the standard, no one copies Sander. No one copies Sander because it is extraordinarily difficult to play like Sander. And when Sander shows up again, it’s likely that he’ll be rocking a different version of control than we’ve seen from him so far. In the Trudeau household, we’re all hoping to see more from him this year.

I was a guest on a podcast!

The whole episode drops next Tuesday - you’ll see it in the links email - but I’m sharing a little snippet of it here to whet your appetite.

That clip is a bit more on the education side, but there is soooooo much game talk in the whole episode. It was a fun experience.

But also, a funny little behind the scenes that might lead you to listen to it all in a very different light: I recorded that interview in - no kidding - July of 2023 while staying in 1 of 2 AirBnBs in rural Payette, Idaho for one of our first pilots of FutureShock. When the hosts reached out to me to let me know the episode was dropping this month, I had completely forgotten I did it in the first place.

By way of context, in July of 2023 I hadn’t even kicked off the Pokemon Travelogue yet. We hadn’t decided to play a whole season of tournaments. There is this whole world in which I am now deeply enmeshed that I had only a vague understanding of less than 18 months ago. I think it’s very funny to listen to the episode while thinking in the back of your mind, “This is a person who within 6 weeks is going to go about a dozen layers deeper down the Pokemon rabbit hole.”

Coming up:

• I'm on a loooong flight to California on Monday, so the next edition of the links should be ready to go on Tuesday.
• Last weekend we were in Gdansk, Poland, for the 3rd regional tournament of the season. The write up of that will be coming your way in time for Thanksgiving, hopefully even a bit earlier.
• In less than a month, we'll be in Stuttgart, Germany, for our 4th tournament. That tournament will mark the debut of the new Surging Sparks set - and in Nerd Notes next month, I'm going to go in depth on how I'm changing my deck for that new set.

BTW, if you have idea of things you want me to write about or questions you think I should ask, you can always drop me a line.