Hypersplit is like Neal's Infinite Craft but turned upside down. Instead of crafting your way up from the four elements, you start big and go small, down to quarks.<p>It's using GPT-4 to split the words, but a big difference from Infinite Craft is that I can precalculate the possible paths since they are limited (sooner or later you hit a quark or a loop). This means that I can ship it with the whole dictionary locally instead of hitting an LLM every split.<p>While it started with just the splitting, I quickly got the urge to split things faster... which turned the app into the incremental clicker it is now. It's pretty short and starts lagging a lot once things get crazy with a lot of chain reactions (turning off Effects helps a bit), but I plan to focus on optimizations next.<p>I also wanted to try and build something without any graphical assets. The whole game is all just emojis and CSS. A big gotcha is that you can't rely on emojis looking the same on different platforms (for example, on Windows, the explosions are misaligned).
Love to automate these types of things. Unfortunately there isn't much benefit to automating the shopping/leveling, as this performs far better than any passive bonus those provide<p><pre><code> let stop = false;
let play = () => {
const button = document.querySelector('[data-word-button]');
if (stop || !button) return;
button.click();
setTimeout(play, 0);
}
play(); // to play
stop = true; // to stop
localStorage.clear(); location.reload(); // to restart</code></pre>
The Infinite Craft game was sending queries up to ChatGPT to make the new items, if I recall correctly. This one is happening so quickly it must be happening locally. Is that right? Is it using a local Javascript LLM?
Game idea for someone with more time than me: an LLM powered "roguelike".<p>You fine-tune the LLM to spit out RPG skills based on the player's chosen class. (Dwarven princess = summon dwarf minions, has a rare bloodline magic, whatever). The classes could get _real_ weird, like a school bus character could summon school children minions and skill up to flying a la The Magic School Bus. (Or that could be a prestige class haha.)<p>A second model would match the RPG skills to a known/finite set of game mechanics supported by your game. This is ultimately a simple classifier - not really an LLM (though the skills on the input side are language, okay pedants). Determining the set of game mechanics to support shouldn't be too hard - how many different ways of doing damage and (de)buffs could there be? (Okay there can be a shit-ton.) Balancing the game could be done live via reinforcement learning. To constrain the search space, it'd follow the old Zelda format of limiting you to one room/screen until the enemy is defeated (or like Binding of Isaac for those too young.)<p>It'd have to use 2d/sprite based/placeholder graphics because I don't think the rigging/animation AI is currently good enough... but bad graphics hasn't stopped fun gameplay before. Imagine Charizard vs Death Star. Good luck with the DMCAs.<p>Another problem is that with simple graphics and <i>literally infinitely variable enemies</i>, there's no learning on the player's side so there's no real "skill" accumulation to improve on for the next run. There would be a lot of reading on the player's part if you decide to just show a stats-list for an enemy type. I suppose you could turn it into a zombie game where new enemies are slowly introduced in a horde of other enemies the player's familiar with.