# RoyaleMind > RoyaleMind (royalemind.com) is the home of LarryZero, one of the strongest Clash Royale AIs, and you can play against it in your browser. LarryZero plays today's game, with current cards, evolutions and heroes. It learned from millions of top players' moves and keeps improving by playing itself in an offline simulator. Its best network beats FirstLight General, a leading open-source Clash Royale AI, 0.595 over 100 duplicate pairs. Anyone can try it at https://royalemind.com/play, with no account or download. It is an independent fan research project, not affiliated with, endorsed, sponsored or approved by Supercell. LarryZero never connects to the live game or plays on the ladder. ## Try it - **Play against LarryZero:** open [royalemind.com/play](https://royalemind.com/play), pick your deck and LarryZero's deck from the current top decks, and play a full battle in your browser with a mouse, touch or keyboard. The game runs in the game's own engine on the project's host and streams to you. One player at a time; the page shows when the host is busy or offline. - **Watch it play:** the [overview](https://royalemind.com/) plays back a recent game move by move, with LarryZero's own estimate of its chance to win and where it is thinking of playing next. - **Who it is for:** players who want a sparring partner to practise against with current decks, and anyone curious how a game AI learns. ## How strong it is - **Against FirstLight General** (FirstLight CR's public 12M-weight network, the outside yardstick): LarryZero's p8 network scores 0.595 over 100 duplicate pairs (range 0.532 to 0.658), about +67 rating. It is the first network of the project's own to beat it. - **Still climbing:** the newer p16 network beats p8 0.568 over 200 pairs (about +47 more). - **The goal:** an AI that plays today's game better than the best human players. The play page runs a preview version of the model while the main line keeps training. Every version is tested in a gauntlet of rated matches, and the site shows the newest test results. LarryZero is the model; RoyaleMind is the project and the website. The name nods to AlphaZero, whose recipe (self-play, with a look-ahead search as the teacher) the project follows. larryzero.com redirects to royalemind.com, and royalmind.org is the project's former address. ## What LarryZero is - **The game:** today's Clash Royale (game data 16.635), simulated offline by [crforge](https://github.com/voonhous/crforge), an open-source simulator of the current game: cards, evolutions, heroes and timings. - **The decks:** the most popular decks among top-ranked players, each with two evolution slots and usually a hero. Matches pair random top decks, mirror matches included. - **The network:** a transformer that looks at every unit on the board and picks a card and a tile, or waits. - **How it learns:** first by imitation, copying top players from 3.6 million plays in real battle replays; then by self-play against itself, frozen past versions of itself and hand-written bots (a league). A look-ahead search that plays out candidate moves and teaches the network, AlphaZero-style, is being tested. - **How it is tested:** a gauntlet of duplicate-pair matches (each pair of games swaps sides of the arena with fixed seeds) gives every version a rating, with the scripted bot at zero. A crashed or unfinished game never counts as a win. - **The hardware:** three ordinary home computers. One trains the network on its GPU; all of them play games for it. ## Pages - [Overview](https://royalemind.com/): what LarryZero is, a recent game played back move by move with its own estimate of its chance to win, the decks it plays, and how it learns. - [Play](https://royalemind.com/play): a streamed preview in the game's own engine where anyone can play a full battle against a preview version of LarryZero, free and with no account. One player at a time. ## Public data - [Summary](https://royalemind.com/api/public/summary): JSON with training totals (games and decisions), the current run, and its latest test results against each opponent. - [Recent games](https://royalemind.com/api/public/replays): JSON list of a few test games along the current line of training, from the first to the newest. ## Rules the project keeps - Offline only: training and demonstration games never connect to Supercell's servers, and LarryZero never plays on the live ladder. It is meant as a sparring partner for practice. - Game art and names belong to Supercell and are used in a fan project, with the disclaimer on every page.