A history of games for AI/ML

By Archie Chaudhury, co-founder of LayerLens. Follow him on X at @ArchChaudhury.

A couple of weeks ago, LayerLens announced the Stratix Cup, a competition to determine the best LLM, at least for playing soccer. Inspired by the World Cup, we created a custom environment that enables LLMs to write code to manage their teams, dictate different strategies, and analyze their opponent.

This approach is somewhat different from other similar benchmarks or evals, which usually setup a custom agentic workflow where models call tools in real time to interact with the environment: this could correspond to moving a character in Minecraft or opening a box in a 2D scroller.

However, adaptive, deterministic games constrained by a small set of moves are not completely new. Historically, deep learning models, from which modern day LLMs are derived, have leveraged games both as a testbed and a way to show progress.

Chess before deep learning

Games were the original AI benchmark. Claude Shannon published Programming a Computer for Playing Chess in 1950, before the term "artificial intelligence" existed. Arthur Samuel's checkers program, described in his 1959 IBM Journal paper, learned from self-play and gave us the phrase "machine learning." The appeal was practical: games have fixed rules, a clear definition of winning, and no need to model the messiness of the physical world.

The two milestones of the 1990s pointed in opposite directions. IBM's Deep Blue defeated Garry Kasparov in 1997 using brute-force search and a hand-tuned evaluation function. It learned nothing. Five years earlier, Gerald Tesauro's TD-Gammon had reached near world-class backgammon using a neural network trained purely through temporal-difference learning and self-play. Deep Blue got the headlines, but TD-Gammon was the prototype for everything that followed.

Atari and the deep RL era

The modern chapter starts in 2013, when DeepMind showed that a single convolutional network could learn to play Atari games directly from pixels. The DQN paper, published in Nature in 2015, reported human-level or better performance on 29 of 49 games using one architecture and one set of hyperparameters. Games turned out to be the ideal substrate for deep reinforcement learning: simulation is cheap, rewards are unambiguous, and scores are reproducible across labs.

AlphaGo and AlphaZero

Go was the holdout. Its branching factor is roughly 250 moves per position, which made Deep Blue-style search useless, and most researchers expected another decade of work. In March 2016, AlphaGo beat Lee Sedol 4-1. The system combined two deep networks, a policy network to propose moves and a value network to evaluate positions, with Monte Carlo tree search. It was bootstrapped on 30 million positions from human expert games, then improved through self-play. Move 37 of game two, a shoulder hit that professional commentators initially called a mistake, became the canonical example of a learned system producing strategy outside human convention.

DeepMind then removed the human data entirely. AlphaGo Zero (2017) started from random play, knowing only the rules, and after three days of self-play defeated the version that beat Lee Sedol by 100 games to 0. The follow-up, AlphaZero, generalized the same algorithm to chess and shogi. With nine hours of training, it beat Stockfish 8, then the strongest chess engine in the world, winning 155 games and losing 6 out of 1,000. The original preprint is worth reading for the chess results alone: AlphaZero evaluated roughly 80,000 positions per second against Stockfish's 70 million, compensating with a far better evaluation function. Its willingness to sacrifice material for long-term positional pressure has since filtered into top-level human preparation. MuZero (2019) pushed one step further and learned the rules themselves, matching AlphaZero without ever being told how the pieces move.

Imperfect information and real time

Board games are fully observable and turn-based. The next wave attacked games that are neither. DeepMind's AlphaStar reached Grandmaster rank in StarCraft II in 2019, and OpenAI Five beat OG, the reigning Dota 2 world champions, the same year. Both systems dealt with hidden information, continuous play, and team coordination, and both consumed enormous amounts of compute to get there.

Why games still matter for LLMs

LLMs inverted the recipe. AlphaZero learned a single game from scratch; an LLM arrives pretrained on most of human text and gets evaluated on games it was never optimized for. That makes games newly useful as evals rather than as training grounds. Static benchmarks saturate and leak into training data. A game generates fresh states every match, the rules verify every outcome, and a model that memorized opening theory still has to survive the middlegame.

This is now an active area: lmgame-bench evaluates LLMs on classic video games, TextArena covers 57 competitive text games with a public leaderboard, BALROG tests agentic reasoning in roguelikes including NetHack, and Google DeepMind's Game Arena on Kaggle ran a public chess bracket between eight frontier models in 2025, then added poker and Werewolf in early 2026 to test imperfect-information play.

The Stratix Cup sits in this lineage with one deliberate difference. Instead of asking models to select moves turn by turn, we ask them to write the code that plays, then adapt it between matches. Seventy years of game-playing AI measured how well a system plays. We think the more interesting question for LLMs is how well they strategize, and games remain the cleanest place to measure it.

Follow the bracket and live matches at Stratix Cup Season 1.