Dave Palmer, a former Microsoft engineer and Tempest world record holder, built an AI that ultimately surpassed his own high score on the notoriously difficult arcade game by leveraging direct game state extraction, expert system scaffolding, and a neural network tailored to the game’s unique geometry. His project highlights how aligning AI input with a game’s true structure and focusing its attention on relevant threats enables superhuman performance, even outpacing top human players.
Dave Plamer, a retired Microsoft software engineer and Tempest world record holder, recounts his journey building an AI to play the classic 1981 Atari arcade game Tempest. Despite a power outage, Dave is determined to share his story, which blends arcade nostalgia, stubborn engineering, and the humbling experience of being surpassed by his own creation. Tempest is a notoriously difficult game, especially on its “Extreme” setting, and Dave had long prided himself on his mastery. However, after a year of development, his AI not only matched but exceeded his official world record, demonstrating superhuman performance on the hardest levels.
The project began with Dave extracting the internal game state directly from the Tempest arcade emulator (MAME) using Lua scripting, bypassing the traditional approach of training an AI on raw pixels. By accessing the game’s memory, he could provide the AI with the same information a human player sees—enemy positions, types, player location, and more—encoded as normalized floating-point values. To help the AI learn, Dave first built a hand-coded expert system in Lua that played competently, if not exceptionally, generating useful training data for the neural network to imitate and learn from.
The AI’s architecture is based on a variant of the Rainbow Deep Q-Network (DQN), which incorporates several advanced reinforcement learning techniques. The system uses a replay buffer to store millions of frames of gameplay, allowing the AI to learn from both recent and surprising experiences. The neural network predicts not just the best action, but a distribution of possible outcomes, and is structured to handle Tempest’s unique action space—firing, zapping, and rotating the spinner dial. Training was accelerated by running multiple emulator instances in parallel on powerful hardware, including a high-end Dell workstation with dual Nvidia GPUs.
A major breakthrough came when Dave restructured the AI’s input representation to match Tempest’s polar, circular geometry, rather than treating the game as a flat grid. By encoding lane positions as sine and cosine values and introducing an attention mechanism, the AI could focus on the most relevant threats—such as nearby enemies—rather than averaging all inputs. This change made the problem more tractable for the neural network, allowing it to develop advanced strategies, anticipate threats, and make decisions that even Dave’s expert system hadn’t discovered.
Ultimately, Dave reflects that the key to success wasn’t simply more computational power or a larger neural network, but making the game’s structure legible to the AI and scaffolding its learning process. By aligning the AI’s perception with the true physics and geometry of Tempest, and giving it the ability to focus on what matters, he enabled it to achieve superhuman performance. The project serves as a testament to the importance of thoughtful representation and empathy in AI design, and as a reminder that even the best human players can be outpaced by a well-crafted machine learner.