Local Agentic Theory For Mobile Games — Shafik Quoraishee & Joanne Song, The New York Times

Shafik and Joanne from The New York Times emphasized that their puzzles are created by human designers, not AI. They explicitly state there is no AI or AI features within the games themselves, including things like Wordle Bot which is a separate tool. This commitment to human-created content distinguishes their approach from AI-generated game development.

The discussion traces the history of AI in gaming back to the 1980s, citing Pac-Man as an early example utilizing finite state machines. These systems used conditional logic to dictate ghost behavior based on Pac-Man's actions, representing a basic form of symbolic AI that has evolved significantly over time.

The presentation highlights the benefits of running AI computations locally on devices, rather than relying on cloud infrastructure. This approach reduces latency by eliminating round trips to servers, improves privacy as data remains within the device's security zone, and enables functionality even without an internet connection – crucial for scenarios like subway rides.

Shafik introduces EfficientZero and EfficientZero V2 as state-of-the-art reinforcement learning models. These models are characterized by their 'sample efficiency,' meaning they learn quickly with less data compared to traditional methods, representing a significant advancement in training AI agents for complex games.

The agent acts as a 'live layout auditor' constantly measuring the interface and dynamically resizing controls to accommodate user needs. This goes beyond traditional accessibility features, actively responding to issues like shaky taps or difficulty with targets. For example, it can catch violations and rewrite the layout live, adapting to the human rather than forcing the human to adapt to the game.

Shafik uses a tug-of-war analogy to illustrate how the system dynamically tunes the game to the player, rather than maintaining a fixed state. This means accessibility and challenge are not treated as separate concerns but become two ends of a single dial constantly adjusting for optimal user experience. The goal is to create a responsive and empathetic gaming environment.

To truly understand games, local agents need significant improvements in speed – decisions must be made within a 16ms frame to prevent stuttering. They also require predictive models to anticipate the impact of layout changes before implementation. Crucially, long-term memory is needed to learn an individual's unique habits and needs over time, leading to highly personalized experiences.