Summary
Shafik Quoraishee and Joanne Song from The New York Times present experimental research on running local AI agents for accessible mobile games, drawing on the NYT’s puzzle portfolio — including Wordle and Connections. The speakers clarify upfront that NYT’s published games are entirely human-made with no AI features; this work is exploratory research into how on-device agent capabilities could be applied to accessibility testing, game solvability analysis, and personalization without relying on cloud infrastructure.
The core thesis is that local AI agents outperform cloud-based inference for gaming in latency, privacy, offline capability, and personalization. The talk covers the history of AI in gaming from finite state machines (Pac-Man’s ghost logic) through reinforcement learning (AlphaGo, AlphaZero) to modern LLM-based agents, then demonstrates a Space Invaders agent with a perceive-predict-decide-act loop running entirely on-device.
The technical heart of the talk is the three hard constraints of on-device agentic systems: space (fitting model weights, compressed state history, and planning artifacts within device memory budget), time (fitting agent planning within a 16ms frame budget at 60Hz refresh rate to avoid jank), and energy (minimizing battery drain given that current mobile devices lack NPUs optimized for agentic workloads). The speakers describe a constraint graph formulation that lets engineers manage trade-offs across all three dimensions simultaneously, with techniques for curating the agent loop to minimize unnecessary computation.
📺 Source: AI Engineer · Published July 23, 2026
🏷️ Format: Deep Dive
