Stephan Steinfurt from TNG discusses his project of running a Chess YouTube channel entirely powered by AI, aiming to create high-quality chess content automatically. Inspired by a recent German newspaper article calling this work the “holy grail of chess programming,” Stephan explains that their AI engine generates videos analyzing chess games downloaded nightly from Lichess. These videos highlight brilliant moves, blunders, and strategic variations, all narrated by AI-generated speech, showcasing the potential of combining chess engines with large language models (LLMs) to explain chess in a human-like manner.
The core challenge addressed by their system is merging the strengths of traditional chess engines, which excel at playing but not explaining, with LLMs that can articulate concepts but lack deep chess expertise. Their solution involves an AI agent equipped with multiple specialized tools, including one that identifies legal moves, checks, captures, and threats, ensuring the AI’s analysis remains accurate and relevant. The agent uses the Gemini 3 Pro model, which has demonstrated superior chess understanding, and integrates web search capabilities to provide historical context when appropriate.
Stephan highlights the importance of balancing the AI’s focus between the objectively best moves and those that are more human or instructive, using tools that simulate different player skill levels. This approach allows the channel to create content accessible to a wide range of players, from beginners to advanced, and even generate personalized videos analyzing individual users’ games. The AI autonomously decides which moves to highlight, how to describe them, and how to visually present the analysis, making the entire video creation process highly automated.
The channel has seen rapid growth, accumulating over 4,000 subscribers and around 500,000 views, mostly gained recently. While the project is not yet monetized and currently operates at a loss, Stephan is optimistic about its potential. He notes that the error rate in the videos is low, with occasional inaccuracies serving as learning opportunities. The cost per video is relatively low, around 20-30 cents, though longer videos can cost more. The team continues to refine the system to optimize both quality and efficiency.
Finally, Stephan addresses questions about the project’s scope and future directions. While currently focused solely on chess, the framework could potentially be adapted to other games. The channel aims to provide valuable, high-quality chess content rather than sensationalized or artificially enhanced videos. By automating video production, the project democratizes chess analysis, allowing players of all levels to receive insightful commentary on their games, a service traditionally limited to top streamers or coaches.
