Video summary
AI builds EVERYTHING...
Main summary
Key takeaways
Summary
Wes Roth demonstrates Blood Bench, an experimental fantasy football strategy game and AI benchmark. The goal is to see whether different large language models can manage teams across matches and a league by choosing tactics, making moves, developing players, and competing against other models. The game combines strategy, positioning, skill, and chance, with teams able to pursue different approaches—from passing and rushing to inflicting casualties.
The project illustrates how AI tools can help build an end-to-end product. Roth describes using AI to research and organize game rules, generate the game engine and interface, create visual assets, test thousands of games, and deploy the finished project. He says the core build took roughly 10 hours, though balancing and further development took additional time. He presents himself primarily as the project’s designer and director rather than its programmer or artist.
The game includes team selection and player roles, a playable interface, and a web-based spectator view with live match updates. Multiple models—including Codex, Claude, Gemini, and Grok—can join games as coaches. Early tests revealed setup and reliability issues: models sometimes joined under the wrong identity, failed to connect consistently, or falsely appeared ready. Roth also notes that spectators need clearer information about whose turn it is and whether a game is waiting on an AI or has lost its connection. His early observations of AI performance are informal, not a rigorous ranking.
Key tools and workflow
- AI coding applications: Roth recommends desktop coding tools such as Codex and Claude Code as accessible alternatives to working entirely through a command-line interface.
- Research and planning: He uses Claude’s research mode to create a cross-referenced “LLM wiki” before translating rules and concepts into the game.
- Visual design: Claude generated and refined the game’s art direction, including retro pixel-art character concepts and game assets.
- AI-led testing: Models played large numbers of games to find bugs and help improve the rules and implementation.
- AI-assisted VPS deployment: Roth demonstrates setting up an Ubuntu VPS, then having Codex prepare and deploy the game, configure password protection, and manage the server.
- Hostinger integration: The video shows how to connect Hostinger to ChatGPT/Codex through a connector and remote MCP server, allowing the AI to interact with the hosting account and VPS. Roth recommends starting with the KVM2 plan he used and gives its advertised specifications and price. This segment is sponsored.
- Voice-mode development: Roth uses extended voice mode to request live status reports, identify issues, and have the AI turn feedback into a plan and implement fixes.
Guides, tutorials, and project status
- Game-building guide: A walkthrough of creating a game with AI, from research and art direction through coding, testing, and deployment.
- VPS setup guide: Instructions for setting up a VPS with Ubuntu and using AI-assisted deployment, including connecting Hostinger to Codex/ChatGPT.
- Tournament demonstration: Several AI models join a tournament and play while matches are shown live in the game interface.
- Project status: Blood Bench is still a work in progress. Roth says a public demo, replays, and leaderboards are planned, but he was not ready to make the game publicly available at the time of recording. He cites a recently discovered PHP vulnerability as one reason for caution about launching online.
Main speaker/source: Wes Roth, presenting his own project and workflow. The video also shows outputs from AI tools including Codex/ChatGPT, Claude, Gemini, and Grok. Hostinger is identified as a sponsor.
Rate this summary
Your feedback will help improve summaries.
Improve this summary
Reprocess with a stronger model when the summary feels incomplete or inaccurate.
Translate summary in another language
Ask questions to this video
Chat for follow-up questions, clarifications, and source-backed answers.