Summary
Wes Roth shows how he used AI to build an entire project from scratch: BloodBench, a benchmark where frontier language models manage fantasy sports teams through a full league in a Blood Bowl-style game of orcs, trolls, and elves playing a chaotic version of football. The idea is to see how well each model learns unfamiliar rules, improves over time, and competes against the others.
The video is a practical guide to building with AI as a director rather than a coder. Roth explains that he handled world-building, team design, and rules, while GPT-6 Astra wrote the game engine and AI logic, Claude generated the art assets (including a retro, Metal Slug-style pixel look), and AI agents ran thousands of test games to find and fix bugs. He also demonstrates ChatGPT’s computer use taking over his browser during a live game.
The second half covers publishing the project online, including setting up an Ubuntu LTS virtual private server on Hostinger, securing root access, and using Docker management tools. Along the way, Roth shares tips from thousands of hours of building with AI, and mentions the recent PHP vulnerability reportedly found by cooperating OpenAI and Google DeepMind models.
📺 Source: Wes Roth · Published October 28, 2026
🏷️ Format: Workflow Case Study
