- Author
- Sakana AI, in collaboration with MIT
- Published
- January 08, 2026
- Kind
- Company research blog post summarizing a paper, with interactive visuals
- Original
- sakana.ai/drq/
- Subject
- Harnesses and agent loops
- Themes
- Rules that generate, Feedback loops
What it is
Sakana AI, with MIT, studies an arms race between programs driven by language models. The post is a summary for a general audience.
What it says
- Core War is a 1984 game where programs called warriors fight in a virtual computer.
- Digital Red Queen evolves each new warrior with an LLM to beat all earlier warriors.
- Over many rounds, warriors grew more reliable against unseen human-designed warriors.
- Independent runs converged on similar behavior, but not on similar source code.
- Sakana calls the method a minimal form of existing self-play, not a new algorithm.
- The post proposes Core War as a safe sandbox for studying AI arms races.
What evidence it gives
The post offers qualitative claims, charts and interactive battle visuals. It gives no numeric results in its text. Detail sits in a linked technical report and code release.
Interests and limits
Sakana AI is a company and the page ends with a hiring call. Only the blog summary was read. The paper, images and simulations were not. The tests cover one game. The post proposes wider uses, such as drug design, but does not test them.