Digital Red Queen: Adversarial Program Evolution in Core War with LLMs
pub.sakana.ai · 6,936 words · saved by 1 readers
A self-play algorithm that uses LLMs to evolve adversarially competing programs in Core War
--> This page requires Javascript. Please enable it to view the website. Visualization of Core War battles. In this work, we study Digital Red Queen (DRQ), a self-play algorithm that uses large language models to evolve assembly programs called "warriors" which compete against each other for control of a virtual machine in the game of Core War. Symbols indicate instruction opcodes, and colors denote the warrior that last modified each address. Code and data share the same address space, enabling self-modification and creating a volatile environment for warriors. --> --> Examples of discovered
saved by
related reading
- Guardian Angels: LLM Personalization for Productivity and Security · Gwern.netgwern.net
- Core Waren.wikipedia.org
- PopuLoRA: Co-Evolving LLM Populations for Reasoning Self-Playvmax.ai
- Harness Engineering for Self-Improvement | Lil'Loglilianweng.github.io
- LLM Powered Autonomous Agents | Lil'Loglilianweng.github.io
- The Waluigi Effect (mega-post) — LessWronglesswrong.com
- [2206.08896] Evolution through Large Modelsarxiv.org
- As Rocks May Think | Eric Jangevjang.com
- Scaling Self-Play with Self-Guidancearxiv.org
- Andrej Karpathy — AGI is still a decade awaydwarkesh.com
- Evolution as Backstop for Reinforcement Learning · Gwern.netgwern.net
- Mechanize War, Inc.mechanizew.ar