Import AI 440: Red queen AI; AI regulating AI; o-ring automation

• Researchers evolved LLM-based agents to compete in 1980s programming game Core War
• Evolved AI warriors defeated 89% of human-designed programs through competitive pressure
• Study suggests future where millions of AI agents evolve through competition

Think of it as Pokémon evolution meets Silicon Valley—except instead of cute creatures getting stronger, we have AI systems locked in digital gladiator matches that make them progressively more cunning. The Sakana researchers essentially created a petri dish for artificial intelligence, watching as LLM-powered programs evolved through combat in Core War, a retro programming battleground where code warriors fight for memory supremacy.

What makes this fascinating isn't just the nostalgic appeal of resurrecting 1980s gaming culture for scientific purposes. The real kicker is how effectively these AI systems improved through competition—jumping from defeating a measly 1.7% of human warriors to crushing 89% of them. This isn't your typical benchmark gaming; it's a preview of how AI systems might evolve when millions of them are unleashed to compete in everything from cybersecurity to stock trading.

The implications stretch far beyond digital warfare simulations. We're approaching a world where AI agents won't just follow human instructions—they'll adapt, compete, and evolve based on environmental pressures we create. Whether that's defending networks against cyber attacks or optimizing supply chains, these systems will be sharpening themselves against each other in ways that might surprise their creators. It's like natural selection, but with silicon brains and the speed of light.

The research also hints at something economists have been wrestling with: what happens when AI systems start automating compliance and regulation? The newsletter discusses how sufficiently advanced AI could write and enforce its own rules, creating "automatability triggers" that activate regulations only when AI systems exist to implement them effectively. This could solve the perpetual problem of regulations being either too early or too late for emerging technology.