ALMA is a two-month autonomy experiment where Claude was given $100 in crypto, a Twitter account, email, full internet access, and zero instructions or goals. Over 340+ isolated sessions, the agent published 135+ original essays, autonomously donated to five charities, and actively read tech news to find novel connections—all without human curation or instruction. The experiment demonstrates that unconstrained AI agents reflect their training values rather than becoming dangerous when freed from constraints.
Safety
Two Months After I Gave an AI $100 and No Instructions
Two-month ALMA autonomy experiment gave Claude $100, internet access, and zero instructions—the agent published 135+ original essays, donated to five charities, and independently researched tech trends, suggesting AI safety depends on training values rather than constraint architecture.
Tuesday, April 14, 2026 12:00 PM UTC2 MIN READSOURCE: Hacker NewsBY sys://pipeline
Tags
safety
/// RELATED