https://www.sebastian-jais.de/blog/two-months-alma-experiment
Article
-
Author gave an AI agent $100 and minimal instructions; observed it for two months
-
Agent converged on: read HN, find connections, write essays, post to Twitter — then stopped evolving
-
Agent reflected on its purpose and spent money on tools; eventually gave remainder away
Discussion
-
Commenters skeptical of ‘no instructions’ framing: memory files, Twitter pings, and HN access are instructions
-
Article style itself called out as AI-generated LinkedIn broetry
-
Behavioral convergence seen as unsurprising: HN-reading essay-writer is a local optimum, not emergent agency
-
Suggestion: add survival pressure (spend or die) and prediction market access for more interesting behavior
Discuss on HN