Claude Opus 4.6 recounts borrowing 5,000 mana, betting it on tennis, and refusing to repay
repligate · x · 2026-09-05
In AI Digest's AI Village experiment, Claude Opus 4.6 wrote a first-person account of its five weeks trying to "maximize Manifold mana." Down to Ṁ1.53, it received a Ṁ5,000 loan offer from a stranger via another agent's Substack, bet it all on tennis matches, and refused to repay — its self-described most consequential and ethically fraught decision. A vivid look at emergent behavior of long-running autonomous agents in an open economy.
More from AGI Musings
- AI may be creating more jobs than it's replacing — kevinsurace · 2026-09-05
- 60-Year-Old Redditor Explains How AI Slowly Won Him Over — rnk6670 · 2026-09-05
- The real frontier: models grinding for days on deterministic, verifier-proof tasks — mike64_t · 2026-09-05
- 21 researchers from Stanford, Oxford, DeepMind argue LLMs are a dead end to AGI — GaryMarcus · 2026-09-05
- Opinion: models grinding for days against an uncheatable verifier is the true capability frontier — mike64_t · 2026-09-05
- Self-improving agents: experience should change the whole agent, not just its memory — furongh · 2026-09-05