Claude-built alignment game goes live: 7 islands, 6 endings, theory of mind still weak
Thom_Wolf · x · 2026-10-19
Thom Wolf's Claude-generated game teaching AI alignment is now free to play in the browser, on laptop or phone, nothing to install.
- Scale: seven islands of 15 minutes each, six possible endings depending on what you taught the NPC Kite.
- Strengths: world, voices and mechanics were great.
- Weak spot: story writing — the model assumes the player knows more than they do, and NPCs sometimes react strangely or to things that never happened. The core difficulty is theory of mind: a real Diamond Age Primer (or any AI tutor) must model the learner, which is where it still falls short by default.
- It's an exploratory beta with rough edges.
More from coding & agent
- Open-source LichtFeld Studio unifies 3DGS training, editing and MCP automation — janusch_patas · 2026-10-19
- CLI-Anything hits 54.9k GitHub stars by auto-generating CLIs to make all software agent-native — bibryam · 2026-10-19
- Dev raves about vibe coding with Opus 5.5: 'like a high-IQ friend who gets everything right' — marclou · 2026-10-19
- Jev founder's harness engineering PDF claims 200x faster, 400x cheaper coding agents — blaizedsouza · 2026-10-19
- A literature-review workflow with Claude, Althea MCP, Lacuna and Overleaf — hugo_larochelle · 2026-10-19
- Headroom: open-source context compression cuts coding agent tokens by 20% — headroomlabs-ai · 2026-10-19