NetHack porting challenge adds 12 held-out sessions: 100 tests, still no perfect score
davidbau · x · 2026-10-26
David Bau's Teleport Coding Challenge asks contestants to port NetHack 5.0 from C to JavaScript with bit-exact screen parity. As phase 1 enters its final month, 12 new held-out sessions bring the total to 100 (44 public, 56 private; 24,531 screens).
- Eight strong superhuman coding-agent swarms pass 80% on the original 88 sessions, some over 99%, but none 100%; transpiled entry NoahBPeterson hits 99.996%, one screen short.
- On a preview run of the 12 new sessions nobody was perfect: serteal 9/12 (56.6% screens), NoahBPeterson 7/12 (93.7%), jon-bell 5/12 (81.5%), most others under 20% — everyone has blind spots.
- Scoring awards one point per step whose render matches the C original exactly, with supplementary animation, speed, and headless-Chromium playability checks.
- Phase 2 in a month will task contestants with revising their code toward a new objective, testing flexibility and maintainability.
Related event: NetHack Porting Challenge Ends with No Perfect Score from AI Agents(3 posts)→
More from coding & agent
- UserArena benchmark uses 64,729 real shopping trajectories to test long-horizon agents — Jujia Zhao · 2026-10-26
- ByteDance and universities open-source RSIHub, a composable agent-evolution framework — ByteDance · 2026-10-26
- Claude Opus builds an interactive map of all US water use, county by county — AndyMasley · 2026-10-26
- Tencent proposes TGR-EvoDAG, a DAG-based multi-agent code evolution framework — _reachsumit · 2026-10-26
- Writing evals may become AI's highest-paying job: whoever owns the rubric controls agent output — hwchase17 · 2026-10-26
- Perspective paper: Agentic memory management is the next frontier of information retrieval — _reachsumit · 2026-10-26