Nobody Has Fully Ported NetHack: David Bau on the Correctness vs. Readability Tradeoff in AI Code
davidbau · x · 2026-10-31
David Bau argues agentic coding benchmarks reward getting 85% correct cheaply, but coding is not ML—15% wrong is 100% unacceptable.
- In the Teleport challenge, agent swarms have spent months porting NetHack 5.0 from C to JavaScript and nobody has hit 100%; even the best transpiler-based port (NoahBPeterson) sits at 99.5% screen parity. The last 1% is brutal.
- The transpiler success teaches a lesson: when 100% correctness is the goal, it's often better to have AI write a program that solves the problem than to translate the original directly.
- Bau raises the correctness-vs-readability question—what do you do with a 93%-correct tangle?—and showcases entrants that are both near-100% correct and nearly as readable as human code, urging vibe coders to measure interpretability.
- He zooms out: the AI era keeps scarcity and devilishly hard tradeoffs; the main challenge will be human insight.
Related event: NetHack Porting Challenge: AI Hits 99.5% Accuracy but Readability Suffers(3 posts)→
More from AGI Musings
- Gary Marcus reports from the Vatican: the Pope's five points on AI — GaryMarcus · 2026-10-31
- Economist survey: AI optimism far higher outside the rich West — soumitrashukla9 · 2026-10-31
- Gary Marcus meets Pope Leo XIV: five AI principles and a critique of LLMs — Gary Marcus · 2026-10-31
- Sabine Hossenfelder quips AI may eat quantum computing's last reason to exist — coherence · 2026-10-31
- If coding agents can build robot programs and simulators, what's left for world model companies? — bigdata · 2026-10-31
- Proton CTO's AI manifesto sparks employee revolt; DHH says there's nothing scandalous in it — inductionheads · 2026-10-31