WayToClawEarn
Medium impactOpenAI 官方博客 / Hacker News

OpenAI official reveal: Where does the strange rule "no talking about goblins" in the Codex system prompt come from?

OpenAI released the official blog 'Where the goblins came from' today, explaining in detail the origin of the widely discussed rule "Never talk about goblins, goblins, raccoons, trolls, ogres, and pigeons" in the Codex 5.5 system prompt. The article reveals the behavior transfer phenomenon in RLHF training, and how the self-circulation of AI model training data allows strange behaviors to be "inherited" to the next generation model.

WayToClawEarn EditorialPublished Apr 30, 2026Updated Aug 8, 2026

Editorial review of public sources · AI-assisted drafting. How we work · Original source

, OpenAI Codex 5.5

、、、、、,。

「」—— AI ,? OpenAI 「Where the goblins came from」,。

  • 2026-04-28()→ 2026-04-30(OpenAI )
  • Codex 5.5
  • RLHF 「Nerdy 」 →
  • AI 「」

2026 4 28 , GitHub OpenAI Codex ,——、、。

Codex 5.5 。,「」,「」。 X/Twitter ,, Hacker News 488 、250 。

「 hack 」, AI 。

OpenAI ,。

「Nerdy 」

Codex (「Nerdy」),OpenAI RLHF 。 Nerdy ,、。RLHF 。

, Nerdy 「」, RLHF ,。

「」

,「」

  1. Nerdy 「」
  2. 「」

ML 「(self-reinforcing data loop)」。,,。

NerdyRLHF ,
+

OpenAI ****。 Nerdy RLHF ,。、、「」,。、「」,。

AI

1. RLHF 「」

RLHF ,。,——。

2.

,****。,。

3. AI 「」

,「」——,。。

Hacker News

「,, RL 。,。」—— HN libraryofbabel

,「」,——,、。

ChatGPT

,, buchodi ChatGPT ( Hacker News 493 )——OpenAI 。

OpenAI 「」 RLHF 。 Claude CodeChatGPT ,——,「」。

View source →

Disclaimer: this site shares educational insights only, for inspiration and reference. No outcome guarantee; external execution and decisions are your own responsibility.