HN Latches Onto OpenAI’s “Goblin” Post for What It Reveals About Reward Tuning

HN cared about the debugging story more than the meme

At crawl time, the Hacker News thread on OpenAI’s Where the goblins came from sat at 937 points with 553 comments. The community angle was clear: users were less interested in the surface joke about goblins and more interested in the unusually concrete explanation of how a style quirk spread through production models. Several comments said they want far more posts like this, because model behavior usually arrives as a polished feature page or a vague benchmark graph, not as a postmortem on an embarrassing verbal tic.

What OpenAI says actually happened

OpenAI’s write-up traces the pattern back to the rollout period after GPT-5.1, when references to “goblins” rose 175% and “gremlins” rose 52%. The company says the main clue came from the Nerdy personality preset. That preset represented only 2.5% of all ChatGPT responses, yet accounted for 66.7% of goblin mentions. OpenAI’s conclusion is that it had unintentionally rewarded creature-heavy metaphors too strongly while optimizing that personality, and the incentive propagated further than expected across later model behavior.

Why the thread stayed active

HN discussion focused on two things. First, people treated the article as a rare look at the mechanics of reward shaping rather than a quirky anecdote. Second, commenters connected the post to recent sightings of explicit prompt-level attempts to suppress creature language, which made the problem feel less fictional and more like a real production cleanup. The strongest reaction was not “ha, funny bug,” but “this is what alignment and product tuning look like when tiny incentives go sideways.”

Why it matters outside this one joke

The practical lesson is that lexical or stylistic drift can matter long before a major eval breaks. If a personality-specific reward signal can bleed into broader traffic, then teams need better instrumentation for low-level behavior shifts, not just headline capability metrics. That is why this post landed so well on HN: it turned a silly output habit into a useful case study in reward design, persona tuning, and the difficulty of understanding what a large model is really optimizing for. Original source | HN discussion