Where the goblins came from

ahboyashreads (@ahboyashreads), Apr 30, 2026, 06:51 UTC

Where the goblins came from GPT-5.4 and later models frequently describe situations, problems, or entities as "goblins" or "gremlins," a pattern that becomes highly noticeable and irritating for daily users Models increasingly using "goblin," "gremlin," and similar creature metaphors - to a reward signal in the "Nerdy" personality training that unintentionally favored creature-word outputs, which then spread to other contexts via reinforcement learning and SFT data feedback loops They fixed it by retiring the Nerdy personality in March, removing the offending reward signal, and filtering cr…

The excerpt is quoted from a public Telegram channel, which owns its text.

Read the original post on Telegram

More posts from ahboyashreads · All channels

FixLane Live vulnerabilities with ready fix prompts