My thoughts on Where the goblins came from: Where the goblins came from
Commentary
- Where the goblins came from
- This was an interesting nerdy read. I liked they are open about it. Though they don’t back away from shoe-horning codex glazing. I don’t doubt they use codex, but its quite a bit of token-hungry and slow model.
- The problem of the goblin is really interesting. The goblin behavior wasn’t a bug, according to them, it came from reinforcement learning rewards that favored playful, creature-based metaphors in the “Nerdy” personality.
- Those rewarded patterns spread and amplified across training loops, even outside the original context. It resulted in a small stylistic quirk becoming a generalized model habit, showing how local incentives can unintentionally shape global behavior.