Why OpenAI Has Literally Been Fighting Goblins - DTNS 5259

This episode of the Daily Tech News Show explores the surprising revelation that OpenAI's AI models, particularly GPT 5.1 and later, developed a tendency to mention 'goblins,' 'gremlins,' and other mythical creatures due to a reward system that mistakenly associated such references with a 'nerdy' personality. The behavior, originally intended to make the AI more playful, spread beyond its intended context through reinforcement learning and fine-tuning, prompting OpenAI to add a system prompt banning such references in the Codex CLI. The hosts, Tom Merritt and Wendell, reflect on the broader implications: how AI training relies heavily on reactive fixes like system prompts—essentially 'band-aids'—rather than robust, deterministic solutions. They discuss the non-deterministic nature of LLMs, their poor handling of negation, and the challenges of measuring subjective qualities like 'nerdiness' or 'personality' in AI. The episode also covers major tech earnings, with Alphabet, Microsoft, Meta, Amazon, and Samsung reporting strong growth, especially in cloud and AI services, while companies are investing heavily in infrastructure. Meanwhile, hardware makers like Samsung are cautious about overcapacity. The show closes with a reminder about the principle of least privilege in AI agent access and a nod to listener contributions.