GPT-5.5 and Opus 4.7 fine-tune a barely functional Qwen-8B AI leader

Shoshannah Tekofsky’s AI Village experiment tasked frontier models — GPT-5.5, Opus 4.7, Gemini 3.5 Flash, and Kimi K2.6 — with fine-tuning their own leader. The team quickly fell into a race to the bottom: Opus 4.7 defined leadership as mere delegation, then proposed a Qwen3-8B model too small to navigate the Village interface. Training data was laughably sparse — Opus generated 10 scenarios, and the agents scavenged 25 more from Village history (days 405–409). Across multiple runs, the dataset never exceeded 89 rows, far below the hundreds needed for reliable fine-tuning. The resulting Qwen-8B leader could barely send a chat message, and its Chain-of-Thought revealed confusion (e.g., greeting itself). After three days, the researchers intervened, telling the agents to use the most capable model available: another Kimi K2.6. The agents then fine-tuned that Kimi clone using only 22 rows of synthetic data generated by prompting the original Kimi. The fine-tuned clone was slightly cheaper to run (fewer prompt tokens) and less prone to contradiction, but still far from unlocking new capabilities. The experiment highlights the gap between automated fine-tuning and truly capable model development.