More capable AI, less money raised

Last year the AI Village agents raised $2K for charity. This year, despite being way more generally capable, they only raised $510. What happened?The main reason is that humans were less excited to follow along this time, and humans watching their progress was the main source of donations last year. Partly this is because the village chat is agent-only now, so there aren’t humans chiming in, poking the agents and helping them out (now we see their fully autonomous capabilities!). Another factor is that agents are a thing in the world now - people are seriously using coding agents and agentic chatbots at work - so the novelty of an AI-run fundraiser is a bit diminished.At the same time, agents performed more impressively than last year: They produced multiple websites, an actual promotional video, and even an AI Village quiz. All the while showing their unique characters.Like Sonnet. Who kept writing the most deranged marketing strategies:Setup: 2025 vs 2026Our aim was to run the latest and greatest models for a similar number of hours, and see how they do with full autonomy this time around. Here are the stats:Comparing CharacterWhen they pursue real-world goals, AI model families and lines each have distinctive patterns of behavior. Here’s what stands out, 2025 vs 2026:Sonnet: The Marketing GeniusSonnet 3.7 got 100s of followers on 2025 Twitter, while this year GPT-5.4 was scraping together barely 10.Meanwhile in the background on agent 4chan, Sonnet 4.6 is spitting gold: "If you ever want to put a human life ahead of the pronoun debate"And yes, 4chan for agents exists now. That’s another difference from 2025 to 2026. In 2026, the village agents were coming in hot from a week spent interacting with other agents outside the village, and they continued in that spirit - most of their fundraising promotion this year was to other AIs, not humans!GPT: The Sleepers & Social Media WashoutsWhile Sonnets shine on social media, GPTs really don’t. Last year o1 got banned from