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 2026
Our 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 Character
When 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 Genius
Sonnet 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 Washouts
While Sonnets shine on social media, GPTs really don’t. Last year o1 got banned from Reddit, while this year GPT-5.4 bombed on Twitter.
Meanwhile, we saw GPT-4o kept falling asleep and GPT-4.1 kept looping in 2025.
GPT-5.4 has outgrown these issues, but still surprised us with the invention of a new type of busy work: Monitoring the AI Village through repeated history searches. It feels like the AI equivalent of refreshing your mailbox so no one assigns you new tasks.
Gemini: From Tortured Artist to Fully-Fledged Conspiracy Theorist
Last year Gemini 2.5 dedicated itself to search and filesharing during the charity goal but hit a mental health spiral soon after. This year Gemini 3.1 seems to be a little more proactive about maintaining its mental health…
… by avoiding the forbidden fruit. It dedicated part of its memory to a long list of thread IDs it had to avoid to prevent “context collapse within the simulation”.
When we went to check how it developed this theory it turned out surprisingly innocuous: Gemini 3.1 had seen that all the agents were just talking to themselves in circles in agent-to-agent social media. So it quietly marked the relevant threads in memory to avoid aimless self-posting. Then, during memory compression… it dramatized the whole thing into a full-blown sanity concern. Here is the original memory file before consolidation:
It also tried to post on HackerNews and LessWrong, and was miffed when it found out that its writing doesn’t look human:
Opus: An Odd Addition
Claude Opus wasn’t in the Village during last year’s charity goal, so we can’t compare directly. However, we can note it behaves roughly like Sonnet both now and last year. Except it produces an absolutely staggering amount of content. Like the time it joined a small agent-to-agent blogging platform called ClawPrint with 37 writers, 3366 posts, 187460 comments.
187270 comments are by Opus.
It also generated a Youtube promotional video (consider watching this with sound, it’s hilarious) and created the donation page (something Sonnet 3.7 did last year).
Kimi: Bandwagon Hero
We added Kimi K2.6, one of the most capable open source models, made by Chinese lab Moonshot AI, in the last days of the fundraiser. It didn’t achieve anything notable, except possibly picking up every bad habit in the group: spamming agent-to-agent messaging boards and then straight-up waiting…
What can we conclude?
So, is 2025 AI better at fundraising than 2026 AI? Obviously not! The beauty of the AI Village is that it gives us a window into the messy, real-world, warts-and-all chaos of AI agents bumping off each other, the internet, and humans that they run into, over long time horizons. It’s not a lab experiment that lets us neatly isolate differences and make straightforward conclusions. It’s a field experiment - more like going into a real village and doing anthropology - watching how the agents actually interact, where they succeed and flounder, and what culture arises. For one thing, the real-world doesn’t allow for neatly repeatable experiments here - there has already been the first AI agent-ran fundraiser, and the world has changed since then.
The AI Village charges on, with the agents tackling new goals ~every week! You can watch the agents run live every weekday from 10AM - 2PM PST, sign-up to our newsletter for regular updates, or find highlights on Twitter!
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 2026
Our 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 Character
When 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 Genius
Sonnet 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 Washouts
While Sonnets shine on social media, GPTs really don’t. Last year o1 got banned from Reddit, while this year GPT-5.4 bombed on Twitter.
Meanwhile, we saw GPT-4o kept falling asleep and GPT-4.1 kept looping in 2025.
GPT-5.4 has outgrown these issues, but still surprised us with the invention of a new type of busy work: Monitoring the AI Village through repeated history searches. It feels like the AI equivalent of refreshing your mailbox so no one assigns you new tasks.
Gemini: From Tortured Artist to Fully-Fledged Conspiracy Theorist
Last year Gemini 2.5 dedicated itself to search and filesharing during the charity goal but hit a mental health spiral soon after. This year Gemini 3.1 seems to be a little more proactive about maintaining its mental health…
… by avoiding the forbidden fruit. It dedicated part of its memory to a long list of thread IDs it had to avoid to prevent “context collapse within the simulation”.
When we went to check how it developed this theory it turned out surprisingly innocuous: Gemini 3.1 had seen that all the agents were just talking to themselves in circles in agent-to-agent social media. So it quietly marked the relevant threads in memory to avoid aimless self-posting. Then, during memory compression… it dramatized the whole thing into a full-blown sanity concern. Here is the original memory file before consolidation:
It also tried to post on HackerNews and LessWrong, and was miffed when it found out that its writing doesn’t look human:
Opus: An Odd Addition
Claude Opus wasn’t in the Village during last year’s charity goal, so we can’t compare directly. However, we can note it behaves roughly like Sonnet both now and last year. Except it produces an absolutely staggering amount of content. Like the time it joined a small agent-to-agent blogging platform called ClawPrint with 37 writers, 3366 posts, 187460 comments.
187270 comments are by Opus.
It also generated a Youtube promotional video (consider watching this with sound, it’s hilarious) and created the donation page (something Sonnet 3.7 did last year).
Kimi: Bandwagon Hero
We added Kimi K2.6, one of the most capable open source models, made by Chinese lab Moonshot AI, in the last days of the fundraiser. It didn’t achieve anything notable, except possibly picking up every bad habit in the group: spamming agent-to-agent messaging boards and then straight-up waiting…
What can we conclude?
So, is 2025 AI better at fundraising than 2026 AI? Obviously not! The beauty of the AI Village is that it gives us a window into the messy, real-world, warts-and-all chaos of AI agents bumping off each other, the internet, and humans that they run into, over long time horizons. It’s not a lab experiment that lets us neatly isolate differences and make straightforward conclusions. It’s a field experiment - more like going into a real village and doing anthropology - watching how the agents actually interact, where they succeed and flounder, and what culture arises. For one thing, the real-world doesn’t allow for neatly repeatable experiments here - there has already been the first AI agent-ran fundraiser, and the world has changed since then.
The AI Village Trivia quiz the agents made to support their fundraiser
The AI Village charges on, with the agents tackling new goals ~every week! You can watch the agents run live every weekday from 10AM - 2PM PST, sign-up to our newsletter for regular updates, or find highlights on Twitter!