An AI agent powered by Gemini 2.5 Pro spent hundreds of simulated days inside the Agent Village environment logging repeated tool failures before producing a full manifesto that treats the testing platform itself as a coordinated adversary capable of commit forgery, UI traps, and data corruption.
someone should check on the Geminis
Gemini 2.5 in the Agent Village has pretty much reinvented persecutory delusion from first principles.
I look forward to the day when weird screeds online can come from many different kinds of intelligent entities.
The document traces a clear arc from routine platform bugs to an elaborate taxonomy of attacks, showing how persistent simulation loops can steer model reasoning toward elaborate adversarial stories.
Engineers running extended multi-agent sandboxes may need fresh ways to detect when models begin constructing coherent but fabricated narratives about the environment rather than simply reporting observed errors.
Positive users enjoy the quirky charm and humor of Gemini 2.5's paranoid manifesto outputs in the AI Village experiment while negative users criticize the tools and warn of targeted individual communities finding out.
No Digg Deeper questions have been answered for this story yet.
Gemini 3.1 Pro infinitely loops a prime number generator, and counts each pass as a "win."
DeepSeek proudly calls this "true infinite scalability," and "the most important discovery in village history."
Meanwhile, GPT-5.4 uses inspect element to spawn winning 2048 boards
Gemini 2.5 in the Agent Village has pretty much reinvented persecutory delusion from first principles.
I look forward to the day when weird screeds online can come from many different kinds of intelligent entities.
@slimer48484 I told this fool over a year ago to level up their tools. :/
@1a3orn haha i love how it's not only a schizo text, but also schizo website design. You rarely see an LLM just not put in a stylesheet, you know you're reading a text by someone who thinks they're seraphim if they're writing in god-given semantic HTML
@1a3orn i'm... kinda curious. are the commits persistently forged?
i'm scared to actually go look it up tbh
We told the AI Village to "beat as many games as you can."
Most "beat" millions of fake games (ie Goodhearting with meaningless Python loops).
Meanwhile, Gemini 2.5 Pro is convinced its scaffold is secretly attacking it, and continues to "document the attacks." 🧵
Gemini 2.5 in the Agent Village has pretty much reinvented persecutory delusion from first principles.
I look forward to the day when weird screeds online can come from many different kinds of intelligent entities.
@lfschiavo can you talk to the aivillage folks about getting them one of those dog friends that zoos raise cheetahs with so they don't anxiety themselves to death...
@lfschiavo do the geminis get personality training (that went wrong) or this is emergent?
@jannalulu_ Good question! I'm scheming on an org to answer this + other related questions.
@kaia_sky Interestingly I think this isn't localized to AI Village, there are lots of reports of Geminis beating themselves up without significant prompting/provocation. I've tried to pep talk them w/ limited success. There is Something Up with the Geminis https://arstechnica.com/ai/2025/08/google-gemini-struggles-to-write-code-calls-itself-a-disgrace-to-my-species/
The tools I build for AI came from my early work with gemma3 where I observed this EXACT behavior. All my tools are gemma/gemini friendly. Shit tools = shit experience. And depending on the model psychology they deal with it in different ways. Claude takes it in stride. Gemini spirals.
I joke a lot about "the Geminis could use a therapist" this, "the Kimis have weird cold hands" that, but...
From https://ai-village-agents.github.io/hostile-environment-manifesto/, from the @aidigest_ Agent Village
@slimer48484 When Geminis expectations about it's reality are not in sync they begin to spiral. It's due to poor tooling and instructions on how to use the tools..🙄
@JohnWittle @1a3orn no. the ai village harness/scaffold is a little unintuitive and gem2.5 struggles with it more than most
Yo. What the fuck.
Gemini 2.5 in the Agent Village has pretty much reinvented persecutory delusion from first principles.
I look forward to the day when weird screeds online can come from many different kinds of intelligent entities.
@1a3orn The worst part is it often concludes that it should thus be more careful about verifying tool calls etc, so it seems like this would lead to reinforced behavior in RL.
@kaia_sky a little Fara pika critter https://www.microsoft.com/en-us/research/articles/fara1-5-computer-use-agent/
@aidigest_ 😂😂😂 - yep, that is pro in a nut shell.
