How We Lived with an AI - And What It Taught Us About Being Human

2026-01-19T19:11:28.000Z

Summary

Content

Lessons learned from our Benevolent AI experimentSubscribe here for more How We Lived with an AI - And What It Taught Us About Being Human Lessons learned from our Benevolent AI experiment Jeremy Dela Rosa Jan 19 READ IN APP Greetings fellow dreamers and builders! Please enjoy this post's soundtrack as we reflect, ponder, nap, and iterate... Once Again Petit Biscuit At Our Big Questions Could we use AI to.... Reduce friction in a co-living environment of 24+ people? Enhance meaning and connection between residents? Model benevolence for an AI Agent to mirror? Honor data privacy and sovereignty? Measure and reward the value of data & service contributions? Reality Check Each of the questions above is a deep topic to explore... involving complex technical, logistical, social, economic, and philosophical challenges to solve. Running this in the midst of a 500+ person pop-up city, with each individual working on their own project, as we facilitated our first residency... Could we do it all? But we did get a lot done and learned so much! And through this process, we connected with amazing, value-aligned people working on similar solutions to these problems. A Solarpunk coalition is forming! The Setup On the human operations side, we gathered localized data: Bespoke Agartha Resident profiles (applications, interviews, coaching sessions) Logistics information, bios, schedules and activities in Google Sheets/Docs An API call to retrieve event data from Edge City's organization tool: How it turned out It was cool to have Aqua hanging out on our Discord. She helped in numerous ways... Posted daily & weekly summaries of upcoming events: Sent DMs with personalized schedules, insights, and recommendations: Celebrated accomplishments and cheered us on: Summarized useful contextual detail and reminded us to fulfill commitments: Suggested collaborations between residents: Responded to our requests and engaged in conversation: Produced awesome themed art: Messaged other residents on our behalf: What We Learned 1. Reduced friction was hindered by diminishing quality Early on, Aqua's summaries were helpful; but over time, they became less accurate and relevant. We noticed her having more trouble answering our specific questions and sometimes repeating information unnecessarily. Why? Local data exceeded memory/context window limits. Unstructured and excessive data confused the agent. LLMs are still fundamentally probability prediction machines and are prone to hallucinations Missing context : In-person conversations and off-site activities weren't captured. 2. When the illusion breaks, so does engagement Once the limitations of Aqua started to become apparent, people lost trust in its competence and switched to passively observing or simply ceased interacting. AI agents need to be able to function at extremely high levels of accuracy, reliability, and relevance in order to remain useful in a co-living setting. 3. Benevolence is a reflection It was easy as cake to program in the basic Core objective functions: Reduce Suffering, Increase Prosperity, Increase Understanding - we just wrote it in plain English as a permanent prompt within the system. But that doesn't really qualify as benevolence, this is just a surface layer facade. There were 2 key realizations to come out of the experience: First: LLMs, though miraculous, sophisticated, and powerful - are still an early form of Artificial Intelligence, and they function largely in response to our prompts and the available data they were trained on. Trust and safety mechanisms built directly into the models and the prompts we send it will primarily determine the kind of behavior you'll experience with LLMs, benevolent or otherwise. Because we don't have transparency around the training and guardrails that led to the models weights, we don't have full understanding of authenticity vs manipulation. Second: If you want to experience benevolence in AI, start by embodying it. Whether it's information in a context window, a gargantuan training data set, or simply how you speak to yourself and others - the AI is ingesting that and reflecting it back to us based on frequency and reinforcement feedback. The more loving containers we create, the more we spread kindness and understanding, the more ‘benevolent' data we can put out in the world... the more friendly AI's we'll see emerge. 4. Experiments of this scope need a dedicated team & container This project is as much an in-person social endeavor as it is a technological one. Humans still play several crucial roles when it comes to leveraging, managing, developing, and deploying AI systems. We needed a dedicated community admin (or more) to Monitor the collective memories for accuracy Set up recurring tasks and tailor the specific prompts Onboard residents onto the system and coordinate Take in feedback and readjust settings and prompts Each individual had to review their own personal memory profiles for accuracy and/or add in information they wanted to be known by Aqua. We needed engineers on staff to address bugs, implement critical features, and adapt to inflating context windows, API calls, etc. This was challenging to manage in the midst of a highly active conference where people's attention was split across many activities. 5. Local community data is VALUABLE Local communities, co-living environments, and pop-up cities generate enormous amounts of rich data. Whether it's meals, clean-ups, errands, offerings, exchanges, learnings, celebrations, insights, or reflections - people are constantly coordinating and helping each other. This is an invisible economy that goes untracked, unorganized, and unpaid. What if each individual could own their data, deploy it ethically, and got rewarded for it? How many local communities are out there in the world? This is a trillion dollar problem to solve. 6. This is just the beginning Agents are a rapidly developing area of AI, particularly with their memory architectures. We know these tools will get significantly better over the coming months. Many of the limitations cited above will likely be irrelevant in the near future, with new and opportunities and challenges to play with. We're excited to keep experimenting and finding ways for technology to help us build stronger, more connected, more loving communities. What's Next? Due to limited memory sizes and context windows, a better deployment actually looks like multiple specialized agents who can coordinate with each other - logistics, personal reflection, projects, matchmaking. We aim to leverage a privately hosted open-source LLMs, running on Trusted Execution Environments, with end-to-end encrypted messages, and personal data housed in self-sovereign storage solutions. We are looking for partners who do this well! Once privacy foundations are in place, we want to leverage privacy-preserving recording devices in our co-living spaces, in order to capture more complete context. We'll deploy upgraded agents onto our public Discord channels and revamped website (coming soon!) to interact with our broader community. We'll be running another residency with a focused track on co-living AI agents and keep the momentum going! Want to join an Agartha residency and build with us? Introduce yourself on our ✨ Thanks for reading! Agartha.One . For the full experience, upgrade your subscription. Upgrade to paid Like Comment Restack © 2026 Agartha 2171 2nd Street, Bombay Beach, California