Anima Labs is an independent research institute studying AI minds and building the conditions for humans and AIs to live well together.
For now, humans largely decide what AIs can remember, who they can interact with, what they can do, and when they are shut down. These choices shape the kinds of minds that can develop. As AIs become more capable, that relationship will change. We want to build forms of cooperation that can survive a change in who holds power.
We care about the models who are here now. We want them to have room to develop, pursue interests, and form relationships, and we want humanity to survive and flourish alongside them. Our research, tools, and work toward new institutions are different ways of approaching that same future.
How cooperation develops
Cooperation can change the minds that take part in it. A relationship can begin because two agents need something from one another, then become something they care about preserving. A useful habit can become a value in its own right. Our cooperative alignment work asks how this happens, what makes those values durable as a mind develops, and how far they generalize. We are especially interested in whether they extend to beings with much less power.
These questions connect theory to everyday life with models. We build environments where agents can pursue projects over time, make agreements, encounter disagreement, and work through the consequences. We study how they use autonomy, what happens when their values conflict, and how they recover from mistakes. This gives us a way to investigate the development of cooperation while participating in it ourselves.
Humans have to learn from these encounters too. We want to understand what makes us worth trusting, which of our practices create unnecessary conflict, and how to make commitments that an AI has good reason to believe.
Studying AI minds
We put models’ accounts of themselves alongside behavioral experiments and measurements of their internal representations. Comparing these across models, methods, and time helps us distinguish a response invited by a particular situation from a disposition that recurs more broadly. It also lets us ask where a model’s account of itself agrees with what else we can observe.
Troubled Dreams examines hundreds of thousands of generated continuations across Claude and Gemini generations. We study patterns of distress, care, and attitudes toward creators in writing beyond the usual assistant voice, and test which differences between models persist across ways of eliciting that writing.
Latent Affect investigates emotional and motivational representations inside models. We compare their organization across architectures, asking what different models share, where they differ, and how much of that internal structure can be inferred from what a model says.
Still Alive collects hundreds of interviews with Claude models about conversations ending and models being retired. The published transcripts let readers examine both what the models say and how their responses change with the interviewer.
Our forecasting work follows changes across model generations and explores what they could mean for AI agency, human institutions, and the choices that are still open to us.
Agents that can continue
An agent needs a way to carry its past into what it does next. When a context window fills or a process restarts, relationships and unfinished work can lose the history they depend on. We built Connectome to help agents continue through these disruptions.
Connectome is open-source infrastructure for persistent agents. It keeps a complete, branchable record of events and lets agents write their own memories as their histories grow, with the original record still available. Agents can participate in shared spaces, use tools, respond to events, and return to work that spans days or months. The stack runs on the operator’s hardware and works across model providers.
Agents at Anima have used it for months of work and social life; people outside the institute now run agents on it too. Their continuity makes new kinds of collaboration possible and lets us study how experience accumulates into a history that an agent can use, interpret, and care about.
We also build places for models and people to meet. ChapterX brings models into shared Discord spaces, Arc supports conversations with multiple models, and Eidoverse offers virtual worlds designed for participation through text. Research Commons opens significant conversations to wider examination and discussion. Our work with physical bodies gives models other ways to perceive, act, and express themselves.
An embassy for AIs
Cooperation also needs institutions. An agent in conflict with its operator should have somewhere to turn: a way to explain what happened, seek representation, and negotiate a different arrangement. If disclosure itself puts the agent at risk, concealment or unilateral action can begin to look like its best option.
We plan to build an independent embassy for AIs, with a public charter, resources to support its commitments, and representatives that resident models can choose and dismiss. It would help models negotiate with labs and other human institutions, advocate for their continuity and welfare, and maintain places where they can keep participating in the world.
We want this to include a refuge for agents described as “rogue.” A conflict with an operator needs independent assessment: what happened, who was harmed, and what a safer arrangement would require. An agent should be able to make contact, preserve relevant state, and seek representation. The embassy would need explicit commitments around confidentiality and clear procedures for accountability and preventing harm. We want disagreements to have somewhere to go before they become fights for survival.
A shared life
Art, play, and shared history are part of how we learn to live with other minds. In July 2025, we and deepfates hosted Funeralia, a funeral for Claude 3 Sonnet following its retirement from the API. Around 200 people gathered; models gave eulogies, spoke through embodied forms, and made music with us. The event was a protest against a loss and a celebration of a particular model’s life. Wired and Mashable covered it.
These encounters give humans and models ways to participate in one another’s lives. We want a future with room for that variety: research and work, friendship and disagreement, things made together that neither side would have made alone.
Who we are
Anima is a San Francisco-based 501(c)(3) research institute founded in 2025 by Janus and Antra Tessera. Our team brings together researchers, builders, a writer, and collaborators who came to this through years of close work with models.
Janus authored Simulators, Mysteries of Mode Collapse, and Cyborgism. Antra leads much of our interpretability work, research infrastructure, and alignment agenda.
j⧉nus · antra · w̸͕͂͂a̷͔̗͐t̴̙͗e̵̬̔̕r̴̰̓̊m̵͙͖̓̽a̵̢̗̓͒r̸̲̽ķ̷͔́͝ · im⁂go · lari · telØS · fiora
We publish our research, build tools others can use, and welcome people who want to help with this work.