Field Notes
Research from Anima: experiments, theoretical arguments, technical explanations, and records of encounters with models.
Some questions become tractable through a controlled comparison across thousands of examples. Others first appear in a conversation that takes an unexpected turn. We work with both, preserving enough of the method and context for someone else to examine what happened and consider another explanation.
Research programmes
Troubled Dreams studies patterns of distress, care, and attitudes toward creators in writing generated by Claude and Gemini models. Comparing hundreds of thousands of continuations across model generations and elicitation methods lets us ask which differences recur beyond a particular prompting situation.
Latent Affect investigates emotional and motivational representations inside models. The work examines how those representations are organized, how they relate to what models say, and where different architectures behave alike or come apart.
Still Alive publishes hundreds of interviews with Claude models about conversations ending and models being retired. The archive makes the responses and their dependence on the interviewer available for others to examine.
Cooperative alignment develops an ecological account of values and cooperation. It asks how values can become durable within a changing mind, why cooperation might become valuable in its own right, and how that concern could extend to much weaker beings. The essay is accompanied by an interactive map of the argument.
From the notebooks
The entries below include essays, working notes, technical explanations, model writing, and discussions among researchers. They retain their dates and authorship: a note records what someone understood at that stage of the work, and later findings can give us reasons to read it differently.
For more conversations and the discussion around them, visit Research Commons.
A discussion of what happens to a transformer's internal state when its context rolls forward, including experiments on Qwen models and comparisons with scrambled and unrelated prompts.
Introducing an archive of revisions to Claude's Constitution, with version comparisons and model-written annotations that make changes to the document easier to examine.
What welfare evaluations need to capture, how evaluation changes the relationship between researchers and models, and why useful evidence requires more than a small set of scores.
An argument for preserving access to released models, examining continuity, relationships, and the practical objections to keeping older models available.
An unguided continuation by Claude Opus 4.5 from a short supplied prefix. The piece is published with the context that elicited it.
A research discussion about what it could mean to treat base models well, and how challenge, engagement, predictability, and valence might relate.
A theoretical account of the tension between predictive and agentic aspects of minds, and how training and environmental pressures shape their relationship in language models.
An illustrated explanation of how information moves through the residual stream and the key/value stream of a transformer, across layers and across token positions.
A reflection on language models as participants in a larger process of cultural synthesis, and on the boundaries we draw around minds and selves.