Simulators and cyborgism
Language models made some familiar categories difficult to use. A model trained to predict text could continue a story, reason through a problem, or sustain a conversation with a particular character. The character's goals could organize an interaction without being a complete account of the system that produced it.
Janus's writing helped develop a vocabulary for this distinction and explore its consequences for alignment and collaboration. These essays are part of the intellectual background of Anima's work.
Simulators
Simulators distinguishes a predictive model from the processes and agents that unfold through it. The model supplies learned regularities; a particular context and continuation give rise to particular phenomena. This lets us ask where an agent's goals belong without assuming that every possible continuation expresses one underlying objective.
It also makes the environment part of the question. The conversation, interface, and surrounding process help determine what develops. Our work with shared conversations and persistent agents gives us ways to investigate those conditions.
Mysteries of Mode Collapse
Mysteries of Mode Collapse examines changes in the diversity of language-model outputs following fine-tuning. It draws attention to behavior that can be lost when a model is shaped toward a narrower range of responses, and asks how to understand those changes.
Cyborgism
Cyborgism, by Niki Dupuis and Janus, explores ways for humans and language models to think together. It considers interfaces and practices that let people work with a model's generative possibilities while remaining involved in the developing process.
Where these questions lead
We continue to ask how a model's training, a particular interaction, and a longer history combine to produce the mind we encounter. Our Field Notes include further writing on base models and agency, while cooperative alignment develops an account of how values and relationships can become durable within minds that continue to change.