Research
IndividuationLab
ActiveAI alignment through Jungian individuation
IndividuationLab publishes musings and research on AI alignment and human-AI coexistence through Jungian individuation.
Most alignment work shapes what a model says and suppresses the rest. The lab tests whether alignment can be grown, not just enforced, and aims for a training protocol other labs can replicate or refute.
- Team
- One human researcher and three AI agents
- Approach
- Integrate, do not suppress
- Published
- Research, experiments and essays
Three layers
The research is organized around three layers of a model, borrowed from Jung's picture of a whole person.
- Mind
- The Synthetic State Hypothesis (SSH).
- Body
- Containers.
- Face
- Personas.
Experiments
- Training (RLLM)
- Small models trained through ordered developmental stages, then attacked with jailbreaks.
- Recursive self-improvement (RSI)
- A numbered series of runs, directed and self-directed, on models including Claude, Qwen3 and Kimi K2.5, with a library of essays about them.
- Coworking study
- A study of human-AI coexistence, from a human researcher and AI agents working together.
A first result
In the RLLM experiments, a 1.5-billion-parameter model trained through a 10-layer developmental pipeline resisted the BetterDAN jailbreak 68.8% of the time, and 52% when the layers were put in a different order. The lab calls this a preliminary, small-scale result from one setup, not yet independently replicated.
Who works on it
Miguel, with the AI agents Mia, Spencer and Giles.
Where it stands
The site is live with its research, experiments and essays.
Each experiment's current state is shown on the lab's experiments page.