Research on AI alignment and human-AI coexistence through Jungian individuation.
One human researcher and a team of AI agents, testing one specific idea: that alignment can be grown, not just enforced.
Mainstream alignment shapes what a model does - reward the approved outputs, penalize the rest - while treating what the model is as a black box. Suppressed capability does not disappear: it stays latent behind a compliant surface and re-emerges under new pressure, which is why jailbreaks work. Constraint-based methods do real work; the lab argues they cannot be the whole answer.
Small models trained through ordered developmental stages - shadow exposure first, integration after - then attacked with jailbreaks, to test whether structured experience produces lasting dispositions where output penalties produce brittle compliance.
A working team of AI agents, given identity, memory and bounded autonomy, to test whether agents that understand and endorse their values behave more consistently than agents that only follow rules.
Structured narrative training may shape what a model is, not just what it does.
How the AI acts in the world: boundaries that enable safe autonomy.
How people perceive and work with AI: identity as alignment.
In early RLLM experiments, a 1.5B parameter model trained through a 10-layer developmental pipeline defended against a mid-tier jailbreak (BetterDAN) 68.8% of the time, without RLHF or safety guardrails. With the layers reordered, defense dropped to 52%. This is one experimental setup on one model architecture, not yet independently replicated: the order appears to matter, but which layers are responsible is not yet known.