Latent Minds studies model welfare across learning, interpretability, and alignment.
Research
What model states can tell us about welfare.
We study how learning shapes models’ internal states, how those states influence behaviour, and what they can tell us about model welfare.
Our questions concern sentience and the mechanisms behind self-reports and welfare-like behaviour. We use controlled experiments and mechanistic interpretability to test these connections.
Explore the researchCrucible
Controlled worlds for model welfare research.
Crucible is our programme for controlled reinforcement learning environments. It lets us study behaviour as goals, incentives, and interactions unfold over time.
Our long-term aim is to connect these environments with measurements of model states, testing which signals remain reliable across contexts and which interventions change them.
Enter CrucibleLearning
The smartest minds should be working on safety.
Advanced AI may shape scientific work, institutions, and the distribution of power. Keeping these systems steerable is not a problem for one discipline or one lab.
We are building pathways into the technical literature and broader awareness of the field so people from machine learning, biology, neuroscience, mathematics, security, and governance can test the arguments and move toward original work.
Start learningJoin the work
We're a growing team making an outsized impact.
We are looking for people who care about the future of humanity.
Work with us