Mindscapes: World Building for AI Safety

Project summary

Mindscapes is an interdisciplinary research group based in Oxford, building rich worlds and characters for understanding and shaping the psyches of LLMs.

What are this project's goals? How will you achieve them?

The primary aim of Mindscapes is to develop a science of LLM psychology, character training and multi-agent interaction dynamics from the ground up. In particular, I am interested in:

a) crafting characters with deeply internalised morality (as opposed to reward sycophants)
b) character training methods based on self-distillation that can scale in a robust way alongside capabilities RL
c) creating rich, diverse simulated worlds using fine-tuned base models for training and observing the behaviour of different LLM characters
d) qualitative, non-prescriptive character evaluations that are incentive-compatible even under the assumption of eval awareness

I will be assembling a team drawing from both technical and non-technical fields such as experimental psychology, anthropology, literature, philosophy, geography etc. We will be designing character and worlds, observing these characters in diverse environments, and using these observations as the basis for generating testable theory. The project will be mostly part time until 5 December (Oxford end of term) and a smaller group will be selected to work full-time during the holidays.

How will this funding be used?

Minimal funding:

Compute + API costs: $26,000
Recruitment + outreach: $4,000

At this level, we would focus on post-training 7-14B models, delivering research blog posts and an open-source library and guide for character training and evaluations.

At $60k, we could scale up to multiple runs on llama 70B base and larger datasets, begin investigating more complex multi-agent dynamics.

At $120k, we could do experiments in fine-tuning large MOE models such as Kimi and Deepseek via Tinker, and scale up the number of fellows. We could also pay research managers and support with travel and accommodation for those moving to Oxford.

Who is on your team? What's your track record on similar projects?

I (Samuel Ratnam) am a Computer Science and Philosophy undergraduate at Hertford College, Oxford and will be leading this project. Previously, I have worked at Workshop Labs (https://www.workshoplabs.ai/) on model personalisation and on Alignment Pretraining (https://arxiv.org/abs/2601.10160) in collaboration with Geodesic Research and UK AISI. I also co-founded and help run the Idealists Collective (https://idealistscollective.org/) - a community of over 150 philosophers, artists and technologists for imagining and creating better futures. I am also mentoring SPAR projects on generalisation engineering and model welfare.

Advisory Board:
- Anders Sandberg (Institute for Future Studies, formerly FHI)
- Jai Dhyani (Separatrix)
- Arthur Roing Baer (Moving Castles)
- Celeste (computer-10)
- David Africa (Resolution)
- Antra Tessera (Anima Labs)

Note: We also have an application for a $100k seed grant in progress from bluedot impact