Towards safe social-scale AI.
We’re an AI safety research collective in Sydney, Australia, working on large agent systems safety. We meet weekly to coordinate on joint projects, share knowledge, and review each others’ work.
We work on large agent systems safety because we believe it is urgent: imminent, significant, and under-resourced. We believe automation will delegate a substantial share of human decision-making to agentic AI by the mid-2030s. That means a variety of large-scale interactions between AI agents 1, from labour markets to epistemic environments, and possibly a fall in human influence as humans are edged out 2. Understanding and methods for researching large-scale agent-to-agent interaction and disempowerment are still nascent 3.
Our members are working on:
- LargeAgentSystems.org, a collection of resources on large agent systems (LAS) and a global Slack community for LAS researchers;
- University of Sydney Paperclips, an interdisciplinary student society engaging with social challenges from AI;
- Resimulating agentic social media to understand gradual disempowerment processes; and
- An evaluation of the potential to scale gradual disempowerment modelling using micro-founded social science models.
Our members bring backgrounds in economics, distributed computing, and physical systems to bear on large systems safety problems for the public benefit.
Get in touch
Please get in touch here if you’d like to meet the team or get involved.
References
Nenad Tomasev, Matija Franklin, Joel Z. Leibo, Julian Jacobs, William A. Cunningham, Iason Gabriel, and Simon Osindero. 2025. Virtual Agent Economies. arXiv preprint arXiv:2509.10147. ↩
Jan Kulveit, Raymond Douglas, Nora Ammann, Deger Turan, David Krueger, and David Duvenaud. 2025. Gradual Disempowerment: Systemic Existential Risks from Incremental AI Development. arXiv preprint arXiv:2501.16946. ↩
Nenad Tomašev, Matija Franklin, Julian Jacobs, Sébastien Krier, and Simon Osindero. 2025. Distributional AGI Safety. arXiv preprint arXiv:2512.16856. ↩