What you'll do
We prove the concept of an automated pipeline for reimplementing social sciences models with a mix of human and agentic AI-derived parameters, giving insight into gradual disempowerment. We'll collaborate with domain experts to for model selection and analysis. We'll run the first simulation of gradual disempowerment with agent behaviour based observed AI behaviour on a large multi-agent system.
Current status
Stage 1 - Proof of concept.
- Empirically grounded GD modelling - replicating a parameter measurement pipeline Amin et al. (2026) applied to MoltBook data. We'll validate the method in an OASIS resimulation and use the method to simulate a mixed human-AI system, for studying AI influence at varying proportions of humans and AI. This is not an economic model since we've pivoted the overall programme since starting, but is sufficient for a POC in empirically-grounded agent-based modelling (common in economics).
- Automated implementation and execution - adapting and improving ReplicatorAgent to automatically implement a model from Korinek and Suh (2024)'s analytic economic modelling of gradual disempowerment.
- Collaboration with domain experts - discussing collaborations with computational social sciences and AI departments at several universities.
Remaining work in Stage 1, not started:
- Model-parameter matching tool - a method and tool which, given a model specification, provides a valid parameter measurement method to extract parameters from some dataset.
Stage 2 - Scaling the POCs.
If Stage 1 shows that the full automation pipeline is feasible, we'll proceed. Otherwise, we'll focus on the biggest blocker.
Who's involved
Stephen Elliott - Founder, Research and Operations Lead. CS/Economics graduate, UNSW Sydney. Started Gigascale Labs during the Sydney AI Safety Fellowship 2026 to model large systems in the AGI transition. Recruited 6 volunteers for pilot studies. Previously co-founded an events aggregator startup and ran an AI study group.
Shruti Tamilselvan - Member of Technical Staff. University of Sydney Masters student, Data Analytics. Built scalable data pipelines at USyd and conducted market research at IFZA Dubai. Stakeholder management experience as Regional Lead Oceania, Principles for Responsible Management Education.
Juan David Lopez - Member of Technical Staff. SEng student at USyd. Projects in ML and low-level programming. Achievements include Vice Chancellor's International Scholarship, Dean's List, Dalyell Scholarship.
Dr Deborah Bunker - Project mentor. Professor Emeritus of Systems and Information, USyd Business School. Formerly Chief Science Officer, Natural Hazards Research Australia (until June 2024). Chair, National Committee for Information and Communications Sciences (NCICS), Australian Academy of Science.
Berkeley Existential Risk Initiative - fiscal sponsor, TBC once funding confirmed.
The output (Stage 1)
1.1 Empirically grounded GD modelling (underway):
- Finish the replication of Amin et al..
- Publish it in an OSS library and blog.
- Validate the method with an OASIS resimulation.
- Preprint + conference submission on the varying human/AI influence experiment described in "Current Status".
1.2 Automated implementation and execution (underway):
- Benchmarking ReplicatorAgent on Korinek and Su - see "Current Status" (complete).
- Improving ReplicatorAgent to successfully implement Korinek and Su (complete).
- PR to the ReplicatorBench repo if our method is SOTA thereupon.
- Blog on the above.
1.3 Collaboration with domain experts (underway):
1.4 Model-parameter matching tool (not started):
- Schema for matching LLM parameter measurement methods to given models.
- Implementation on a small benchmark pool of models and parameter extraction methods.
- Preprint describing the method, submitted to a conference.
Stage 2: TBD based on Stage 1 outcomes.
