I work on cultural evolution online, and on what ranking functions
do to it. The methods are complex systems and computational social science. The
normative side is grounded in democratic theory. What a ranking function does to a
population leads straight to what it ought to do.
The constructs I measure are discourse quality, media framing, the structure of
belief in a population, and the effect of a ranking function on the people under
it. The field argues about these and rarely operationalises them. So I write the
codebooks, run the annotation, and report the agreement.
The same instruments point at language models. Part of the work is on models
directly: when theory of mind appears over training, and how they handle generics
and default reasoning. Part is on populations of them, as a model system for
cultural evolution and as a platform on which we vary the recommender.
Current
Measuring the quality of political discourse on Reddit
2026
- Three democratic theories that disagree: Habermas on deliberation, Mouffe on conflict, Young on inclusion.
- A 56-variable codebook the theory constrains.
- About 75,000 comments across 653 subreddits.
- Annotated by language models, validated by seven coders.
How news frames the victims of conflict
2026
- Who is named, who is counted, and which perpetrators are left implicit.
- Do those choices track the event's severity or the outlet's geopolitics?
- About 1.36 billion articles over ten years, in many languages.
- Most of the difficulty is matching articles to events.
Does false-belief reasoning emerge the way it does in children?
2026
- Children acquire it along a consistent curve. Nobody has charted it in a language model.
- 41 checkpoints across the training run of an open model.
- Scored by contrasting teacher-forced log probabilities.
- Predictions registered before the runs.
Simulating social media with language model agents
2026
- Agents post, read, and respond under a recommender we control.
- Vary the recommender, measure the discourse.
- The open question: which results come from the agents, not the ranking.
Sandboxing cultural evolution with LLMs
2026
- A model system for cumulative culture: design no individual worked out.
- Language model agents on a hidden fitness landscape with a known optimum.
- A detection battery: accumulated through transmission, or worked out alone?
- Freeze-and-branch replay that cuts the peer channel mid-run.
Belief networks, and how a population's attitudes hold together
2025–2026
- Attitudes as a network: beliefs are nodes, correlations are edges.
- Decades of the General Social Survey.
- The claim: that structure is the environment any new belief has to fit.
- Liberals and conservatives appear to differ in how their beliefs connect.
Pulling knowledge graphs out of text
2026
- Can a model turn an ontology and text into a graph you'd trust?
- Extraction, then coverage checking, then entity normalisation.
- Benchmarked on Text2KGBench and CS-KG-3600.
Convening
What platforms are for
14–16 Dec 2026
Earlier
Timid walks and prudent walks
honours work
Swinburne
- Self-avoiding walks: lattice paths that never cross themselves.
- Timid and prudent walks: restricted variants that trade generality for tractability.
- My honours year, supervised by Nathan Clisby.
- Where I learned to do research.