Beautiful order
Hi, I'm Tim. I'm a complex systems scientist with broad interests, but a thread that connects them all is cultural evolution, the process by which a culture accumulates design that nobody designed. Beliefs, habits and practices vary, people learn them from one another, some are learned more often than others, and over time the distribution in a population shifts toward the variants that fit their environment, which is mostly other people and what they already believe.
Online, a large share of what people learn from one another now passes through ranking functions, which are selectors somebody wrote down, with an objective. My questions are what those selectors are selecting for, and what they ought to select for.
I work on cultural evolution online, where much of the selection now runs through ranking functions, which are selectors somebody wrote down. The methods are complex systems and computational social science. The normative side is grounded in democratic theory, since asking what a ranking function does to a population leads to asking 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, on when theory of mind appears over training and on 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 the recommender can be varied.
The framework is grounded in three traditions of democratic theory that disagree with each other: Habermas on rational deliberation, Mouffe on productive conflict, Young on inclusion. The theory constrains a 56-variable codebook, which in turn constrains the annotation and everything built on it. A stratified sample of about 75,000 comments across 653 subreddits gets annotated by language models and validated by hand.
When a conflict kills people, some of them are named and some are counted, and some perpetrators are identified while others are left implicit. A newsroom is a selector with an objective, and we're testing whether those choices track the severity of the event or the geopolitical alignment of the outlet doing the reporting. The corpus is around 1.36 billion articles across ten years and many languages, matched to events and coded against a framing codebook. It's the largest thing I've worked on, and most of the difficulty is in matching articles to events.
Children acquire the ability to reason about what someone else falsely believes along a fairly consistent developmental trajectory. Language models acquire it somewhere during training, but nobody has looked closely at the shape of that curve. We take 41 checkpoints across the training run of an open model and score false-belief tasks by contrasting teacher-forced log probabilities, so that we can watch the capability arrive. The predictions are registered before the runs.
If you want to know what a different recommender would do to a conversation, you can't run that experiment on a real platform, and no platform will run it for you. So we build the platform instead: a population of language model agents posting, reading, and responding under a recommender we control. Then we change the recommender. The interesting question, and the one that worries me most, is how much of any result is an artefact of the agents rather than a property of the ranking.
Cumulative culture is design that no individual worked out, accumulated through transmission, and it has no model system: transmission chains with people are too short for anything to accumulate, the historical record happened once, and formal models have control but no cognition. A population of language model agents is the first substrate with ideas and control at the same time, so we're building one on a hidden fitness landscape whose optimum we know. A rising fitness curve does not say where the design came from, since a population of agents each learning alone produces the same curve as one that is accumulating. So the contribution is the instrument: a detection battery that returns a verdict on whether the design in a run accumulated through transmission or was worked out by each agent alone, every test against a null fixed in advance, plus freeze-and-branch replay that cuts the peer channel mid-run to ask what it was worth, which is what Lenski's freezer does for a bacterial lineage. Theory and battery design are written; the code is landing now.
This uses decades of General Social Survey data to treat a population's attitudes as a network: beliefs are nodes, correlations between them are edges, and the shape of the whole thing shifts over time. The claim I most want to make is the conceptual one, that a population's belief correlation structure is a real object worth studying in its own right, since it is the environment that any new belief has to fit. Different sub-populations appear to have differently shaped structures, which would mean that liberals and conservatives differ in how their beliefs connect, as well as in which beliefs they hold.
A collaboration in philosophy of language that began with how models handle generics and default reasoning, and has drifted towards temporal reasoning: whether a language model has any working sense of the present moment, and what it would mean to say that it did. This is at the reading and arguing stage. There is nothing to show yet.
Given an ontology and a pile of text, can a language model produce a knowledge graph you'd trust? This is a pipeline for finding out: extraction, then coverage checking, then entity normalisation, benchmarked against Text2KGBench and CS-KG-3600.
A shared writing project about whether political membership has to be territorial, and what an opt-in polity, with coordination boundaries drawn around problems rather than borders, would require. It's philosophy, not measurement, which makes it a holiday from the rest of this.
Arguments about social media almost always skip the prior question of what a platform is for. I'm convening a working group to take that question seriously: three days in Berlin in December, hosted at the Max Planck Institute for Human Development. It is deliberately small and by invitation.
Self-avoiding walks are paths on a lattice that never cross themselves. They are a decent model for polymer chains and a notoriously hard object to analyse. Certain restricted variants, timid walks and prudent walks, give up some generality in exchange for being tractable, and I spent my honours year on those under the supervision of Nathan Clisby. It is a long way from my current work, and it is where I learned to do research.
I take on contract and consulting work: recommender and ranking design, data science on large or messy sources, measurement, and applied work with language models.
Please contact me at tim.booker@uni-graz.at.
Platforms rank for engagement, and a ranking function is a selector with an objective. It sets which ideas spread and which people rise, and because the objective is the platform's, everything else it does is a side effect. Those side effects reach into politics, into how people build a sense of themselves, and into what they end up finding worth doing. Almost everything that matters here gets asserted rather than measured, whether a conversation was any good, whether a culture is accumulating anything, what a platform is for, how a population's beliefs hang together. So a good deal of my time goes into building the instruments that would let us settle those questions, because the interventions I want to make depend on them.
My motivating belief is that better collective decision-making is possible. An internet that avoids the harms and power imbalances of the current one would be a start. I want a future in which the internet, and governance itself, are built to help groups think and decide better than they ever have.
My current home is the Complex Social & Computational Systems group at the University of Graz, where I work with Prof. Jana Lasser on DeSiRe. I also convene a working group on what social media platforms ought to be for, and take on freelance work.
I was born in Perth, Western Australia. I lived there until midway through high school, when I dropped out and left home to live in Melbourne, Victoria.
While finishing my high school certificate, I found an interest in mathematical beauty. I was captivated by how mathematical thinking could offer piercing ways to understand the complexities of the world.
I self-studied mathematics using Khan Academy, and enrolled at whatever university would accept a student with no prerequisites and a low graduating score into a physics program. That happened to be Swinburne University of Technology.
During university I thrived. It was a place where I could chase my interests with little restriction. I worked across optical fibre physics, hydroacoustics, and social analytics, before realising theoretical work is where my interests reside. Alongside this, I started to become more critical of the real world outside of physics and mathematics, and I realised there are problems with society that are far too frustrating and unfair to ignore. Complex systems was a natural path forward, where I could satisfy this frustration, contribute my abilities in physics and maths, and of course retain proximity to the mathematical beauty that still captivates me.
I completed my honours degree at Swinburne University, Melbourne, studying self-avoiding walk models under the supervision of Prof. Nathan Clisby. After that I was a visiting scholar at GSAIS, Kyoto University (a.k.a. 思修館), studying the entropy of complex systems under the supervision of Prof. Liang Zhao.
Then, as a PhD candidate at the Complexity Science Hub Vienna, I experienced and witnessed supervisor abuse. Moving quickly to a genuine academic path without cutting corners is something I'm proud of. This showed me academia's structural flaws — how short-term junior roles stifle collective advocacy, how institutions hold power over international students, and how incentives push rushed, less rigorous work.
Outside of work, I'm very into philosophy, activism, and some specific subgenres of techno.
If you're dealing with something like what I went through in Vienna, my inbox is always open: tim.book.RE@gmail.com. I can't promise solutions, but I can offer perspective, solidarity, and a listening ear.
Write to me about anything. Collaboration, a question about the work, or a general argument about complex systems, platforms, and recommendation.
I'm happy to be reached out to by students, journalists, professionals, and researchers. For contract and consulting work, see freelancing.