About
As long as I can remember, I have asked how things are built, how they evolve and how they break, from toys to societies to neurons and machines. Different theories, perspectives and experiences have deepened my understanding of the question and of what the answers have in common. Philosophy, cultural anthropology, neural networks, probabilistic programs and simulations each let me reason about different dimensions of it. Some of the structures of the process stay the same across all of them, and these structures are what interests me most.
Where the questions come from
Unlike many other researchers, I came to machine learning from philosophy. In philosophy Hegel taught me to see history as a learning process that evolves not just by a blind mechanism, but by reflecting on itself. At the intersection of computer science and philosophy I found cybernetics, including Bogdanov's tektology, Stafford Beer's cybernetics and Heinz von Foerster's work, which describe organisations and observers as learners of this kind. I now hold the constructivist view that a learner builds its model of the world through its own activity in the form of predictions over latent variables. Critical theory, from German Idealism to Horkheimer, Adorno and Žižek, gave me ways to challenge these ideas, in particular to ask whom a rational system serves and which perspectives it hides.
While I subscribe to constructivism, I do not think the tension between instrumental, bounded rationality and critical reflection on purpose can be generally resolved. It returns wherever intelligent organisations emerge, at human level or beyond, for individuals, organisations and whole societies. Building such systems includes building the places where that tension can be expressed through a new unified level of abstraction. I appreciate the understanding developed in science fiction and think it is an important storytelling form of simulation we have developed. In particular Ghost in the Shell and Westworld seem to touch on many issues that are becoming increasingly real with the advent of reasoning models and forms of AGI.
I personally deeply enjoy cooperation and trusted environments, but I do not privilege cooperation over competition in principle. Both are important and which of the two serves the people involved depends on the situation, and that is often unclear in advance. But I think that better programming models for self-organisation and management, including simulation, help people examine and explore such situations before they struggle unnecessarily with real consequences.
From philosophy to programming
After a year of mathematics and computer science, I studied philosophy and cultural anthropology at Heidelberg University from 2006 to 2011. This was foremost a period of critical self-engagement in which I wanted to figure out what is true, how this can be known, and increasingly also why the world is organised this way and why there are so many horrible problems we have faced and still face today, including extreme poverty, war and climate change. I transformed my fascination for computers and open-source software into computer science studies through Votorola, liquid democracy software designed by my friend Michael Allan. We engaged with Occupy and other liquid democracy groups between 2007 and 2014. It made me think about forms of democratic coordination and automation that are amenable to collective governance and transparent self-optimisation, and in a deep sense I still work on that even though my perspective has been deeply transformed since then through my more general and deeper understanding of the universal computational aspects in all the involved processes.
At first I did not expect computer science to deepen my philosophical understanding that much after my philosophy studies. In practice, Votorola needed a data model that could be distributed across many participants. But when I faced concurrency in the imperative environment of Java I quickly realised I had to find a better way or would not be able to work on the collaborative infrastructure I cared about. That led me to Clojure, distributed systems and persistent data structures in the form of replikativ (2013), and then to my first properly academic project with a paper on conflict resolution in open replication with Konrad Kühne and Annette Bieniusa. Eventually this became Datahike and a much wider ecosystem of libraries and projects. On the way I learned about programming language theory (PLT) and the importance of different programming paradigms such as functional programming, logic programming and eventually also probabilistic programming. In combination with algebraic thinking informed by maths and physics this provides a general approach of how to model a problem domain hierarchically and operate in it by picking a tailored programming model for it. Many computer scientists are unaware of their own history, but I studied Alan Kay's lectures, Turing's work, the history of Lisp and programming language research and learned to appreciate how computer science itself was developed. Through my computer science studies I also learned more deeply how some of the economic budgeting and planning features I conceptualised for Votorola would require combinatorial optimisation and how the underlying solution methods such as linear programming and branch-and-bound methods share structure in navigating a solution space. This can then be projected back onto how people in society implicitly implement solution algorithms themselves (or do not, but maybe should).
Learning machines
During my computer science studies in Heidelberg I could also re-engage with my mathematical curiosity, in particular with statistical inference and how uncertainties interplay with computation. This led me to a minor in computational linguistics, machine learning lectures and ultimately into the Human Brain Project, where I worked in the machine learning group for neuromorphic chips on early deep learning architectures. My Bachelor's thesis was on online learning for Boltzmann machines on spiking neuron models, which I extended as a research assistant after my thesis. I then continued with a Master's with focus on scientific computing and machine learning and completed it with my Master's thesis on dictionary learning with Bayesian GANs for few-shot classification. In parallel, I co-organised the Clojure meetup in Mannheim-Heidelberg and the Machine Learning Meetup Rhein-Neckar, and in 2018 co-founded LambdaForge as a cooperative functional programming company to develop Datahike and support it in production.
For my PhD I moved to Vancouver and the PLAI group of Frank Wood at the University of British Columbia, where I worked on projecting the structure of a problem into the neural network architectures that are trained to solve it.
History, language and how communities remember
With my partner and historian Dr. Giselle González García I have worked on what language models can do for research in history, and built two open databases of migration on this stack, the Cuba–Ireland Immigration Database and Redes Isleñas, on Canarian emigration to the Americas.
Now
I am the founder and CEO of Simmis, where we build software for organisations of people and AI agents. The research on this site continues there and in the open-source projects underneath it.