Christian Weilbach

Christian Weilbach

I study how learning systems can improve themselves and stay correct. I am most interested in the organisations around them, since the structures that check a learner have to learn and grow together with it.

I did my PhD in machine learning at the University of British Columbia with Frank Wood, on how the structure of a problem can be translated into the neural networks that solve it. Before that I wrote my Bachelor's and Master's theses at Heidelberg University, on online learning in spiking neuron models and on dictionary learning with Bayesian GANs, and studied philosophy and cultural anthropology. I am now the CEO of Simmis, where we build software for organisations of people and AI agents. The open-source stack underneath it covers versioned memory, forkable runtimes, probabilistic inference and a language whose kernel checks its own optimisations. More about me.

Projects and papers since 2006

The figure shows what I have worked on and which project grew out of which. Some branches ended, and a few came back years later. Scroll to replay the years.

My research as a family tree of ideas, 2008 to 2026 Organisations and agents Inference and simulation Learning structure Languages, runtimes and proofs Memory 2008201320162019202220252026mid 2026
Each line is an idea or project that grew out of earlier ones. Ochre dots are papers, squares are companies, diamonds are talks. Grey branches ended; dashed lines are ideas that lay dormant for years and came back. Hover for names, select for code and writing; a small dot above marks projects with writing. A slow pulse marks projects whose repository changed in the last two months, stronger for more recent changes.

How the projects relate

My projects form a stack, every level supports the same operation. A person, an agent or a program makes an attempt on a fork and compares the result with evidence. The attempt is adopted only after a check. All levels share one versioned memory, so forking is cheap.

  1. What should we become, and who decides?

    • Simmis workspace, branches and review
    • Kontor double-entry accounts
  2. On whose behalf does this work happen?

    • Dvergr rooms, attempts in forks, review
  3. What is worth knowing?

  4. What counts as the same?

    • Klara versioned worlds, staging, checked admission
    • Ansatz Lean 4 compatible kernel
    • Wandler checked rewrites
    • Raster numerics to devices
  5. What may run, and at whose cost?

  6. What must not be forgotten?

Predictions run down the levels (blue) and evidence runs up (ochre). The dashed line stands for the question of what each level is for, which comes up again at every level.

Current work

  • Klara A language in which a running program keeps earlier versions of itself and tries optimisations on a fork. A small kernel checks each optimisation before the program adopts it. Research prototype.
  • Foerster Probabilistic inference that forks a running simulation instead of re-running it. Beta.
  • Datahike A Datalog database that keeps every past state queryable. A branch shares all unchanged storage. Stratum, Proximum and Scriptum apply the same model to columnar SQL, vector search and full text. Datahike is used in production and close to 1.0.
  • Raster Numerical computing for the JVM that compiles the same code for different devices. It runs the Stuttgart simulation, and a LeNet-5 training step on CPU ran 1.7 to 2.4 times faster than JAX in our measurements. Beta on the JVM, alpha on GPUs.
  • Simmis A workspace in which people and AI agents share one memory. Work happens on branches and joins the shared state after review. Early access.

I am looking for people who want to work on self-improving systems, in programming languages, probabilistic inference or systems engineering. If that is you, write to me at christian@weilbach.name or book a 30-minute call.

Writing

  • Who checks a system that improves itself?A system that improves itself has to predict its own future, think through changes on forks and decide which ones to keep, while what it values changes with it. A programming model for this, built across a full stack from versioned memory to organisations of people and AI agents.

Selected papers

All research, with figures and code

Talks

  • Rationalizing a Society of Learners Max Planck Institute for Intelligent Systems, Tübingen, October 2026.
  • Programming as and for inference London Clojurians, May 2026.
  • Partially evaluating higher-order probabilistic programs without stochastic recursion to graphical models LAFI workshop at POPL, 2025.
  • Flexible diffusion models for videos, molecules, and inference in structured simulators Google DeepMind, London, 2024.
  • Video and graphically structured diffusion models OpenAI, San Francisco, 2023.
  • Graphically structured diffusion models Mila, Montréal, 2023.
  • Anglican: Turing-complete functional probabilistic programming as a domain-specific language Trends in Functional Programming, 2019.
  • Datalog as a distributed systems language Clojure Meetup Vancouver, 2019.
  • Let's replicate state together clojureD, Berlin, 2017.
  • Machine learning workshop CubaConf, Havana, 2017.