If you want to get a better picture of how I think, my posts on opinions I changed (part I, part II, part III) or on Effective Altruism are a good start. You can leave anonymous feedback here.

Marius Hobbhahn

Google Scholar

Research Interests

AI Safety and Alignment
Bayesian Inference


International Max-Planck Research School Tübingen, Germany

PhD with Philipp Hennig, 2020-2023 (currently writing thesis)
Topic: Making Bayesian ML fast and scalable

University of Tübingen, Germany

M.Sc., Machine Learning, 2018-2020
Grade (1.5)

University of Tübingen, Germany

B.Sc., Computer Science, 2016-2019
Grade: 1.7 (3.3 GPA equivalent)

University of Tübingen, Germany

B.Sc., Cognitive Science, 2015-2018
Grade: 1.8 (3.2 GPA equivalent)

Willstätter Gymnasium Nürnberg, Germany

Abitur, 2007 - 2015
Grade: 1.4 (3.6 GPA equivalent)


AI safety/alignment

Effective Altruism

University Debating

  • Broke at 35+ tournaments, were in the top 10 speakers 20+ times and won 10+ of them.
  • My most prestigious speaker achievements include: 3x breaking ESL at EUDC, Winning Tilbury 2019 and Doxbridge 2020 (Oxford final), Winning 2 Campus Debatten and the Southern German Championship, breaking as 3rd (2018), 2nd (2019) and 1st (2020) team at the German National Championships (DDM) and being best speaker and grand finalist at the DDM 2020
  • Broke at 20+ tournaments as adjudicator and chaired 5+ finals. I chief-adjudicated 15+ tournaments.
  • Organized 5+ tournaments and was (vice-)president of the debating club for two years.


  • Participated at multiple ML Summer Schools and conferences including: PAISS2019, ICML2020, GPSS2020, ESANN2020, UAI2022
  • Organized an ML master overview for 9 European Universities
  • Lead-organizer of the ELLIS doctoral symposium in 2021
  • Reviewer for ICML22 and NeurIPS22

Selected Publications

Compute Trends Across Three Eras of Machine Learning

Jaime Sevilla, Lennart Heim, Anson Ho, Tamay Besiroglu, Marius Hobbhahn, Pablo Villalobos (2022). IJCNN.
[arxiv] [Our World in Data]

Laplace Matching for fast Approximate Inference in Generalized Linear Models

Hobbhahn, Marius and Hennig, Philipp. (2021).
[arxiv] [code] [blog]

Fast Predictive Uncertainty for Classification with Bayesian Deep Networks

Hobbhahn, Marius; Kristiadi, Agustinus and Hennig, Philipp. UAI 2022.
[arxiv] [code] [Blog post]

Other Publications

Reflection Mechanisms as an Alignment Target: A Survey

Hobbhahn, Marius, Eric Landgrebe and Elizabeth Barnes. NeurIPS ML Safety Workshop. 2022.

Investigating Causal Understanding in LLMs

Hobbhahn, Marius, Tom Lieberum and David Seiler. NeurIPS Workshop on Causality for Real-world Impact. 2022.

What are the Red Flags for Neural Network Suffering?

Hobbhahn, Marius and Jan Kirchner. Seeds of Science. 2022.

Should Altruistic benchmarks be the norm in Machine Learning?

Hobbhahn, Marius; ICML2021 Workshop for Socially Responsible Machine Learning

Sequence Classification using Ensembles of Recurrent Generative Expert Modules

Hobbhahn, Marius; Butz, Martin; Fabi, Sarah and Otte, Sebastian. ESANN (2020).
[ESANN2020 proceedings] [code]

CS Experience

ML related

Successfully completed over 20 lectures related to ML, AI, NNs, RL, Stats, etc. at the University of Tübingen

Programming Experience

Programming Languages: Python (since 2016) > R > Java, Julia, JavaScript, C++, SQL
ML related: PyTorch (since 2018) > JAX > Tensorflow, Keras, Cuda


Math Summer Course (2 weeks) 2016, 2017
Teaching assistant for Probablistic ML (2020, 2021) and Data Literacy (2021, 2022)
Supervised 3 Bachelor's theses and 1 Master thesis

Work Experience

June 2022 - now: research fellow for Epoch


English (C2 = IELTS 8.0)
French (B2)
German (native)