“Techno-Cautious-Optimist”
Professional Experience
Independent AI Research
Independent Research Project “TraDe” (September 2023 - Oct 2024)
Built an automated interpretability technique called Transformer Decompiler (TraDe) from the ground up.
Funded by a Research Grant from Open Philanthropy / Hasler Foundation
In Collaboration with the Pattern Recognition Research Group (University of Bern)
Employed as a Researcher (Jan-May 2024) at the University of Bern
Paper was presented at ANNPR 2024
Collaboration with Jeremy Scheurer of Appollo Research (November 2023 - July 2024)
Generated a dataset of interpretability testbeds.
Created an evaluation tool for LLMs that investigates their ability to code in a novel programming language.
Paper was presented at the ICML - Mechanistic interpretability workshop 2024
Bühler
Quarterly AI Consulting (February 2024 - Present)
Advised on the long-term strategy in and application of AI
Zetcom
AI-Intern (June - September 2022)
Set up an evaluation framework for deep learning APIs in NLP and computer vision.
Integrated the APIs in the existing code base.
Consulted on the future use of various AI-based technology.
Filmmaking
Self-employed at Lyght (2019-2023)
Produced videos for big and small companies
Advised on advertisement and communication strategy
Project manager at Video Factory (2017 - 2019)
Oversaw and implemented the production of Video Advertisements
Played a pivotal role in shaping the company's overall strategic direction.
Swiss Military
Mandatory Service (2018)
Completed the 18-week Swiss military school as a reconnaissance specialist.
Projects
AI safety fundamentals
Founded the organisation “Bern AI Safety” BAIS to facilitate six-week courses on the fundamentals of AI safety with a focus on the alignment problem.
GPT-2 Interpretability
Took part in an Appart Research Interpretability-Hackathon and won 2nd prize for examining GPT-2 to localize the circuits responsible for its ability to handle brackets. (Project Report)
Evolutionary training and hyperparameter search
Created a custom RNN from the ground up in Python and performed hyperparameter optimisation through evolutionary training (bachelor thesis).
Reddit-like student platform
Won first prize in the University of Bern software development competition with a typescript-based forum platform for students.
Adversarial Interpretability Experiments
Conducted a series of experiments which showed that transformers could learn to be more resistant to SAE-based interpretability methods. Included large hyperparameter sweep and adversarial training setups.
AutoBench
Implemented a framework to quickly and automatically produce and apply model evaluations that probe an LLMs values along an axis of the users choosing and immediately compile and visualize the results. (Blog Post)
Education
BSc in Computer Science at University of Bern September 2020 - August 2023
Minor in mathematics, philosophy and information systems.
Teaching assistant in “programming for scientists”
ARENA - Alignment Research Engineer Accelerator - September 2024
Skills & Hobbies
Programming Languages: Python, Java, Javascript/Typescript, Dart
Frameworks: TensorFlow / Keras, PyTorch, Flutter, Angular
Languages: English, German (both equally fluent)
Hobbies: Programming, Fitness, Landscape Photography, Paragliding
Interests: Philosophy, Politics, History of Technology, Data Visualisations
Why AI safety?
A few yers ago AI safety seemed like a topic only relevant in sci-fi movies. I still have a lot of sympathy for this position. Read more about it and why i now care so deeply about AI safety in this blog post i wrote in 2022:
