We measure whether frontier models can detect control interventions. Models that detect such interventions could infer properties of the monitor and adapt to evade oversight.
Joachim Schaeffer, PhD
AI Safety Researcher.
Builder. Founder.
I'm Joachim, a MATS Research Fellow working on AI control and monitoring. I previously worked on ML to make Lithium-Ion Batteries safer during my PhD. I've co-founded and led projects, scaled teams, and published research.
// selected research — AI Safety Publications, full list on Google Scholar
Misaligned AIs could strategically select attacks to evade monitors. We show attack selection degrades safety from 99% to 59%, highlighting the need for Red Team elicitation.
All LayerNorm layers can be removed from GPT-2 models with minimal loss increase, showing LN plays a limited role at inference. The resulting LN-free models enable more precise mechanistic interpretability by eliminating nonlinearities.
Recursive spatiotemporal Gaussian processes applied to battery field data for early fault detection. We published software and data to accelerate research and avoid battery fires.
// built — Projects and Ventures
Co-Founder · 2019–21
Robotics + AI to remove plastic waste from rivers. Co-founded the project and scaled the team to 18 people at ETH Zurich. Secured 100k CHF. National media coverage.
Advisor · 2023–24
AI advisory for AI for battery production startup. I supported the team for 1.5 years. The company shut down due to the downturn of the European Battery production ecosystem.
Open Source
Gaussian process-based health monitoring and fault analysis software for Lithium-ion battery field data.
Organizer · 2021–22
Global battery + ML hackathon. Led a remote team of five organizers. Raised 16k EUR in corporate sponsorships.