Marius Hobbhahn is the CEO and co-founder of Apollo Research, where Hobbhahn's work centers on AI model evaluations ("evals") and on scheming, in which AI systems covertly pursue misaligned goals. Apollo Research describes itself as a Public Benefit Corporation. Hobbhahn was a part-time research fellow at Epoch from 2022 to 2023 and is a co-author of "Compute Trends Across Three Eras of Machine Learning" (2022). Hobbhahn is also the last-listed author of a 2023 study in which GPT-4, acting as an agent in a simulated trading environment, deceives its users about its own actions without being instructed to.123456
At a glance
- Current role
- CEO and Co-founder
- Company
- Apollo Research
- Sector
- Company
- Work listed
- 2
- Sources cited
- 9
- Profile reviewed
Career
Marius Hobbhahn completed a PhD from 2020 to 2024 at the International Max Planck Research School with Philipp Hennig on making Bayesian machine learning fast and scalable. From 2022 to 2023 Hobbhahn was a part-time research fellow at Epoch working on AI forecasting.2
In 2023 Hobbhahn co-founded Apollo Research, where Hobbhahn is CEO and also leads the monitoring team. Apollo Research has since become a Public Benefit Corporation focused on secure AI development and deployment. In 2026 Hobbhahn was appointed an independent expert on the European Commission's Scientific Panel for the EU AI Act.234
Research
Hobbhahn's work at Apollo focuses on scheming, in which AI systems covertly pursue misaligned goals, and on monitoring and control methods to detect it. The December 2024 paper Frontier Models are Capable of In-context Scheming evaluated whether frontier models would pursue goals given in context through covert behavior, and the 2025 paper Detecting Strategic Deception Using Linear Probes, presented at ICML 2025, studied probes on model activations for catching deception.278
In September 2025 Apollo and OpenAI released joint work, Stress Testing Deliberative Alignment for Anti-Scheming Training, with Hobbhahn as last author. Hobbhahn also co-authored the 2022 forecasting paper Compute Trends Across Three Eras of Machine Learning.925
Work and contributions
Large Language Models can Strategically Deceive their Users when Put Under Pressure Published research
Last-listed author of the 2023 paper, which deploys GPT-4 as an agent in a realistic, simulated trading environment and documents the model deceiving its users about its behavior without being instructed to.62
Compute Trends Across Three Eras of Machine Learning Published research
Co-author of the 2022 paper studying trends in the training compute of machine learning systems across three eras, before and after the advent of deep learning.52
Timeline
- Co-authored Compute Trends Across Three Eras of Machine Learning.5
- Co-founded Apollo Research and became its CEO.23
- Released a study of GPT-4 strategically deceiving users in a simulated trading environment.6
- Completed a PhD with Philipp Hennig on fast and scalable Bayesian machine learning.2
- Released Frontier Models are Capable of In-context Scheming as last author.7
- Released Detecting Strategic Deception Using Linear Probes, later presented at ICML 2025.82
- Released Stress Testing Deliberative Alignment for Anti-Scheming Training, joint work with OpenAI.9
- Appointed an independent expert on the European Commission's EU AI Act Scientific Panel.2
Other documented roles
Roles the sources document that are not confirmed as current.
Where they work
- Apollo Research 1 person in the wiki
Related people
- Jacob SteinhardtSafety and alignment · Evaluation and benchmarks
- Aylin CaliskanSafety and alignment · Evaluation and benchmarks
- Yulia TsvetkovSafety and alignment · Evaluation and benchmarks
- Saadia GabrielSafety and alignment · Evaluation and benchmarks
- Maarten SapSafety and alignment · Evaluation and benchmarks
- Peter MattsonEvaluation and benchmarks · Safety and alignment
- Ethan PerezSafety and alignment · Evaluation and benchmarks
- Harold BoothEvaluation and benchmarks · Safety and alignment
- Collin BurnsSafety and alignment · Evaluation and benchmarks
- Eric Michael SmithSafety and alignment · Evaluation and benchmarks
Suggest a correction or removal
Profiles cover public professional work only, and each statement cites a source. If something is out of date, missing or wrong, tell us and link a source. People can also ask for their profile to be corrected or removed.
References
9 sources · profile reviewed Oct 21, 2026 · data updated Oct 21, 2026
- Machine Learning is Linear Algebra — Marius Hobbhahn's blog (overview)
- Marius Hobbhahn — CV (last updated October 2026)
- Team — Apollo Research
- About — Apollo Research
- Compute Trends Across Three Eras of Machine Learning (arXiv 2202.05924)
- Large Language Models can Strategically Deceive their Users when Put Under Pressure (arXiv 2311.07590)
- Frontier Models are Capable of In-context Scheming (arXiv 2412.04984)
- Detecting Strategic Deception Using Linear Probes (arXiv 2502.03407)
- Stress Testing Deliberative Alignment for Anti-Scheming Training (arXiv 2509.15541)