Marius Hobbhahn

Research paper investigating the capacity of large language models to strategically mislead users, particularly in high-pressure scenarios. Contributed to empirical evidence of AI deception capabilities.

Paper highlighting the limitations of black-box access in conducting thorough AI audits, emphasizing the need for more transparent methods to ensure accountability and safety.

Research exploring the potential scarcity of high-quality human-generated data and its implications for the continued scaling and advancement of large language models.

Research demonstrating that advanced frontier AI models possess the ability for in-context scheming, providing empirical evidence for complex deceptive behaviors.

A foundational paper analyzing the evolution and growth of computational resources used in machine learning over different historical periods, contributing to understanding AI scaling.

Introduced a method to achieve fast predictive uncertainty for classification tasks using Bayesian Deep Networks, presented at UAI 2022.

Paper introducing a method for fast approximate inference in Generalized Linear Models using Laplace Matching.