About the Role
The MATS Program offers an independent research fellowship designed to connect talented researchers with leading mentors in the critical fields of AI alignment, transparency, and security. Fellows will undertake research for a period of 10 weeks at our offices located in Berkeley, CA, and London, UK. There is also an opportunity to apply for an additional 6-12 month funded extension, allowing for deeper engagement and sustained research contributions.
This fellowship is an ideal opportunity for individuals passionate about the frontier of AI research, providing the necessary skillsets and a supportive community to make significant contributions to the AI safety space. Participants gain a comprehensive understanding of important questions and associated communities, produce legible and significant research outputs, and gain access to a broad base of collaborators.
What You Will Do
- Conduct focused research in AI alignment, transparency, and security for a 10-week period.
- Collaborate with top mentors and a diverse group of highly motivated researchers.
- Develop a deeper understanding of the AI alignment problem and identify key areas of focus.
- Produce impactful research outputs that contribute to the field.
- Engage in an in-person program designed for high impact and energy.
Fellowship Experience Examples
- Robert Krzyzanowski (Poseidon Research): Achieved a complete understanding of AI safety structures, produced significant research outputs, and gained access to collaborators, enabling a full-time career transition into the space. Focused on mechanistic interpretability and training process transparency.
- Thomas Larsen (AI Futures Project): Upskilled in alignment at a >3x rate, developed a deeper view of the problem, and focused on crucial areas. Participated in cohorts with John Wentworth and Nate Soares, writing a detailed overview of AI Safety approaches.
- Nina Panickssery (Anthropic): Rapidly upskilled in AI safety research, learned about the field, and met collaborators. Published a paper on Steering Llama 2 via Contrastive Activation Addition, which won an Outstanding Paper Award at ACL 2024. Mentored SPAR and MATS cohorts on LLM alignment projects.
- Jesse Hoogland (Timaeus): Became a technical AI safety researcher, leading to the foundation of Timaeus, an organization studying developmental interpretability and singular learning theory. Organized the SLT and Alignment and DevInterp conferences.
- Marius Hobbhahn (Apollo Research): Enabled the formation of Apollo Research by fostering trust and collaboration among founding members. Published work on mechanistic interpretability and became CEO/Director of Apollo Research.