MATS Winter 2027 Applications Open with Focus on Open Archetype Training
geoffreyirving · x · 2026-08-20
MATS Winter 2027 is now accepting applications for a fully-funded, 12-week fellowship ($6.4k/mo) targeting AI alignment, interpretability, security, and governance researchers. Mentor David Africa outlined specific research streams focusing on model motivations, character, open-ended environments, and new forms of misalignment.
Key project topics include:
- Open Archetype Training: Exploring model archetypes beyond the assistant (e.g., sages, oracles) to find forms more conducive to flourishing.
- Failure Modes of Tiling Personas: Investigating how beneficent characters train successors to uphold values.
- Model Behavior in Open Environments: Deploying agents on the web for tasks and clustering their behavior trajectories to profile outcomes.
Related event: MATS Winter 2027 Applications Open with Full Funding(3 posts)→
More from Research
- OpenAI Researcher Departs to Focus on Formal Verification of Frontier AI Systems — HanchungLee · 2026-08-20
- Latent On-Policy Self-Distillation Enables Self-Improving Agents — burkov · 2026-08-20
- Recirculation Shows Perplexity Reduction Promise Across Five Model Families — xennygrimmato_ · 2026-08-20
- Recirculation Architecture Reduces Perplexity Across Gemma3 Models — xennygrimmato_ · 2026-08-20
- Daphne Koller on AI in Drug Discovery: Structure Over Compute — ziv_ravid · 2026-08-20
- SemaPLC: Verification-Gated Agent Harness for PLC Code Generation Beats Baselines — Midea-AIRC · 2026-08-20