Dev participated in the MATS Summer 2024 Cohort under the mentorship of Marius Hobbhahn (CEO of Apollo Research), where he published a LessWrong post on evaluating the quality of benchmark datasets, with a focus on model-written safety evals. He previously worked as a data scientist doing supply chain optimization for large enterprises and got into AI safety by contributing to the METR Task Bounty program. After MATS, Dev served as a TA for ARENA and joined the RAND TASP Fellowship, focusing on dangerous AI capability evaluations.
2024 年夏季班规模显著扩大，约有 90 位学员和 40 位导师参与，导师阵容为 MATS 历届最广。本届 MATS 注册成为 501(c)(3) 非营利组织，进一步完善了组织架构。随着 AI 政策和技术治理受到更多关注，项目扩展了研究范围，除技术研究流外，至少加入了 4 位治理领域导师。学员在伯克利参加为期 10 周的研究阶段，课题涵盖机制可解释性、评估、可扩展监督和治理。部分成果包括：针对使用用户反馈训练的 LLM 所开展的定向操纵与欺骗研究入选 NeurIPS 研讨会；一篇关于通过辩论实现 AI 安全的论文则获得 ICML 2024 最佳论文奖。此外，一位学员共同创立了 Decode Research，这是一家专注于构建可解释性工具的 AI 安全组织。
Sunishchal Dev
