Joan Velja, Lennie Wells - Building an empirical Science of AI Debate [MATS Symposium | Winter 2026] - YouTube
MATS fellows Joan Velja and Lennie Wells give a spotlight talk at the Winter 2026 MATS Symposium. Project Title: Building an empirical Science of AI DebateMe...
MATS fellows Joan Velja and Lennie Wells give a spotlight talk at the Winter 2026 MATS Symposium. Project Title: Building an empirical Science of AI Debate Mentors: Shi Feng, Jacob Pfau Brief Abstract: AIs are getting quite good. Their reliability and autonomy makes oversight harder, so much that soon humans will only be spot checking their actions. We use this window to ground protocols for scalable oversight before they're load-bearing. Apply to MATS and learn more: https://www.matsprogram.org I'm not used to microphone. Thanks for coming. Um I'm Lenny. I've been working with Yohan…
saved by
related reading
- Jo Jiao, Boyd Kane - Limited-parameter finetuning [MATS Symposium | Winter 2026]youtube.com
- What will be left for us to work on?normaltech.ai
- Oversight Assistants: Turning Compute into Understandingbounded-regret.ghost.io
- AI researchers debate how close we are to recursive self-improvementyoutube.com
- The limits of AI safety via debate — LessWronglesswrong.com
- [2604.22082] Removing Sandbagging in LLMs by Training with Weak Supervisionarxiv.org
- [2506.13609] Avoiding Obfuscation with Prover-Estimator Debatearxiv.org
- Automated Alignment Researchers: Using large language models to scale scalable oversight \ Anthropicanthropic.com
- Neel Nanda on Mechanistic Interpretability: Progress, Limits, and Paths to Safer AI — EA Forumforum.effectivealtruism.org
- Agentic Misalignment in Summer 2026alignment.anthropic.com
- My personal cruxes for working on AI safety — EA Forumforum.effectivealtruism.org
- Thoughts on AI in academiatheinfinitesimal.substack.com