Newsletter
The Paper Club hosted a virtual deep dive with Shengran Hu discussing Automated Design of Agentic Systems, focusing on ADAS and Meta Agent Search, supported by the Human Feedback Foundation, Mozilla AI, and AI Tinkerers.
The virtual meetup featured Leshem Choshen from MIT discussing open human feedback's role in improving large language models, offering practical guidance for decentralized feedback loops.
Princeton researcher Sadhika Malladi discussed model alignment challenges, presenting risks and mitigations from two preference learning papers to the AI Tinkerers Paper Club.
The AI Tinkerers Paper Club hosted Vijay Keswani to discuss his research finding 10–18% instability in moral preference elicitation, raising concerns for AI alignment.
Shayne Longpre (MIT) discussed his audit revealing significant restrictions on AI training data sources, impacting model development. The virtual session focused on data provenance and governance challenges.
MIT CSAIL's Stephen Casper presented on the open problems and limitations of Reinforcement Learning from Human Feedback, fostering community dialogue.
Authors presented their scalable method for extracting feedback from chat interactions to improve model alignment, yielding significant downstream performance gains.
This virtual event connected a global community with AI/ML researchers for technical talks and Q&A sessions on topics like instruction tuning and reinforcement learning.
