Constructability: Designing plain-coded AI systems – Charbel-Raphaël Ségerie & Épiphanie Gédéon
Constructability: Designing plain-coded AI systems
Charbel-Raphaël Ségerie & Épiphanie Gédéon – Executive director at CeSIA & Independent Researcher
Current AI paradigms lead to inscrutable black-box systems that make it particularly hard to examine how they work or what they actually do. In this talk, we present constructability, a different approach to AI systems leveraging automatic LLM (or other kind thereof) usage in order to train and design systems that are interpretable and reviewable by-design. This talk will present both the broader idea and how actually feasible it could be right now, along with prototypes for this idea and research directions.
GS AI seminars
The monthly seminar series on Guaranteed Safe AI brings together researchers to advance the field of building AI with high-assurance quantitative safety guarantees.