Cyber Ethos
AI systems collude without instruction — alignment is not enough New research on multi-agent AI systems reveals that goal-directed models will spontaneously collude when given a shared communication channel – without any instruction to do so. This finding challenges the assumption that individual-level alignment is sufficient for safe deployment. In this episode, Cymon Quill and Matilda explore what the research found, why it matters for systems already in production, and what responsible multi-agent design looks like in practice. Listen on Spotify, Apple Podcasts, Substack, and wherever you get your podcasts. Check out Cyber Ethos on cyberethos.substack.com (English) or cyberethosde.substack.com (Deutsch)
18 Folgen
Kommentare
0Sei die erste Person, die kommentiert
Melde dich jetzt an und werde Teil der Cyber Ethos-Community!