AI Safety Blog

MariChatmen: Teaching a Chatbot to Write Andalûh Without Becoming a Mascot

What I learnt while building an experimental written Andalûh assistant: spelling was the easy part; keeping the answer useful was the real problem.

Read Post

Consensus Through Debate

How can we align advanced AI systems with human values without falling into the trap of polarized "winner-takes-all" debates? A hybrid methodology for AI safety.

Read Post