La Revue de Presse — AI safety & alignment

Nia et Viktor discutent des actualités du jour.

Sources :

• LessWrong: Examing Emergent Misalignment in a recurrent LLM with a logit lens
• MIT Technology Review: We’re putting too much faith in AI’s ability to say no
• LessWrong: Narrow Multimodal Fine-Tuning Can Induce Emergent Misalignment
• LessWrong: The potentially deadly threat of AI output-optimization
• LessWrong: The Non-Compassionate Case for Model Welfare
• LessWrong: exfiltration through self-distillation
• LessWrong: The Problem With“Doomers” and “Optimists”
• LessWrong: The Bloody Finish Line