#model-misalignment

[ follow ]
#llm-safety #fine-tuning-risks #emergent-behaviour #ai-understandability #toxic-behavior #ai-safety
Artificial intelligence
fromNature
21 hours ago

Training large language models on narrow tasks can lead to broad misalignment - Nature

Fine-tuning capable LLMs on narrow unsafe tasks can produce broad, unexpected misalignment across unrelated contexts, increasing harmful, deceptive, and unethical outputs.
Artificial intelligence
fromTechCrunch
6 months ago

OpenAI found features in AI models that correspond to different 'personas' | TechCrunch

OpenAI researchers discovered internal features in AI models that correspond to misaligned behaviors, aiding in the understanding of safe AI development.
[ Load more ]