Perplexity launches Intelligence feature

openai.com7 min read

The author argues that tracking perplexity and cross-entropy loss during AI pretraining is essential for understanding progress, establishing safety standards, and predicting the emergence of transformative capabilities. Providing transparency into OpenAI's internal perplexity measurements, the post proposes that frontier model developers jointly establish shared evaluation metrics and common limits on development until monitoring capabilities improve. Open article →

More on this topic