DigiNews

Tech Watch by Johan Denoyer

← Back to articles

Perplexity and Intelligence

Quality: 8/10 Relevance: 9/10

Summary

OpenAI argues that perplexity (cross-entropy loss) should be tracked as a core metric for AI progress to benchmark model developers and guide safety standards. The post traces the evolution from GPT-2 to GPT-4, highlighting how reductions in perplexity accompanied emergent capabilities and discusses transparency and shared norms for monitoring advanced systems. It also notes the role of pretraining, prompting, and lightweight elicitation in shaping modern AI behavior and safety considerations.

🚀 Service construit par Johan Denoyer