Announcement
Perplexity CTO casts Decider as an “intelligent grep” for AI training
Denis Yarats says the decision model helps monitor reinforcement-learning runs, including one with 15,000 parallel rollouts.
TLDR
In his OpenRouter AMA recap, Perplexity CTO Denis Yarats says Decider grew from classifiers the company has used internally for years. Beyond monitoring reinforcement-learning runs, he discusses Decider as a reranker and argues benchmarks don’t tell the full story for decision models. He also describes mixed-model teams working together rather than routing between models, plus search-ranking models that retrain autonomously.
Combined views
832
2 Sources, first seen ago
Comments