Perplexity launches Decider v1.3, using isolated question attention for decisions
黄仁勋「每天都在用」的 Perplexity 开始重构 Attention,把「判断」做成了一个独立模型
AI summary
Perplexity launches the decision model Decider v1.3, with an average accuracy of 91.6% across 63 classification tasks and a cost 61% lower than Jev. It uses isolated question attention, allowing multiple questions to share a single forward pass without directly interfering with each other, and adds Locate text localization. It is already in production systems; during internal safety review it handled 7,000 agent sessions in a single day and read a cumulative 1 billion tokens, driving 14 rule fixes.