那个教 ChatGPT 说话的人，做了一个「哑巴」模型

TypeSafe AI, founded by ex-OpenAI researcher and ChatGPT co-inventor Diogo Almeida with $40M funding, launched "Jev," a non-text-generating System One model that skips autoregressive token production entirely, outputting only structured decisions (classifications, scores, booleans) in 70 - 500ms latency - 40 - 200x faster than frontier models. Jev costs $0.042 per million input tokens with free output (~$0.0004 per decision, roughly 76x cheaper than GPT-5.6 Terra and hundreds of times cheaper than Claude Opus 5), and matched GPT-5.6 Terra's accuracy across four enterprise workflows (67.8% vs 67.9%) using its proprietary RLCD training method; however, benchmarks are internal-only, the model offers no explainability for audit, and sustainability at scale remains unproven.

why now

TypeSafe 发布 Jev 模型，快 200 倍且便宜，Agent 执行层新选择。

topic

AI Tech & New Models

source

极客公园