Jev beats Gemini 2.5 Flash Lite on classifier eval quality and speed
Malte Ubl ran TypeSafe AI's Jev against an existing classifier eval, where it saturated the eval on quality and was 6x faster than Gemini 2.5 Flash Lite.
Measuring Jev: 18,000 judgments and probability calibration checks
The author spent a week measuring Jev: 18,000 judgments across a full HTTP spec for 8 cents, then hand-labelled 59 sentences to check whether the probabilities mean anything. Every bucket came in below its prediction.
Everyone posted a Jev demo last week. I spent the week measuring it instead. 18,000 judgments across a full HTTP spec. 8 cents. Then I hand labelled 59 sentences to check whether the probabilities mean anything. Every bucket came in below its prediction. 🧵
· 0 likesOpen on X
More like this
Fast row classification in DuckDB with Jev
The author released a DuckDB extension that uses TypeSafe AI's Jev to quickly classify rows in CSV/Parquet files or DuckDB tables, taking about 10 seconds for 1k rows—faster than an LLM and more ergonomic than a classifier.
Classifying 1,018 AI Papers with Jev for $0.08
The author summarized each paper with DeepSeek V4 Flash, then sent the title, summary, and 24 candidate topics to Jev for classification, processing 1,018 AI papers at $0.08 total cost and 256ms median end-to-end latency per paper.
