CHOOSEBusiness workflows
Jev Model Testing
Kevin Ly@kevinlysocial
0:29This project tests Jev, a new model from TypeSafeAI, which uses Reinforcement Learning for Calibrated Decisions (RLCD) to interpret commands and return probabilities, demonstrating faster and cheaper performance on structured workflows.
TakeIt chooses from defined options and returns probabilities.
Pattern↑ Jev AI decision models
I’ve been testing Jev, the new model from @typesafeai , co-founded by ChatGPT contributor @CompleteSkeptic Diogo Almeida.
TypeSafe reports 193.6x faster and 444.6x cheaper on its own structured workflow benchmarks. Those are the company’s results; I haven’t benchmarked my full app.
Jev uses RLCD: Reinforcement Learning for Calibrated Decisions. It chooses from defined options and returns probabilities.
I connected it to a photo review tool for the pile of images after a site inspection. Local Whisper transcribes my speech, Jev interprets review commands, and the app saves my notes against the photos.
“Next photo” worked. The note saved. “Keep the previous one” came back unclear. I’m still making the inspection judgements.
The appeal is fast decisions that let me keep my attention on the photos.
Next I’d like to try this at my electronics workbench: a camera and vision model feeding observations to Jev, helping flag steps worth checking while my hands are occupied. Jev currently accepts text only, so that vision layer will need to be added.
With a model this fast and cheap... there are some interesting use cases coming up.