CHOOSESecurity & safety
Jev: Decision Intelligence Model
Alcides Ticlla@AlcidesTicllaCh
5:56Jev is a decision-making AI model that provides typed answers with probabilities and confidence scores, trained using RLCD for reliable probability interpretation, and is significantly faster and cheaper than traditional LLMs for specific tasks.
SignalYou send a state plus typed questions (Choice, Score, Noul) and get decisions back, each with a probability on every option and a confidence score.
Pattern↑ Jev AI decision making
$0.000081 in 0.114 seconds. The LLM doing the same job: $0.013880 in 8.566 seconds.
That gap is what Jev is. TypeSafe AI's first model gives up text generation entirely: no chat, no code, no prose. You send a state plus typed questions (Choice, Score, Noul) and get decisions back, each with a probability on every option and a confidence score. Trained with RLCD so probabilities mean what they say: claims of 0.2 happen about 20% of the time. The founder co-invented RLHF and helped build ChatGPT, then spent four years asking where all the automation went. His answer: computers speak a different language.
Input is $42 per billion tokens. Output is free. 70-500ms end to end, because a parallel pass over a fixed answer space has no token-by-token chain. Vercel swapped an LLM safety classifier for it: 5-18x faster, more accurate. Fine print kept in view: 193.6x/444.6x claims are self-published workflow evals refereed by GPT-6 Astra + Fable 5.1 averages, the architecture is undisclosed, and the subsidy question is open. The name is the thesis: Jev, for Jevons. Cheaper intelligence, far more intelligence.
Paper: typesafe.ai/blog/introducing-system-one-mo
Series: youtube.com/playlist
Audio deep dive: youtu.be/ij8S3w7OWJs
#Jev #TypeSafeAI #DecisionModels