Technology

New AI Promises Fast Answers Your Apps Can Use Directly

Martin HollowayPublished 3month ago2 min readBased on 1 source
Reading level
New AI Promises Fast Answers Your Apps Can Use Directly
source:typesafe.ai

TypeSafe AI announced Jev, its first public System One Model, in early access on Sept. 15. The company describes System One Models as a new kind of advanced AI for fast decisions that software can use directly, according to TypeSafe AI.

Jev was built with a training method TypeSafe AI calls Reinforcement Learning for Calibrated Decisions (RLCD). Most chat AI writes answers one piece at a time. Jev produces all its answers at once in a single query, with a method the company calls parallel sampling.

Its answers are limited on purpose. Jev does not write free text. Like filling in a short form instead of writing an essay, it returns only answers in a fixed format set in advance, with confidence scores meant to show how likely each answer is to be correct.

TypeSafe AI states Jev is about as capable as current leading models on System One tasks, while being about 100 times faster and more efficient. The company claims total response time of 70ms to 500ms, which it describes as 40x to 200x faster than leading models for System One-shaped queries. TypeSafe AI lists Jev pricing as $0.042 per million input tokens, or pieces of text sent in, with output tokens free.

Looking at system design, the change is simple. Writing one piece at a time makes waiting time grow with answer length. Filling everything in at once removes that wait when the answer shape is already known.

In my view, fit matters more than speed alone. Software needs clean answers to make choices, check results, handle errors and link steps together. Free text forces developers to clean up and recheck. Fixed-format answers fit straight into that flow, and reliable confidence scores could let apps sort, filter or pass hard cases along without extra work, if those scores hold up in daily use.

Looking at what this enables, the idea is to split the work. Send simple, structured questions to a fast decision model and save general AI for open-ended work. That only works if the narrow-task ability is truly comparable, as TypeSafe AI states, and if speed and cost allow the model to run inside live steps. A 70ms to 500ms budget fits there, and free output tokens help since cost depends only on inputs.

Worth flagging for testing is what the announcement does not settle by itself. The claims of similar capability, 100-times efficiency gains, and 40x to 200x speedups are company statements about System One tasks and System One-shaped queries. Outside tests will need to define those tasks clearly, compare against named rivals under the same conditions, and check confidence scores when real data changes. Early access is the right time for that checking.

The broader context here is hopeful. Faster, cheaper answers in a usable format make it easier to try AI where delay or messy text once ruled it out. If the promise holds, developers get a tool that works more like a function call and less like a chat.