Cole Medin walks through his experiments with Typesafe AI’s Jev, then surveys community projects and addresses the objection that it is just another classifier. His practical argument: reserve language models for generating and reasoning, and use a decision-focused model for the many small choices between those steps.

A decision interface, not a chat interface

  • Jev takes a situation and a set of questions with constrained answers. It returns decisions and confidence scores rather than free-form prose.
  • In Medin’s support example, the same customer complaint feeds both a department-routing question and a frustration assessment. The routing result favors billing with 64% confidence; the sentiment result indicates a frustrated but civil customer.
  • Typesafe calls this a System 1 model and describes its training as reinforcement learning for calibrated decisions, or RLCD. Those are the company’s framing and training claims, not an independently established description of the architecture.
  • The intended role is a classification or decision step inside an application—not a replacement for a model that writes code, explains something, or holds a conversation.

The speed and cost pitch

  • Medin cites comparisons claiming decisions are 20–200 times faster and 40–1,000 times cheaper than the compared language models. These are reported comparisons, not benchmarks reproduced for this capture.
  • He also cites a zero-failure claim for structured output. Avoiding malformed output is a narrower property than always choosing the right answer.
  • He acknowledges that some competing models make better decisions when given more time. His interest is in the combination of low cost and low latency, especially when many decisions must happen rapidly.

Playing a game to test a game

  • Medin describes a local game he built and demonstrates Jev choosing actions such as attacking, moving toward enemies, and dodging while he keeps his hands off the keyboard.
  • The model receives candidate actions and makes repeated choices with confidence scores. The appeal is fast feedback while a coding model is developing the game.
  • The division of labor matters: Jev does not create the game. It interacts with it; a generative model remains responsible for writing the code.
  • This is a demonstration of one game-testing setup, not evidence that Jev can play arbitrary games or discover every gameplay bug.

Pull-request triage in Archon

  • In his open-source workflow tool Archon, Medin uses Jev for classification and routing before a pull-request review.
  • Those early decisions determine what level or kind of review a pull request needs. A language model then performs the review itself.
  • The saving is not just the cheaper classifier: routing can avoid assigning an unnecessarily deep review to every change.
  • His workflow calls a Python classification script and accesses Jev through OpenRouter. He also points to direct access through Typesafe.

Picking the model for the request

  • A second experiment uses the incoming user query as the situation and a small menu of model categories as the choices: strong, coding, open, or fast.
  • Examples route a deeper question to a stronger model, a Bash-to-PowerShell conversion to a coding model, and a simple conversion question to a cheaper model.
  • For the small set displayed, Medin reports a total routing cost of four-tenths of a penny and an average decision time of two-tenths of a second. He explicitly says the display is only a small subset of his testing.
  • These numbers describe his example workload; they do not establish universal routing accuracy or production latency.

Community projects broaden the possible uses

  • Doom: he shows another developer’s game-playing demonstration, with decisions displayed alongside gameplay.
  • Browser automation: he points to browser-use/jev-ultrafast, framing the current page layout as a situation and possible browser actions as the choices.
  • Pong: a comparison demo adjusts game speed to what each model can handle. Medin uses it to illustrate responsiveness rather than presenting a controlled performance study.
  • Other applications: Awesome Jev collects projects around routing, agent decisions, guardrails, calibration, games, and finance. This is a source of ideas, not validation of every listed application.

Is this just classification with new branding?

  • Medin agrees that classification is not new. His distinction is generality: traditional classifiers he has built were trained for a specific task and dataset, whereas Jev accepts different situations and user-defined choices through the same interface.
  • His deliberately subjective example asks which country has the coolest buildings. It illustrates flexibility, but neither the chosen country nor its confidence score supplies objective evidence about architectural quality.
  • His stronger examples are operational questions: which review level is needed, which model should receive a request, and which game action should happen next.
  • The takeaway is complementary infrastructure: use fast classification where the possible decisions are defined, and retain generative models where the application needs open-ended output.

What this capture establishes

  • This outline follows the full video transcript and YouTube description. The Firecrawl sponsor segment and channel promotions are omitted.
  • The examples, timings, and savings above are Medin’s reported demonstrations. No API tests, accuracy evaluation, or independent verification of Typesafe’s performance claims were performed for this post.