Jev: System 1 models for instant answers ⚡️🧠
Jev from Typesafe AI: fast probabilistic models for classification, interfaces, games and search.
Twitter is a madhouse right now: everyone is testing the Jev framework and discovering some wild use cases.
Here’s the idea: most modern LLMs operate on the principle of System 2 thinking (slow, deep reasoning). When all you need is to extract a status from some text or classify a request, a traditional model starts generating tokens one by one, spending hundreds of milliseconds and burning through resources.
The folks at Typesafe AI have released Jev, a framework for building and working with so-called System 1 Models (“fast, intuitive thinking”).
What’s the big deal?
- No text generation: The model does not generate sequences of words at all.
- Probabilities in a single pass: Instead of an autoregressive token-generation loop, the model runs the input through the neural network just once and immediately returns a probability matrix for predefined schemas.
- Microsecond latency: Without a heavy generation loop, the response is practically instantaneous (tens or hundreds of times faster than a standard GPT call).
- 100% type safety and determinism: The result goes straight into strictly typed structures, without the risk of the model getting the syntax wrong.
This architecture is a great fit for tasks where ordinary LLMs were overkill or unsuitable because of latency. You load a dataset into the model and specify your criteria, and it immediately returns probabilities.
What people have already built on Twitter:
- Instant classification and moderation: Filtering spam and scams, identifying toxicity or tags. Feed it a comment and a set of criteria, and probabilities come back immediately (“spam: 99%”, “bot: 95%”, “negative: 12%”).
- Controlling UIs and devices: Jev handles interfaces very well. The model does not need to parse the entire DOM or read a screenshot through a heavy vision model: in one pass, it estimates the probabilities of which action or on-screen element is the target right now.
- Game AI at breakneck speed: People are already running the model in Snake, Tetris and Flappy Bird. In a single pass, the network returns probabilities for the best next move in milliseconds.
- Ultra-fast tooling and filtering: Querying data in natural language. For example, a hypothetical
jev(people, 'с русскими именами')can filter an array of objects almost instantly without prompting a heavyweight LLM to construct an SQL query. - Probabilistic search: Instant search over a finite list. You can feed it a sentence, and the model will pick the most suitable emoji or category in microseconds.
There are countless scenarios, and new use cases appear in the feed every hour. Combining fast probabilistic System 1 models for immediate reactions with traditional LLMs for complex logic opens the door to some very interesting architectural patterns. Let’s keep watching how this develops.