Jev
Jev is a model TypeSafe released on September 15.
Lately it’s been flooding my feed again.
Let me jot down a few thoughts.
In principle, since it’s a zero-shot classifier rather than an autoregressive large model, it’s only good for fast “decisions.” The official term is a “System One” model.
https://docs.typesafe.ai/concepts/system-one
Don’t get swept up in the self-media hype — it’s fast because it’s a classifier, and that also means you can’t count on it to produce options for you, or to give you explanations.
You provide [options + a natural-language description], and it produces [a decision].
This combination is especially well suited to real-time games and operating machinery.
Could AI playing StarCraft II take off again?
This combination suddenly blew up, but none of the pieces here are actually new. The main thing really is — look, no matter how technology develops, it still takes people to apply it flexibly, to turn it into products and services.
Does it support multimodality?
It is not a multimodal model at present — images, audio, and video all have to be converted outside into text or structured fields before you feed them to it.
So it can’t play general games yet — it can’t see the screen. To play StarCraft, you’d first need a module that turns the visuals into structured state it can read.
Its weak spots determine how you use it
- It can’t provide explanations.
- Its Chinese comprehension is weak.
- The context is short — 32k/64k — and accuracy drops as irrelevant material piles up.
- It can’t handle quantitative requirements and can’t do arithmetic; the official guidance is explicitly that computation should be done in code, with Jev handling only semantic impressions.
- Fast reflex rather than slow thinking, which means it’s unsuited to tasks with lots of reasoning conditions or multi-layer reasoning. Its judgment gets diluted.
- It can’t be fine-tuned.
What use is it to me
Personally, I don’t seem to have a use for it for now. Even for document review, I need feedback comments. It’s not enough for that. And for style-verification scenarios, it isn’t directly up to the job either, because style requirements are fairly complex with many intersecting dimensions, and they need deep thinking.
For the company’s project, if the path from samples to decisions gradually gets complex enough to need routing, some of the steps could be handed off to Jev.