Every third post in HN has this same complaint comment. Everybody knows.

I stopped reading at “The honest caveat”…

I thought it was a fine informative readme, starts with the problem, outlines the core of the solutions, and some limitations. Everything I want to know in the first few paragraphs. No need to spend human time to improve it.

Well nothing about the Doom demo or this entire model is new new either, is it? I don't even think Typesafe themselves are claiming anything novel, they say that they're focusing on practicality instead of chasing big numbers and AGI. Classifiers are older than generative models and are used everywhere. Fast classifiers are used in sampling machinery of every big model and for automation in agentic game plugins for years, except they're usually small and finetuned for the task, not general-use.

I think many people wondered why non-generative models are so underused on a big scale, well here's a long overdue attempt to market that which evidently goes well with people being interested in this again. The field has been captured by the vibe coding and valuation-goes-up hype and a bit. AI has a ton of low hanging fruits that are much more practical than using one tool that gets most attention for everything.

I had a long talk with chat gpt about this today as well. I think its duable and prolly not too hard either, also you could do lotsa funky stuff with stitched frames of a video in one 4x4 grid for example and send that as one image for analysis. that way temporal understanding can be had for fractions of a second by jev... also because vlm works in pixel space you can get around the whole state machine issue as well, so many possibilities...

[flagged]

This probably throws a spanner in the wheels there:

> Every piece of reasoning the frontier model does for free has to be rebuilt here as deterministic state.

EDIT: Not to shit on this though. I totally believe that some smart mixture of LLM-reasoning + Jev-style + determinism is going to be pretty amazing.

Did you do the follow up questions? I was invited within a few hours of joining today.

It's also now on openrouter and cloudflare

Curious about people’s experience here. I am working on a small model, verify by jev, and escalate to big model. Some cases, the small model is not a model but some regex.

cheap-confirm-escalate

Using jev as the confirm step.

It's a quick proof of concept of computer use powered by Jev, a new general-purpose classifier model. Here it takes the description of the interface (it can't do images yet) and outputs commands. It's faster and many times cheaper than using frontier generative models like GPT to do the same.

Jev docs gives a good explanation https://docs.typesafe.ai/concepts/system-one

Also see recent discussion https://news.ycombinator.com/item?id=49717558#49718350