I hooked Doom to a basic 187M param zero shot sequence classification model to do similar 10x+ faster by classifying in a single forward pass plays doom, but also works w other biz tasks too: - intent detection - prioritizing tasks - customer support routing - scoring - etc

demo built by @agnt_gg try AGNT today 👇

Nice, I was considering speed running "OpenJev". I feel like parallel sampling just got a good marketing run and this stuff has been possible the last 10 years

100%, still so handy tho if you know the solutions before hand and are trying to optimize for speed and compute the "LLM shaped hammer" wastes a lot of resources for tasks better suited for classifiers and such

I think it's smart to make a company around this and market it tbh. 99% of company building is about about spoonfeeding technology to people and teching them how to use it. I deployed something like this at a startup 3 years ago and considered building a product around it, but it just doesn't excited me much

Would this not also work for a lot of robotics policies? Fun to think about.

oh yes!! the inference speed is ~30ms to first token so it's incredibly fast and cheap to run

much needed sane post.

But does it have a cool name like Jev

so you're saying it's better than jev?

give me a benchmark and i'll show you an outcome - Charlie Munger, probably

and I gave it access to a whole falcon-9 class rocket and it was able to launch and land itself as well

The single-forward-pass framing is the key: it turns policy decisions into a cheap classifier call instead of a long generative loop. I would watch calibration drift as the task mix changes, since routing errors are harder to spot than missed Doom shots.

This is so beyond my understanding, but it looks amazing, keep it up

@NathanWilbanks_ @grok does this use vision , if so how?

how do i even get started to try to make something like this?

like a sandwich: small bites, one after the other - learn about perceptrons - start thinking about how to codify the states of things - once you have turned some state into a numerical representation, think about how you can loop through, reason over, and change that state

GLiNER can do the same agentic decision making tasks to play video games like Jev while being local, free and open weights. it's so lightweight it'll run on a laptop.

bart-mnli?

it uses deBerta v3 as the encoder, with GLiClass's label scoring architecture and training on top:

i found if you offer a few different move options (slight nudge left, veer left, etc.) the model will use them for more precise aim

you could also stack instances of this model for different tasks one for aiming/shooting one for movement/dodge one for pathfinding etc

That's what I was questioning myself. We do have GliNER as the mid point between high eng effort, small models and low eng effort, big models. Is the inference speed around the same?

since its output is a decision matrix, not text like a llm, its inference speed is VERY fast. ~30ms to first token vs seconds for most LLMs to first token