Last night my X feed was bombarded with Type Safe AI's Jev model and their technique of RLCD. I am glad to find models which can do things with less money and compute power but I honestly couldn't understand what did they do except making a very good JSON model?
Honestly, a 2 to 4 billion parameters model which I can run on my current laptop can do those things, the only bottleneck I may have is the confidence score and I beileve that also can be solved with a simple pytorch script.
So, what makes it very special? I personally think the whole hype is first because the founder of Type Safe AI was participating in ChatGPT (which is not really a little thing of course) and second, it could reduce the cost of something as simple as making smart choices.
And I am trying hard to recreate the procedure to find out what can happen. If you could recreate this model, I'd be happy to see your results as well.
[link] [comments]