Thread

I built non-autoregressive decision models with RL a year ago

1363 points · 319 comments · nandakishor_ml

dcow · 1789828104 · focus · HN ↗
I can understand why the author feels bitter but it still feels juvenile to me. Certainly both Jev and Laya are based on the research of countless prior papers and academics. Diogo decided to build a product out of the concept. The author didn't. Publishing research papers and model weights is probably part of the problem--it feels academic. If you look at the author's profile they focus on applying AI to healthcare. Not selling general AI type safety to AI pilled companies and devs. There's a big difference there. Whether that's good or bad you can argue all day. But for the author to expect otherwise is pretty weird. I do applaud them for not stewing too much on it and trying to do something about it, though.
1. operaopera · 1789828352 · focus · HN ↗
  I believe his qualms were with the "hype" in Jev's announcement: specifically calling this kind of model a breakthrough, without crediting previous art, and keeping everything closed source.
  1. prodigycorp · 1789828521 · focus · HN ↗
    And how is laya previous art? The project was vibecoded and posted yesterday.
    <a href="https://github.com/NandhaKishorM/laya/commits/main/" rel="nofollow">https://github.com/NandhaKishorM/laya/commits/main/
    <a href="https://huggingface.co/convaiinnovations/laya/commits/main" rel="nofollow">https://huggingface.co/convaiinnovations/laya/commits/main
    prometheus1992 · 1789830960 · focus · HN ↗
    @prodigycorp - reading your comments here on this post - you seem pretty hurt by this post.
    prodigycorp · 1789831126 · focus · HN ↗
    Yeah, the reason why I am annoyed by it is because a person (who felt like a burner account of the laya creator) yesterday was haranguing me for saying that projects like this were vibe coded, posting the link to this project.
    I evaluated this project yesterday and found its claims un-credible. It's literally nothing like jev. That's some context behind why, a day later, I find it annoying that this is somehow the top story on HN.
    <a href="https://news.ycombinator.com/item?id=49752902">https://news.ycombinator.com/item?id=49752902
    sreekanth850 · 1789834568 · focus · HN ↗
    I don't really see the breakthrough in Jev. Classification, scoring, routing and returning probabilities over predefined choices are all established problems. We implemented category routing in our own retrieval system in a slightly different way: embed the incoming query, compare it against category profiles and route to the highest cosine-similarity. Obviously Jev isn't similarity based, but the underlying task of making a constrained decision from predefined choices isn't novel. TypeSafe says Jev has a new architecture and RLCD training, but Jev's actual architecture, weights and training details aren't public. So we can't even claim Jev is specifically a BERT classifier, but also don't see enough public technical evidence yet to call the underlying idea a breakthrough. Atleast they should publish a technical paper to prove their idea is breakthrough.

Open on Hacker News to reply ↗

Unofficial Hacker News client; not affiliated with Y Combinator.