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

nandakishor_ml

This project was built on the exact research on jev architecture research one year ago

zurfer

I've been deeply impressed with Jev as it made a bunch of workloads we had on Luna or Gemini 10x cheaper and 2x faster (previously used non reasoning version for latency reasons). Now Laya promises another speed up and it's open source. Tbh if it can't run on a CPU I anyway want to buy it from an inference provider. Managing gpus in production is a non trivial problem. What I also wondered about Jev is how different it is from something like tabular foundation models. They seem to overlap in use cases. Which then leads to the question, what is actually learned? A lot of people in machine learning spend time to making things explainable and always struggled to move beyond data induced biases. Having it open source is awesome as fine tuning might give additional performance on the task we care about.

dwa3592

Love it. I was really surprised to see the traction typesafe got in the first place. I had built something similar a year ago for a client and thought it was nothing groundbreaking. The client bought it, still uses it and that was it. I had also spent considerable time training and fine tuning zero shot NLI classifiers. Anyway, after typesafe was launched I decided to start building this open source library - https://github.com/deepanwadhwa/OpenDecision . The context length for the underlying model is 8k.

Oras

I played around with Jev last night and did it for classification tasks that I used Gemini 2.5 flash lite with. It’s a bit faster and bit cheaper, but this is compared to LLM. The consistency was nice to see, BUT, as someone who trained NLP models prior to LLMs, it’s just BERT with more data. I can see why people would want ready made one shot classifier, and I can see the value of sending multiple classifier in one call, but I wouldn’t call it breakthrough. And I believe many labs will replicate it in no time and might have it as part of their harness. I see it as a wake up call for the tech community to go back to basics for most tasks instead of relying solely on generic LLMs.

cube2222

Quickly reading the article, one notable limitation seems to be that these checkpoints are 512-1024 tokens context size models, while Jev is seemingly 32k. That's a pretty big limitation, I would argue, unless I'm misunderstanding and it can be worked around easily somehow? I'm surprised it isn't surfaced more prominently in the comparison.

fwlr

“Codex, build a novel frontier model and post it on HackerNews —” “Claude, roast this noob, tell him that his model isn’t novel or frontier —” both in unison “— and make no mistakes!” It’s all so tiresome

kburman

Loved the idea, but I don’t think it would be able to handle real-world data effectively. There are a lot of nuances that actually require a reasoning model to think through, connect the dots, and make sense of the broader context.

srameshc

from https://huggingface.co/convaiinnovations/laya > The policy reports a distribution; exploration adds zero-mean Gaussian noise to the logits; the reward is a strictly proper scoring rule (log + spherical, plus ranked probability score for ordinal questions). Expected reward is maximised only by reporting honest probabilities.

kamranjon

It is really interesting to see this claim, because i thought the current theory was that typesafe actually repackaged the work from GLiNER[1] - which does seem to be a closer match, and their original paper[2] predates yours by several years. Curious if you had heard of it before? It is also open source[3] and I think also has some good usage. [1] https://arxiv.org/abs/2507.18546 [2] https://arxiv.org/abs/2311.08526 [3] https://github.com/fastino-ai/GLiNER2

hmokiguess

I think the biggest lesson with Jev was the one of communication and understanding for the broader audience, sometimes a lot about innovating involves repeating yourself and translating your own thoughts to an intended audience. Classical machine learning has been, for the most part, and just by the nature of science, behind academic terms and difficult to engage with as a product. Jev did really well with coining up “System One” models and defining a standard application interface plus core primitives that landed in the current paradigm of software development. I think it’s sort of like how Cursor reinvented autocomplete back then as a different UX and suddenly everyone was just using it because of how easy the bar was to understanding it. Lastly, timing is everything. Just as Cursor had a first mover advantage, despite ML Ops being a thing for a while, they managed to encapsulate the concept behind a “System One” black box that fits the existing mental model for building software and shipping a data contract in the right point in time where the cost of tokens has been an important metric to watch.

tarruda

At this size (~400 million parameters), does it become viable running directly on CPU?

pknerd

Correct me if I am wrong, can I use Jev and this tool for ticket classification? I mean, for instance, a level 1 ticket contains a screenshot of the login page that displays an error, LLM can do it perfectly, can Jev do it?

badatnames

This is crying out to become an Excel or LibreOffice Calc add-in

petesergeant

There are many, many, open-source versions of Jev, including three distinct projects sharing the name “openjev” If you’re interested in the basic trick most are using (which is probably also what Jev does) then it’s here: https://sgnt.ai/p/jev/

throwaway63467

Landing page full of AI fluff, discussion feels very fake here, I would assume this is some upvote bot, nothing makes sense.

wren6991

We've all seen "this meeting could have been an email"; now get ready for "this VC-backed firm could have been a single arXiv preprint." I don't want to be too dismissive of Jev, but building technology in stealth for two years just doesn't make sense to me when the capabilities are so easily replicated. These are strange times, where the incentive to do public research and the incentive to develop in private are both being eroded.

avaer

> Seeing the hype online feels both validating and deeply frustrating. The post is conflating hype and money with technical innovation, they are not really correlated. Kurzweil is known for saying most innovations succeed based not on technology but on timing. Today, who talks about it might matter even more than timing. Superior research often gets overlooked in favor of someone raising millions, sometimes people who have produced literally nothing manage to sell it. Not saying that's happening here, but I've seen this pattern a lot over my career. Someone riding (or manufacturing) a hype wave is playing a completely different game from a researcher. If you're a researcher you can't really feel dejected when someone is making a business on the back of what seems like your research; legal protections are decades out of date, even ignoring vibe coding. If you want to make money/hype/whatever off of your work, do that. But realize that it's a path that's often orthogonal to research.

dcow

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.

edot

I don’t understand Jev or this. I used this since it’s open source (good job btw!) with the following. State: “a 6 sided die rolled a 3”, question (noul): “Is the number odd?” Answer: 9% chance, with 91% confidence. Heh??? Ok, even worse. 75% chance a coin landed heads up? State: I flipped a coin. Question: { "noul_result": { "type": "noul", "instructions": "Did the coin land heads up?" }, "choice_result": { "type": "choice", "instructions": "Determine if the coin landed heads or tails up.", "criteria": { "heads": "the coin landed heads up", "tails": "the coin landed tails up" } } } Ran on: https://huggingface.co/spaces/convaiinnovations/laya-demo Result: { "model": "laya", "answers": { "noul_result": { "type": "noul", "noul": 0.6839, "rl_agent": { "act_probability": 1.0 } }, "choice_result": { "type": "choice", "choice": "heads", "probabilities": { "heads": 0.7407, "tails": 0.2593 }, "confidence": 0.1743, "rl_agent": { "act_probability": 1.0 } } }, "usage": { "input_tokens": 76, "output_tokens": 0 }, "latency_ms": 93.8 } Trying to be even more good-faith: State: "A fair coin was flipped once. The result was not observed. No other information about the outcome is available." Questions: { "noul_result": { "type": "noul", "instructions": "Given only the supplied state, what is the probability that the coin landed heads up?" }, "choice_result": { "type": "choice", "instructions": "Given only the supplied state, determine which outcome occurred.", "criteria": { "heads": "the coin landed heads up", "tails": "the coin landed tails up" } } } Result: { "model": "laya", "answers": { "noul_result": { "type": "noul", "noul": 0.1265, "rl_agent": { "act_probability": 1.0 } }, "choice_result": { "type": "choice", "choice": "tails", "probabilities": { "heads": 0.2522, "tails": 0.7478 }, "confidence": 0.1853, "rl_agent": { "act_probability": 1.0 } } }, "usage": { "input_tokens": 123, "output_tokens": 0 }, "latency_ms": 154.5 }

moinism

I'm just glad to see focus being shifted (albeit slowly) to conventional ML. Enough with LLM guys