After chatgpt everything in AI mostly became LLMs and building wrappers around them. It's like people forgot how to do ML.
To those of us who actually trained models back in the day, its kind of cute to see people wowed by a classifier. Yes, this is 0 shot and doesn't need training (most people wanting this would've used structured output, this is cool because it's cheaper and faster). But anyone with basic ML knowledge could've built this in a few hours.
The question is mostly why wasn't this productized. And it's interesting indeed that it took this long to become a finished product.
There's probably other areas where people are using LLMs where a more tailored ML solution might work better.
Probably because doing it wrong (using an llm in place of a classifier) is more profitable? (For the people selling inference.)
LLMs still are better than Jev at the task, just across the board slower.
Anyone who had a reason to try this already tried it (ads/recommendations) - back in 2023/2024 during the first fine tuning wave and it was accurately determined that it was not worth the effort, the results were more bogus than just using CoT, so frankly parallelism meant nothing if bogus * parallel = bogus.
So thrown into the dumpster and nobody really cared to revisit because it was already tried.
Pretty much sometime between then and now it somehow became the state where the tradeoff makes sense now.
Maybe it's just a case of the idea just now crossing the threshold into working just good enough to be worth it
There have been tons of applications for this. People were using earlier LLMs like BERT for classifiers long before LLMs became viable chatbots.
People didn’t know they wanted classifiers until OpenAI gave them a taste.
I think the hype with Jev is just that, while structured generation is great, LLM judges tend to kind of suck for precise classification. And the more powerful the base model, the more accurate they can get, but they get increasingly expensive/impossible to finetune. It was specifically the latency/price point Jev offered vs. the general accuracy it claimed that generated all the excitement. Plus the promise of cheap calibration (tuning).
"Jev exists because LLMs exist" is kind of a truism, as Jev apparently is literally a Transformer model.
In early 2016, Clarifai's core product was basically just an image classifer exposed via an API. And at that point, they'd raised $40 million--the same amount as TypeSafe.ai/Jev--and they were experiencing viral growth among developers + signing contracts with a bunch of flashy logos. The demand was so high that Amazon launched Rekognition and Google launched their similar APIs to compete.
The AI hype cycle and the number of people thinking about using AI certainly puts more wind at Jev's back, but even in an alternative universe where we don't have contemporary LLMs, Jev's core claims would be remarkable and there would be a big market for it.
I have 80,000 voice recordings to classify. Very few are in English, and the classes are not in English. Many of the classes are project names or other proper nouns. How could a system not trained on data specific to the problem possibly be expected to work?
and you might find the scores it gives back about its confidence are useful for escalating to a more expensive classifier.
but also, there's no requirement to use it as a zero-shot classifier, you can provide it as many examples as you like, and prioritise giving it examples it had previously gotten wrong. and with input caching it might be economical.
I'd be surprised if they or others don't start offering a fine tuning API for models like this, like openai does for some models (or used to, I haven't checked in a long time).
As are, honestly, most projects I work on.
There are already Jev-shaped open weight local models coming out that can be fine-tuned on specific data. We'll see which of them the community settles on.
I adhere to the idea that this is software's "Tower of Babel" moment where everyone just fundamentally ships things in completely diverging architectures, because creating a ground up architecture is no longer something that needs to be avoided for an economically viable business mode that in the past two decades would have otherwise incentivized people into industry standards. In a world where "taste" is the focus, single ingredients in the recipe aren't enough.
This particular "innovative" concept already has a rich, open research background. What Jev appears to have done is scale that up a bit and isolate good training data, which results in a great product but not really something impossible to imitate. The only major difference currently is that the open source decision models need to be fine-tuned as they're not trained off of the entire internet yet.
My thoughts too. It sounds like a minor feature being framed as a whole new business.
Then again, Dropbox and Docker are too.
It is definitely not the MNIST classifier you had trained in 2019.
The difference is that you only train it once and the modern LLM machinery sort of takes care of that with large contexts.
It's great that Jev proved this is a viable product. I'd expect a great many research innovations coming from making this work better/faster/cheaper, and around interfacing modern agents with it.
Anyone fitting MNIST in 2019 was already outdated by several years at the very least. GPT-2 was 2019! We already had zero-shot classifiers then. In fact, the paper for GPT-3 was literally
"Large Language Models are Zero-Shot Reasoners"
These kinds of zero-shot classifiers were already developed and used in-house for many years. They just weren't commercialized as a separate product, because anyone who could use an LLM proper could build layers around it to fulfill any classification task like this.
GPT2 was absolutely unusable as a classifier. Using GPT3 as a classifier cost a few order of magnitude more than what this thing is priced at, and the context window was a few thousand tokens.
> These kinds of zero-shot classifiers were already developed and used in-house for many years
This is like Google's favorite coping mechanism for falling behind at AI. "We had everything inhouse for several years, we didn't release it for $reasons."
> anyone who could use an LLM proper could build layers around it to fulfill any classification task like this
You missed the part where it costs more than two orders of magnitude lower :)
I'm not claiming there are major architectural innovations, but that's not the point. Once you prove there's a market, there's a cambrian explosion of innovations.
There already was a market for this, that's why people continued developing them in house for their own purposes. I trained people for this. It was just silently used at various companies without the loud marketing fanfare. The fact that other groups are able to replicate and beat Jev within days should tell you something. That's not "the new normal". That's an example of a product built on marketing hype alone.
Doesn't mean Dropbox was not innovative.
LLMs better than chatgpt existed at Google two years before chatgpt. Doesn't mean chatgpt was not innovative.
Not all innovation is about groundbreaking research. More often than not, it's figuring out what people will pay for. The research to make it work well often comes after that.
Now all you need is to give it more context along with your query.
There's no scenario where you are training BERT online to give you an answer.
Rank System Score Public / sealed accuracy Evidence
1 decider-4b v2 64.13 83.5% / 34.7% Evaluator-run, offline
2 Jev 1.13 63.29 86.6% / 36.7% Evaluator-run API
3 JevK5 v0.2 62.04 85.3% / 33.1% Evaluator-run
4 Cygnet 12B 61.76 87.9% / 33.8% Evaluator-run, offline
5 Hopper 59.43 82.3% / 34.1% Evaluator-run
28 Kev 4B 36.14 66.2% / 22.4% Evaluator-run
41 Laya 421M 30.25 58.4% / 30.8% Evaluator-run
It's so easy that I question why I would ever pay for JEV when eventually I'll have done enough random things that I will also have a large corpus and likely a general model as well.
It's a fast classifier you can use out-the-box, ~1.5bn tokens is about $40 (I've been hammering it)
It just works ... a whole bunch of low-level/low-importance workflow stuff that was getting farmed out to small/fast LLM models now has a competitive alternative ... and bits that hadn't even been considered to go into some external descision/classifier service can be tested/deployed at ~$0.00003/req
I don't get this wall of negativity on it, it's genuinely innovative/useful tech ... would expect HN to be more positive, regardless of whether it's the absolute best execution
Starting from a business POV one should inflate terminology, hack together an MVP, and see if the market demands it before doing hardcore R&D.
But starting from technical/craftsman POV all you see is a hack and a lot of big words, so it’s easy to become jaded.
I think the point being made is that Jev is great but it has no competitive moat, and open source versions will very soon catch up if their secret sauce is just synthetic data.
(Whether or not that is true, I don't know.)
I've been unable to find a good use case for now.
I'm building a woodworking app and I've managed to create an autopilot that can take a simple instruction ("get me 5 2x4s", "cut the middle 2x4 into 4 equal pieces", "move the 2x4 3 feet left") and the action instantly happens with next to no lag. There is already an llm but now it can share an intent, and the geometry system shows jev the various actions and jev chooses the action that gets it closer to the goal until it has found a state that matches the intent or gives up. The result is the llm can think "higher level" and let the cheap fast model grind out the options in a relative blink of the eye, without the 30s of reasoning the llm would have done about the various operations it could try.
But I'm guessing people will find the right training regime and data mix soon to close the gap.
But big things I see are instability and inaccuracy - like pick a random problem.
I mean, anyone here could have a decision maker just return a random number. Fast and easy.
The question really is how GOOD is it?
I’m sure this has been a gradual and long decline. Maybe it even started with the dot com boom and accelerated with crypto. With AI it seems to have got worse.
Edit: One is that jev/laya are tuned to have better probabilities, but a reranker can be fine tuned to do that as well. And jev/laya use RLCD?
Main difference is that laya/jev/et-al give you a zero-shot classifier that requires no training. You can prompt engineer your way to a quick fairly reliable cheap enough decision engine that you can use to iterate quickly (by prompt engineering).
Right now a lot of people are doing this with LLMs and it's too slow and expensive.
Imo the right iterative approach to productionizing these systems is something like:
1. Build it with an LLM. Iterate on the prompt
2. Start building a real-world dataset
3. When the prompt works, turn it into a clear rubric for Jev or similar
4. Keep iterating until desired accuracy achieved
5. Use the real-world evals you've built to train a custom classifier fine-tuned to your needs
OpenAI has a section on their embeddings model api page for zero shot classification. Of course you can choose an open weights embedding too if you’d like.
https://developers.openai.com/cookbook/examples/zero-shot_cl...
I think Jev wins on marketing and convenience. Most SWEs don’t want to talk about embeddings, cosine similarity, or precision/recall tradeoffs. They want something which plausibly works and is easy to use.
Like, their example is of classification for a support interface.... `refund_requested`. Pretty convenient bool given the example is about a refund- what if 99% of submissions don't ask about a refund? Also, is that user not a `churn_risk`? What could possibly qualify as a churn risk if not a user asking for a refund?
https://ollaya.dev/library/laya The examples suffer the same problem of why I'd prefer to use a string column vs an enum. Changing an enum means you need to update the db, using a string you can do whatever.
I'm not trying to be negative, I genuinely want to know about some practical examples (that don't require tons of backwards maintenance).
Really I think "smart grep" is a pretty good one ('look for an error looking vaguely like this'). Also I think sql-based shell history + decision model is quite good to make the last 'which one of those choices is best fit given users past few commands' etc.
But with Jev you're just paying for input (prefill) which is really fast, and in case of Jev specifically costs 50% of Deepseek V4.1 Flash (which has famously really cheap input token pricing).
I put 250MB / 1M lines of logs through Grev and it cost ~$10USD, DSv4.1 would be at least 10x that and much, much, much slower. With Jev/Grev that 1M requests took 10 mins
Edit: completely misread your question - yeah you could finetune specialized models to do that, probably based on some decent pretrained llm base, that is true for roughly any Jev-shaped problem. Do you want to bother doing that, also having to deal with having to host a zoo of specialized models?
I very much appreciate your to-the-point, non-vibed README as well, ty for that.
But if that were solved, I could see giving ollaya/grev to LLMs themselves, giving LLMs their own massive token-saver.
On the readme I'm so sorry to tell you that, but it's 100% written by Opus 5.5 with zero "pretty please don't write slop" prompting, it's just how slop is going to look like from now on. I've been writing code for 15 years or sth like that and the code is also what I'd call pretty reasonable..
And if you have not been, it’s for when you have to extract the context from text. When you have numbers or fixed options, it’s just a matter of code.
So if you find yourself having to decide if a given user comment is a refund_request, that’s for that.
It’s not perfect, you still have to fine-tune (or calibrate) using examples you have (and keep those examples updated over time). But it’s way better than trying to parse text with regexes.
>example is a text classification task instead of a decision
decision model = classifier
system one model = small non-reasoning LLM
noul = boolean
confidence = f(probabilities)
It's sad to see how gullible engineers are today.
that laya is even a thing is further evidence, people took that author at face value, the paper contents are incomplete and describe something that does not sound like Jev at all
this was the period of arxiv history that led to the new vouching system, laya author contributed to that imo
That said, Typesafe false marketing caused Laya to fit perfectly into pretty much every advantage that they are claiming: "system one decision model", cheap, fast, no hallucinations, structured, confidence output, parallel, calibrated. Their BS is their own demise.
I think Laya's author genuinely bought their BS and thinks he built the same thing. Unlike Typesafe, I don't think he's intentionally misleading people.
The only unique thing about Jev is that it's a general purpose classifier. Funny enough, they were so busy spreading marketing bullshit that they forgot to mention the only real thing that makes Jev unique.
I suspect most people only read the blog post, and thought it was great how a VC company "stole" an idea and was "outdone" by a rando... without actually checking the facts. Confirmational reading bias, we live in a post-truth world with dysfunction media ecosystem
But like you said, at the end of the day he's just a rando.
He's not asking for $40m, not saying "I made ChatGPT, but i hate it, so I built the next big thing". Not claiming to co-invent RLHF.
Laya is just noise. Jev's bullshit affects me today - I see people injecting it into the codebases where it has no place.
Just how to "make fkn $500k ARR fast?"
https://news.ycombinator.com/item?id=49674396
too much LI/Xitter influencer consumption
To be fair, he's just asking how to get customers. And the post is 2 days before Jev's launch date? I don't think he's trying to sell Laya there (though he probably will at this point).
example ranking shared in comment on this HN post
Stepping back from Laya, there are dozens of Jev-a-likes now, most built on LLMs, whereas it is speculated that Jev likely uses transformers, but is not LLM like by generating a token at a time. Most likely they are reaching into the logits to get the details and assembling the json object deterministically. This is what the various LLM spins are doing, which is probably close but not the same as Typesafe.
This is the main PR (merged) for using DiffusionGemma (diffusion generally) for Jev-a-likeness in vLLM. https://github.com/vllm-project/vllm/pull/57250
if you use gateways, GoModel support the S1 endpoints, my favorite feature is the virtual models, stable name, I can swap out the backing model(s)
https://gomodel.enterpilot.io/docs/getting-started/quickstar...
(the "kev" in the docs is my fault, I should have said Jev / System1 in my feature request)
Also curious, it seems from looking at the accuracy scores you gave that it seems to be NLI > Gliclass > Laya (for Bert types)? Why do you seem to feature/recommend Laya more - is Laya better in some way?
Smarter move if you have an eval set is to just train a classifier and call it a day.
top open one is trained by perplexity cto for $3k, kinda cool https://x.com/denisyarats/status/2102252088067850507
Bro is writing off the H200 lol
On a sidenote I really can't stand the term "swarm" and definately plays into AI doomerism.
Very small context window, but for some existing small llm work I was doing, it was a drop-in replacement and it makes me happy I can get use out of old hardware I have running.
> It is an independent project, not affiliated with Ollama.
(already merged)
GoModel (gateway) already supports Jev like endpoints too
We've stumbled into general differentiable models..