The link rgbrgb posted is a good overview. The best open ones are close to Jev now, but they're big models. And I agree, if you have an eval set for a fixed task, a trained classifier is the better choice.
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
My understanding is that Laya (or whatever it was called in 2025) was yet another fine-tuned classifier, not a general purpose one.
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 am fairly confident if Jev-style decision models are seen as prominent (which, they seem to be), Ollama will support them. Surprised the team hasn't implemented this already.
Depends on the model. The small ones I support today are well below Jev on harder queries, but fine for simple, well-defined questions. The open models that get close to Jev are bigger, and I'm adding support for those next.
Has anyone actually seen better or the same results with Laya compared to Jev? From my experience, Laya performs significantly worse. It's less confident and often makes wrong decisions with more complex queries.
Developer here. You're right, Laya is a lot weaker than Jev, especially on harder queries. It's a small model, so it's fast, but that's the trade-off. The open models that get close to Jev are much bigger, and running those is what I'm working on next.
Probably Kev and/or the decider models. Kev is trained on one of the 4B qwen models, similar for decider but it ranges from 0.8B through to the 35B-A3B model so far I believe.
Yes. JEV generalizes better because they probably have an enormous corpus and trained on it for a long time. Laya's out of the box model is much weaker. However, in the age of LLM's it's incredibly easy and cheap to generate large datasets to fine tune laya for your task, and the training loop is pretty quick and cheap too.
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.
Isn't the point of Jev that it generalises better?
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
I didn't see any negativity in the post you replied to.
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.
I think your point is valid but many are annoyed that it is presented as groundbreaking, revolutionary, novel frontier tech when it is a known classification system. It’s the hype that feels undeserved. Honestly it was one of the best marketing campaigns I’ve seen.
You should try building something with it, the hype is what it is, but the model is crazy useful.
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.
Nothing yet. Unfortunately it sometimes feels like our industry has been overrun by grifters and chancers.
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.
Did anyone else notice the huge gap between scores on private vs public for ALL Jev-like models compared to LLMs (such as GPT Luna)? Doesn't it mean those models aren't generalizing so not very useful on data they haven't seen?
one day, perhaps people will click through to the laya author's arxiv paper content and the why may become clearer, you won't have to read it, a skim will suffice
This is just anecdotal and I might be doing it wrong but I made jev and laya versions of a simple semantic grep tool (https://github.com/lgastako/jevplay) and played with them a bit, and at first it seemed like laya was comparable (eg on queries like "this is a mans name" or "this is a womans name" on names.txt) but the more I played with it, eg. "this is a vegetable" on foods.txt the further the gap widened in favor of jev. Then I started trying variations of the query eg simply "mans name" and for the most part laya just fell apart and didn't return anything useful for a lot of stuff. I was hoping to find that laya was competitive because it's much faster to have the model running locally but it's just not, yet.
Calibrated probability across multi task with zero shot I guess. A reranker is single task and tuning it make it even more narrow. And I guess some piping to make multiclass efficient since you cannot mask logprob for independent questions in the same output space without throwing calibration away.
If you mean that I can compare the probilities between different tasks fair. This is a thing you can not do with a reranker. This sounds cool but not the amount of hype we got cool
> difference between an instruct based re-ranker and laya/jev I just don't see it
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
You now have a system that has produced useful results in production from the very beginning and by the end it's a reliable super cheap classifier that can make thousands of decisions per second.
I don’t think that’s it. I sincerely doubt most developers are doing side by side comparisons of calibration quality.
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.
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.
Exactly. Just think about a discord or twitch moderation bot that screens messages in real-time. When the streamer starts playing a game, the whole context switch in an instant and comments like «kill them» will suddenly have a whole different meaning.
Yea but a re-ranker can do that too if instruct trainrd. You update the instruction on which it is ranking the the values, a.k.a it is classifying them.
I'm not sure what this means for AI startups if their innovations can be copied by OSS so quickly (what, like 2 weeks?). There's "consumer surplus" for everyone, to borrow an economic concept. But we do ideally want some of the surplus to flow to the innovator, too. I know there were precursors, but that's fine - it's hard to have a totally novel idea in such a popular field. I don't know what the end game is for TypeSafe - they'd need to demonstrate perpetually better results, or compete in another axis: UX, support, custom solutions, etc. So much of the time, someone proving a concept, or it simply getting enough publicity, is enough for a "Cambrian explosion" of follow-ups and copies. Famously, that was true for "Attention is All You Need", and the general idea of "next-token prediction" being so powerful.
We've stumbled into general differentiable models..
Presumably the training recipe and training dataset itself cannot be easily copied in a week or two. So if they want to shut down these competitor models they need to make it obvious how they are better than them.
It a paradox when the article is claiming the prior art is absurd, but then goes on to analyse the one side and compare it to another for which most of the values (except scaling the concept) are unknown. And even for scaling, it uses the first, pre-laya instance to judge the limited schema, while overlooking that Laya is just doing this scaling. Important to note that prior art is not having built the exact same thing.
There is definitely more to the story. There is a huge financial interest for each side to discredit the other. Fact of the matter is, we still don't know who will prevail. These are cutting edge tech stacks and they were just released.
Because what they did is kinda trivial. Its basically like the Dropbox comment really[0], except here you don't need petabytes of storage and infinite VC pockets.
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.
I think GP has a point, though. BERT models have existed for a while but OpenAI made classification via LLM convenient and accessible for regular developers.
People didn’t know they wanted classifiers until OpenAI gave them a taste.
I don't know if that first part is true? Classifiers were/are one of the dominant applications of classic ML and neural networks, especially in production. Even today, image classification, object recognition, language detection, segmentation models etc are still super common.
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.
I think the answers are "a lot" and "a lot more"? But what I'm saying is that Jev's virality isn't because people didn't have access to classifiers before it. Jev's general claims about capability and performance vs. cost would have been a very big deal 5 years ago too. In a vacuum, the idea that you can get a general classifier that is very accurate across any domain and on any modality with minimal latency and a very low price point is wild.
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.
The answer for why it wasn't productized might just be pretty straightforward.
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.
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?
you could help it along by providing clues about the domain in the prompt. including what various proper nouns might indicate (assuming they're not chosen completely at random, and encode at least some meaning).
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).
If it's not trained on my data, it's going to be something on the order of a 500-shot prompt. This is a complex application with more corner cases than I'd like.
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.
It's not magical and you won't get good results if you ask a question that is that vague. Instead you will want to decompose it in to decidable questions that make up "is urgent" and if you find a case that corresponds to "is urgent" you add a new question to the panel that surfaces that.
Theorizing, but I'm guessing that before LLMs, people weren't using anything for situations like these. As people started using LLMs, people's use cases grew, LLMs got slower and more expensive, and people lost the wow-factor and are now worrying about price. Great timing to launch a product like this, where certain use cases can be distilled down into something faster and cheaper.
There's probably other areas where people are using LLMs where a more tailored ML solution might work better.
I am looking forward to some hotshot implementing a jev-style model and then looking like a cost cutting, performance boosting wizard by replacing it with a logistic regression.
The moat is the RL synthetic data pipeline they set up to train jev. Open sourcing that would be the coup, not the model architecture and training scripts, which are trivial.
Today we might need to evaluate “innovation” in a new standard, and have a different expectation for what innovator would be awarded. Getting public attention in such an era where innovation happens every a few days could’ve already been something precious. And that attention would allow TypeSafe to be heard easily next time. Like OpenAI, Anthropic, or any others, they launch frequently but still each time they launch something new, that would hit headlines. I think that’s the “surplus” flown to innovators today.
I think the idea of a "feature startup" is dead. What used to be a niche subscription business is now an individual Epic level of work. The smallest viable business becomes what two or three years ago was a mid tier enterprise. It is no longer "look at this tool I maintain", but "we take this specific approach using these hundreds of tools merged together to solve a problem in a specific way that nobody is going to compete with. Not because they can't compete if they wanted to, but that the competitions approach diverges in fifty different chosen ways that they are targeting a different market segment essentially."
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.
> I'm not sure what this means for AI startups if their innovations can be copied by OSS so quickly
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.
Laya is pretty easy to set up on its own without ollaya. I just did that and replaced my current jev API usage to laya running on a GTX 970 with 4GB of vram.
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 would be good to list 1) zero-shot accuracy and 2) latency on the models page . The LLM-based models' latency is probably much higher than the BERT approaches I would assume.
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?
I installed it, I tried the examples, it works.... But forgive my lack of imagination... what is this useful for?
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).
I have a lot of semi-practical examples of how you can use this model wrapped in unix-ish tools - https://github.com/aurorainfra/grev (readme links to docs of each tool with some more or less practical examples)
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 gives users past few commands' etc.
Ok, those are pretty decent examples, and clears up the utility a bit: speed and tokens. Some of it's still a bit iffy (e.g. `cutv 'email address' 'phone number' < examples/users.csv`, csv is already in columns), but I can see using it for some niche queries. Neat tool.
I very much appreciate your to-the-point, non-vibed README as well, ty for that.
Yeah, speed is the one, I believe the default TypeSafe API quota is 1.5-2k queries per second (batched in bigger requests).
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..
I still think that `churn_risk` above is incorrect and unacceptable (perhaps there are sensible fixes, but saying "no churn risk" about a refund, in a leading example on their homepage, flabbergasting).
But if that were solved, I could see giving ollaya/grev to LLMs themselves, giving LLMs their own massive token-saver.
Is for when you want an AI to make a decision. If you have been using gpt or claude or open source models for that, than it’s a way cheaper alternative.
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.
If you don't want to spend a lot and want low latency, e.g. home automation. "It's cold and dark in here, do something about it", it will then turn on the lights and heater nearly instantly.
For everyone dismissing Jev's innovation as being trivial, no it's not.
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.
Just because it is zero-shot does not mean that Jev is an architectural innovation, especially in the year 2026.
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.
To get a classifier that worked in 2019, you had to train a model. Either from scratch or from a starting point.
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.
Of course it costs less and is more accurate. There have been several years of language model training, distillation, and reinforcement learning to get there. None of this is thanks to the Jev team.
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.
LLMs better that 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.
125 comments
[ 1.4 ms ] story [ 53.5 ms ] threadSmarter move if you have an eval set is to just train a classifier and call it a day.
>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.
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
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.
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.
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
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.
https://benchmarkheaven.com/jev-models
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?
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)
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:
You now have a system that has produced useful results in production from the very beginning and by the end it's a reliable super cheap classifier that can make thousands of decisions per second.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.
> It is an independent project, not affiliated with Ollama.
[1]: https://ollaya.dev/docs/faq
We've stumbled into general differentiable models..
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.
[0] https://news.ycombinator.com/item?id=9224
Probably because doing it wrong (using an llm in place of a classifier) is more profitable? (For the people selling inference.)
There have been tons of applications for this. People were using earlier LLMs like BERT for classifiers long 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.
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
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.
There's probably other areas where people are using LLMs where a more tailored ML solution might work better.
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.
My thoughts too. It sounds like a minor feature being framed as a whole new business.
Then again, Dropbox and Docker are too.
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.
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.
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?
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 gives users past few commands' etc.
I very much appreciate your to-the-point, non-vibed README as well, ty for that.
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..
But if that were solved, I could see giving ollaya/grev to LLMs themselves, giving LLMs their own massive token-saver.
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.
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 that 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.