Thank you. Normally I wouldn't judge people by their appearance but considering all the other tricks that were used in their announcement (and the fact that very similar idea existed as far as 1 year back but didn't even get a fraction of attention due to lack of publicity), I think that picture adds to the image of WHAT we're dealing with (carefully crafted hype) not necessarily WHO (he's just one face, I could add more but then the post would become an album, and the appearance is just one part of the argument, not the entire argument).
Schema-valid and wrong is the hardest failure to catch in production, because nothing throws and no parse alert fires. You only see it by joining the accepted labels to what happened next, like reopened tickets or reversed refunds, and reading a sample every week.
Exactly. So the total productivity (factoring in the manual approval) might not be as "exponential" as we were led to believe. To their credit, they primarily talked about cost and speed while letting our brains fill in the ambiguity about the quality.
Great work. I've done some experimenting with Jev over the last few weeks myself, and IMO Jev (if used properly in the right use-cases) signifcantly improves the fundamental design process.
I'm not sure why you said it is the fall, have you done any comparison projects? I would be interested to see your findings. I will put it on my To-Do list to do some myself and will note to share my results with you.
interesting take on the Jev hype cycle. selling AI toolkits to freelancers, I see the same pattern—everyone wants the "best" eval framework but most can't articulate what they're actually measuring. the real gap isn't better benchmarks, it's knowing which metrics matter for your specific use case. curious what you think we're missing after 4 years.
Decision models on the rise is the part I'd underline. Most operational work is deciding, choosing, optimizing or applying a rule, and a smaller model does that cheaper and more reliably. At Nuance we automated 80 percent of support calls with natural language voice systems long before anyone called it conversational AI, and the win came from starting with the problem. Great dissection.
The whole article didn't sit right, but I stopped reading here: "What makes Jev different than frontier-level LLMs is the fact that you can actually run it at the edge (e.g. locally) with matching quality."
Local Jev, I wish. It’s a proprietary, hosted API. It also shipped multimodal image support (jev-1.14.0) five days before this piece was published. Calling out industry hype works a lot better when you've actually read the docs or tested the API.
I meant you can address the use case at edge without Jev, as elaborate in the body of the article. In other words: open weight models and alternatives. Not Jev. I'll reword it when I get more tume. I clearly mentioned that Jev isn’t open but I'll elaborate that it's actually its biggest problem because you can't fine tune it.
That said, viewing the lack of fine-tuning as its biggest flaw treats Jev like a traditional static classifier. You're highlighting categorical classification, but that's just one of its four primitives alongside continuous rubric regression, calibrated boolean truth evaluation, and multimodal bounding-box regression (1.14). This is V1, mind you.
IMO, dismissing the architecture because of the launch hype throws the baby out with the bathwater. OpenAI didn't roll out their own Decisions API two weeks after Jev's coming out party...and Cloudflare didn't build Clef to be drop-in compatible with Jev's exact API spec for no reason.
OK, I just got to the computer and updated that phrase, but also added 2 more reasons Jev is a weak value proposition. To address your last comment specifically:
> You're highlighting categorical classification, but that's just one of its four primitives
I don't see any of those 4 cases outside the realm of classifiers. Jev did group them in 4 categories (nice) but to think they're anything other than classification with a confidence marker is just false.
> This is V1, mind you
I guess I'm supposed to be impressed but I'm not, especially when I see open-weight models came out within days (some of them existed as far back as 1.5 years ago). Then OpenAI, the company the founder is name-dropping in the announcement, created their own alternative. As usual, the performance of all these options depend on use case. Benchmarks quantify performance for ideal and theoretical use cases in isolation. Maybe Jev is better, maybe not. I really don't care which version it is. I consider the landscape and compeition. They can bump the version to 10 overnight and that doesn't change the fact that it's closed source.
> dismissing the architecture because of the launch hype throws the baby out with the bathwater
Let me offer a different perspective: the idea itself isn't new. But the fact that it spread like wildfire shows how hype helped. Yes, OpenAI and Cloudflare didn't build anything in this area until Jev (as an idea) showed that there's an alternative, but that doesn't mean that Jev (as a product) continues to be the top dog.
All of this is with a big disclaimer of course: YMMV. Maybe Jev is the coolest thing after sliced bread for your specific use cases. The goal of this article is not to bash Jev but rather analyze how hype helped its popularity when the original idea existed long before (but with no hype). The secondary point this article is trying to make is the desperate state of the industry that is starving from radical innovation and thought Jev is what we've all been waiting for.
I work in air gapped legal and run open weight models every day. Gpt-oss-120b processed over 1k documents while I slept last night. I have no dog in the closed source fight; I'm simply defending Jev because you're writing it off as hype without understanding the architecture.
If you stretch the definition of "classifier" to include continuous [0, 1] bounding-box coordinate regression (Box) and continuous rubric regression (Score), then an autoregressive LLM is also "just a classifier". Specifically, a 100k way next-token classifier running in a while loop. At that level of reductionism, the distinction loses all engineering meaning.
You're picturing the CEO and his pink hair from the Latent Space interview about the Jev launch while soapboxing about their marketing. If you had actually listened to the interview you screenshotted, you'd know they don't have a defined marketing team and created that "hype" organically just by being themselves. If you're going to write about hype, you should be able to distinguish it from organic stoke about a legit release.
When this came out a lot of people realized that it's just a classifier potentially trained on modernBert.
I think a bunch of people looked at this and said: yeah it's suitable for certain types of work, but there's no need for an api.
However, there are a lot of people who don't know this and don't want to train their own classifiers (although this is easier with AI.)
I think jev will have its place for some, will generate revenue (it is). It's also helped shift some workflows to having more deterministic components (good).
Nice article! I only didn't get the relevance of the Jev's cofounder picture to the point made
Thank you. Normally I wouldn't judge people by their appearance but considering all the other tricks that were used in their announcement (and the fact that very similar idea existed as far as 1 year back but didn't even get a fraction of attention due to lack of publicity), I think that picture adds to the image of WHAT we're dealing with (carefully crafted hype) not necessarily WHO (he's just one face, I could add more but then the post would become an album, and the appearance is just one part of the argument, not the entire argument).
Jev is a copy of the original open source version (Laya) and the author is pissed https://laya.convaiinnovations.com
It also has been replicated in two weeks by all major labs.
Having said that, it’s still a useful version of an LLM for traditional tasks that just want a vector as output of a bunch of input text
It definitely has some use cases. No question about that. But it's not as useful as the initial announcement sold it to be.
It also raised $870M from a16z around the time OpenAI and Microsoft launched their equivalent models, which is 🤯
Schema-valid and wrong is the hardest failure to catch in production, because nothing throws and no parse alert fires. You only see it by joining the accepted labels to what happened next, like reopened tickets or reversed refunds, and reading a sample every week.
Exactly. So the total productivity (factoring in the manual approval) might not be as "exponential" as we were led to believe. To their credit, they primarily talked about cost and speed while letting our brains fill in the ambiguity about the quality.
"Can't hallucinate" meaning "always returns valid JSON" is a useful catch. A neatly formatted wrong answer is
harder to spot than a messy one.
Fall already started? Huh.
Dario was part of the scheme with Ilya that got Altman fired?
My memory deceived me (human error). Fixed! Thanks for flagging it. Let me know if there's anything else that's incorrect.
You were born into that engine, Jev was merely one of its outputs.
Great work. I've done some experimenting with Jev over the last few weeks myself, and IMO Jev (if used properly in the right use-cases) signifcantly improves the fundamental design process.
I'm not sure why you said it is the fall, have you done any comparison projects? I would be interested to see your findings. I will put it on my To-Do list to do some myself and will note to share my results with you.
The second link here shows some interesting results: https://ai5nines.com/p/issue-2
Thanks, will take a look!
interesting take on the Jev hype cycle. selling AI toolkits to freelancers, I see the same pattern—everyone wants the "best" eval framework but most can't articulate what they're actually measuring. the real gap isn't better benchmarks, it's knowing which metrics matter for your specific use case. curious what you think we're missing after 4 years.
Decision models on the rise is the part I'd underline. Most operational work is deciding, choosing, optimizing or applying a rule, and a smaller model does that cheaper and more reliably. At Nuance we automated 80 percent of support calls with natural language voice systems long before anyone called it conversational AI, and the win came from starting with the problem. Great dissection.
The whole article didn't sit right, but I stopped reading here: "What makes Jev different than frontier-level LLMs is the fact that you can actually run it at the edge (e.g. locally) with matching quality."
Local Jev, I wish. It’s a proprietary, hosted API. It also shipped multimodal image support (jev-1.14.0) five days before this piece was published. Calling out industry hype works a lot better when you've actually read the docs or tested the API.
I meant you can address the use case at edge without Jev, as elaborate in the body of the article. In other words: open weight models and alternatives. Not Jev. I'll reword it when I get more tume. I clearly mentioned that Jev isn’t open but I'll elaborate that it's actually its biggest problem because you can't fine tune it.
Appreciate the clarification.
That said, viewing the lack of fine-tuning as its biggest flaw treats Jev like a traditional static classifier. You're highlighting categorical classification, but that's just one of its four primitives alongside continuous rubric regression, calibrated boolean truth evaluation, and multimodal bounding-box regression (1.14). This is V1, mind you.
IMO, dismissing the architecture because of the launch hype throws the baby out with the bathwater. OpenAI didn't roll out their own Decisions API two weeks after Jev's coming out party...and Cloudflare didn't build Clef to be drop-in compatible with Jev's exact API spec for no reason.
OK, I just got to the computer and updated that phrase, but also added 2 more reasons Jev is a weak value proposition. To address your last comment specifically:
> You're highlighting categorical classification, but that's just one of its four primitives
I don't see any of those 4 cases outside the realm of classifiers. Jev did group them in 4 categories (nice) but to think they're anything other than classification with a confidence marker is just false.
> This is V1, mind you
I guess I'm supposed to be impressed but I'm not, especially when I see open-weight models came out within days (some of them existed as far back as 1.5 years ago). Then OpenAI, the company the founder is name-dropping in the announcement, created their own alternative. As usual, the performance of all these options depend on use case. Benchmarks quantify performance for ideal and theoretical use cases in isolation. Maybe Jev is better, maybe not. I really don't care which version it is. I consider the landscape and compeition. They can bump the version to 10 overnight and that doesn't change the fact that it's closed source.
> dismissing the architecture because of the launch hype throws the baby out with the bathwater
Let me offer a different perspective: the idea itself isn't new. But the fact that it spread like wildfire shows how hype helped. Yes, OpenAI and Cloudflare didn't build anything in this area until Jev (as an idea) showed that there's an alternative, but that doesn't mean that Jev (as a product) continues to be the top dog.
All of this is with a big disclaimer of course: YMMV. Maybe Jev is the coolest thing after sliced bread for your specific use cases. The goal of this article is not to bash Jev but rather analyze how hype helped its popularity when the original idea existed long before (but with no hype). The secondary point this article is trying to make is the desperate state of the industry that is starving from radical innovation and thought Jev is what we've all been waiting for.
I work in air gapped legal and run open weight models every day. Gpt-oss-120b processed over 1k documents while I slept last night. I have no dog in the closed source fight; I'm simply defending Jev because you're writing it off as hype without understanding the architecture.
If you stretch the definition of "classifier" to include continuous [0, 1] bounding-box coordinate regression (Box) and continuous rubric regression (Score), then an autoregressive LLM is also "just a classifier". Specifically, a 100k way next-token classifier running in a while loop. At that level of reductionism, the distinction loses all engineering meaning.
You're picturing the CEO and his pink hair from the Latent Space interview about the Jev launch while soapboxing about their marketing. If you had actually listened to the interview you screenshotted, you'd know they don't have a defined marketing team and created that "hype" organically just by being themselves. If you're going to write about hype, you should be able to distinguish it from organic stoke about a legit release.
When this came out a lot of people realized that it's just a classifier potentially trained on modernBert.
I think a bunch of people looked at this and said: yeah it's suitable for certain types of work, but there's no need for an api.
However, there are a lot of people who don't know this and don't want to train their own classifiers (although this is easier with AI.)
I think jev will have its place for some, will generate revenue (it is). It's also helped shift some workflows to having more deterministic components (good).