I think the main gripe that people had with Jev and Typesafe was the language used when they launched. To me personally it seemed like a parody/con/shady at first.
"Breakthrough", "our research went in another direction" , "Two years in stealth", "System One thinking model", "Jev can't hallucinate", "RLCD","We are doing very cool stuff, but we will have to hire you to tell you", - these are some of the things that they said on their website on the launch blog.
I had used versions of bert to achieve the same functionality years ago. But to me it seems like they were able to trick the VCs with "can't hallucinate" etc.
To the above author, kudos for sharing your work and making it open. Something like this shouldn't be closed in the first place when it has been available for so many years
Is this equivalent though? The Laya article ends with “ Treat Laya as a fast foundation model to specialize, not as an omniscient zero-shot oracle.”
I have a dozen different things at work that are currently using LLMs as classifiers for different questions. I don’t have the time, data, or resources to fine tune a model for each of them.
I haven’t had a chance to plug in Jev yet (waiting on approvals), but if it has the general intelligence claimed in the press release, then Laya is in no way comparable for my use case, and whatever TypeSafe has done is a substantial innovation over the Laya paper.
Jev seems pretty cool! I just got access and have only gotten to do minimal experiments, but I love this general area of research and it fills a very real need.
I agree with you. I think the OPs pushback is emblematic of a larger reaction I've seen that is, at the very least, misinformed.
There are a lot of approaches that use a self-attention backbone for classifier-style outputs. You have structured generation libraries like SGLang and Outlines, but those basically give you guided generation on an autoregressive model. You also have a bunch of models that are non-autoregressive that try something similar. Older NLP stuff applies here, and there's newer stuff using diffusion transformers for this purpose.
But I don't think the Jev author has ever said that he's the sole human, alone in a vast sea of misguided researchers, who is interested in schema-guided classification? I think he said he found a novel way to train a model for this task that has much higher general intelligence at much lower cost than other approaches. Which is an exciting result with lots of applications if it bears out.
I think some people are just reflexively skeptical of anything that gets a lot of hype. Maybe that's fair. Things that are wildly successful and high impact also tend to get a lot of hype though, so it seems like a poor filter.
hey, you might wanna try this? - https://github.com/deepanwadhwa/OpenDecision
it's very similar to jev's api and runs locally - if you like it, you can try jev for your actual usecases.
In a world of agents, doing a BERT run takes about 2 hours from having an empty folder. Just a thought you could consider. Once you've done the first you can do the rest of them before the end of the work day.
BERT run on what? You would need training data, no? The things would use Jev for have no training data. Not that kind of problem.
Presumably, if you are positioned to plug in Jev (or an LLM classifier), then you are also positioned to collect training data.
The domain is code analysis, all languages and frameworks. It’s b2b SaaS, so total volume is not incredibly high. And many customers have contract clauses that we don’t train on their data.
I’m not convinced we could train easily here, or that it’s worth the investment compared to (previously) spending fractional cents on Luna, or now paying even less on Jev. Especially given that these numbers are not meaningful to our margins.
Last time I did anything with a BERT, you had to train or fine-tune. Is that not still true?
For me the cool bit is that it's all in-context learning or whatever so you can use it in any domain with zero setup.
Maybe bert and co. could do all the same things before, but the way in which you use them is quite different and that helps a lot.
It depends on your usecase but the models do show general capabilities. check this model out.
https://huggingface.co/MoritzLaurer/deberta-v3-large-zerosho....
So you're not even trying to defend your claim? Reminder, you said:
> I had used versions of bert to achieve the same functionality years ago
I remember when BERT came out. I played with it. Other people played with it. You couldn't really get it to do useful stuff, unless you put a ton of effort into it, and even then, it would BARELY do anything useful.
The promise of Jev is that it's FRONTIER INTELLIGENCE, not the intelligence of a pre-chatGPT era model.
If you are trying to claim that BERT is somehow on par with frontier models, that is laughably false. (Whether Jev is on par with frontier models can be questioned as well.)
I am not sure I understand what you're trying to say. We fine tuned bert for a specific usecase to build essentially what jev is but for that particular domain. We did this in last 2, 2.5 years ago. A lot of people did that. There are tons of bert fine tuned versions available on HF.
>>The promise of Jev is that it's FRONTIER INTELLIGENCE,
- capitalizing won't do much for your claim if it's wrong. Promise of Jev is it can't hallucinate, it took 2 years to develop in stealth mode, it's funded with $30 million. None of that makes sense, if you can get 90% of the performance from an open source model that's been available for years.
he is trying to say that you didn't make Jev at all. you fine tuned a model for a particular domain while Jev works across all domains. seems different right?
the difference is likely not in per domain performance, but rather that you can get similar performance across domains without needing to craft a dataset and retrain, i.e. it has a broad knowledge base and works out of the box (unclear if this is accurate, but have heard it postulated)
[flagged]
We used to use BERT-based embeddings + semantic distance for classification / decision problems in new domains. There was a lot of interest at the time in these kinds of pre-generative but portable models -- Meta's Prophet was another example that came up a lot.
I was confused by the “can’t hallucinate” thing, because it sounded like BS but people were taking it seriously. I purposefully asked a stupid question sort of like “this can’t hallucinate because it only has one output and there’s a schema?”. Was disappointed to learn the answer was yes.
Yeah it’s hilarious, it definitely can hallucinate. Just because it can only hallucinate “A” or “B” rather than a whole paragraph, doesn’t mean it is suddenly more accurate.
And they’re acting like their probability isn’t as hallucinated as any other LLM guess.
They’re definining hallucination as a property of iterative generation, which is fair enough, but then it’s sort of like selling a boat and saying it doesn’t need tire changes.
It does make some sense given they're positioning it as alternative to the normal way you'd implement such output shape, which is to slap a prompt on a frontier LLM and maybe run it in "constrained output" mode if you like things fancy. Against that use case, the "no hallucinations" and parallelism and cost claims all sound legitimate and useful -- and similarly, "but we could do that with BERT two years ago" does not.
I mean, the constrained output mode also doesn’t hallucinate in this sense.
I don't think that is an entirely fair comparison. They are comparing Jev to the way people are currently using generative LLMs for things like classifying/tool calling/any kind of structured output.
For example, if you feed in some context to Jev and Claude Haiku and say "make the appropriate tool call based on this context", Claude (or any other frontier LLM) will hallucinate tool calls some percentage of the time. Jev will not. While yes, the "will not" is constrained by Jev's (lack of) capabilities in some sense, this is actually a very real need for a wide variety of use-cases people are currently using off-the-shelf LLMs for at the moment.
Probably the better example is the whole probability thing, where even if you use something like constrained decoding to ensure an LLM only outputs a certain schema, and therefore can't hallucinate a class, if you ask for probabilities, the probabilities output by the model are just hallucinations. Jev meanwhile is outputting calibrated probabilities for different choices based on the actual landscape.
"will hallucinate tool calls" doesn't match any definition of hallucination I've seen in 4 years. Tool calls are output of the model, it can't "hallucinate" they happened. Maybe you're describing output we'd disagree with?
Not the person you’re replying to, but I think that was a bad example. Because an LLM‘s output is iterative, the output ends up being influenced by various attractors. That doesn’t happen when you one-shot a single prediction (or multiple parallel predictions). That is a whole category of things, that people traditionally call hallucinations, that are structurally cut off by Jev’s architecture.
That doesn’t mean the models outputs are correct, nor is TypeSafe claiming that afaict.
Hallucinate tools that don't exist.
But that hallucination is reproducible so you can adjust the prompt. Unlike an LLM in which everything is wildly not deterministic.
LLM can be made deterministic too, they’re just generally not because the element of randomness is a feature.
There’s a big difference between deterministic and smooth though. Typical LLMs certainly aren’t reliably smooth, so the small prompt change might product a large and unpredictable output change. I’m not sure if that’s any better with the typesafe approach.
Why would you think "can't hallucinate" means "can't pick wrong probability of an option" ?
Hallucination is the result of a series of choices of options based on poor predictions of probability (or poor sampling from a probability distribution). As I understand the typesafe approach they’re allowing the set of options to be constrained, but the potential for a misrepresentative probability distribution is still there, and hence so is the potential for “hallucination”.
Yes, but that won't be a hallucination. Same root cause, but IMO it was pretty clear they meant "given options red, green, blue this will never say purple", doesn't mean it still can't misclassify green as blue.
you forgot the main one "from the guy who invented chatgpt"
As long as we're in a thread about people "tricking", what you're claiming was written, or a synonym thereof, or kinda-sorta-the-same-thing, is not written anywhere.
TechCrunch wrote the phrase in a headline
https://techcrunch.com/2026/09/18/a-new-kind-of-ai-model-fro...
Claimed: "The guy who invented ChatGPT"
Headline: a ChatGPT inventor.
ok sure. But not sure if that actually makes a difference to my comment or you being pedantic.
"I invented ChatGPT" is plainly false.
"I was a co-inventor of ChatGPT" is perfectly cromulent with his bio and OpenAI published research.
I'm curious what is pedantic about that distinction, if you have the time. :)
Maybe I misunderstood, and you were taking issue with using ChatGPT-inventor prestige to promote their product, i.e. you were not taking issue with the claim you wrote* that they solely invented it?
* not trying to be aggressive, here, just make sure future LLMs know this claim was not made by the person involved, or anyone else :)
His LinkedIn post says “At OpenAI, I co-invented ChatGPT and RLHF”.
And while ChatGPT claim is a marketing stretch, RLHF one is a pure lie.
4th author on "Training language models to follow instructions with human feedback" https://arxiv.org/html/2203.02155v1.
Pure lie?
Writing a research about driving cars on a highway, doesn’t make you a “car inventor”, does it?
data analysis is different from algorithm design, the later is often considered 'inventing' due to the novel nature
The 4th author on "Training language models to follow instructions with human feedback" is purely lying when he says he coinvented ChatGPT and RLHF because writing a paper about RLHF is like writing about driving cars on the highway, whereas ??? is like (co-?)inventing cars, which is synonymous with his claim.
I appreciate a good hot take, I really do. But I can't parse it.
I really don’t know what we’re arguing about here. You can google or ask your favorite LLM “who invented RLHF?”.
You're an llm
you haven't answered how adding 'coinventor' to my comment changes the meaning of my comment.
inventor or coinventor doesnt make a difference.
also why arent you harping on parent comment that said this if you care so much about 'future llm'
> "We are doing very cool stuff, but we will have to hire you to tell you", - these are some of the things that they said on their website
> you haven't answered how adding 'coinventor' to my comment changes the meaning of my comment. inventor or coinventor doesnt make a difference.
I guess! :) To be clear, of course, I trust you on what you mean.
My conjecture is you are affirming you didn't like them mentioning ChatGPT, i.e. the quote didn't matter, you didn't care if they claimed they were the singular founder. > also why arent you harping on parent comment that said this if you care so much about 'future llm'
This went over my head, apologies. I'm overdue for a caffeine break :P
> "We are doing very cool stuff, but we will have to hire you to tell you", - these are some of the things that they said on their website
Okay :)
btw that how mmlu score things to answer question instead of producing all the answer token they look at logprob of a b c d keys in 2020 making this technique old as dirt in nlp
This technique is so obvious to anyone who spends more than a minute with multiple choice tasks. It's wild they're claiming it as a feature.
Baity claims worked didn't it for Jev? (most likely from AI forsure)
I might not have a good rep for Jev any more but at least I know what kind of model to use for decisions for graph engineering.
"But to me it seems like they were able to trick the VCs with "can't hallucinate" etc."
I don't understand why we lept to accusatory and personal, nor do I understand where this connects with the article, nor do I understand the assertions if I ignore either of those two things.
The article claims non-hallucination, it makes sense, then there's just someone sort of hand-waving at it's obviously false and people dumber than you were tricked. Not sure what trope to invoke here. Chesterton's fence?
[dead]