I saw a lot of ppl think about what jev could use under the hood and could someone explain why this can't just be an embedding model where we just embed all the input + decisions and give back the cosine (or whatever) similarities?
if you compare an embedding model to something like Jev which asks 100 questions and use the answers as the embedding you will be able to get move mileage out of the latter. especially because you don't need to train any classifiers for your task, you can work directly on the answers.
that said, I don't understand the hype. I have been doing what Jev does for 2 years now by just forcing json tokens onto an LLM. you can even get the LLM to think. and you can ensemble multiple LLMs.
I suppose the appeal of Jev is how cheap and fast it is, but then it's entirely unsuitable for anything but the most cursory extraction. using it to play games seems like a waste of time especially when most of those games will be played better by an algorithm written by an LLM (just give it the state and ask it to write a bot).
> I suppose the appeal of Jev is how cheap and fast it is
Ok, so you do understand the hype. I mean cost & speed are reaaaaally big issues for normal LLMs. If you can fill a specific use case way cheaper and way faster, that’s a fantastic development. We need more niche tooling that’s more efficient and better for niche use cases. Not everything has to be general purpose.
Not clear to me that playing games is the point of jev. This post is just a random fun experiment someone did because they wanted to.
But I can ask 100 causally masked questions against common prefix, and get 100 answers, all in a single PP pass using any existing "classical" attention transformer model? Like, I had the impression that is what everyone was doing for classification already?
Is the difference "we did RL to tune logit distribution"? Because I really do not see anything new there. What is the difference?
Embeddings just convert the tokens to a vector that represents the text in an abstract semantic space. JEV goes a step further and actually processes the instructions/meaning of those embeddings to produce output, just not the usual series-of-tokens output we expect from an LLM.
My day job is improving data quality in big enterprises (we build products for this), and I see where this article is going, it's a bit watered down, but I see this every day at our customers, most companies just don't have a simple way to query their own data, create metrics, get ai to actually do stuff cause data is spread between 3-10 systems where 4 are legacy and don't even have an API and even if they can access they have like zero context 100 custom fields and suffer from no first time right inputs from the last 30+ years.
(ofc. I see only the ppl who need this kind of service so I'm biased, and maybe it's not a wide spread issue)
I have used the flash model for over 3b tokens and ofc. I saw some hallucinations and premature termination (I also get this on Astra - way more often than with deepseek v4 flash), but I never had a infinite loop (using the copilot as harness).
I just want to put this out there, smolmachines is a wonderful program to solve this, I use this mostly for stuff needing docker socket / docker in docker (example strix and agents). (I'm using podman on my host)
Me and my wife share calanders both ways between Google cal and iCloud cal its an absolutely shit show to setup, and I also rember having to use the Web app for certain settings how can we not have solved this by now?
reply