Non-autoregressive System 1 decision engine. Typed choice, score and yes/no decisions over any text in a single forward pass, in 100+ languages, with a router that picks the right checkpoint per re...
Isn’t “hallucinating”, just a term for an unjustified error? So why does the author think that only outputting numbers, and selecting a prewritten choice make it immune to hallucinations? It seems like if I had an LLM classify apples and oranges, it would be perfectly capable of calling an apple an orange.
This article implies that the Jev approach is being used to help decide between a set of choices. This goes back to the old heuristic approach of earlier AI, with a dose of loose matching thrown in.
Where it feels like it would be handy is when you enter a free-form request, but the response is bounded to a finite set of responses. Hallucination doesn’t enter into it because the system is told to pick from a set of weighted choices.
I’m still trying to figure out the limitations, but it smells like a return to old-timey heuristic-based AI and fuzzy logic.
LLMs look for the next highest probability token to output. It doesn’t matter if it’s ‘true’ in the knowledge sense. That’s how you end up with a ‘hallucination.’ It’s whatever is in the training data set. Throw in a small degree of ‘randomness’ each time so the answers aren’t always the same, and you could end up with a good answer, a slightly wrong answer, or a wildly wrong one.
RAG embeddings limit the set of possible responses to what’s in a document set. So there’s a better chance you’ll get something anchored in data you’ve provided, but it still relies on next-token probabilities so it can stick whatever token has the highest priority next.
What this Jev business does (AFAIK) is limit the response set. So the answer can be, say, one of four possible options, based on training data. It’s much closer to how old expert systems worked, where you trained it on a large knowledge set and it tried to find the best path through heuristics. Here, since you’ve limited the answer space, it can also come back much faster, burning less GPU power, memory, etc.
I’m still trying to figure out the applications and pros and cons. Everyone should do their own research.
Looks interesting. Background article: https://dev.to/nandakishor_m_6cc0adfde9f/i-built-non-autoregressive-decision-models-a-year-ago-then-a-frontier-lab-called-it-a-18me
Isn’t “hallucinating”, just a term for an unjustified error? So why does the author think that only outputting numbers, and selecting a prewritten choice make it immune to hallucinations? It seems like if I had an LLM classify apples and oranges, it would be perfectly capable of calling an apple an orange.
https://www.forbes.com/sites/josipamajic/2026/09/19/jev-cuts-ai-decision-costs-100x-and-vercel-cloudflare-rushed-to-add-it/
This article implies that the Jev approach is being used to help decide between a set of choices. This goes back to the old heuristic approach of earlier AI, with a dose of loose matching thrown in.
Where it feels like it would be handy is when you enter a free-form request, but the response is bounded to a finite set of responses. Hallucination doesn’t enter into it because the system is told to pick from a set of weighted choices.
I’m still trying to figure out the limitations, but it smells like a return to old-timey heuristic-based AI and fuzzy logic.
Okay, I haven’t kept up with the jargon for LLMs does hallucinate only refer to text outputs?
LLMs look for the next highest probability token to output. It doesn’t matter if it’s ‘true’ in the knowledge sense. That’s how you end up with a ‘hallucination.’ It’s whatever is in the training data set. Throw in a small degree of ‘randomness’ each time so the answers aren’t always the same, and you could end up with a good answer, a slightly wrong answer, or a wildly wrong one.
RAG embeddings limit the set of possible responses to what’s in a document set. So there’s a better chance you’ll get something anchored in data you’ve provided, but it still relies on next-token probabilities so it can stick whatever token has the highest priority next.
What this Jev business does (AFAIK) is limit the response set. So the answer can be, say, one of four possible options, based on training data. It’s much closer to how old expert systems worked, where you trained it on a large knowledge set and it tried to find the best path through heuristics. Here, since you’ve limited the answer space, it can also come back much faster, burning less GPU power, memory, etc.
I’m still trying to figure out the applications and pros and cons. Everyone should do their own research.