Jev 1.13

“I think their blog post is actually really really good! The strongest, most positive description of this is probably something like "it can understand a generic context like an LLM can, but will reliably spit out a decision within a defined range based on those inputs". Also it's pretty fast.”

“I’ve been running Jev inside Grandmaster Spas for about five hours. Two things immediately stand out: speed and cost transparency. The cost sits right next to the decision. That sounds simple, but it matters. Many US AI models still hide the real cost of a task—or bury it on a pricing page you can’t interrogate directly from the chat. With Jev, I can see what the intelligence is doing and what that decision costs. That’s how AI should work in production. Congratulations to the team behind Jev. You’ve built this the way production software needs intelligence to behave.”

“5 hours into TypeSafe / Jev. Not another chatbot. A decision engine that tells me the answer and the bill. Latency in milliseconds. Cost on the same screen as the result. No “it depends”. No hunting through pricing pages. No asking the model what it just charged me — and getting a paragraph back. Most US labs still hide that. Or bury it. Or make you calculate tokens after the fact. This is what production AI should have looked like from day one: fast, typed, priced in the open. Congratulations to the TypeSafe team. You built the layer the chat models keep pretending they are.”

“Jev classified 1,700 emails for 18 cents. I needed 28 minutes to classify what Jev actually is. The core idea is beautifully simple: give it context + a fixed set of possible decisions → get probabilities back, insanely fast and cheap. Basically a probabilistic decision layer for your software, while LLMs handle the deeper reasoning/generation. Great demos, but the explanation took so many scenic routes that Jev itself could probably have clipped the best 5 minutes out of this video.”

“That's probably one of the best videos I've seen so far on the topic. Thank you so much. This was incredibly well explained and illustrated. Very cool examples. making deterministic or or near deterministic tool calls by using JEV to create a proper API call that was just awesome.”

“first impression on Jev is that it's worth taking seriously, this really is fast like a freak”

“They have various benchmarks, e.g. how much time it takes them to do wikipedia page -> page games. Jev seems to take the same or fewer hops but in ~10x less time and for ~10x less money. It's totally reasonable to compare against LLMs doing chain of thought if it gets comparable performance.”

“i should be happy enough about the pricing, like for as intelligent and general as it is, it's unbelievably cheap, basically free. after seeing the doom demo i got excited about the speed too, but that 100ms seems to be the floor for jev. i have things i'd like to do with a jev lite when available.”

“I'm biased but I wouldn't call it misleading - generating text is super awesome and flexible, (we describe that in the blog post - and I personally use string models all the time) but it's true you pay a high tax for autoregressive generation > Also "can't hallucinate" seems wrong? Sure, it can't emit an invalid type, but it can still emit a completely wrong valid value. that is likely true of all ML! perhaps we could debate semantics, but I don't think it's fair to say a random forest "hallucinates" in the way LLMs do”

“Imagine what if we integrate both the jev and exiting llm, if i'm thinking right, the jev can decide which function functions needs to be called, with which parameter and we can just fed the whole prompt and the tool calls result to the llm, imagine how much cost it would reduce.”

“I saw this but it just feels like a nothing burger. It’s super dumb can’t do text it’s his a classifier. It’s a defussion model that’s can only make decisions but it’s not smart at all”

“Bare in mind: policy explicitly says “We will not train or fine tune any AI/ML models on your prompts or other Input”, but retention is standard: “as long as reasonably necessary”. I also did some testing and outside extremely fast webui browsing, it didn't do much when it comes to triaging, rag, or filtering. The confidence scores were unreliable and in many cases unrealisticly high or low.”

“I've not seen a single video on Jev where it actually writes some code, or fixes bugs. It's all silly examples of multiple choices answers with a score. That is totally useless to me as a coder.”

“I’m going to be honest. This is overhyped x1000 because of the seed funding . The issue is that there are already very specialized small models that do the same thing, way cheaper (free), and faster. Take a look at these: small models executing this locally are ⁠MoritzLaurer/DeBERTa-v3-xsmall-mnli-fever-anli-ling-binary⁠, ⁠answerdotai/ModernBERT-base⁠, and ⁠cross-encoder/ms-marco-MiniLM-L-6-v2⁠.”

“What day-one hype usually gets wrong Benchmarks are the vendor's, at version 0.01. Nobody has run it on out-of-distribution inputs yet, and market-state text is very out of distribution compared to "which support team?" The demo's 0.114 s is their latency. Fast and cheap makes it easy to call it a thousand times an hour — which makes a badly calibrated answer more dangerous, not less, because you'll act on it constantly.”

“Initially I was as enthusiastic about this as the next guy. But if you think about it (and try it out), isn't this just 3 types of Machine Learning models (Clustering and 2 Classification), tight together behind an API? That's why it is super fast and cheap and that is why this isn't an LLM - because this is ML.”

“these realtime demo are fake in a sense that the real simulation is paused every time a request is sent to jev to let jev reply after 200ms, also the examples are simplified to the core to make it possible to jev to function, they are demo for the sake of doing a demo, not to showcase real world capabilities.”

“I'm pretty disappointed with Jev in terms of intelligence. I've tried to make many demos and the decisions were often wrong. But a few of those are actually legit: the chat from CJ is a tool call only chat and he advertise that as such. The autopilot is obviously just making decisions based on state but that's advertised as that. But yeah a lot of demos focus on the velocity but they are way inaccurate”

“Praticamente, siccome non mi dà nessuna spiegazione, mi devo fidare ciecamente dell'output. Peggio, se sbaglia, è impossibile capire dove e perché...e che faccio? Lo butto via? Mah! Utile? Insomma...”

“Me parece una tontería, para empezas si utilizas un modelo estadounidence que vale 4000 dólares por token para ese tipo de tareas significa que sos un completo ignorante. La IA ya genera salida estructurada... eso queire decir que Jev no trae valor. Especialmente cuándo hay modelos locales que pueden hacer eso sin problema.”