Rohan Paul

4 things to do when you get access to Jev: 1. don't replace your model, put Jev in front of it Jev makes the small choice first, and your expensive model only runs when it's really needed one email tool swapped 2 AI calls per email for 1 Jev call, and kept all their normal safety checks on a 120 ticket test: 42 seconds with Jev in front → 22 minutes without it cost: $0.0003 → $0.059 the person who ran that test said his slow version was slowed down even more by retry errors, so the real gap is smaller how to do it: - look through your code for AI calls that only choose something and never write text - write the list of possible answers in your own code, don't let the model invent them - send that list to Jev, and keep your old call as a backup - put it behind an on/off switch so you can undo it in one line - run both for a week and compare them before you delete the old one you're not rebuilding your product, you're replacing one call 2. ask 5 small questions instead of 1 big one someone tested Jev on 2,000 phishing emails ask it one big question and it gets 89.4% a simple 2 line text rule gets 91.8% Claude Haiku 4.5 asked the same question gets 94.2% so when it has to give the final answer alone, Jev loses even to a text rule then he asked 5 small questions instead, and added the answers up in his own code 95.0%, the best score in the whole test SAME MODEL, SAME EMAILS how to do it: - take the big question you were about to ask - write down the 5 things a person checks before they answer it - ask each one separately, all in the same request, they run at the same time and cost almost nothing extra - add the answers up in your own code, with your own weights - test those weights on 100 examples where you already know the right answer - change the order of your options and run it again, reordering 4 options changed 7 answers out of 120 Jev is good at noticing things and bad at making the final call so keep the final call in your code 3. never ask Jev if it can answer on 120 test tickets, any question like "do you have enough info?" said yes on 85% of them a simple "need more info" flag said yes on 119 out of 120 tickets, and it then answered those tickets about 87% correctly it doesn't know what it doesn't know how to do it: - delete any question like "can you answer this" or "is there enough context" - make every option a real action your code can run - always read the confidence score, not just the answer - choose your limit by risk: low for reading data, 0.85+ for anything you cannot undo - send everything below the limit to a human or to your big model in that same test, using 0.8 as the limit passed 30 cases to a human, and 93% of them were passed for the right reason 4. run it quietly next to what you already have there's already a langchain package published and a pydantic-ai adapter being reviewed, so this is closer to a settings change than a rebuild how to do it: - leave your current system in charge, it still makes every decision - send the same input to Jev too, and throw its answer away - save both answers plus Jev's confidence score into one table - after a week, look only at the rows where the two disagreed - switch over only for the cases where Jev was right that list of disagreements becomes your test set, and your normal traffic builds it for free and don't use it when you already have labelled data if the question never changes and you have examples to train on, a small model you host yourself beats Jev on speed and price, and needs no API key at all Jev wins when you have no labelled data and the question keeps changing so use it where the list of answers is short and the question is boring that's most of your agent anyway everything above comes from other people's public tests, not from production, because the model is only 4 days old i'm just sharing what i find while i test this and try to make the work in my own company faster and cheaper tomorrow i'll show you what happened when i put Jev in front of the meta ads work we do for one big client and most of you are still on the waitlist anyway so start with step one, because it needs no access and no API key at all: find the calls in your code that were never writing tasks in the first place gl https://t.co/pCaKGTaxrY

前OpenAI研究员Diogo Almeida发布了一款新型模型：Jev，一种“放弃生成文字只输出类型化概率决策”的模型 称速度快40–200倍，输入成本便宜5–238倍，输出接近免费 就是说把大模型从会聊天会写长文的生成器，改成软件可以直接调用的概率决策函数 用并行概率决策替代逐token字符串生成，主打结构化、可校准、低延迟、低成本 创始人Diogo Almeida称曾在OpenAI参与指令跟随/对话方法相关研究，他的问题是：模型聊天能力早已超过人，但真正可嵌入代码的自动化仍缺关键一环 Jev背后是一套全新栈：新架构、并行采样器以及新训练方法RLCD 像前沿智能函数调用，输入非结构化状态/程序状态，输出预先定义好的类型化概率决策。放弃自由字符串生成，换来无类型错误、schema内不生成非法/幻觉式输出，并自带置信度 它抢的应该是海量的微决策场景，每一次请求背后几十上百个小判断的基础设施层 可塞进业务系统里做判断器，作为分类、路由、打分、抽取，实时应用里的快速判断，工作流里的“智能if”等 如果真能把每次判断压到毫秒级、接近零输出成本、附带置信度，那这个还是有价值的 #Jev #新型模型

我快速看了Jev的一些公开帖子和官方文档说明以及例子，终于有人AI模型关心速度，幻觉和成本，而且Jev不是一个大语言模型，这是它能做到这么极致的原因！ 官方也给出了好几个例子，比如模型路由和LLM安全护栏。 所以在大语言模型狂奔3年后，AI模型的新的分叉出现了： 不是大尺寸，不是语言模型，极低的幻觉，成本，极快的速度和准确率，没有太多的推理。 几乎每一条都是和LLM反着来的。 不得不多，这个方向真的非常对，希望早点体验到它。 下面这个视频来自官方demo一个并行化的例子，真的快如闪电，成本极低！ 这几篇来自官方的文章一定要看看： https://t.co/bm5dYnjZ9O https://t.co/6ZXGhuISMh https://t.co/iFBnlbp1XI https://t.co/HId6Dkb5se https://t.co/M5lBmItSON

Jev shipped from stealth last week and nobody's talking about it. Setup: One of the original ChatGPT co-authors spent two years building an AI that cannot generate text. Surface Story: It is not a cheaper chatbot. It is a decision engine. You ask it a natural-language question - 'does this ticket sound fraudulent?' - and instead of writing a paragraph it returns a number. 0.8 means 80 percent yes. You can define your own categories too: annoyed, irritated, outraged, and it returns a probability per label. Think of it as a smart if-statement for your business logic. Twist: The interesting part is the architecture. It does not generate token-by-token, so it answers in milliseconds. TypeSafe claims 20 to 200x faster and 40 to 400x cheaper than an LLM call. Pricing is $0.042 per million input tokens, output is free, and there are official Python and JavaScript TypeScript SDKs. Parallel questions run simultaneously. You stop treating an LLM as a data extractor and start treating it as a classifier with calibrated confidence. Hidden Reality: The real use case is agent guardrails. Before your agent fires a tool call or spends money, Jev returns a probability that the action is wrong. That is the bet - code plus AI, where the model scores options and your code decides how to proceed. Not another chat wrapper that rambles and breaks your pipeline. AI Reveal: The math is the story. $0.042 per million input tokens with free output against frontier LLM pricing means a million classification calls cost less than a single GPT-4o prompt at scale. At typical ticket-routing volume - say 50 million decisions a month - that is roughly $2.10 in inference. Compare that to even Haiku-class pricing and the gap is structural. The tradeoff: it is waitlist-only, demos are limited to video-game control and Wikipedia link selection, and benchmarks are missing. No public evals means the calibrated-confidence claim is still unverified. Irony Closer: Wild that we live in a timeline where a ChatGPT co-author's most interesting work in years is a model that refuses to talk, and the market is already pricing OpenAI as a 2 percent favorite to hold the best model by end of September on Polymarket, below Anthropic at 97 percent - while the actual product they cannot ship is the one making decisions in milliseconds. Call it. https://t.co/MzLLsSgj62