Matija Sosic

关于 Jev 的 45 秒视频 核心理念其实非常简单优美 不止于game、trade 在速度、文本理解、排序的基础上可以衍生出很多：https://t.co/HpbXiO33fb https://t.co/CRLpFPi4NC

This Lab is changing everything we know about how AI is being built into software. I got a chance to look into TypeFace’s new model, Jev, and I genuinely think these guys are onto something interesting. For context: TypeSafe is a San Francisco AI lab co-founded by Diogo Almeida, one of the researchers behind the work that helped make ChatGPT possible. They’ve just emerged from two years in stealth with $40 million in seed funding and their first public model. And the interesting part about this model is you can’t even chat with it. You can’t ask it to write an essay or generate code either. Jev is built to make decisions inside software, and the way it does this is pretty interesting. Here’s all you need to know: TypeSafe calls Jev a new class of AI models, System One Models. Imagine you’re building a customer support app. Someone complains about an order, threatens to cancel, and asks for a refund. You want your app to figure out how urgent the complaint is, which department should handle it, and whether a human needs to intervene. With Jev, you provide the message, define those questions, and specify the possible answers. It evaluates them and returns structured decisions with probabilities your software can immediately act on. And here’s how it differs from conventional LLMs: ⟢ It doesn’t generate text: You define the possible answers beforehand using three primitives: Choice, Score, and Noul (yes/no). ⟢ Its structured outputs are native: Instead of generating JSON token by token like a conventional LLM, Jev produces typed decisions directly. ⟢ It evaluates questions in parallel: You can ask 20 questions about the same document and receive all the decisions in one request. ⟢ It’s trained to express uncertainty: TypeSafe developed a method called RLCD,that’s aimed at producing calibrated probabilities, so software can identify uncertain decisions and escalate them. And the numbers are pretty crazy. According to TypeSafe, Jev can be 20 to 200x faster on comparable decision making tasks. Their workflow evaluations report 193.6x faster performance and 444.6x lower costs, and It costs just $0.042 per million input tokens, with output tokens free. First of all, that’s crazy. And in independent testing, they used Jev to evaluate 777 judgments across 37 documents in under 0.7 seconds, for roughly a quarter of a cent. Of course, Jev can still make incorrect decisions, and labs might need broader independent testing. But I think there’s a pretty big opportunity here. Imagine software that can afford to make hundreds of intelligent decisions without calling an expensive reasoning model for every single one. Checking claims, routing requests, verifying agent actions, and deciding what happens next. I like to think that’s the direction TypeSafe is exploring. And if they can maintain useful accuracy at these speeds and prices, I can see why they’re betting on it. Plus they even used it to play a game. Ridiculous! There’s so much stuff we can do with Jev running behind it, and I’m trying to build something cool myself. Will lyk when I come up with something. https://t.co/z4Yqsj4t45

1 useful AI thing today: Jev because ChatGPT's co-inventor just shipped an AI that refuses to write a single word Jev, it just decides You don't talk to it, you hand it a message, a ticket, or a document and ask a question. It sends back a decision your software can use immediately: this queue, this score, this yes/no. 6 things you should know: 1. The person behind it already helped invent the chatbot era. Diogo Almeida worked on the methods behind InstructGPT and ChatGPT, then spent 2 years asking a different question: if models are already great at talking, why isn't software fully automated? His answer was TypeSafe and a new model class they call System One. Jev is the first public one. 2. It only answers 3 kinds of questions. Choice: which team should get this ticket? Score: how urgent is this, from 0 to 2? Yes/no: does this customer want a refund? You can ask several at once. It answers them in parallel. 3. The speed and price are INCREDIBLE. Typical replies: 70 to 500 milliseconds. Input: about $0.042 per million tokens. Output: free. That is why people are putting it in places ChatGPT struggled One Minecraft bot ran 2 minutes for about a cent. A browser agent found flights in 7 seconds for $0.0039. One developer ran about 5,000 requests for roughly two dollars. 4. The useful part is the confidence score. High confidence: automate. Low confidence: send it to a human or a smart model. That is how you get cheap automation imo 5. The early use cases are "boring": route the ticket, score the risk. Jev is a smart if-statement that understands messy language. 6. limitations: it cannot write, explain itself, do reliable math, or look at images. 1 useful thing from this: keep ChatGPT for writing. Use Jev for deciding. so keep the "normal AI models" for the hard thinking and the writing. Use Jev for the rapid-fire decisions in between. so what would you automate first if the judgment cost almost nothing? https://t.co/sm8b8v8qBC

A video reflecting the core idea about the JEV model. It might not be obvious right off the bat, but all the decision-making procedures have been packed into Tadle's later-revealed reasoning. https://t.co/RrM12k8uoW