Diogo Almeida

卧槽，今天 AI 圈最火的模型，应该就是 JEV 了。 现在所有大模型都在拼更会说话、更会写、更像人，JEV 直接反着来，它根本不生成文字，只负责做判断和决策。 我已经迫不及待用JEV帮我决策自动交易股票和加密货币了😂 官方给的数据非常夸张：20—200 倍更快，40—400 倍更便宜，输入每百万 Token 只要 0.042 美元，输出 Token 直接免费。 我们今天用 GPT、Claude、Gemini 做自动化，经常是让模型先写一大段话，再让程序从里面抠出 JSON、标签、分数和下一步动作。 JEV 干脆把“说废话”这一步整个删了，直接告诉软件： 选 A 还是 B、概率多少、置信度多少。 客服分流、内容审核、Agent 路由、推荐系统、风控、交易信号过滤……现实世界里大量 AI 调用，本来就不是为了让它写小作文，而是为了让它每秒做成千上万个判断。 这也是我觉得 JEV 最值得关注的地方。 目前 JEV 还在 Early Access，感兴趣的同学可以先去加入白名单等体验。

Jev plays doom in realtime and it looks like an actual player. their numbers: 20 to 200x faster, 40 to 400x cheaper, output tokens free. Two things we might see now: - the opponent we're up against in a game might be a realtime AI. - a TA, or anything else we'd hire a real person for, can run on this instead. 10 calls a second is 36,000 calls an hour, and that costs around 7 dollars. anything that has to keep up with a person was off the table before this. can't wait to play against one of these. https://t.co/y0qXEyWPv9

ChatGPTにつながるRLHF研究の主要メンバーが、2年のステルスを経て新型AI「Jev」を発表。 RLHFは簡単に言うと、人間の評価を使って「人にとって自然で役立つ回答」をするようLLMを調整する手法。ChatGPTが今の会話型AIになった重要な技術の1つです。 今回のJevはその発想を「会話」ではなく「自動化の中の判断」に振っている。 ・開発元はTypeSafe AI ・創業者は元OpenAIのDiogo Almeida氏 ・新しい学習法「RLCD」を採用 ・特定タスクで40〜200倍高速 ・入力100万token $0.042、出力tokenは無料 ・$40Mのシード調達 普通のLLMが文章を返すのに対して、Jevは「解約確率87%」「人間対応91%」のように、決められた型で確率付きの判断を返す。 なのでChatGPTの代替というより、AIエージェントの裏側で大量の判定を高速にさばくモデル。 ↓補足

THE GUY WHO CO-INVENTED CHATGPT JUST DROPPED A NEW MODEL THAT REFUSES TO CHAT Diogo Almeida helped build RLHF at OpenAI. Now he thinks generating text is a massive bottleneck for agentic workflows. His new company @typesafeai just dropped Jev. It is a ruthlessly efficient System One model that acts as a pure API for structured decisions. Ask a question and get back typed answers with exact confidence scores. By stripping out generative prose entirely they hit wild 150ms response times: → 200x faster than standard models → $42 per billion input tokens → output tokens are free forever IMO this is an incredibly smart way to route logic without paying for words you do not need 👀

A ChatGPT co-inventor just released Jev after 2 years in stealth, a new frontier AI model that is 20 to 200x faster and 40 to 400x cheaper with output tokens free, optimized for decisions rather than conversation. https://t.co/wGkR4wSQZh

🤯Jev模型游戏实测太离谱了，直接把“实时决策”这件事干穿了！ 不是聊天模型，是专为高频决策设计的System One： 1️⃣实时打Doom，每秒约10次决策，1小时成本才$7 2️⃣Minecraft实测：2分钟自动生存只要1分钱，夜间自己躲僵尸，零额外提示 3️⃣延迟70-500ms，一次前向传播就能并行回答多个问题 4️⃣价格$0.042/百万输入token，输出几乎免费 5️⃣只输出结构化结果（Choice/Score/概率+置信度），零类型错误，幻觉极低 6️⃣天生适合游戏Bot、Agent循环、客服分流这类「每秒都要做选择」的场景 这才是能真正跑进游戏里的AI大脑！ #Jev #TypeSafeAI #AIAgent #Doom #Minecraft #SystemOne​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​ https://t.co/T8BAFqW6cN https://t.co/yTdRpandwJ

Jev launched three days ago Here are the best demos to come out 🧵 https://t.co/NCPoyOCObD

Jev is an AI model built for fast, low-cost decisions that went viral yesterday. It plays doom on its own because of low cost/latency decision. This will be the SOTA for World of Warcraft bots. https://t.co/gUunMcDvI8

Jev is an AI model built for fast, low-cost decisions that went viral yesterday. It plays doom on its own because of low cost/l... via @kimmonismus #AI #MachineLearning #GenerativeAI #Innovation #AItools https://t.co/Km5iGD4kDf

diogo almeida helped build the methods behind chatgpt at openai. today he launched something that isn't a chat model at all. jev never generates a single word of text. feed it unstructured state plus a set of typed questions, and it returns a probability-scored answer to every question at once, in one parallel pass, not token by token the way every frontier chat model works. 70 to 500 milliseconds end to end. frontier conversational models on comparable tasks run 3 to 329 seconds. $0.042 per million input tokens, output entirely free, since there's no text being generated to bill for in the first place. his own framing: "think of jev as a frontier-intelligence function call, unstructured state in, typed probabilistic decisions out." trained on a method he calls rlcd, reinforcement learning for calibrated decisions, a genuine departure from the rlhf and rlvr most labs currently use. the honest limit worth naming directly. "hallucination-free" here means something narrower than the phrase suggests. bounded, typed outputs literally can't hallucinate text, since there's no text. but picking the wrong option within a valid schema is still entirely possible. this isn't a replacement for a chat model. it's the classification and routing layer that sits in front of one, cheap enough to run before and after every actual llm call. typesafe ai emerged from two years of stealth with $40 million in seed funding led by dcvc. almeida's actual question, the one that started this: why haven't superhuman chat models led to agi. his answer isn't a bigger chat model. it's a different category entirely, one optimized for the boring, high-volume decisions that sit underneath every real automation pipeline, and that decomposing those decisions might matter more for building genuinely useful ai systems than compressing everything into one more capable conversation. https://t.co/fcp2AfOZSl

An early architect of ChatGPT just shipped an AI that refuses to write a single sentence Diogo Almeida, who was a primary author on the InstructGPT paper that led into ChatGPT and later worked on GPT-4, spent two years in stealth building something deliberately different. His new company, TypeSafe AI, just launched with $40M in seed funding led by DCVC, and its first model is called Jev, a nod to Jevons Paradox. Jev doesn't generate text token by token. You feed it unstructured state plus typed questions defined in advance, and it returns typed answers your code can use directly: pick a category from a list, assign a score, estimate the probability something is true. All the answers come back in parallel instead of one word at a time. There's no hidden chain of thought either. TypeSafe calls this a "System One" model, trained with something they call Reinforcement Learning for Calibrated Decisions, and complex behavior is supposed to be assembled by developers out of many small typed decisions using regular code, not by letting the model reason freely inside a black box. That constraint is exactly why it's fast. TypeSafe claims 20 to 200x faster and 40 to 400x cheaper than comparable frontier models. Input runs 4.2 cents per million tokens, and output is free, cheap enough to compute that the company says metering it isn't worth the trouble. TypeSafe is also calling it hallucination-free: Jev can't return anything outside the options you defined, and it can't break the response structure. It can absolutely still pick the wrong option, it just can't go off-script while doing it. This isn't a GPT-6 Astra competitor and it's not another chatbot. It's built for the kind of automation where software never needed prose in the first place, just a fast, reliable decision in a fixed format.

一个模型发布技术帖，居然把半个 AI 圈的 big shots 和 X 上的科技网红都炸了出来，纷纷跑到评论区报到。短短几个小时就爆款。 因为发帖的人不是普通创业者。 Diogo Almeida 是 RLHF 和 InstructGPT 的共同发明者之一，也就是参与创造 ChatGPT 的那批人。现在，他却提出了一个颇具颠覆性的判断： Chat 可能不是 AI 实现大规模自动化的正确形态。 过去两年，他一直在秘密开发另一类模型。今天，他们发布了第一个 System One Model：Jev。 它不聊天，不写文章，也不逐字生成答案。 Jev 接收一段非结构化信息，直接输出预先定义好的结构化决策，以及每个选择对应的概率和置信度。所有结果一次并行生成，不需要像 LLM 那样一个 token 接一个 token 地“说话”。 换句话说，它更像一个带有前沿模型智能的 function call： 输入业务状态，输出可以直接交给软件执行的 typed decisions。 TypeSafe 声称，在这类任务上，Jev 可以达到与前沿 LLM 接近的智能水平，但速度快 20 至 200 倍，成本低 40 至 400 倍，延迟只有 70 至 500 毫秒。输出 token 甚至免费，因为它根本不生成字符串。 过去，我们一直试图把会聊天的 LLM 塞进软件，再用 JSON Schema、重试、验证器和 guardrail 约束它。Jev 则从模型架构开始就放弃自由文本，只允许输出合法类型。 当然，“不会产生类型错误”不等于“不会做出错误判断”。目前的数据也主要来自 TypeSafe 自己设计的 workflow eval，仍需要独立验证。 但它指出了一个很有意思的方向： Agent 可以继续使用通用大模型思考和规划，但真正进入生产系统的大规模自动化，可能需要另一类更快、更便宜、更受约束的决策模型。 大模型负责想，System One Model 负责在软件里每秒做成千上万个决定。 AI 的下一次效率跃迁，也许不是生成得更快，而是不再生成。 https://t.co/onRTD8mFjO