The Rundown AI: TypeSafe Launches Jev, Salesforce Debuts Koa, and China Outlines Recursive Self-Improvement

The Rundown AI: TypeSafe Launches Jev, Salesforce Debuts Koa, and China Outlines Recursive Self-Improvement The Rundown's newsletter highlights major new developments across AI, including TypeSafe's zero-hallucination decision model Jev, Salesforce's synthetic-trained Koa reasoning model, and a Chinese research paper framing a five-level path to self-improving AI. Former OpenAI researcher Diogo Almeida's TypeSafe emerged from stealth with Jev, a system priced at $42 per billion input tokens that makes preset judgment calls without generating open text. Salesforce introduced Koa, an in-house CRM reasoning model trained entirely on synthetic data that achieved 3x fewer errors than top models on internal benchmarks. Over 30 Chinese researchers published a five-tier roadmap on recursive self-improvement culminating in systems that rebuild their own training processes. Quick briefs highlight Google's Gemini 3.8 Live, Odyssey 3's universal world model, Meta One subscriptions, and a report on OpenAI's 'Project Lily' chat reviewers. TypeSafe Unveils Jev for In-Software Judgments Created by former OpenAI researcher Diogo Almeida, TypeSafe's Jev is designed to live inside software and make background decisions rather than generate text. Running at $42 per billion input tokens with free output and latency between 70 to 500 milliseconds, the system claims zero hallucinations because it only selects among preset options. Applications include routing requests, record scoring, and filtering jailbreaks. Salesforce Launches Koa Reasoning Model Salesforce built Koa on Nvidia's open Nemotron 3 Super architecture to handle high-volume sales and support agent tasks. Trained completely on synthetic multi-industry persona simulations rather than client data, Koa reduced errors threefold against top models on an internal benchmark. It runs hosted within Salesforce's infrastructure alongside AIforce and Claudeforce integrations. Chinese Paper Explores Recursive Self-Improvement A paper titled 'The Last AI Built by Humans', authored by researchers across ByteDance, Tsinghua, and Shanghai AI Lab, outlines five levels of recursive self-improvement leading to AI designing its successors. Assessing 491 published papers, the authors found 75% at Levels 1 or 2 and under 6% at Level 5, identifying coding as the quickest near-term pathway for testing autonomous iterations. Industry Briefs and Releases Additional updates include Google launching Gemini 3.8 Live voice models capable of thinking while speaking, Odyssey previewing its multi-domain robotics model Odyssey 3, Meta debuting Meta One subscription tiers from $2.99 to $499 monthly, and a 404 Media investigation reporting on OpenAI's Project Lily contract chat reviewers.

AI行业动态简报：ChatGPT联合开发者推无幻觉决策模型Jev，Salesforce自研推理模型Koa 本期通讯涵盖多项AI前沿动态：前OpenAI研究员创立的TypeSafe推出专职后台决策的Jev模型；Salesforce推出针对CRM场景的自研推理模型Koa；中国研究团队发布AI递归自改进五级路线图；此外还包含OpenRouter应用指南及Google、Meta等厂商的最新动态。 前OpenAI研究员Diogo Almeida创立的TypeSafe推出Jev模型，专用于预设判断，宣称零幻觉且十亿输入Token仅售42美元。 Salesforce发布基于英伟达Nemotron 3 Super打造的企业推理模型Koa，在CRM基准测试中错误率降至竞品的1/3。 中国30余位AI学者联合发布论文，系统阐释AI递归自改进（RSI）的五个发展阶段。 Google推出支持边思考边说话的Gemini 3.8 Live语音模型，Meta则上线Meta One多档位订阅服务。 TypeSafe推出决策型AI模型Jev 前OpenAI研究员Diogo Almeida创立的TypeSafe结束隐身状态并推出Jev。该系统专门内置于软件中执行预设判断，不生成文本，仅在预先设定的选项中做出带有置信度评分的选择，宣称具备零幻觉特性。Jev响应速度在70至500毫秒之间，十亿输入Token收费42美元且输出免费，适用于请求分类、记录评分以及安全过滤等后台流程。 Salesforce发布自研推理模型Koa及配套生态 Salesforce基于英伟达开源的Nemotron 3 Super模型，推出了针对销售与客服场景的自研推理模型Koa。该模型采用完全合成的数据集训练，涵盖十余个行业的模拟对话。内部CRM基准测试显示其错误率比主流模型低3倍。该模型由Salesforce自主托管以保障数据不出域。同时，Salesforce还发布了AIforce并将ClaudeForce面向所有客户开启公测。 中国研究团队绘制递归自改进（RSI）五级路线图 来自字节跳动、清华大学和上海人工智能实验室等机构的30多位研究人员发表了题为《人类构建的最后一个AI》的论文，提出递归自改进的五级演进框架：从第1级人类主导优化，到第2级自我诊断，第3、4级的自主学习与部署后适应，直至第5级彻底重构自我改进流程。研究指出代码领域是RSI最清晰的落地路径，并统计分析了491篇论文，显示目前仅不到6%的研究触及第5级。 行业快讯与实用动态 Google推出能够在思考同时持续发声的语音模型Gemini 3.8 Live；Odyssey发布可控制机械臂、人形机器人和自动驾驶的世界模型Odyssey 3；404 Media披露OpenAI曾通过“Project Lily”由数百名外包人员阅读并评估ChatGPT真实对话记录；Meta上线月费2.99至499美元不等的Meta One订阅服务；OpenRouter发布接入多模型优化开发成本的实操方案。

[Read Online](https://therundownai.beehiiv.com/p/chatgpt-co-creator-launches-a-new-kind-of-ai) ^| [Sign Up](https://www.therundown.ai/subscribe?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-1013) ^| [Advertise](https://therundownai.typeform.com/to/kraZ1TSO?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-1013) **Good morning, {{ first\_name | AI enthusiasts }}, and welcome to our 4,056 new readers.** Diogo Almeida helped build the methods that taught AI to talk to people — the research behind ChatGPT. His next model can't generate text at all… By design. TypeSafe's Jev is a new type of AI system that lives inside software and makes judgment calls, doing the background decision-making work that chatbots aren’t always the right tool for — at a wild speed and price point. *NEW: Our Community AI Workflow Hub now takes questions. Stuck on a prompt, a tool, or a workflow that won't cooperate? Ask the community and get answers from people who've already figured it out.* **In today’s AI rundown:** ChatGPT co-creator launches a new kind of AI Salesforce trains its own reasoning model OpenRouter 101: Add models to Codex, Claude Code Chinese researchers detail 'the last AI built by humans' **LATEST DEVELOPMENTS** Image source: TypeSafe **The Rundown:** Ex-OpenAI researcher Diogo Almeida’s TypeSafe just [emerged from stealth](https://typesafe.ai/blog/introducing-system-one-models-and-jev?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-1013) with Jev, a new kind of AI that answers preset questions inside software with confidence scores attached, claiming zero hallucinations and extreme speed and price. **The details:** Jev runs at $42 for a billion input tokens, and output is free (!), a rate TypeSafe estimates at 238x below Claude Fable 5.1's pricing. Responses also output in just 70 to 500 milliseconds, between 40-200x faster than today’s LLMs. TypeSafe says Jev “can’t hallucinate,” because it only chooses between options set in advance, calling the system a “frontier-intelligence function call.” TypeSafe’s Jev use cases include quick judgment calls inside apps, like sorting requests, scoring records, or screening another AI's outputs for jailbreaks. **Why it matters:** This isn't a system to stack up against LLMs, with Almeida calling it "more like a database than a coworker." That's where the nod to Jevons paradox (the cheaper something gets, the more it gets used) comes in — if Jev really is this fast and cheap while staying reliable, it is likely to become a standard part inside software. **The Rundown:** Search results go stale, so teams bolt on an API, add a scraper, wire up a parser, and hope it holds. Yeah, it works in the demo, but then it breaks under production load six months later. This guide from [You.com](https://you.com/resources/ydc-apis?utm_campaign=28104480-The%20Rundown%20Q3&utm_source=external-newsletter&utm_medium=email&utm_term=rundown_primary_715&utm_content=rundown_primary_715) explains why treating research as simple retrieval is the wrong frame, and what a purpose-built stack looks like instead. **What you'll learn:** Why general-purpose search hits 93% SimpleQA accuracy but still fails high-stakes workflows The difference between an API that returns results and one that answers questions How teams reclaim 6–12 months of time with a single endpoint What finance, legal, and compliance workflows need beyond web search Stop treating research as retrieval. [Download the guide](https://you.com/resources/ydc-apis?utm_campaign=28104480-The%20Rundown%20Q3&utm_source=external-newsletter&utm_medium=email&utm_term=rundown_primary_715&utm_content=rundown_primary_715). Image source: Salesforce **The Rundown:** Salesforce just [introduced](https://www.salesforce.com/agentforce/koa?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-1013) Koa, a new in-house reasoning model for its sales and support agents built on Nvidia’s open Nemotron 3 Super model and adapted for business tasks. The training set is fully synthetic, simulating personas like irate support callers and reps closing deals across over a dozen industries, without using customer data. Koa made 3x fewer errors than top models on an internal CRM benchmark, matching or beating them on tasks like updating deals or routing tickets. The company hosts Koa itself, keeping customer requests inside its systems instead of sending them to an outside model provider. Salesforce also launched AIforce, which pushes CRM data and permissions into outside AI tools, and opened its ClaudeForce version to every customer as a beta. **Why it matters:** Salesforce is in a unique spot with the AI labs, acting as customer, distribution partner, and now competitor all at once. Koa now keeps the high-volume agent reasoning in-house on a model tuned to its own workflows, while options like Claudeforce stay as ways to help the frontier labs reach their customers. **The Rundown:** In this guide, you will learn how to give your AI agent more models to work with using OpenRouter. We'll set up an API key, compare models, and test cheaper options for making ad mockups. **Step-by-step:** Sign in to OpenRouter and browse the model catalog. Filter for your task, then sort by low to high price. Open a model page to check its capabilities and cost Open API Keys, choose New key, and name it after the project or agent you'll use it with. Set a spending limit and an expiration date before creating it Ask your agent to help you set up OpenRouter. Store the key privately using its setup instructions, then give it the model name and a small task. We asked our bot to make more hat ad mockups with a cheaper image model Compare the results and check usage in OpenRouter. Our mockups were useful for direction, but they didn't match the real hat or the GPT Image 2.5 version **Pro tip:** Tell your agent to create a skill to use OpenRouter to make cheap, bulk variations for you when you're brainstorming. We will go over this in part 2! **The Rundown**: AI transformation fails when you DIY. StackAI gives you one platform to build, deploy, and scale, backed by a team of dedicated AI experts who know what works. IT retains 100% control over access, tokens, and usage. **Why enterprises choose StackAI:** No-code platform for rapid building and deployment 300+ native integrations, no LLM lock-in Global governance and advanced analytics Image source: arXiv **The Rundown:** 30+ Chinese AI researchers, including those from ByteDance, Tsinghua, and Shanghai AI Lab, [published](https://arxiv.org/pdf/2609.11873?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-1013) “The Last AI Built by Humans,” a roadmap detailing five levels of recursive self-improvement, ending with AI that builds its successors. Level 1 has the AI carrying out upgrades humans designed, while Level 2 lets it diagnose its own weak spots and decide how to fix them. Levels 3 and 4 hand over what the model learns next and how it adapts after launch, and Level 5 lets it completely overhaul the improvement process itself. The authors say coding has the clearest path to RSI, since fixes can be tested instantly, while robotics, science, and medicine face slower, costly feedback. The researchers also sorted 491 existing papers onto the ladder, with 75% landing at Level 1 or 2, and under 6% making it to Level 5. **Why it matters:** Level 5 is the stage many of the Western AI labs have been warning about, with OpenAI, Anthropic, and Google all listing AI automating its own R&D as a risk in their safety frameworks. But China doesn't seem as worried about RSI (or, as we saw yesterday, a slowdown in general), treating it as a milestone rather than a threat. **QUICK HITS** ▸ **A Rundown reader built an emergency vet coordinator with ChatGPT** Today’s workflow comes from an **anonymous Rundown reader:** “I used ChatGPT as an emergency care coordinator for my dog, Mopsy. My 10-year-old dog developed a dental abscess that caused severe swelling around her eye. An ophthalmologist told us that the pressure was damaging her eye and had partially torn her retina, so treating the infected tooth quickly became urgent. I started calling vet offices myself, but I repeatedly hit the same obstacles: emergency hospitals did not perform dental surgery, specialty dentists were booked, and most primary-care vets required a consultation, bloodwork, and a separate surgical appointment. After over a dozen calls, I was overwhelmed and asked ChatGPT to help. ChatGPT researched vet practices within an hour's drive and sent urgent, individualized emails from my Gmail. We ultimately contacted 20+ practices. At the same time, it created a live outreach tracker containing every clinic we had contacted. In half an hour, one vet responded that they could evaluate Mopsy the next morning, run her bloodwork on-site, and take her straight back for surgery if she was cleared.” *See the full workflow here. How do you use AI? Tell us for a chance to be featured.* 🧠 [Koa](https://www.therundown.ai/tools/koa?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-1013) - Salesforce’s CRM reasoning model built for enterprise work 🗣️ [Gemini 3.8 Live](https://www.therundown.ai/tools/gemini-3-8-live?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-1013) - Google's new voice models that think while they talk 🎧 [StepAudio 3](https://www.therundown.ai/tools/stepaudio-3?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-1013) - Five-model audio suite for voice agents, transcription, music ⚙️ [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-1013) - TypeSafe's early-access AI for software **NEW: Ask the Community** - Workflows are great when they work. When they don't, you can now ask a question in the Hub and get help from readers who've been there, plus The Rundown team. [Post your question](https://app.therundown.ai/community?sort=new&compose=question&utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-1013), or jump in and answer someone else's. **Google** [debuted](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-gemini-3-8-live-extended-thinking/?utm_source=x&utm_medium=social&utm_campaign=&utm_content=) Gemini 3.8 Live, voice models that can continue speaking while they think, with the Extended Thinking variant topping AA's speech-to-speech quality ranking. **Odyssey** [introduced](https://odyssey.systems/introducing-odyssey-3?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-1013) Odyssey 3, a world model that can control robot arms, humanoids, self-driving cars, drones, and games, set to release in the coming weeks. **A 404 Media report** [detailed](https://www.404media.co/inside-project-lily-the-humans-reading-your-chatgpt-chats?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-1013) OpenAI’s ‘Project Lily’, finding that hundreds of human contractors read and rate real ChatGPT chats, often without the user’s knowledge. **Meta** [rolled out](https://about.fb.com/news/2026/09/introducing-meta-one-subscription-service-more-features-ai/?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-1013) Meta One, a new paid tier across its apps running from $2.99 to $499 a month for extra Meta AI usage and creator tools. Read our last AI newsletter: [Trump, China both shoot down the AI slowdown](https://www.therundown.ai/articles/trump-china-both-shoot-down-the-ai-slowdown?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-1013) Read our last Tech newsletter: [U.S. confirms weapons are in orbit](https://tech.therundown.ai/p/us-confirms-weapons-are-in-orbit?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-1013) Read our last Robotics newsletter: [Japan’s latest answer to home robots](https://robotnews.therundown.ai/p/japan-latest-answer-to-home-robots?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-1013) Today’s AI tool guide: [OpenRouter 101: Add models to Codex, Claude Code](https://app.therundown.ai/guides/openrouter-101-add-new-ai-models-to-codex-and-claude-code?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-1013) RSVP to next workshop on Sept. 30: [Turn Cowork into your chief of staff](https://app.therundown.ai/live/claude-cowork-201-turn-your-chat-into-a-chief-of-staff?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-1013) See you soon, *Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown*