去年创办的人工智能公司Recursive开始名副其实。 该公司正在试图建立能够自我改善的AI，发布于星期三推出其第一个模型弗林特。 弗林特是一个可以运行在笔记本电脑上的免费编码模型，主要由AI帮助建造。
Recurive的自动化研究系统，它被称为Eureka Machine，几乎帮助了Flint开发的每一个步骤，即称公司。 这让一个小团队在不到四个月的时间里从零开始建立模型。
弗林特(Flint)是对递归自我改进的早期公开测试：即AI可以建立更好的AI的想法，这反过来又会建立更完善的AI. 也成为该行业最受争议的想法之一。 一些研究者担心AI改善自身可能脱离人类控制。 现在，Recursive的版本比较有限：一个AI研究助理，由制作一个足够小的模型供笔记本电脑的人监督。
该公司于5月以46.5亿美元估值从隐形公司(包括GV,Nvidia，和AMD Ventures)等背书商获得6.5亿美元的资金。 递归由Salesforce前首席科学家理查德·索彻(Richard Socher)领导。
因为弗林特运行在用户自己的计算机上，而不是公司的云服务器上，开发者可以将其离线，保持代码私密，并避免AI云服务根据模型处理多少文本收取的使用费。 该型号可自由下载在Hugging Face上，这是共享AI型号的热门网站，根据Apache 2.0许可证，是允许自由使用和修改的开源许可证。
弗林特是一个专家的混合模型，意思是它只用一小片自己来完成任何特定的任务：一次只有其400亿个设置中的25亿个，被称为参数。 这使得它足够轻的普通笔记本电脑，并且远小于Meta的Muse Spark Open, Muse Spark Open使用了8160亿个参数，需要至少一个千兆字节的内存。
弗林特也快了点在Recursive的测试中，它用Apple的M4 Pro芯片在MacBook Pro上每秒生成约131个令牌(满文)——大约是类似型号在同一台机器上管理的每秒58至69个令牌的两倍，其中包括Google的Gemma 4和Nvidia的Nemotron 3.5闪电。 在Nvidia的RTX Spark上，弗林特每秒超过150个令牌。
在SWE-Bench Verifed上，一个测量AI能在多大程度上修复现实世界软件bug的基准测试，弗林特在500亿参数下比其他模型表现优异。 在Recursive自己的平均值上，经过6个测试，只有一个型号，Alibaba的Qwen3.8 27B，得分高于弗林特。
公司承认弗林特需要比大付费编码助理更多的指导，并建议将大任务分成小步骤。 然而，早期审查是积极的。 Wired's Will Knight使用带有48GB内存的Mac上的弗林特来构建几个小程序，包括一个待办列表应用程序和一个用于搜索的浏览网络的工具。 他发现较小的编码模型经常会遇到错误或比Claude或Codex慢，但弗林特每次都钉住他的编码请求。
Eureka机器提出一些想法，进行实验，并利用结果来决定下一步要尝试什么。 根据该公司的发布录影带，它为建造弗林特的几乎每个部分做出了贡献，从培训前到培训后和推断。
Eureka贡献的称递减包括一个更稀疏的模型形状，低位训练，将弗林特缩小71%，质量和自定义内核几乎没有损失，这些都是加快基础计算的专门程序。
Eureka机器救了研究人员多少时间，上个月，OpenAI 称，到8月中旬，它的AI agents为人类研究者每工作一次记录3.1 workdays.
如果Recursive的赌注有回报，每个新模式都应该能够在没有人类帮助的情况下承担更长的工作。 该公司先前告诉TechCrunch，它在6月对Amazon Web Services的4.1亿美元计算交易是“少了人头数，多了agent的计数”。
更长远的目标，索彻告诉Wired，是让Eureka机器实现完全的递归自我改进，由人类监督这个过程，如果出问题就能够阻止它。 递归称每款新型号应使Eureka机器成为更好的研究者，帮助它更快地建造下一款。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Recursive, an AI startup founded last year, is starting to live up to its name. The company, which is trying to build AI that can improve itself, released its first model, Flint, on Wednesday. Flint is a free coding model that can run on a laptop and was built largely with the help of AI.
Recursive’s automated research system, which it calls the Eureka Machine, helped with nearly every step of Flint’s development, the company says. That let a small team build the model from scratch in under four months.
Flint is an early public test of recursive self-improvement: the idea that AI can build better AI, which in turn builds even better AI. It’s also one of the most debated ideas in the industry. Some researchers fear that AI improving itself could slip out of human control. For now, Recursive’s version is more limited: an AI research assistant overseen by people that produced a model small enough for a laptop.
The company emerged from stealth in May with $650 million in funding at a $4.65 billion valuation, from backers including GV, Nvidia, and AMD Ventures. Recursive is led by CEO Richard Socher, Salesforce’s former chief scientist.
Because Flint runs on a user’s own computer rather than a company’s cloud servers, developers can use it offline, keep their code private, and avoid the usage fees that AI cloud services charge based on how much text the model processes. The model is free to download on Hugging Face, the popular site for sharing AI models, under the Apache 2.0 license, an open-source license that allows free use and modification.
Flint is a mixture-of-experts model, meaning it uses only a small slice of itself for any given task: just 2.5 billion of its 40 billion settings, known as parameters, at a time. That makes it light enough for an ordinary laptop and far smaller than Meta’s Muse Spark Open, an open model built for Macs that was released last week. Muse Spark Open uses 816 billion parameters and needs at least a terabyte of memory.
Flint is also fast. In Recursive’s tests, it generated about 131 tokens (chunks of text) per second on a MacBook Pro with Apple’s M4 Pro chip — roughly twice the 58 to 69 tokens per second that similar models managed on the same machine, including Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning. On Nvidia’s RTX Spark, Flint topped 150 tokens per second.
On SWE-Bench Verified, a benchmark that measures how well AI can fix real-world software bugs, Flint outperformed other models under 50 billion parameters. On Recursive’s own average across six tests, only one model, Alibaba’s Qwen3.8 27B, scored higher than Flint.
The company acknowledged that Flint would need more guidance than the big paid coding assistants, and suggested breaking large tasks into smaller steps. Still, early reviews were positive. Wired’s Will Knight used Flint on a Mac with 48GB of memory to build several small programs, including a to-do list app and a tool that scours the web for research. He found that smaller coding models often ran into errors or were slow compared to Claude or Codex, but Flint nailed his coding requests every time.
The Eureka Machine comes up with ideas, runs experiments and uses the results to decide what to try next. According to the company’s launch video, it contributed to nearly every part of building Flint, from pre-training to post-training and inference.
Recursive said that Eureka’s contributions include a sparser model shape, low-bit training that shrank Flint by 71% with almost no loss in quality and custom kernels, which are specialized programs that speed up the underlying computations.
Recursive didn’t say how much time the Eureka Machine saved its researchers, but it’s not alone in leaning on AI to do research. OpenAI said last month that by mid-August, its AI agents were logging 3.1 workdays for every one worked by a human researcher.
If Recursive’s bet pays off, each new model should be able to take on longer stretches of work without human help. The company previously told TechCrunch that its $410 million compute deal with Amazon Web Services in June was “less about headcount and more about agent count.”
The longer-term goal, Socher told Wired, is for the Eureka Machine to achieve full recursive self-improvement, with humans overseeing the process and able to stop it if something goes wrong. Recursive says each new model should make the Eureka Machine a better researcher, helping it build the next one faster.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.