Anthropic 向投资者披露,其 Claude Mythos Preview 模型在几乎无需人工干预的情况下,仅用约三天时间便完成了将 AI 模型适配至 AMD MI355X GPU 的软件工作 1。这一效率显著优于传统方式所需的六至十二个月周期,显示 AI 自动编写底层内核代码的能力正在削弱 Nvidia CUDA 平台的护城河优势 1。据报道,Claude Mythos 移植或编写了“几乎 100% 的内核”,实现了高度自动化 1。
尽管该进展由 Anthropic 提供数据支持,且该公司作为芯片采购方及即将 IPO 的企业具有较高公信力,但报道也指出了当前的局限性 1。目前尚未披露解码(decode)步骤的具体耗时,也未对比 AMD 与 Nvidia 平台的具体运行速度 1。此外,Agentic 编码流程仍需大量 GPU 资源进行测试验证,存在硬件测试瓶颈 1。相关关键人物包括 Anthropic 联合创始人 Tom Brown、AMD AI 软件副总裁 Anush Elangovan 以及 AMD CEO Lisa Su 1。在商业背景方面,Anthropic 已与 AMD 达成高达 2 吉瓦的协议,而包括 Nvidia 和 AMD 在内的七家公司以 IPO 价格购买了总计 260 亿美元的 Anthropic 股票 1。
Anthropic has disclosed to investors that its Claude Mythos Preview model completed the software work required to onboard AI models onto AMD MI355X GPUs in approximately three days, a task that traditionally takes between six and twelve months 1. The AI system reportedly wrote or ported nearly 100% of the necessary kernel code with minimal human intervention 1. This development suggests that AI’s ability to generate low-level kernel code is beginning to erode the competitive moat historically held by Nvidia’s CUDA platform 1.
Key figures involved in this shift include Anthropic co-founder Tom Brown, along with AMD AI Software Vice President Anush Elangovan and AMD CEO Lisa Su 1. While AMD executives have previously made similar claims regarding their software ecosystem, data provided by Anthropic—a chip purchaser preparing for an initial public offering (IPO)—carries greater credibility 1. Commercially, Anthropic has reached agreements worth up to 2 gigawatts with AMD, and seven companies, including Nvidia and AMD, purchased $26 billion in Anthropic stock at IPO prices 1. Despite these advancements, limitations remain, as the report did not disclose the time required for decode steps or provide specific performance comparisons between AMD and Nvidia runtimes 1. Additionally, agentic coding still requires significant GPU resources for testing and validation 1.
评论
还没有评论,欢迎留下第一条。