Amnon Shashua 的创业公司 doubleAI 推出了 Maxwell，这是一种在运行软件内部工作的 AI 模型，能够使软件使用更少的计算能力。Shashua 共同创立了这家位于特拉维夫的公司，并担任首席执行官。该公司在周一宣布了 Maxwell，并开放了早期访问。
Amnon Shashua’s startup doubleAI has launched Maxwell, an AI model that works inside running software to make it use less computing power. Shashua co-founded the Tel Aviv company and is its chief executive. The company announced Maxwell on Monday and opened it for early access.
Shashua 向《华尔街日报》记者 Belle Lin 阐述了公司的计划，该报在公司正式宣布前一小时报道了该发布。公司称 Maxwell 为一种新型 AI 的第一款产品，命名为代码执行模型。
Shashua set out the company’s plans to Belle Lin of The Wall Street Journal, which reported the launch an hour before the announcement. The company calls Maxwell the first of a new kind of AI, which it names code execution models.
Maxwell 的工作原理 编码代理为人们编写源代码。而 Maxwell 则是在程序运行时进行监控，并重写其底层的机器代码，公司如此表示。它会将该代码适配于所见的硬件和数据。源代码保持完全不变。Maxwell 可在 CPU 和 GPU 上运行。
How Maxwell works Coding agents write source code for people. Maxwell instead watches a program while it runs and rewrites the machine code underneath, the company says. It fits that code to the hardware and the data it sees. The source code stays exactly as it was. Maxwell works on both CPUs and GPUs.
代码运行期间所做的更改不会经过人工审查。公司表示，Maxwell 形式上证明了每项更改都不会改变程序的行为。更快的版本可能依赖于某些条件。如果这些条件不再成立，一个守护机制会将控制权交还给原始代码，正如 Maxwell 网站所解释的那样。
Changes made while code runs get no human review. The company says Maxwell formally proves that each change leaves the program’s behaviour unchanged. A faster version may rely on certain conditions. If they stop holding, a guard hands control back to the original code, Maxwell’s website explains.
“AI 改变了软件的编写方式，但未改变其运行方式。”——Amnon Shashua，doubleAI，在一份声明中表示。根据公司的说法，在真实工作负载下，Maxwell 使数据管道的速度提升了超过 38 倍。它使 AI 推理内核的速度提升了超过 4.5 倍。该初创公司告诉《华尔街日报》，它为一位早期客户降低了 67% 的计算成本。
“AI has changed how software is written, but not how it runs.”Amnon Shashua, doubleAI, in a statement On real-world workloads, Maxwell made data pipelines more than 38 times faster, according to the company. It made AI inference kernels more than 4.5 times faster. The startup told the WSJ that it cut one early customer’s compute costs by 67%.
《华尔街日报》报道称，构建开放权重多模态模型的 LTX 发现，Maxwell 使其模型推理速度翻倍。
LTX, which builds open-weight multimodal models, found that Maxwell doubled the speed of its model inference, the WSJ reported.
“将GPU时间减半立刻转化为您的服务成本和利润率。” Zeev Farbman，LTX，对《华尔街日报》表示。从节省的费用中，doubleAI计划向客户收取他们在算力上节省费用的一部分。Shashua未透露具体比例。他预计，算力账单巨大的AI实验室，如OpenAI和Anthropic，将受益最大。对他来说，即使节省10%，也可能价值数亿美元，他对《华尔街日报》表示。
“Halving the GPU time immediately translates to your serving costs, your margins.”Zeev Farbman, LTX, to The Wall Street Journal Paid from the savings doubleAI plans to charge customers a share of what they save on compute. Shashua did not say how large a share. He expects AI labs with huge compute bills, such as OpenAI and Anthropic, to benefit most. For them, even a 10% saving could be worth hundreds of millions of dollars, he told the WSJ.
租用AI算力成本更高。《华尔街日报》援引SemiAnalysis的数据报道称，一年期合约下Nvidia H100芯片的小时价格在一年内上涨了60%。该初创公司大约将一半的资金用于购买1000块Nvidia B300芯片以满足自身需求，《华尔街日报》报道。
Renting AI computing power costs more. The hourly price for Nvidia H100 chips on a one-year contract has risen 60% in a year, the WSJ reported, citing SemiAnalysis. The startup spent roughly half its funding on 1,000 Nvidia B300 chips for its own needs, the WSJ reported.
“你不能仅仅依赖租用算力。价格正在上涨。” Amnon Shashua，对《华尔街日报》表示。该公司得到Nvidia和Dell的支持。根据《华尔街日报》的报道，这家成立两年的公司约有40名员工，累计融资2.2亿美元。其公告称，公司完成了由Lightspeed和Bessemer Venture Partners共同牵头的2亿美元A轮融资。Nvidia、Dell Technologies Capital、Greenfield、Pitango和BRM也参与了投资。
“You cannot rely just on renting compute. Prices are going up.”Amnon Shashua, to The Wall Street Journal Backed by Nvidia and Dell The two-year-old company has about 40 employees and has raised $220m in total, according to the WSJ. Its announcement says it raised a $200m Series A co-led by Lightspeed and Bessemer Venture Partners. Nvidia, Dell Technologies Capital, Greenfield, Pitango and BRM also invested.
Shashua也是Mobileye的首席执行官，Mobileye是他帮助创立的自动驾驶公司。7月，Mobileye表示，一旦确定继任者，他将卸任。
Shashua is also chief executive of Mobileye, the autonomous driving company he helped launch. In July, Mobileye said he would step down once it names a successor.