财联社7月29日讯（编辑 马兰）一份内部文件显示亚马逊重新设计了人工智能语音助手Alexa以减少其对Anthropic模型Claude的依赖从而降低Alexa的运营成本。

A specific aspect of the new design is that it redirects the replies of Alexa to use Amazon's own in-house artificial intelligence models to a greater extent, thereby reducing the calls to Claude and improving the utilization rate of each GPU. These measures are expected to result in each computing unit handling more than four times the amount of tasks.

A specific aspect of the new design is that it redirects the replies of Alexa to use Amazon's own in-house artificial intelligence models to a greater extent, thereby reducing the calls to Claude and improving the utilization rate of each GPU. These measures are expected to result in each computing unit handling more than four times the amount of tasks.

This also reflects the next focus of the artificial intelligence competition. With the improvement of the performance of the latest models, the key point of competition is shifting from building smarter artificial intelligence to finding more economical operating methods.

除了亚马逊之外，谷歌正通过Gemini Flash推动低成本人工智能的发展，而OpenAI和Cursor等公司则在推出能够将简单请求直接定向到成本更低的模型的技术。

与之前的Alexa版本相比，最新版本的Alexa+依赖大规模语言模型来生成大量回复，这些模型运行在利用GPU的云服务上，这使得原本相对较低的成本明显增加。

根据早些时候的内部预测，升级后的AI版Alexa+，其云成本预计在今年达到约17亿美元，几乎是上一年的三倍。Alexa+每位活跃用户每月的云成本将比亚马逊的目标高出约60%，这意味着Alexa+将无法达到亚马逊的财务利润目标。

Meanwhile, Amazon has also expanded its service scope in the United States this year, further increasing its cloud costs. The pressure on costs has also led Amazon to consider deferring projects with excessive expenditures. Earlier reports indicated that the project Project Moonraker for enhancing the artificial intelligence agent capabilities of Alexa might be postponed in some parts, in order to lower costs.

An insider also revealed that Amazon is reorganizing its artificial intelligence department with resources gradually shifting from its existing Nova models to a new frontier model research project led by researcher Pieter Abbeel and beginning to phase out most of its in-house flagship Nova models.

此前，亚马逊的人工智能通用技术部门进行了裁员，其2024年成立的研究小组AGI Lab也已被关闭。这一举动令该公司内部十分意外，因为前沿模型研究人员一直是亚马逊最宝贵的技术人才种类之一。

种种迹象表明，亚马逊正在对人工智能业务实施集中，试图将优质资源配置到最关键的应用之上。这在一定程度上也意味着亚马逊希望更加精准地控制人工智能业务，并创造更加高效的回报模式。

来源：亚马逊开始控制AI成本：减少调用外部模型 资源转向前沿研究 | 财联社