“PC需要被重新发明。”
当地时间10月21日，英伟达CEO黄仁勋在柏林GTC大会上宣称，以往“输入指令，电脑执行”的范式已经走向终结，PC只有成为个人Agent，理解用户意图并主动执行任务，才符合AI时代的需求。
英伟达成立于1993年PC革命前夕，起起伏伏30余年后，成为了AI革命的最重要推手，也从“在公寓办公”“吃了上顿没下顿”的创业团队变成了全球市值最高的上市公司。
而在2022年底AI浪潮爆发后，英伟达在赚得盆满钵满的GPU业务之外，也开始积极布局机器人、自动驾驶、量子计算等前沿方向，力求押中下一个时代风口。在这一过程中，黄仁勋也没少打PC的主意，而在今年年初宣布“Agentic AI拐点已至”后，相关进程明显加速。
在柏林GTC大会上，黄仁勋也展示了英伟达“为AI时代重新设计的PC”RTX Spark，还称为了这台设备，公司与微软秘密合作了四年，投入了数千个“工程师年”。
不过，PC真的能成为Agent的最佳宿主吗？对于这个“旧瓶”是否真的能让“新酒”更美味，目前还有很多不同看法。
Agent引领AI本地化浪潮
人机交互的方式正在快速变革，快到大部分人还无法适应。
今年以来，从“小龙虾”OpenClaw到OpenAI的Codex，再到腾讯的WorkBuddy、Meta的Muse，新的Agent层出不穷，掀起一波波热浪。但在部署、使用方面，除了对普通消费者仍有一定门槛外，更为突出的问题就是隐私、安全和成本问题。
在这一背景下，模型、设备端也在同步做出调整，较为突出的表征就是小模型、本地模型和AI PC、AI手机的涌现。目前来看，业界考虑的核心逻辑，还是用消费者最熟悉的设备，搭载能在安全、隐私和成本上做得更好的模型，来让更加个性化的Agent能够跑在本地。
从英伟达目前公布的信息来看，其产品逻辑也与此一致。
相关信息显示，在英伟达与微软的合作中，前者提供GPU、AI软件栈和本地模型运行能力，后者则试图从Windows操作系统层面，为Agent提供执行、隔离和权限管理机制。
今年5月，微软公布了面向Agent的Windows安全能力；英伟达则通过OpenShell等工具，尝试让本地Agent在可控的权限范围内运行。英伟达官网文章称，双方希望解决的不只是AI能否在PC上运行，更是它能否在真实工作环境中持续执行任务。微软还在推进执行容器等机制，限制Agent对文件、网络和系统资源的访问范围。
而RTX Spark目前给出的最大卖点，则是从英伟达的芯片能力出发，将数据中心级的AI计算能力压缩进本地个人设备。
根据英伟达公布的信息，RTX Spark采用融合CPU与GPU的芯片设计，提供最高128GB统一内存和最高1 petaflop的FP4 AI算力，用户得以运行更大规模的本地模型，相比云端模型，能够用更小的Token消耗，让Agent完成更复杂的开发、设计、游戏制作等场景。在大会现场，黄仁勋也让自己颇为钟爱的Hugging Face MicroDuck“小黄鸭”机器人走上舞台。并介绍称这些机器人的“大脑”正是在RTX Spark上构建，并由Agent驱动它们完成行走和指令响应，由此证明其新一代PC的强大AI能力。
黄仁勋还设想称，十年后人们对PC的理解会完全不同，就像今天人们还沿用“手机”的称呼但很少用这个设备“打电话”。在他看来，未来每个家庭都可能会配备一台超级AI电脑，全天候执行Agent任务。
不过，在有关此次GTC的讨论中，仍有不少声音提到“PC只是权宜之计”，认为Agent爆发未必能催化AI PC走向主流，未来能让Agent真正落地个人设备、本地场景的，可能是其他模式。
为Agent找“宿主”，还是让Agent“当家做主”？
目前，端侧、本地智能应用时间较长、验证较多的是智驾座舱、工业传感器等垂类应用场景。而在消费端的普及上，仍以手机、PC为主力，AI眼镜、智能手表以及智能手环、戒指甚至帽子等可穿戴设备也屡屡“上新”。
今年以来，最大的变化当属盒子产品的走红，包括家庭、工作场景，都冒出了不少新的盒子产品。而这些盒子最主要的特点，就是7x24小时地在本地跑Agent，这实际上正迎合了想改变“看屏幕-手动输入”这一主流人机交互方式的想法。此外，具身智能的爆发，也让机器人真正走入人类生活的设想，变得离现实更近了一步，但那更多还被视为Agentic AI后的下一个阶段，即物理AI时代。
上述种种设备，都自有其拥趸，也都在主打Agent概念，但同时也各自面临瓶颈。而值得注意的是，目前对Agent“最佳宿主”的讨论中，越来越多声音都提及，即使在本地场景，更重要的也不是为Agent找到单一载体，而是如何让Agent连接尽可能多的可信设备。换言之，就是让Agent在经授权范围内作主导。
上文提及的盒子模式，包括工作站、边缘服务器等，也是这种思路下的一种实现方式。
其核心是把Agent放到后台实现“无感”“常开”，保证Agent在场，并形成记忆和智能，而手机、PC或其他设备，更多作为入口层和呈现层，即使设备被更换，甚至未来某类设备被另一类替换，也不影响Agent本身。
有端侧芯片从业者介绍称，公司今年量产的芯片原本是针对AI PC设计的，但在出货上，反而是边缘服务器、盒子类产品采购订单大增，其增速远快于AI PC。他认为，这其中既有主流消费电子受存储涨价影响销售走弱的因素，也体现出新品类释放的想象力和迭代速度要快于传统品类。
另外，据称随着Muse引发的个人智能体爆发，目前越来越多的盒子产品也在探索本地Agent和设备中联网的端侧Agent更好协同的路径。
除了盒子类产品，还有一种被讨论较多的方式是“虚拟主机”的模式。
以往，PC兴起的核心优势之一就是将CPU、内存、存储和软件都集中在一台机器上，在硬件、软件能力过线后，更便于各类场景的使用。但Agent恰恰可以突破这种集中式结构，用户的旧电脑、游戏主机、家庭服务器，甚至其他具备计算能力的设备，都可能成为分布式的算力节点：一台设备运行本地模型，另一台处理视觉任务，主力电脑则继续承担日常工作。目前越来越多聚焦本地场景的厂商，在提供将个人网络中的多台PC连接起来，根据任务需求分配AI推理工作的解决方案。
当然，更有可能的是，那个更好的承载Agent的模式还没有被真正想到或看见。
上述端侧芯片从业者称，人们习惯科技变革中沿着延长线看问题、做预测，找到下一个风口或解决方案，但历史经验已经证明，真正的“爆款”和“答案”，永远在延长线之外。（本文首发于钛媒体APP，文｜飞向Tai空）
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。
"PC needs to be recreated."
On October 21st, local time, the CEO Hoang In-hoon declared at the GTC conference in Berlin that the past paradigm of “input instructions, computer execution” had come to an end and that PC would only meet the needs of the AI era if it became an individual Agent who understood the intent of the user and took the initiative to carry out the mission.
Founded on the eve of the 1993 PC revolution, Yin Weida became the most important driver of the AI revolution after more than 30 years of ups and downs and downs, and turned from the “employer team in the apartment” into the highest-market listed company in the world.
Following the outbreak of the AI wave at the end of 2022, in addition to the lucrative GPU business, Young Weidar began to actively deploy the front lines of robots, autopilots, quantum computing, and so on, in an effort to keep the next era in motion. In-hoon Wong did not make little use of the PC in this process, which was significantly accelerated by the announcement earlier this year of the “Agentic AI Junction”.
At the GTC conference in Berlin, Hoang In-hoon also showcased Young Weidar's “Re-engineered PC” RTX Spark, also known as the device, and the company worked secretly with Microsoft for four years, investing thousands of “years of engineers”.
But can the PC really be the best host for Agent? There are many differences of opinion as to whether the old bottle really makes the new wine better.
Agent leads the localization wave of AI
The mode of human interaction is rapidly changing and is fast becoming unadapted for most people.
From OpenClaw to Codex in OpenAi this year, to WorkBuddy in Tencent and Muse in Meta, the new Agent has sprung a wave of heat. However, in terms of deployment and use, in addition to a certain threshold for ordinary consumers, the more prominent issues are privacy, security and cost.
In this context, models and end-of-equipment adjustments are being made simultaneously, with the more prominent manifestations being the emergence of small models, local models and AI PC and AI mobile phones. At present, the core logic of industry considerations is to use the best-known equipment of consumers to carry better models that can be done in terms of safety, privacy and cost, so that the more individual Agent can run locally.
The product logic is consistent with that of the information that is currently being published in Weida.
The information shows that, in collaboration with Microsoft, the former provides GPU, AI and local modelling capabilities, while the latter seeks to provide Agent with implementation, isolation and permission management mechanisms at the Windows operating system level.
In May of this year, Microsoft released the Windows security capability for Agent; and, through tools like OpenShell, we tried to get local Agent to operate within a controlled permission. According to YBIN, what the parties want to resolve is not only whether AI can operate on the PC, but also whether it can continue to operate in a real working environment. Microsoft is also advancing mechanisms such as the implementation of containers to limit Agent ' s access to documents, networks and system resources.
The biggest selling point currently given by RTX Spark is to compress the data centre-level AI computing capacity into local personal equipment, starting with the British chip capability.
According to the information published by Ying Weidar, RTX Spark uses a chip design that integrates CPU and GPU to provide up to 128GB unified memory and up to 1 petaflop FP4 AI calculator to enable users to run larger-scale local models that can be consumed with a smaller Token consumption than the cloud model, allowing Agent to complete more complex developments, designs, games, etc. On the floor of the General Assembly, Wong In-hoon also put on stage his beloved Hugging Face MicroDuck “Duck”. It is on RTX Spark that the “brains” of these robots are constructed and are driven by Agent to complete the walk and command response, thereby demonstrating the strong AI capabilities of the new generation of PCs.
Huang In-hoon also envisaged that a decade later the understanding of PC would be completely different, as today people still use the term “cell phone” but rarely “call” with this device. In his view, every family in the future could be equipped with a super-AI computer to perform the Agent mission all day.
However, in this GTC discussion, there are still a number of voices referring to “PC as a stopgap measure” and suggesting that the Agent outbreak may not be a catalyst for the mainstream of AI PC, and that there may be other models in the future that will allow Agent to actually land on personal equipment, local scenes, and so on.
To find a host for Agent, or to get Agent to take over?
At present, end-side, local intelligence applications are longer and more highly validated in niche applications such as the smart drive module and industrial sensors. On the consumer side, mobile phones, PCs, AI glasses, smart watches and wearable devices such as smart bracelets, rings and even hats are still dominant.
Since this year, the biggest change has been in the red-hot form of box products, including home and work scenes. And the most important feature of these boxes, which is 7x24 hours of running local Agent, is actually meeting the idea of changing the “screen-hand-input” mode of mainstream human interaction. Moreover, the idea of a smart outbreak that would actually allow robots to enter human life is a step closer to reality, but it is also considered to be the next stage after Agenic AI, the physical AI era.
All of the above-mentioned equipment, which has its own resources, and which also dominates the Agent concept, also faces bottlenecks. It is noteworthy, however, that in the current discussion of Agent's “best host”, more and more voices are referring to the fact that, even in the local scene, what is more important is not to find a single carrier for Agent, but how to connect him to as many credible devices as possible. In other words, it would be to allow Agent to dominate within the mandate.
The box models mentioned above, including workstations, peripheral servers, etc., are also one way of achieving this in this context.
The core is to put Agent backstage in order to “unsurely” “open” to ensure that Agent is present and to form memory and intelligence, while mobile phones, PCs or other equipment are used more as entrance and presentation layers, even if the equipment is replaced and even if some type of equipment is replaced in the future by another type, without prejudice to Agent itself.
An end-side chip operator described the company ' s production of chips this year as originally designed for AIPC, but on delivery it was a surge in purchase orders for peripheral servers and box-based products, which grew much faster than AIPC. In his view, these included factors that weakened sales owing to increased storage prices for mainstream consumer electronics, as well as reflections of the imagination and iterative speed of the release of new products as compared to traditional ones.
In addition, with the outbreak of the individual AI agent, allegedly triggered by Muse, a growing number of box products are also exploring ways of better synergy between the end of the network in local Agent and the equipment.
In addition to box-type products, there is a more discussed model of “virtual hosts”.
In the past, one of the core advantages of the emergence of the PC has been the concentration of CPU, memory, storage and software on a single machine, which has made it easier to use the various scenarios once hardware, software capabilities have passed the line. But it is precisely through this centralized structure that Agent can be a distributed calculator node of the user's old computers, game hosts, home servers and even other computing devices: One equipment runs local models and the other handles visual tasks, while the main computer continues to perform routine tasks. A growing number of producers focusing on local scenes are providing solutions to connect multiple PCs in their personal networks and to assign AI reasoning to mission needs.
Of course, it's more likely that the better way to carry Agent has not really been thought of or seen.
These end-side chips practitioners say that people are used to looking at problems along the extension line, predicting and finding the next windfall or solution, but historical experience has shown that real “explosives” and “the answer” are always outside the extension line. This post was launched on titanium medium APP, and Ventura flew to Tai
Auto-fetched and translated by AI Atlas. All rights belong to the original author.