NVIDIA Vera CPU Sets a New Standard for Agentic Workloads in AI Factories NVIDIA Vera CPU 为AI工厂中的代理工作负载树立新标准

The AI industry’s current obsession with "agentic" systems isn’t the next revolutionary leap—it’s the desperate, logical endpoint of a scaling law that’s running out of road. We’ve been sold a narrative of continuous, linear progress: first we scaled data and parameters, then we scaled post-training refinements, then we scaled test-time compute for deeper "thought." Now, we’re told scaling the scope of action—letting models chain tools, browse the web, and take real-world steps—is the new frontier. But this feels less like a breakthrough and more like a pivot away from a plateau.

Look at the pattern. When pretraining on web text hit diminishing returns, the field shifted to squeezing more utility from those models via fine-tuning and reinforcement learning from human feedback. When that yielded incremental gains, the clever new trick was "chain-of-thought" and giving models more tokens to "think," essentially brute-forcing better answers with more computational runway. Now, facing the stark reality that bigger models aren’t getting dramatically smarter, the goalposts are being moved from intelligence to utility. The new scaling law isn’t about creating a more brilliant mind; it’s about gifting a passably competent one with more hands, more connections, and a longer leash. It’s a tacit admission that we’re not closing in on AGI through scale alone, so instead we’re building sophisticated workflow automators and branding it as the next epoch.

The term "agentic" itself is a masterpiece of marketing obfuscation. It conjures images of autonomous digital minds pursuing goals. In reality, we’re mostly talking about LLMs wrapped in elaborate if-then loops, calling APIs, and parsing their own outputs with a success rate that would get a human intern fired. The impressive demos are just that—demos. They’re curated paths through a garden of forking possibilities, where failure is edited out. The moment you move from a controlled benchmark to the chaotic, poorly-documented reality of the open web or complex enterprise software, the whole fragile house of cards collapses. Reinforcement learning from real-world feedback sounds powerful, but it’s also dangerously unpredictable. Do we really want models learning from the consequences of their own potentially disastrous actions in live environments?

This shift exposes a foundational contradiction in Silicon Valley’s AI gospel. We were promised thinking machines. We are getting ever-more-capable pattern matchers and probabilistic text generators. The industry’s response to hitting a cognitive ceiling is to reframe the problem: forget profound understanding, let’s focus on operational execution. It’s a pivot from "Artificial General Intelligence" to "Artificial Generally-Useful Workflow-Bot." And there’s big money in the latter, especially for selling to enterprises. But let’s not mistake a lucrative new product category for the dawn of a new kind of intelligence.

The real scaling law of the last two years might be one of diminishing cognitive returns, masked by ever-more-complex engineering. Each new "wave" requires exponentially more compute, data, and human ingenuity to produce marginal improvements in benchmark performance. The agentic turn is the most complex yet—a Rube Goldberg machine of prompts, tool calls, and error handling designed to make the underlying model appear more capable than it is. The next big scaling frontier won’t be found in a GPU cluster. It will be found in the hard, unsexy work of building robust, verifiable systems that can fail safely. Until then, we’re just watching a very advanced autocomplete slowly, and expensively, learn to click buttons.

人工智能行业当下对“智能体系统”的痴迷并非下一次革命性飞跃——而是扩展规律在穷途末路时绝望而合乎逻辑的终点。我们一直被灌输一种持续线性进步的叙事：首先扩展数据和参数，接着扩展训练后优化，然后扩展测试时的计算以实现更深层次的“思考”。如今，我们被告知扩展行动范围——让模型串联工具、浏览网络并执行现实世界的操作——是新的前沿。但这感觉更像是一次突破，更像是从平台期的转身离开。

人工智能行业当下对“智能体系统”的痴迷并非下一次革命性飞跃——而是扩展规律在穷途末路时绝望而合乎逻辑的终点。我们一直被灌输一种持续线性进步的叙事：首先扩展数据和参数，接着扩展训练后优化，然后扩展测试时的计算以实现更深层次的“思考”。如今，我们被告知扩展行动范围——让模型串联工具、浏览网络并执行现实世界的操作——是新的前沿。但这感觉更像是一次突破，更像是从平台期的转身离开。

看看这个模式：当基于网络文本的预训练出现收益递减时，该领域转向通过微调和基于人类反馈的强化学习，从这些模型中榨取更多效用。当那仅带来渐进式收益时，巧妙的新把戏变成了“思维链”，并给模型更多令牌以“思考”，本质上通过更多计算空间暴力破解更好的答案。如今，面对更大的模型并未变得更聪明的严峻现实，评估标准正从智能转向效用。新的扩展规律不再是关于创造更卓越的智慧，而是关于赋予一个勉强合格的模型更多的“手”、更多的连接和更长的控制线。这是一种默认承认：仅靠规模无法接近通用人工智能，因此我们转而构建复杂的工作流自动化程序，并将其标榜为下一个时代。

“智能体”一词本身是营销混淆的杰作，它唤起自主数字心智追求目标的形象。现实中，我们讨论的大多是包裹在复杂条件循环中的大语言模型，调用API，并解析自身输出——其成功率足以让人类实习生被解雇。那些令人印象深刻的演示仅仅是演示——它们是精心策划的、穿越分叉可能花园的路径。