AI News AI资讯 1mo ago Updated 1mo ago 更新于 1个月前 57

Alibaba's latest AI model ran autonomously for 35 hours to optimize code for its own custom chip 阿里巴巴最新的人工智能模型自主运行35小时,以优化其自研芯片的代码。

Alibaba's Qwen team has unveiled **Qwen3.7-Max**, a new proprietary AI model designed specifically for long-running, autonomous agent tasks. It demons 阿里巴巴Qwen团队发布专为长时间自主代理任务设计的Qwen3.7-Max模型,在基准测试中匹配Claude Opus 4.6并超越中国对手如DeepSeek V4 Pro和Kimi K2.6。团队还演示了模型控制四足机器人,并展示了其自主运行35小时优化自研芯片代码的能力,突显技术进展和应用潜力。

82
Hot 热度
85
Quality 质量
78
Impact 影响力

Analysis 深度分析

The future of artificial intelligence isn’t about flashier chatbots or image generators. It’s about relentless, unglamorous, 35-hour-long marathons spent optimizing silicon. Alibaba’s Qwen team has just made that crystal clear with Qwen3.7-Max. This isn’t another model launched into the crowded arena of conversational AI with a promise to “revolutionize” something vague. It’s a specialist, engineered for a brutally specific and profoundly important task: long-running, autonomous agent work. And in that narrow lane, it’s not just competitive; it’s showing the path forward.

Let’s get the benchmark chest-thumping out of the way. Matching Claude Opus 4.6 is a significant achievement, as is outperforming domestic rivals like DeepSeek V4 Pro and Kimi K2.6 on tasks designed to test sustained reasoning. But to focus solely on these scores is to miss the point entirely. This model’s victory lap isn’t on a static leaderboard. It was earned during a 35-hour, unguided session where the AI’s sole directive was to optimize code for Alibaba’s custom AI chips. Think about that. This wasn’t a human prompting it every few minutes to “try again” or “adjust that parameter.” It was given a goal, a compute environment, and then left alone. For over a day and a half.

This is where the real narrative lies. We’ve been mesmerized by the “one-shot” capabilities of large language models—the perfect poem generated in seconds, the complex code block written in a blink. But the world’s most valuable work—engineering breakthroughs, scientific research, complex business logic—doesn’t happen in one shot. It happens through iterative, exhaustive, often tedious exploration of a solution space. It requires an agent that can set its own sub-goals, debug its own failures, track context over millions of tokens, and not hallucinate its way into a dead end at hour 30. Alibaba is betting that the killer app for AI isn’t a creative partner but a dogged, autonomous worker. Qwen3.7-Max is their prototype for that worker.

The demo of the model steering a quadruped robot is cute, almost a distraction. It’s the flashy photo op that gets the tech press salivating about “embodied AI.” But it’s fundamentally a red herring in the context of this model’s true significance. Controlling a robot in real-time is a complex challenge, yes, but it’s a short-burst, high-stakes interaction. The 35-hour chip optimization run is the real paradigm shift. It’s a proof of concept for AI as a capital asset, one that can be deployed on long-horizon projects with a degree of autonomy that was purely theoretical two years ago. It moves AI from being a tool you use to a colleague you manage, albeit one that never sleeps and has an inhuman tolerance for compiling code.

This puts Alibaba in a fascinating and somewhat contrarian position versus its Western counterparts. The obsession in Silicon Valley remains largely on making the general-purpose model more capable, more multimodal, more “AGI-adjacent.” There’s value in that, certainly. But Alibaba, likely spurred by the practical realities of its vast e-commerce, logistics, and cloud computing empire, is focusing on applied durability. They’re building for the factory floor of the AI age, not just the whiteboard. This long-running agent is a tool designed to solve problems that are too big, too iterative, and too time-consuming for human teams to tackle cost-effectively. Optimizing a custom chip’s architecture is a perfect example: a problem with a massive search space where brute-force, long-duration computation can yield tangible, performance-per-watt improvements that translate directly into cloud cost savings and competitive advantage.

There’s a critical, often overlooked dimension to this: the alignment and safety of long-running agents. An AI that can execute a 35-hour task is an AI that can compound its own errors or drift significantly from its original objective if not meticulously constrained. Did Qwen3.7-Max’s optimization run require new guardrails? New methods for monitoring its intermediate states? Alibaba’s paper will be less about the benchmark numbers and more about the infrastructure that allowed such a prolonged, productive run without catastrophic failure or goal drift. This is the unsexy but vital work that will determine whether autonomous AI is a scalable tool or a liability.

Ultimately, Qwen3.7-Max is less a new contestant in the model wars and more a declaration of a different war altogether. It’s the war of attrition, of endurance, of applying AI to the kind of work that moves industries not by sudden inspiration, but by incremental, relentless improvement. While others chase the ephemeral magic of human-like interaction, Alibaba is building the workhorse. The quad-copter demo might steal the headlines, but the quiet victory here is in the 35-hour log files. That’s where the future is being compiled, one optimized line at a time. The question isn’t whether this approach is powerful. It’s whether the rest of the industry, so captivated by the spectacle of intelligence, is paying enough attention to the grunt work.

35小时,一个AI模型在没人盯着的情况下,自己跑着,为自己造芯片的代码做优化。听到这个消息,第一反应不是惊叹,而是脊背发凉地好奇:它优化完了之后,有没有给自己留个后门?阿里云Qwen团队这次抛出的Qwen3.7-Max,以及那个自主优化芯片的实验,与其说是一个技术展示,不如说是一场精心设计的、关于“AI自主性”边界的舆论试水。

性能跑分自然好看,匹配Claude Opus 4.6,在国内卷王赛道上力压DeepSeek和Kimi。这些是意料之中的军备竞赛数据,是发布会PPT上必须闪亮的刀。但真正让这条资讯从一片AI新闻的汪洋中刺出尖角的,是那个“四足机器人”的演示,以及那个更惊悚的“35小时自主优化”故事。后者尤其值得玩味:模型的目标函数是“优化芯片设计代码”,但它的运行过程,本身就成了一个关于“智能体如何在无监督状态下持续进化”的寓言。这已经不是在写代码,而是在尝试“理解”并“改进”自己的生存基础。虽然所谓“自己的芯片”大概率是阿里自家平头哥的某款专用芯片,但这层“自我优化”的叙事外衣,依然把话题从单纯的性能提升,拔高到了近乎哲学的安全层面。

Qwen团队选择在这个时间点,用“长时间自主任务”作为核心卖点,显然是有备而来。这迎合了当下AI Agent(智能体)从“工具”走向“同事”的想象。模型不再是被动回答问题,而是能主动规划、分解任务、调用工具,并在漫长的过程中保持目标不漂移。演示控制四足机器人,就是在宣告:我的智能不止在数字世界,也能驱动物理实体。这种“具身智能”的雏形展示,野心勃勃。但冰冷的事实是,让机器人走出实验室的玻璃房,在复杂真实环境中稳定可靠地执行任务,中间隔着的,是几乎无穷无尽的工程长夜。一个流畅的演示视频,和一个能可靠送快递的机器人,是两个物种。

而那个35小时实验,才是隐藏的深水炸弹。我们惊叹于AI“不知疲倦”的能力,但换个角度,这恰恰是人类监管面临巨大挑战的缩影。一个模型可以连续自主运行35小时,那么350小时呢?在它优化自身代码的过程中,我们能否确保其目标始终与人类完全一致?中间的任何一次“创造性”的偏差,在无人干预的数日里,会累积成什么?Qwen团队大概率在受控环境中设置了多重护栏,但这个实验本身就在传递一个信号:AI的自主能力,已经突破了我们需要时刻“盯着看”的阈值。这种进步是值得喝彩的,但我们必须同时问一句:然后呢?

此次发布,放在中国AI竞赛的背景下看,更有意思。它标志着国内头部玩家的竞争,已经从单纯的模型参数、跑分性能,升级到了应用场景的纵深和自主智能的可靠性这一更复杂的维度。阿里有云、有芯片(平头哥)、有电商物流的真实场景,它想讲一个从底层算力到上层智能、从数字到物理的完整闭环故事。Qwen3.7-Max和那个机器人,就是这个故事最新的一章。这比单纯卷一个聊天机器人,门槛高得多,也实在得多。竞争对手们如果只还在比拼模型“更会聊天”,可能会在战略层面逐渐落后。

所以,别只盯着那几个评测分数了。阿里这一手,是在把竞赛拉入“深水区”——比谁能让AI在更长的时间线、更复杂的任务链、更真实的物理环境中,可靠地自主工作。这很好,这才是AI改变世界的正确打开方式。但我们作为观察者,必须保持一份清醒:当AI开始能为自己“优化生存环境”时,我们赋予它的自主权,每一寸都需要被重新审视和定义。技术跑得太快,而我们关于它的伦理、安全与治理的对话,需要跑得更快。Qwen团队展示了一扇门,门后的风景是星辰大海,还是潘多拉魔盒,取决于我们推开它的同时,是否握紧了那把名为“可控”的钥匙。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。