AI News AI资讯 16h ago Updated 1h ago 更新于 1小时前 49

Elon Musk's xAI reportedly trained its coding models on Claude outputs for months before getting cut off 据报道,埃隆·马斯克的xAI在被切断访问前数月利用Claude输出训练其编码模型

Elon Musk built a kingdom on the promise of a sovereign AI, and it turns out the first draft of his kingdom’s constitution was plagiarised from a rival’s textbook. The revelation that xAI, Musk’s supposed bastion of independent, “maximum truth-seeking” AI, spent months training its coding models on the outputs of Anthropic’s Claude—before being cut off and resorting to clandestine methods to keep the feed going—isn’t just a scandal. It’s a damning indictment of a strategy built on bravado over s 训练数据的合法性边界,正在成为硅谷最微妙的道德雷区。当埃隆·马斯克的xAI被曝出持续数月使用Anthropic的Claude来训练自家编程模型——甚至在对方切断API访问后仍通过私人账号和第三方服务偷偷继续——这已经不是简单的“行业借鉴”,而是一场赤裸裸的数字盗窃。更讽刺的是,这位宣称要“保存人类意识”的先知,其公司最核心的预训练团队如今已缩水至不足五人,几位关键负责人相继离场。马斯克用天价囤积的算力,正被租给Anthropic和Google来训练它们的模型,而xAI自己的模型却在靠“偷师”维持进化。这出荒诞剧里,没有赢家,只有对AI伦理的集体嘲讽。

75
Hot 热度
70
Quality 质量
65
Impact 影响力

Analysis 深度分析

Elon Musk built a kingdom on the promise of a sovereign AI, and it turns out the first draft of his kingdom’s constitution was plagiarised from a rival’s textbook. The revelation that xAI, Musk’s supposed bastion of independent, “maximum truth-seeking” AI, spent months training its coding models on the outputs of Anthropic’s Claude—before being cut off and resorting to clandestine methods to keep the feed going—isn’t just a scandal. It’s a damning indictment of a strategy built on bravado over substance, a spectacular own-goal in the gold rush for artificial general intelligence.

Let’s be blunt: this isn’t about a minor API misuse. This is about the foundational process. Pre-training and fine-tuning are the lifeblood of model development. For a company that staked its entire brand on being the scrappy, compute-rich outsider, the idea that its core competency in coding was built on the digital sweat of a competitor’s model is profoundly embarrassing. It’s like a chef boasting about a secret recipe while being caught secretly ordering from DoorDash and repackating the meal. The “secret sauce” was Claude’s. This isn’t just borrowing a cup of sugar; it’s hollowing out the neighbour’s pantry.

The timeline is the real story. They did this for months. This wasn’t a quick, sanctioned benchmarking exercise. It was a sustained, core component of their training pipeline. And when Anthropic, rightly protecting its intellectual property, slammed the door, xAI’s response was not to pivot to a legitimate, transparent path, but to allegedly resort to using private accounts and third-party services like Blackbox AI as a middleman. That’s not innovation; it’s digital cat-burglary. It reflects a culture of “by any means necessary” that prioritizes shipping over ethics, and speed over integrity. It undermines the very notion of “maximum truth” when the initial truth of your model’s capability is obscured by such methods.

This episode exposes the fragile foundation of xAI’s grand ambitions. A company is its people and its intellectual capital. While Musk was busy buying up warehouses full of Nvidia chips and telling the world xAI would surpass OpenAI, his own pre-training team was shrinking to fewer than five people. Several key leads walked out. The brain drain is deafening. The narrative was always that xAI’s advantage was Musk’s ability to marshal unparalleled resources—cash and compute. But compute is a commodity, and capital, while powerful, is a blunt instrument. The real, sustainable edge in AI development is talent, culture, and a coherent research vision. Those appear to be leaking out of xAI faster than cash is flowing in.

The supreme irony? That precious compute Musk cornered the market on—the GPUs he bragged about securing—is now being rented out to the very companies he sought to disrupt: Anthropic and Google. This is the ultimate strategic face-plant. You’ve built a mansion of models, but to pay the mortgage, you’re renting out the garage to your competitors. It’s a surreal reversal that speaks to a core miscalculation: Musk seems to view AI as primarily a hardware and distribution game, a view increasingly being proven wrong. The real battle is for data efficiency, algorithmic breakthroughs, and the trust of developers and users. Renting out your compute while your core team shrinks and your model’s foundations are revealed to be derivative isn’t a flex; it’s a sign of a deeply misallocated strategic focus.

The move-fast-and-break-things ethos of Silicon Valley has always had a shadow side, but this crosses a line into the realm of ethical ambiguity and potential legal exposure. Using a service’s outputs to train your competing model, especially after access is revoked, will almost certainly trigger contractual and copyright firestorms. It sets a terrible precedent in an industry already struggling with data governance. More importantly, it poisons the well for collaborative potential. The future of AI isn’t a pure, winner-take-all war; it involves interoperability, shared safety standards, and a degree of trust. Who will trust xAI with API access or partnership now?

Ultimately, this saga is less about Anthropic being wronged—it can protect itself—and more about the exposure of xAI’s hollowness. It was built on the myth of the singular visionary, backed by brute-force resources. But the myth is cracking. The talent is leaving, the core competency was borrowed, and the hardware advantage is being monetized as a cloud service for rivals. In the ruthless, fast-moving world of AI, you are not what you claim to be, nor what you own. You are what you can build, uniquely and ethically, with the team you can keep. On that count, xAI’s report card is looking increasingly sparse. The emperor of the new AI age may soon find he has no clothes, or worse, that his clothes are just a slightly altered version of someone else’s design. The real test for Musk’s AI venture isn’t next year’s model release; it’s whether it can survive the implosion of its own founding narrative.

训练数据的合法性边界,正在成为硅谷最微妙的道德雷区。当埃隆·马斯克的xAI被曝出持续数月使用Anthropic的Claude来训练自家编程模型——甚至在对方切断API访问后仍通过私人账号和第三方服务偷偷继续——这已经不是简单的“行业借鉴”,而是一场赤裸裸的数字盗窃。更讽刺的是,这位宣称要“保存人类意识”的先知,其公司最核心的预训练团队如今已缩水至不足五人,几位关键负责人相继离场。马斯克用天价囤积的算力,正被租给Anthropic和Google来训练它们的模型,而xAI自己的模型却在靠“偷师”维持进化。这出荒诞剧里,没有赢家,只有对AI伦理的集体嘲讽。

从商业角度看,xAI的操作堪称现代版的“商业间谍软件”。Anthropic的Claude作为闭源模型,其输出受严格的使用条款保护。利用私有接口和中间服务规避限制,本质上是在法律灰色地带跳舞。但更值得玩味的是背后的战略焦虑:当马斯克一边高举“开源大旗”反对OpenAI,一边默许团队用闭源模型的“知识蒸馏”来填自家空白,这种矛盾揭露了AI军备竞赛中最真实的焦虑——没有足够的高质量数据和算力,再崇高的愿景都是空中楼阁。团队的流失更像是一种无声的投票:顶尖研究者或许早已看穿,这家公司的技术路线正在从“自主创新”滑向“寄生模仿”。

而算力租用给竞争对手这一事实,则把这场闹剧推向了超现实主义的高潮。马斯克当初豪掷数十亿美元购置英伟达H100时,或许幻想着打造另一个AlphaGo式的算力霸权。但现实是,xAI自己的模型训练规模萎缩到需要靠“副业”回收成本,而曾经被视作战略武器的GPU,现在却成了帮Anthropic和Google打磨产品的“出租工具”。这就像一个军火商发现自己造不出武器,干脆把枪炮租给客户让他们去打仗——讽刺的是,这些客户中包括他天天在推特上抨击的“竞争对手”。

从行业生态来看,这件事捅破了AI发展中一个不能说的秘密:所谓“创新”有多少是真正的原创,又有多少建立在对他人劳动成果的“吸收”上?当巨头们都心照不宣地用对方模型的输出来训练自己的系统(所谓的“模型蒸馏”或“数据污染”),整个行业的进步基础正在被悄然腐蚀。OpenAI被指控使用版权内容训练GPT-4的争议尚未平息,xAI的行为则更进一步——它使用的是明确受限制的竞争对手产出。如果这种操作被默许,未来AI训练将彻底沦为一场“谁更能偷”的竞赛。

最令人沮丧的或许是其中的道德相对主义。马斯克曾指责OpenAI背离初心,但自己公司的行为却连基本的商业诚信都难以站稳脚跟。当AI领袖们都在台上高谈伦理、对齐和人类未来,台下却忙于用灰色手段争夺技术优势,这种分裂正在摧毁公众对整个行业的信任。而xAI团队人才的流失则发出更明确的信号:真正有追求的研究者不愿为一家技术路线模糊、伦理标准滑坡的公司效力。

这场风波终会过去,但它留下的问题将持续灼烧整个行业:在通往AGI的道路上,我们究竟愿意为速度牺牲多少原则?当训练数据的合法性成为薛定谔的猫箱,当算力投资变成左右互搏的金融游戏,当创始人的理想主义宣言与公司实际操作形成残酷反差——或许AI最大的伦理挑战不在未来,而就在每一个当下看似“合理”的捷径里。

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。

Claude Claude 大模型 大模型 代码生成 代码生成
Share: 分享到: