摘要：美国“聊天生成式预训练变换器”（ChatGPT）开发者、开放人工智能公司和美国总统唐纳德・特朗普的首席人工智能顾问开启对中国科技公司深言科技（DeepSeek）的抹黑模式，暗示DeepSeek的R1抄袭了开放人工智能的ChatGPT，但却未能披露其所谓的知识产

美国媒体发布的大卫・萨克斯会见特朗普的人工合成照片

美国“聊天生成式预训练变换器”（ChatGPT）开发者、开放人工智能公司和美国总统唐纳德・特朗普的首席人工智能顾问开启对中国科技公司深言科技（DeepSeek）的抹黑模式，暗示DeepSeek的R1抄袭了开放人工智能的ChatGPT，但却未能披露其所谓的知识产权被盗用的具体证据。

特朗普的人工智能顾问大卫・萨克斯周二对福克斯新闻表示：“有大量证据表明，深言科技所做的是从开放人工智能模型中提取的知识。而且我认为开放人工智能对此很不高兴。”

因中国正处春节长假假期，深言科技及其前身对冲基金高毅资产，都没有对CNN的邮件询问立即回复。

开放人工智能在一份声明中称，中国公司“一直在试图从美国领先的人工智能公司的模型中提取知识”，但并未公开具体点名深言科技。

开放人工智能的官方使用条款禁止一种名为“蒸馏”的技术，该技术能让新的人工智能模型通过反复查询一个已训练好的更大模型来学习。该公司一直在与商业伙伴微软合作，识别试图使用其模型进行“蒸馏”的账户，然后封禁这些账户并撤销其访问权限。但微软却对开放人工智能的说法拒绝置评。

开放人工智能表示，它还将“与美国政府密切合作，以最好地保护最强大的模型，防止对手和竞争对手获取美国技术”。

开放人工智能的老板奥特曼

位于旧金山的开放人工智能公司自身也面临多起媒体机构、书籍作者等指控其盗用版权的诉讼，这些案件仍处在美国及其他地方的法院审理中。

科技投资者、康奈尔大学讲师卢茨・芬格周三发表声明称：“‘蒸馏’技术会违反大多数服务条款，但大型科技公司对此提出指责颇具讽刺意味，甚至可以说是虚伪。因为它们用《福布斯》或《纽约时报》的内容来训练ChatGPT，同样违反了它们的服务条款。”

曾在谷歌和领英工作过的芬格表示，虽然深言科技很可能使用这项技术，但很难找到证据，因为它很容易伪装并逃避检测。

深言科技在其公开研究论文中描述了对“蒸馏”技术的使用，并披露了其对脸书母公司元宇宙和中国科技公司阿里巴巴所开发的可公开获取的人工智能模型的依赖。除了在性能对比时，论文中并未提及封闭其模型的开放人工智能。

甚至在周一深言科技的消息震动市场之前，许多试用该公司人工智能模型的人就注意到，它有时宣称自己是ChatGPT，或者出现提及开放人工智能的条款与政策倾向。

前美国国防部官员、现战略与国际研究中心瓦德瓦尼人工智能中心主任格雷戈里・艾伦表示：“如果你问它是什么模型，它会说‘我是ChatGPT’，最有可能的原因是，深言科技的训练数据采集自与ChatGPT的数百万次聊天互动，这些数据被直接输入到深言科技的训练数据中。”

这种宣称并不一定表明存在知识产权盗窃行为，聊天机器人容易编造信息。但是，尽管深言科技将其技术描述为“开源”，却并未披露用于训练其模型的数据。

艾伦说：“我认为做出这种选择有一个相当明显的原因，那就是他们采集了ChatGPT的数据用于训练。”

关于深言科技的很多方面，都让仔细研读这家初创公司关于其新模型 R1 及其前身的公开研究论文的分析师们感到困惑。

令华尔街震惊的细节之一是，深言科技声称其人工智能助手背后的旗舰v3模型的训练成本仅为560万美元，与打造ChatGPT和其他流行聊天机器人所花费的数十亿美元相比，这一数字低得惊人。

深言科技的人工智能的手机应用

深言科技的论文所宣布的这560万美元可能仅包括实际训练聊天机器人的费用，而不包括早期研究和实验的成本。但这个数字，以及深言科技向开发者收取的相对较低的费用，仍让人们对美国在人工智能开发上投入的巨额资金和电力产生是否必要提出质疑。

与此同时，深言科技也是在种种限制下开展工作的，其中就包括美国对最强大的人工智能芯片实施出口管制。该公司表示，其依赖的是加利福尼亚芯片制造商英伟达生产的一款性能相对较低的人工智能芯片，美国尚未禁止该芯片对华销售。但2022年，高毅资产在社交媒体上发布消息称，就在美国限制英伟达更强大芯片对华出口的几个月前，他们已经积攒了1万个此类芯片。

Did the upstart Chinese tech company DeepSeek copy ChatGPT to make the artificial intelligence technology that shook Wall Street this week?

That’s what ChatGPT maker OpenAI is suggesting, along with U.S. President Donald Trump’s top AI adviser. Neither has disclosed specific evidence of intellectual property theft, but the comments could fuel a reexamination of some of the assumptions that led to a panic in the U.S. over DeepSeek’s advancements.

“There’s substantial evidence that what DeepSeek did here is they distilled the knowledge out of OpenAI’s models,” David Sacks, Trump’s AI adviser, told Fox News on Tuesday. “And I don’t think OpenAI is very happy about this.”

DeepSeek and the hedge fund it grew out of, High-Flyer, didn’t immediately respond to emailed questions Wednesday, the start of China’s extended Lunar New Year holiday.

OpenAI said in a statement that China-based companies “are constantly trying to distill the models of leading U.S. AI companies” but didn’t publicly call out DeepSeek specifically.

OpenAI’s official terms of use ban the technique known as distillation that enables a new AI model to learn by repeatedly querying a bigger one that’s already been trained. The company has been working with its business partner Microsoft to identify accounts attempting to distill its models and then banning those accounts and revoking their access. Microsoft declined to comment.

OpenAI said it will also work “closely with the U.S. government to best protect the most capable models from efforts by adversaries and competitors to take U.S. technology.”

The San Francisco company has itself been accused of copyright theft in lawsuits from media organizations, book authors and others in cases that are still working through courts in the U.S. and elsewhere.

“Distillation will violate most terms of service, yet it’s ironic — or even hypocritical — that Big Tech is calling it out,” said a statement Wednesday from tech investor and Cornell University lecturer Lutz Finger. “Training ChatGPT on Forbes or New York Times content also violated their terms of service.”

Finger, who formerly worked for Google and LinkedIn, said that while it is likely that DeepSeek used the technique, it will be hard to find proof because it’s easy to disguise and avoid detection.

DeepSeek describes its use of distillation techniques in its public research papers, and discloses its reliance on openly accessible AI models made by Facebook parent company Meta and Chinese tech company Alibaba. No mention is made of OpenAI, which closes off its models, except to show how DeepSeek compares on performance.

Even before DeepSeek news rattled markets Monday, many who were trying out the company’s AI model noticed a tendency for it to declare that it was ChatGPT or refer to OpenAI’s terms and policies.

“If you ask it what model are you, it would say, ‘I’m ChatGPT,’ and the most likely reason for that is that the training data for DeepSeek was harvested from millions of chat interactions with ChatGPT that were just fed directly into DeepSeek’s training data,” said Gregory Allen, a former U.S. Defense Department official who now directs the Wadhwani AI Center at the Center for Strategic and International Studies.

Such declarations are not necessarily an indication of IP theft -- chatbots are prone to fabricating information. But DeepSeek, despite describing its technology as “open-source,” doesn’t disclose the data it used to train its model.

“I think that there’s a pretty obvious reason for that choice, which is that they harvested ChatGPT for training data,” Allen said.

Much about DeepSeek has perplexed analysts poring through the startup’s public research papers about its new model, R1, and its precursors.

Among the details that startled Wall Street was DeepSeek’s assertion that the cost to train the flagship v3 model behind its AI assistant was only $5.6 million, a stunningly low number compared to the multiple billions of dollars spent to build ChatGPT and other popular chatbots.

The $5.6 million number only included actually training the chatbot, not the costs of earlier-stage research and experiments, the paper said. But the number — and DeepSeek’s relatively cheap prices for developers — called into question the huge amounts of money and electricity pouring into AI development in the U.S.

DeepSeek was also working under constraints: U.S. export controls on the most powerful AI chips. It said it relied on a relatively low-performing AI chip from California chipmaker Nvidia that the U.S. hasn’t banned for sale in China. But in 2022, a social media post from High-Flyer said it had amassed a cluster of 10,000 more powerful Nvidia chips just months before the U.S. restricted their export to China.

来源：读行品世事