上周，Meta暂停了它的人工智能(AI)机器人，就在它向公众发布仅两天后。这个名为卡拉狄加的机器人接受了“1060亿个开放获取的科学文本和数据令牌”的训练。这包括论文、教科书、科学网站、百科全书、参考资料、知识库等等，”该公司告诉 The Daily Beast。
它本应帮助学者和研究人员快速简洁地找到论文和研究，但取而代之的是被大量错误引用著名科学家的信息所淹没。如果文章中错误地引用了科学家的话，他们的声誉可能会受到威胁。华盛顿大学生物学教授卡尔·伯格斯特罗姆告诉CNET，卡拉狄加的问题在于，它被宣传为获取事实和信息的一种方式。但相反，他说，这就像“一个花哨的游戏版本，你开始写半句话，然后让自动完成填充剩下的故事。”
在卡拉狄加上线后的几个小时内，用户开始报告种族主义和不准确的文章，其中一人发布了一张回应语言偏见请求的图片。该机器人的回应错误地声称黑人“没有自己的语言”，移民“不会说一种与他们移民国家的语言不同的语言”。
其他产生的信息包括一项关于吃碎玻璃的好处的虚假研究，以及斯坦福大学研究人员开发一种“同性恋雷达”软件来在Facebook上寻找同性恋者的虚假信息。
一位Meta AI的发言人告诉CNET，“卡拉狄加不是真相的来源，它是一个使用(机器学习)系统来学习和总结信息的研究实验。”他补充说，卡拉狄加“是短期的探索性研究，没有产品计划。”Meta AI首席科学家Yann LeCun告诉媒体，该机器人被移除是因为团队“被推特上的刻薄言论弄得心烦意乱”。
卡拉狄加发布两天后，Meta AI团队暂停了机器人，Meta的首席AI科学家Yann LeCun在推特上说:“卡拉狄加的演示现在离线了。通过随意滥用它来获得乐趣已经不可能了。开心呢?”
Meta AI(之前被称为Facebook人工智能研究)开发了卡拉狄加，作为一种“组织科学”的方式，并浓缩了网上发现的大量科学信息。他们的想法是让人工智能做一些事情，比如解决数学问题、编写科学代码和编写研究摘要。
Meta在卡拉狄加网站上指出，AI模型确实有可能导致“幻觉”的局限性。该网站建议用户验证弹出的任何信息，并继续说“语言模型的输出不能保证真实或可靠，即使是在卡拉狄加这样的高质量数据上的大型输出，”并补充说，生成的文本可能看起来“非常真实和高度自信”，但它仍然可能是错误的。
匹兹堡卡耐基梅隆大学计算机科学教授文森特·科尼策在接受《每日野兽》采访时表示:“我认为，即使这样的系统有许多可预见的缺陷，它也有可取的用途。”
“我的印象是，如果Meta在这次公开发布中投入更多的精力，首先进行更认真的用户研究，让人们注意到系统的可取用途，并诚实和坦率地对待不可取的用途，本可以做得更好。”
原文：
Meta paused its Artificial Intelligence (AI) bot last week, only two days after it went live to the public. The bot, called Galactica, was trained “on 106 billion tokens of open-access scientific text and data. This includes papers, textbooks, scientific websites, encyclopedias, reference material, knowledge bases, and more,” the company told The Daily Beast.
It was supposed to help academics and researchers find papers and studies quickly and succinctly but instead was overwhelmed by vast amounts of misinformation that incorrectly cited reputable scientists.
Related Stories
The Vision Vanish Is the First Laser Projector With Its Own Disappearing Screen
Watch NowHow to Use Apple’s Spatial Audio on AirPods Max and Pro
Acer’s Bike Desk Harnesses Your Anxious Work Energy To Charge Your Devices
Scientists’ reputations could be put on the line when they’re incorrectly cited in the text and Carl Bergstrom, a professor of biology at the University of Washington told CNET that Galactica’s problem is it was promoted as a way to get facts and information. But instead, he said it acted like “a fancy version of the game where you start out with a half sentence, then you let autocomplete fill in the rest of the story.”
ADVERTISEMENT
Within hours of Galactica going live, users started reporting racist and inaccurate articles, with one person posting an image of the response to a request about linguistic prejudice. The bot’s response falsely claimed Black people “don’t have a language of their own” and immigrants “do not speak a language that is different from the language of the country they are immigrating to.”
Other generated information included a fake study about the benefits of eating crushed glass and falsified information about Stanford University researchers creating a “gaydar” software to find gay people on Facebook.
A Meta AI spokesperson told CNET, “Galactica is not a source of truth, it is a research experiment using [machine learning] systems to learn and summarize information.” He added, Galactica “is exploratory research that is short-term in nature with no product plans.” Meta AI Chief Scientist Yann LeCun told the outlet the bot was removed because the team was “so distraught by the vitriol on Twitter.”
Two days after Galactica’s launch, the Meta AI team paused the bot and Meta’s chief AI scientist Yann LeCun tweeted, “Galactica demo is offline for now. It’s no longer possible to have some fun by casually misusing it. Happy?”
ADVERTISEMENT
Meta AI (previously called Facebook Artificial Intelligence Research) developed Galactica as a way to “organize science” and condense an overwhelming amount of scientific information found online. The idea was to have the AI do things like solve math problems, write scientific code, and craft summaries of research.
Meta does note on the Galactica website that the AI model does have limitations that can cause it to “hallucinate.” The site advises users to verify any information that pops up and goes on to say “There are no guarantees for truthful or reliable output from language models, even large ones on high-quality data like Galactica,” adding that the generated text might appear “very authentic and highly confident,” but it could still be wrong.
ADVERTISEMENT
“I imagine that even with its many predictable flaws, there are desirable uses of such a system,” Vincent Conitzer, a professor of computer science at Carnegie Mellon University in Pittsburgh, told The Daily Beast.
“My impression is that Meta would have done better by putting more effort into this public release, by doing more serious user studies first, drawing attention to desirable uses of the system, and being honest and forthcoming about undesirable uses.”
本文由数字化转型网（www.szhzxw.cn）翻译而成，作者：Nikki Main；编辑/翻译：数字化转型网默然。
免责声明: 本网站(http://www.szhzxw.cn/)内容主要来自原创、合作媒体供稿和第三方投稿，凡在本网站出现的信息，均仅供参考。本网站将尽力确保所提供信息的准确性及可靠性，但不保证有关资料的准确性及可靠性，读者在使用前请进一步核实，并对任何自主决定的行为负责。本网站对有关资料所引致的错误、不确或遗漏，概不负任何法律责任。
本网站刊载的所有内容(包括但不仅限文字、图片、LOGO、音频、视频、软件、程序等) 版权归原作者所有。任何单位或个人认为本网站中的内容可能涉嫌侵犯其知识产权或存在不实内容时，请及时通知本站，予以删除。
