AI辩论

有一家公司正在建造一个AI，它占据了巨大的房间，消耗了一个城镇的电力，并且最近获得了令人惊叹的与人交谈的能力。它可以就任何主题撰写文章或诗歌。它可以轻松通过大学级别的考试。它每天都在获得工程师们尚未公开讨论的新能力。然而，这些工程师确实坐在公司自助餐厅里，辩论他们所创造的东西的意义。下周它会学到做什么？哪些工作可能会被它使变得过时？他们应该放慢或停止，以免激怒龙的尾巴吗？但这不就意味着其他人，可能是那些道德观念较少的人，会首先唤醒龙吗？是否有道德义务告诉世界更多关于这个的信息？是否有义务告诉得更少？我——你——将在那家公司工作一年。我的工作——你的工作——是开发一个关于如何防止AI及其后继者造成破坏的数学理论。这里的“造成破坏”可能意味着任何事情，从加速宣传和学术作弊，到提供生物恐怖主义建议，到是的，摧毁世界。[OpenAI 2027 AGI.pdf](https://bytedance.feishu.cn/space/api/box/stream/download/all/OcrebSvWPoO2TQxuhoQcNL2OnOK?allow_redirect=1)

1.22.The concept of AI is not new,but recent advances in data generation and processing have changed the field and the technology it produces.For example,while recent developments in the capabilities of generative AI models have created exciting opportunities,they have also sparked new debates about potential AI risks.39 As AI research and development continues at pace and scale,we expect to see even greater impact and public awareness of AI risks.402.23.We know that not all AI risks arise from the deliberate action of bad actors.Some AI risks can emerge as an unintended consequence or from a lack of appropriate controls to ensure responsible AI use.413.24.We have made an initial assessment of AI-specific risks and their potential to cause harm,with reference in our analysis to the values that they threaten if left unaddressed.These values include safety,security,fairness,privacy and agency,human rights,societal well-being and prosperity.4.25.Our assessment of cross-cutting AI risk identified a range of high-level risks that our framework will seek to prioritise and mitigate with proportionate interventions.For example,safety risks include physical damage to humans and property,as well as damage to mental health.42 AI38 Intelligent security tools,National Cyber Security Centre,2019.39 What is generative AI,and why is it suddenly everywhere?,Vox,2023.40 See,for example,The Benefits and Harms of Algorithms,The Digital Regulation Cooperation Forum,2022;Harms of AI,Acemoglu,2021.41 AI Accidents:An Emerging Threat,Center for Security and Emerging Technology,2021.42 AI for radiographic COVID-19 detection selects shortcuts over signal,DeGrave,Janizek and Lee,2021;Pathways:How digital design puts children at risk,5Rights Foundation,2021.11

为了回应你的答案，AI很容易“产生幻觉”并生成看似合理的事实。它可以生成完全错误的内容，而且非常令人信服。让我强调一下：AI连续且良好地撒谎。它告诉你的每一件事或信息可能都是不正确的。你需要检查所有的东西。你需要全部检查一下。特别危险的是要求它为互联网提供参考、引用、引文和信息（对于未连接到互联网的模型）。因为GPT-4通常更加扎实，因为Bing的互联网连接意味着它实际上可以拉入相关事实。[这是一份避免幻觉的指南](https://oneusefulthing.substack.com/p/how-to-get-an-ai-to-lie-to-you-in)，但它们不可能完全消除。另请注意，人工智能不会解释自己，它只会让你认为它解释了自己。如果你要求它解释它为什么写东西，它会给你一个完全编造的合理答案。当你询问它的思考过程时，它并没有审查自己的行动，它只是生成听起来像它在做这样的文本。这使得理解系统中的偏见非常困难，尽管这些偏见几乎肯定存在。它也可以被不道德地用来操纵或作弊。你对这些工具的输出负责。