The ChatGPT chat screen on a smartphone.Photographer: Gabby Jones/Bloomberg
智能手机上的 ChatGPT 聊天屏幕。摄影师：Gabby Jones/彭博社
By Olivia Solon and Seth Fiegerman
作者：奥利维亚·索伦和塞斯·费格曼
2024年11月16日 at GMT+8 02:43
2024年11月16日 GMT+8 02:43
Save 节省
BPC > disable Dark Reader or enable JavaScript for site
BPC > 禁用 Dark Reader 或为站点启用 JavaScript
When OpenAI launched its ChatGPT chatbot in 2022, it looked as if a new generation of artificial intelligence tools was about to transform every corner of the economy and people’s lives. Fast-forward to 2024 and AI features have been added to countless services, but the impact has arguably not been quite so dramatic.
当 OpenAI 在 2022 年推出 ChatGPT 聊天机器人时，新一代人工智能工具似乎即将改变经济和人们生活的各个角落。快进到 2024 年，无数服务中都添加了人工智能功能，但影响可能并没有那么巨大。
AI has not created vibrant new professions. Neither has it yet laid waste to entire industries. A growing number of high-profile and well-funded startups have bowed out of the race to develop cutting-edge AI models. Some of the leading AI developers are also starting to see diminishing returns from their costly efforts to build more advanced systems that can significantly outperform prior versions.
人工智能并没有创造出充满活力的新职业。它也还没有摧毁整个行业。越来越多备受瞩目且资金充足的初创公司退出了开发尖端人工智能模型的竞赛。一些领先的人工智能开发人员也开始看到，他们花费巨资构建更先进的系统，这些系统的性能明显优于以前的版本，但回报却在递减。
Yet it’s still early days, and the companies are betting on new approaches to push the technology forward. In September, Microsoft-backed OpenAI began rolling out an updated model that the company said can engage in human-like reasoning to work through more complex questions from users, including fielding complicated math and coding problems. OpenAI Chief Executive Officer Sam Altman called it the “beginning of a new paradigm.” Alphabet’s Google is working on similar capabilities for its products.
然而现在还处于早期阶段，这些公司正在押注于推动技术进步的新方法。今年 9 月，微软支持的 OpenAI 开始推出一个更新的模型，该公司称该模型可以进行类似人类的推理，以解决用户提出的更复杂的问题，包括解决复杂的数学和编码问题。 OpenAI 首席执行官 Sam Altman 称其为“新范式的开始”。 Alphabet 旗下的谷歌正在为其产品开发类似的功能。
The concept of reasoning is difficult to define and measure, but the goal for the tech companies is clear: They’re out to prove that AI tools can act as sophisticated “co-pilots” and “agents,” capable of helping people with personal and professional tasks. In doing so, they hope to shore up AI’s long-term potential.
推理的概念很难定义和衡量，但科技公司的目标很明确：他们要证明人工智能工具可以充当复杂的“副驾驶”和“代理”，能够帮助人们处理个人问题。和专业任务。他们希望通过这样做来增强人工智能的长期潜力。
What is generative AI? 什么是生成式人工智能？
It’s a type of software that can carry out complex tasks such as writing a story or creating an image in response to simple written prompts. During training, these systems are fed vast amounts of information (for example, every book available freely on the internet) and are taught how to use that data to craft something new, such as the blurb for a new novel.
它是一种可以执行复杂任务的软件，例如根据简单的书面提示编写故事或创建图像。在训练期间，这些系统会被输入大量信息（例如，互联网上免费提供的每一本书），并被教导如何使用这些数据来制作新的东西，例如新小说的简介。
The systems apply what they learn from such efforts to future endeavors and their responses become gradually more sophisticated and nuanced. The results are unique and — in a sense — original, but are still effectively a complex form of mimicry.
这些系统将从这些努力中学到的知识应用到未来的工作中，并且它们的反应逐渐变得更加复杂和细致。结果是独特的，并且在某种意义上是原创的，但实际上仍然是一种复杂的模仿形式。
How Large Language Models Work, Making Chatbots Lucid: QuickTake
大型语言模型如何工作，使聊天机器人变得清晰：QuickTake
Why the excitement? 为什么兴奋？
Generative AI’s ability to quickly acquire new skills and perform them at high speed means it could help with all kinds of work now done by humans, such as drafting legal contracts or debugging software. Fortune 500 companies are now using enterprise versions of chatbots from OpenAI and its rivals to summarize documents, speed up coding tasks and draft emails. Goldman Sachs Group Inc. said generative AI could do as much as a quarter of the work currently carried out by people. Analysts at UBS Group AG see the market for the technology easily exceeding $1 trillion.
生成式人工智能能够快速获取新技能并高速执行它们，这意味着它可以帮助完成目前由人类完成的各种工作，例如起草法律合同或调试软件。财富 500 强公司现在正在使用 OpenAI 及其竞争对手的企业版聊天机器人来总结文档、加快编码任务和起草电子邮件。高盛集团表示，生成式人工智能可以完成目前由人类完成的工作的四分之一。瑞银集团 (UBS Group AG) 分析师认为，该技术的市场轻松超过 1 万亿美元。
How has generative AI evolved?
生成式人工智能是如何演变的？
Much of the attention back in 2022 was on AI systems that can generate text and images. Today there are standalone AI services that can clone a person’s voice, generate a song in seconds or create short videos. Increasingly, AI companies are focusing on “multimodal” systems that can process and respond to a range of inputs, including text, images and audio.
2022 年的大部分注意力都集中在可以生成文本和图像的人工智能系统上。如今，独立的人工智能服务可以克隆人的声音，在几秒钟内生成歌曲或创建短视频。人工智能公司越来越多地关注“多模式”系统，该系统可以处理和响应一系列输入，包括文本、图像和音频。
For example, an exasperated parent might be able to show an image of a math problem to a chatbot and ask for an explanation on how to solve it. Demis Hassabis, CEO of Google’s AI lab DeepMind, showed off a prototype AI assistant that can respond to video inputs in real time. Generative AI is also being heavily integrated into more popular products. Google has added it to its core search engine. Apple also brought a suite of AI tools to the iPhone in October.
例如，愤怒的父母可能可以向聊天机器人展示数学问题的图像，并要求解释如何解决它。谷歌人工智能实验室 DeepMind 首席执行官 Demis Hassabis展示了一款可以实时响应视频输入的人工智能助手原型。生成式人工智能也被大量集成到更流行的产品中。谷歌已将其添加到其核心搜索引擎中。苹果还在 10 月份为 iPhone 带来了一套人工智能工具。
At the same time, tech companies are investing in so-called agents that can book travel, write code and field other multi-step tasks for users with minimal human oversight. Microsoft launched a set of agent tools designed to send emails and manage records for workers. Anthropic unveiled an agent that can process what’s happening on the user’s computer in real time and take actions on their behalf. And OpenAI is planning to release similar software early next year.
与此同时，科技公司正在投资所谓的代理，这些代理可以在最少的人工监督下为用户预订旅行、编写代码和执行其他多步骤任务。微软推出了一套代理工具，旨在为员工发送电子邮件和管理记录。 Anthropic推出了一种代理，可以实时处理用户计算机上发生的情况并代表他们采取行动。 OpenAI计划明年初发布类似的软件。
“We will have better and better models,” OpenAI CEO Altman wrote in a post during an Ask Me Anything session on Reddit in October. “But I think the thing that will feel like the next giant breakthrough will be agents.”
“我们将拥有越来越好的模型，”OpenAI 首席执行官 Altman 在 10 月份 Reddit 上的 Ask Me Anything 会议上发表的一篇文章中写道。 “但我认为下一个巨大突破将是代理。”
What does it mean for an AI to reason?
对于人工智能来说，推理意味着什么？
For much of the past two years, ChatGPT users have grown accustomed to having the chatbot spit out near-instantaneous responses — some accurate, some not so much — to a wide range of prompts. But OpenAI’s new 01 model will behave somewhat differently. Before responding, the new software will pause for a matter of seconds while, behind the scenes and invisible to the user, it considers a number of related prompts and then summarizes what appears to be the best response. This technique is sometimes referred to as “chain of thought” prompting.
在过去两年的大部分时间里，ChatGPT 用户已经习惯了聊天机器人对各种提示做出近乎即时的响应——有些准确，有些不太准确。但 OpenAI 的新 01 模型的表现会有所不同。在做出响应之前，新软件会暂停几秒钟，同时在幕后（用户看不到）它会考虑一些相关提示，然后总结出最佳响应。这种技术有时被称为“思维链”提示。
In a blog post, OpenAI referred to it as “thinking.” In OpenAI’s tests, this approach made less of a difference for queries about writing and editing, but people found it resulted in better answers for complex questions related to computer programming, data analysis and math problems.
OpenAI 在一篇博文中将其称为“思考”。在 OpenAI 的测试中，这种方法对于有关写作和编辑的查询影响不大，但人们发现它可以为与计算机编程、数据分析和数学问题相关的复杂问题提供更好的答案。
“The next model update performs similarly to PhD students on challenging benchmark tasks in physics, chemistry, and biology. We also found that it excels in math and coding,” the company said. “For complex reasoning tasks this is a significant advancement and represents a new level of AI capability.”
“下一个模型更新在物理、化学和生物学方面具有挑战性的基准任务上的表现与博士生类似。我们还发现它在数学和编码方面表现出色，”该公司表示。 “对于复杂的推理任务来说，这是一个重大进步，代表了人工智能能力的新水平。”
Play 玩Play 玩
2:17
Google and Microsoft are losing the AI race to Mark Zuckerberg, says Bloomberg Opinion columnist Parmy Olson. Here’s why the AI giants could learn a thing or two from the Meta boss.
彭博社观点专栏作家帕米·奥尔森表示，谷歌和微软在人工智能竞赛中输给了马克·扎克伯格。这就是为什么人工智能巨头可以从元老板那里学到一两件事。
What’s the likely economic impact of generative AI?
生成式人工智能可能产生哪些经济影响？
That remains unclear. Federal Reserve Chair Jerome Powell said generative AI could boost productivity growth, but “probably not in the short run.” Some economists foresee a leap in productivity triggering wage gains that will help workers recoup some of the earning power they’ve lost since the 2008 financial crisis.
目前尚不清楚。美联储主席杰罗姆·鲍威尔表示，生成式人工智能可以促进生产率增长，但“短期内可能不会”。一些经济学家预计，生产率的飞跃会引发工资上涨，这将有助于工人恢复自 2008 年金融危机以来失去的部分赚钱能力。
A study by researchers at Stanford University and the Massachusetts Institute of Technology found that staff at one Fortune 500 software firm became 14% more productive on average when using generative AI tools. The lowest-skilled workers became 35% faster. What wasn’t clear was the impact of this on their higher-paid colleagues, whose costlier skills may no longer be required.
斯坦福大学和麻省理工学院的研究人员进行的一项研究发现，财富 500 强软件公司的员工在使用生成式 AI 工具时，工作效率平均提高了 14%。技能最低的工人的速度提高了 35%。目前还不清楚这对他们的高薪同事有何影响，他们可能不再需要昂贵的技能。
In an optimistic scenario, skilled employees whose jobs get automated will find new work in growth areas created by generative AI. However, early signs pointed to a painful bout of upheaval. International Business Machines Corp. plans to stop or slow hiring for thousands of roles it expects to be replaced by AI. Swedish buy now, pay later firm Klarna Inc. said it would freeze hiring as ChatGPT cut down the time required for certain tasks. And language-learning software company Duolingo Inc. chose not to renew about 10% of its contractors, partly due to AI.
在乐观的情况下，工作自动化的熟练员工将在生成式人工智能创造的增长领域找到新的工作。然而，早期迹象表明将出现一场痛苦的剧变。国际商业机器公司计划停止或放缓招聘数千个预计将被人工智能取代的职位。瑞典公司 Klarna Inc.表示，由于 ChatGPT 减少了某些任务所需的时间，因此将冻结招聘。语言学习软件公司Duolingo Inc.选择不再续约约10% 的承包商，部分原因是人工智能。
What are generative AI’s limitations?
生成式人工智能的局限性是什么？
The answers it pieces together from available published information can sound so authoritative that users may assume it has verified their accuracy. Stripped from useful context such as the source of the information, and with few of the obvious imperfections that often signal unreliable material, the content can be a minefield for someone who isn’t sufficiently well-versed in a subject to notice a response that is biased, flawed or simply wrong.
它从可用的已发布信息中拼凑出的答案听起来非常权威，以至于用户可能会认为它已经验证了其准确性。脱离了有用的上下文（例如信息来源），并且几乎没有通常表明材料不可靠的明显缺陷，对于那些不太熟悉某个主题的人来说，内容可能是一个雷区，无法注意到以下响应：有偏见、有缺陷或根本就是错误的。
AI chatbots have repeatedly made errors when asked for news about current events. Generative AI systems also amplify harmful biases and stereotypes found in the data on which they’re built – a significant problem as the technology increasingly gets deployed into sensitive areas like hiring and healthcare. AI companies have struggled to address the bias problem: Google, for example, attempted a technical fix to make its image generator show more diverse outputs, but it resulted in the tool refraining from producing pictures of White people.
当被问及有关时事的新闻时，人工智能聊天机器人屡次犯错误。生成式人工智能系统还会放大其构建数据中发现的有害偏见和刻板印象，随着该技术越来越多地部署到招聘和医疗保健等敏感领域，这是一个重大问题。人工智能公司一直在努力解决偏见问题：例如，谷歌尝试进行技术修复，使其图像生成器显示更多样化的输出，但结果导致该工具无法生成白人的照片。
What does generative AI mean for society?
生成式人工智能对社会意味着什么？
As these tools become more sophisticated, so does the potential for mischief-making.
随着这些工具变得越来越复杂，造成恶作剧的可能性也随之增加。
Some AI platforms have been taken down or tweaked to stop them from regurgitating hate speech, and ChatGPT allows users to flag distasteful responses to improve the system. But users have found workarounds. An AI might not pick up on gender or racial biases that a human would notice in books and other texts.
一些人工智能平台已被关闭或调整，以阻止它们重复仇恨言论，而 ChatGPT 允许用户标记令人反感的反应以改进系统。但用户已经找到了解决方法。人工智能可能不会注意到人类在书籍和其他文本中注意到的性别或种族偏见。
It’s also a potential weapon for deceit. For example, one AI audio deepfake of Joe Biden was used to urge people in New Hampshire not to vote in the US presidential primaries. Russian trolls used a deepfake video of Kamala Harris to try to influence public opinion ahead of the US election, according to researchers. Teachers worry about students getting chatbots to do their homework. Lobbying firms could use AI to fake thousands of different letters supposedly from citizens to influence the decisions of lawmakers.
它也是一种潜在的欺骗武器。例如，乔·拜登的一段人工智能音频深度伪造品被用来敦促新罕布什尔州的人们不要在美国总统初选中投票。研究人员称，俄罗斯网络喷子利用卡玛拉·哈里斯的深度伪造视频试图在美国大选前影响公众舆论。老师们担心学生会用聊天机器人来做作业。游说公司可以利用人工智能伪造数千封据称来自公民的不同信件，以影响立法者的决定。
In an open letter published in March 2023, 1,000 technologists and researchers called for a pause in AI development to allow more time to weigh the “profound risks to society and humanity.” Instead, the technology continues to advance.
在 2023 年 3 月发表的一封公开信中，1000 名技术专家和研究人员呼吁暂停人工智能开发，以便有更多时间权衡“对社会和人类的深远风险”。相反，技术在不断进步。
