作者：盖瑞·格罗斯曼，爱德曼

不要错过 OpenAI、Chevron、Nvidia、Kaiser Permanente 和 Capital One 领导人在 VentureBeat Transform 2024 上的精彩表现。在这场为期三天的独家活动中，获取有关 GenAI 的重要见解并扩展您的网络。了解更多

人类是否即将创造出智力超群的机器？有些人认为我们正处于这一发展的边缘。上周，Ilya Sutskever推出了他的新创业公司 Safe Superintelligence, Inc. (SSI)，该公司致力于构建先进的人工智能超级智能 (ASI) 模型——一种远远超出人类能力的假想人工智能。在关于推出 SSI 的声明中，他说“超级智能触手可及”，并补充道：“我们同时追求安全性和能力。”

Sutskever 有资格追求这种先进的模型。他是OpenAI的创始成员之一，曾担任该公司的首席科学家。在此之前，他与多伦多大学的 Geoffrey Hinton 和Alex Krizhevsky合作开发了“AlexNet”，这是一种图像分类模型，在 2012 年彻底改变了深度学习。这一发展比其他任何发展都更能在过去十年中掀起人工智能的热潮，部分原因是它展示了图形处理单元 (GPU) 并行指令处理对加快深度学习算法性能的价值。

苏茨克弗并不是唯一一个相信超级智能的人。软银首席执行官孙正义上周晚些时候表示，人工智能“将在 10 年内出现，比人类聪明 10,000 倍”。他补充说，实现 ASI 现在是他的人生使命。

超级智能远远超出了通用人工智能 (AGI)，后者也仍是一种假设的 AI 技术。AGI 将在大多数具有经济价值的任务中超越人类的能力。Hinton 认为，我们将在五年内看到 AGI。谷歌首席研究员兼 AI 远见者 Ray Kurzweil 将 AGI 定义为“可以执行受过教育的人类可以执行的任何认知任务的 AI”。他认为 这将在 2029 年实现。但事实上，AGI 并没有普遍接受的定义，因此无法准确预测它的到来。我们怎么知道呢？

超级智能可能也是如此。不过，至少有一位预言家表示，超级智能可能在 AGI 之后不久就到来，可能是在 2030 年。

尽管专家们对此持有不同看法，但 AGI 或超级智能是否会在五年内实现，或者永远实现，仍是一个悬而未决的问题。人工智能研究员 Gary Marcus 等一些人认为，目前对深度学习和语言模型的关注永远无法实现 AGI（更不用说超级智能），他们认为这些技术从根本上是有缺陷的，而且很脆弱，只能通过更多数据和计算能力的强力推动才能取得进展。

华盛顿大学计算机科学教授、《大师算法》一书的作者佩德罗·多明戈斯 (Pedro Domingos ) 认为超级智能只是一场白日梦。他在 X（原 Twitter）上发帖称：“Ilya Sutskever 的新公司一定会成功，因为永远不会实现的超级智能保证是安全的。”

这些观点之一可能被证明是正确的。没有人确切知道AGI 或超级智能是否会到来或何时到来。随着这场争论的继续，认识到这些概念与我们当前 AI 能力之间的差距至关重要。

与其仅仅猜测那些引发股市狂热梦想和公众焦虑的遥远未来的可能性，至少同样重要的是考虑未来几年可能塑造人工智能格局的更直接的进步。这些发展虽然没有最宏伟的人工智能梦想那么轰动，但将对现实世界产生重大影响，并为进一步发展铺平道路。

展望未来，未来几年，人工智能语言、音频、图像和视频模型（所有形式的深度学习）可能会继续发展和普及。虽然这些进步可能无法实现 AGI 或超级智能，但它们无疑将增强人工智能的能力、实用性、可靠性和应用。

尽管如此，这些模型仍然面临几个重大挑战。一个主要缺点是它们偶尔会出现幻觉或虚构，本质上是编造答案。这种不可靠性目前仍然是广泛采用的明显障碍。提高人工智能准确性的一种方法是检索增强生成 (RAG)，它整合来自外部来源的当前信息以提供更准确的响应。另一种可能是“语义熵”，它使用一个大型语言模型来检查另一个语言模型的工作。

随着机器人在未来一两年内变得更加可靠，它们将越来越多地被纳入商业应用程序和工作流程。迄今为止，许多此类努力都未能达到预期。这一结果并不令人意外，因为人工智能的纳入相当于一种范式转变。我的观点是，现在还为时过早，人们仍在收集信息并学习如何最好地部署人工智能。

沃顿商学院教授伊桑·莫里克 (Ethan Mollick) 在他的《一件有用的事》 时事通讯中表达了同样的观点：“目前，无论是咨询顾问还是典型的软件供应商，没有人能够给出关于如何使用人工智能在任何特定行业中开启新机遇的通用答案。”

莫里克认为，生成式人工智能的实施将在很大程度上得益于工人和管理人员，他们将尝试将这些工具应用于他们擅长的领域，以了解哪些方法有效并能增加价值。随着人工智能工具的功能越来越强大，越来越多的人将能够提高他们的工作产出，从而在企业内部形成一个由人工智能驱动的创新飞轮。

最近的进展证明了这种创新潜力。例如，Nvidia 的推理微服务可以加速 AI 应用程序的部署，而 Anthropic 的新Claude Sonnet 3.5聊天机器人据称表现优于所有竞争对手。从教室到汽车经销店，甚至在新材料的发现中，AI 技术在各个领域得到越来越广泛的应用。

苹果最近推出的Apple Intelligence就是这种加速的明显迹象。作为一家公司，苹果一直等到技术成熟度和需求足够高时才进入市场。这一消息表明人工智能已经到达了拐点。

Apple Intelligence超越了其他 AI 公告，承诺在应用程序之间实现深度集成，同时为用户保留上下文，从而创造高度个性化的体验。随着时间的推移，Apple 将允许用户将多个命令隐式地串在一起形成一个请求。这些命令可能跨多个应用程序执行，但会显示为单个结果。另一个词是“代理”。

在 Apple Intelligence 发布会上，软件工程高级副总裁 Craig Federighi 描述了一个场景来展示这些功能将如何工作。据《科技评论》报道，“他收到了一封电子邮件，推迟了一次工作会议，但他的女儿当晚要参加一场戏剧演出。他的手机现在可以找到包含演出信息的 PDF 文件，预测当地的交通情况，并让他知道他是否能准时到达。”

这种让人工智能代理执行复杂、多步骤任务的愿景并非苹果独有。事实上，它代表了人工智能行业向一些人所说的“代理时代”的广泛转变。

近几个月来，业界越来越多地讨论超越聊天机器人，进入“自主代理”领域，这些代理可以根据单个提示执行多个链接任务。这批新系统不仅仅是回答问题和共享信息，还使用 LLM 完成多步骤操作，从开发软件到预订航班。据报道，微软、OpenAI 和谷歌 DeepMind 都在准备 AI 代理，旨在自动执行更困难的多步骤任务。

OpenAI 首席执行官 Sam Altman 将代理愿景描述为“超级能干的同事，他完全了解我的一生、每封电子邮件、我曾经进行过的每一次对话，但又不会让人觉得他只是我的延伸。”换句话说，他是一个真正的私人助理。

代理还将为整个企业用途的应用程序提供服务。麦肯锡高级合伙人 Lari Hämäläinen将这一进步描述为“能够协调复杂工作流程、协调多个代理之间的活动、应用逻辑和评估答案的软件实体。这些代理可以帮助组织中的流程自动化，或在员工和客户执行流程时为他们提供支持。”

专注于企业代理的初创公司也纷纷涌现，比如刚刚走出隐身模式的 Emergence。据TechCrunch 报道，该公司声称正在构建一个基于代理的系统，可以执行许多通常由知识型员工处理的任务。

随着人工智能代理的到来，我们将更有效地加入永远在线的互联世界，无论是个人用途还是工作。这样，我们将越来越多地与无处不在的数字智能进行对话和互动。

通用人工智能和超级智能的发展之路仍充满不确定性，专家们对其可行性和时间表存在分歧。然而，人工智能技术的快速发展是不可否认的，有望带来变革性的进步。随着企业和个人驾驭这一快速变化的格局，人工智能驱动的创新和改进的潜力仍然巨大。未来的旅程既令人兴奋又难以预测，人类和人工智能之间的界限将继续模糊。

通过现在制定积极措施来投资和参与人工智能、提升劳动力技能并关注道德考虑，企业和个人就可以在人工智能驱动的未来中蓬勃发展。

加里·格罗斯曼 (Gary Grossman) 是爱德曼技术实践执行副总裁兼爱德曼人工智能卓越中心全球负责人。

Don’t miss OpenAI, Chevron, Nvidia, Kaiser Permanente, and Capital One leaders only at VentureBeat Transform 2024. Gain essential insights about GenAI and expand your network at this exclusive three day event. Learn More

Is humanity on the brink of creating its intellectual superior? Some think we are on the cusp of such a development. Last week, Ilya Sutskever unveiled his new startup, Safe Superintelligence, Inc. (SSI), which is dedicated to building advanced artificial superintelligence (ASI) models — a hypothetical AI far beyond human capability. In a statement about launching SSI, he said “superintelligence is within reach,” and added: “We approach safety and capabilities in tandem.”

Sutskever has the credentials to aspire to such an advanced model. He was a founding member of OpenAI and formerly served as the company’s chief scientist. Before that, he worked with Geoffrey Hinton and Alex Krizhevsky at the University of Toronto to develop “AlexNet,” an imageification model that transformed deep learning in 2012. More than any other, this development kicked-off the surge in AI over the last decade, in part by demonstrating the value of parallel instruction processing by graphics processing units (GPUs) to speed deep learning algorithm performance.

Handling Today’s Threatscape at Machine Scale

Sutskever is not alone in his belief about superintelligence. SoftBank CEO Masayoshi Son said late last week that AI “10,000 times smarter than humans will be here in 10 years.” He added that achieving ASI is now his life mission.

ADVERTISEMENT

Superintelligence goes way beyond artificial general intelligence (AGI), also still a hypothetical AI technology. AGI would surpass human capabilities in most economically valuable tasks. Hinton believes we could see AGI within five years. Ray Kurzweil, lead researcher and AI visionary at Google, defines AGI as “AI that can perform any cognitive task an educated human can.” He believes this will occur by 2029. Although in truth, there is no commonly accepted definition of AGI, which makes it impossible to accurately predict its arrival. How would we know?

Countdown to VB Transform 2024

Join enterprise leaders in San Francisco from July 9 to 11 for our flagship AI event. Connect with peers, explore the opportunities and challenges of Generative AI, and learn how to integrate AI applications into your industry. Register Now

The same could likely be said for superintelligence. However, at least one prognosticator is on record saying that superintelligence could arrive soon after AGI, possibly by 2030.

ADVERTISEMENT

Despite these expert opinions, it remains an open question whether AGI or superintelligence will be achieved in five years — or ever. Some, such as AI researcher Gary Marcus, believe the current focus on deep learning and language models will never achieve AGI (let alone superintelligence), seeing these as fundamentally flawed and weak technologies that can advance only through the brute force of more data and computing power.

Pedro Domingos, University of Washington computer science professor and author of The Master Algorithm, sees superintelligence as a pipe dream. “Ilya Sutskever’s new company is guaranteed to succeed, because superintelligence that is never achieved is guaranteed to be safe,” he posted to X (formerly Twitter).

ADVERTISEMENT

One of these viewpoints might prove to be correct. No one knows for certain if AGI or superintelligence is coming or when. As this debate continues, it’s crucial to recognize the chasm between these concepts and our current AI capabilities.

Rather than speculating solely on far-future possibilities that are fueling exuberant stock market dreams and public anxiety, it’s at least equally important to consider the more immediate advancements that are likely to shape the AI landscape in the coming years. These developments, while less sensational than the grandest AI dreams, will have significant real-world impacts and pave the way for further progress.

ADVERTISEMENT

As we look ahead, the next several years will likely see AI language, audio, image and video models — all forms of deep learning — continue to evolve and proliferate. While these advancements may not achieve AGI or superintelligence, they will undoubtedly enhance AI’s capabilities, utility, reliability and application.

That said, these models still face several significant challenges. One major shortcoming is their tendency to occasionally hallucinate or confabulate, essentially making up answers. This unreliability remains a clear barrier to widespread adoption at present. One approach to improve AI accuracy is retrieval augmented generation (RAG), which integrates current information from external sources to provide more accurate responses. Another could be “semantic entropy,” which uses one large language model to check the work of another.

As bots become more reliable over the next year or two, they will be increasingly incorporated into business applications and workflows. To date, many of these efforts have fallen short of expectations. This outcome is not surprising, as the incorporation of AI amounts to a paradigm shift. My view is that it is still early, and that people are still gathering information and learning about how best to deploy AI.

ADVERTISEMENT

Wharton professor Ethan Mollick echoes this view in his One Useful Thing newsletter: “Right now, nobody — from consultants to typical software vendors — has universal answers about how to use AI to unlock new opportunities in any particular industry.”

Mollick argues that a lot of the progress in implementing generative AI will come from workers and managers who experiment with applying the tools to their areas of domain expertise to learn what works and adds value. As AI tools become more capable, more people will be able to advance their work output, creating a flywheel of AI-powered innovation within businesses.

Recent advancements demonstrate this innovation potential. For instance, Nvidia’s Inference Microservices can accelerate AI application deployments, and Anthropic’s new Claude Sonnet 3.5 chatbot reportedly outperforms all competitors. AI technologies are finding increased application across various fields, fromrooms to auto dealerships and even in the discovery of new materials.

ADVERTISEMENT

A clear sign of this acceleration came from Apple with their recent launch of Apple Intelligence. As a company, Apple has a history of waiting to enter a market until there is sufficient technology maturity and demand. This news suggests that AI has reached that inflection point.

Apple Intelligence goes beyond other AI announcements by promising deep integration across apps while maintaining context for the user, creating a deeply personalized experience. Over time, Apple will enable users to implicitly string multiple commands together into a single request. These may execute across multiple apps but will appear as a single result. Another word for this is “agents.”

ADVERTISEMENT

During the Apple Intelligence launch event, SVP of software engineering Craig Federighi described a scenario to showcase how these will work. As reported by Technology Review, “an email comes in pushing back a work meeting, but his daughter is appearing in a play that night. His phone can now find the PDF with information about the performance, predict the local traffic, and let him know if he’ll make it on time.”

This vision of AI agents performing complex, multi-step tasks is not unique to Apple. In fact, it represents a broader shift in the AI industry towards what some are calling the “Agentic era.”

ADVERTISEMENT

In recent months there has been increasing industry discussion about moving beyond chatbots and into the realm of “autonomous agents” that can perform multiple linked tasks based on a single prompt. More than just answering questions and sharing information, this new crop of systems use LLMs to complete multi-step actions, from developing software to booking flights. According to reports, Microsoft, OpenAI and Google DeepMind are all readying AI agents designed to automate more difficult multi-step tasks.

OpenAI CEO Sam Altman described the agent vision as a “super-competent colleague that knows absolutely everything about my whole life, every email, every conversation I’ve ever had, but doesn’t feel like an extension.” In other words, a true personal assistant.

Agents will serve applications across enterprise uses as well. McKinsey senior partner Lari Hämäläinen describes this advancement as “software entities that can orchestrate complex workflows, coordinate activities among multiple agents, apply logic and evaluate answers. These agents can help automate processes in organizations or augment workers and customers as they perform processes.”

Start-ups focused on enterprise agents are also appearing — such as Emergence, which fittingly just came out of stealth mode. According to TechCrunch, the company claims to be building an agent-based system that can perform many of the tasks typically handled by knowledge workers.

ADVERTISEMENT

With the pending arrival of AI agents, we will even more effectively join the always-on interconnected world, both for personal use and for work. In this way, we will increasingly dialog and interact with digital intelligence everywhere.

The path to AGI and superintelligence remains shrouded in uncertainty, with experts divided on its feasibility and timeline. However, the rapid evolution of AI technologies is undeniable, promising transformative advancements. As businesses and individuals navigate this rapidly changing landscape, the potential for AI-driven innovation and improvement remains vast. The journey ahead is as exciting as it is unpredictable, with the boundaries between human and artificial intelligence continuing to blur.

By mapping out proactive steps now to invest and engage in AI, upskill our workforce and attend to ethical considerations, businesses and individuals can position themselves to thrive in the AI-driven future.

Gary Grossman is EVP of technology practice at Edelman and global lead of the Edelman AI Center of Excellence.

声明

本文素材来源于网络，人工智能产业链联盟（union）整理编辑，不代表人工智能产业链联盟（union）立场，转载请注明，如涉及作品版权问题，请联系我们删除或做相关处理！

编辑：Zero