AI 本质上就是能执行复杂任务。在这些复杂任务的表现上能达到或超越人类的水平。

AI 的能力边界

OpenAI 泄露的,他们认为的 AI 的 5 个 level ,用于理解AI功能的演变的路线图。

On July 9th, OpenAI internally presented a five-step roadmap to track its progress towards achieving Artificial General Intelligence (AGI). It’s a structured framework for tracking progress in AI development - and it leaked via Bloomberg.

Level 1: Conversational AI

第一级:对话式人工智能

Level 1 represents the current state of AI technology, focusing on systems that can engage in human-like conversations. These AI models can understand and generate natural language, respond to queries, and perform basic tasks based on textual input. 第一级代表了人工智能技术的现状,重点是可以进行类似人类对话的系统。这些AI模型可以理解和生成自然语言,响应查询,并基于文本输入执行基本任务。

Examples: 示例如下:

ChatGPT (OpenAI): A language model capable of engaging in human-like conversations, answering questions, and assisting with various text-based tasks. ChatGPT(OpenAI):一种语言模型,能够进行类似人类的对话,回答问题,并协助各种基于文本的任务。 Gemini (Google): A conversational AI system designed for open-ended dialogues and complex problem-solving. Gemini(Google):一个对话式人工智能系统,专为开放式对话和复杂问题解决而设计。 Claude (Anthropic): An AI assistant proficient in natural language conversations and various text-based tasks. Claude(Anthropic):精通自然语言对话和各种基于文本的任务的人工智能助理。 These conversational AI systems have found applications in customer service, virtual assistants, and content generation. While they demonstrate impressive language understanding and generation capabilities, they are limited to text-based interactions and lack true reasoning or problem-solving abilities beyond their training data. As we move forward, we can expect to see these systems become more refined and integrated into various business processes, enhancing efficiency and user experiences across industries. 这些对话式人工智能系统已经在客户服务、虚拟助理和内容生成中找到了应用。虽然它们表现出令人印象深刻的语言理解和生成能力,但它们仅限于基于文本的交互,并且缺乏真正的推理或解决问题的能力。随着我们的发展,我们可以期待看到这些系统变得更加完善,并集成到各种业务流程中,从而提高各个行业的效率和用户体验。

Level 2: Reasoners 第二层:推理者

Level 2 AI systems, termed “Reasoners,” represent a significant leap forward in AI capabilities. These systems are expected to solve complex problems at a level comparable to humans with doctorate-level education, without relying on external resources. 2级人工智能系统,称为“推理者”,代表了人工智能能力的重大飞跃。这些系统有望在不依赖外部资源的情况下,以与具有博士水平教育的人类相当的水平解决复杂问题。

Examples approaching Level 2: 接近2级的示例:

AlphaFold (DeepMind): An AI system that can predict protein structures with high accuracy, demonstrating problem-solving skills in complex scientific domains. AlphaFold(DeepMind):一个可以高精度预测蛋白质结构的人工智能系统,展示了在复杂科学领域解决问题的能力。 GPT-4 (OpenAI): While primarily a language model, GPT-4 has shown improved reasoning capabilities in tasks such as analyzing complex scenarios and providing logical explanations. GPT-4(OpenAI):虽然主要是一种语言模型,但GPT-4在分析复杂场景和提供逻辑解释等任务中表现出了更好的推理能力。 These examples illustrate progress towards Level 2, but true “Reasoners” that can consistently solve complex problems across various domains without external resources are still in development. As we approach this level, we can anticipate AI systems playing increasingly significant roles in research, data analysis, and decision-making processes across industries. 这些例子说明了向第2级的进展,但真正的“推理者”,可以在没有外部资源的情况下,在各个领域持续解决复杂问题,仍在开发中。随着我们接近这一水平,我们可以预见人工智能系统在各行各业的研究、数据分析和决策过程中发挥着越来越重要的作用。

Level 3: Agents 第3级:代理

Level 3 introduces “Agents,” AI systems capable of acting autonomously on behalf of users for extended periods. These agents would be able to perform tasks, make decisions, and adapt to changing circumstances over several days without constant human oversight. 第3级引入了“代理”,即能够代表用户长时间自主行动的人工智能系统。这些智能体将能够执行任务,做出决策,并在几天内适应不断变化的环境,而无需持续的人类监督。

Examples showcasing aspects of Level 3: 展示第3级各方面的示例:

CrewAI: A framework for orchestrating role-playing AI agents, allowing multiple AI entities to work together on complex tasks, each with a specific role and expertise. CrewAI:一个用于编排角色扮演AI代理的框架,允许多个AI实体在复杂的任务中协同工作,每个实体都有特定的角色和专业知识。 AutoGen (Microsoft): A framework enabling the development of LLM-based conversable AI agents that can collaborate to solve tasks. AutoGen(微软):一个框架,可以开发基于LLM的可转换AI代理,可以协作解决任务。 Constitutional AI (Anthropic): AI systems with enhanced safety features and the ability to follow complex instructions over extended interactions. Constitutional AI(Anthropic):AI系统具有增强的安全功能,并能够在扩展的交互中遵循复杂的指令。 OpenAI’s hide-and-seek agents: AI agents that developed complex strategies through reinforcement learning, showcasing emergent behaviours over extended periods. OpenAI的hide-and-seek代理:通过强化学习开发复杂策略的AI代理,在较长时间内展示紧急行为。 These examples show progress towards autonomous agency, but they are still limited to specific domains and lack the general-purpose capabilities envisioned for Level 3 agents. As we move closer to this level, we can expect to see AI systems taking on more complex, multi-step tasks with minimal human intervention, potentially revolutionizing project management, customer service, and operational workflows. 这些例子显示了自治代理的进展,但它们仍然局限于特定的领域,缺乏三级代理所设想的通用功能。随着我们越来越接近这一水平,我们可以期待看到人工智能系统在最少的人为干预下承担更复杂的多步骤任务,这可能会彻底改变项目管理、客户服务和运营工作流程。

Level 4: Innovators 第4级:创新者

Level 4 systems, described as “Innovators,” would be capable of developing original ideas and solutions, potentially driving breakthroughs in various fields. These systems would not only solve existing problems but also identify new challenges and create innovative approaches to address them. 4级系统,被称为“创新者”,将能够开发原创性的想法和解决方案,有可能推动各个领域的突破。这些系统不仅将解决现有问题,而且还将确定新的挑战,并创造解决这些挑战的创新办法。

Examples showcasing aspects of innovation: 展示创新方面的例子:

DALL-E 2 (OpenAI): An image generation model that can create unique and creative images based on text descriptions, showcasing a form of visual innovation. DALL-E 2(OpenAI):一种图像生成模型,可以基于文本描述创建独特而富有创意的图像,展示了一种视觉创新形式。 AlphaGo Zero (DeepMind): An AI that developed novel strategies in the game of Go, surpassing human knowledge and demonstrating innovative gameplay. AlphaGo Zero(DeepMind):一个在围棋游戏中开发新颖策略的人工智能,超越人类知识并展示创新的游戏玩法。 AI-Powered Drug Discovery: Systems like Atomwise use AI to innovate in pharmaceutical research, predicting new drug candidates and accelerating the discovery process. 人工智能驱动的药物发现:Atomwise等系统使用人工智能在药物研究中进行创新,预测新的候选药物并加速发现过程。 These examples demonstrate AI’s potential for innovation in specific domains, but they fall short of the general-purpose innovative capabilities envisioned for Level 4 systems. As we progress towards this level, we can anticipate AI playing a more significant role in research and development across industries, potentially leading to breakthroughs in science, technology, and business strategy. 这些例子证明了人工智能在特定领域的创新潜力,但它们没有达到4级系统所设想的通用创新能力。随着我们向这一水平迈进,我们可以预期人工智能在各行业的研发中发挥更重要的作用,可能导致科学、技术和商业战略的突破。

Level 5: Organizations 第5级:组织

The pinnacle of OpenAI’s classification system, Level 5 “Organizations,” represents AI systems capable of performing the work of entire organizations. These systems would manage complex workflows, make strategic decisions, and optimize operations across various departments and functions. OpenAI分类系统的顶峰,第5级“组织”,代表能够执行整个组织工作的人工智能系统。这些系统将管理复杂的工作流程,做出战略决策,并优化各个部门和职能的运营。

While I think no existing AI systems come close to Level 5 capabilities, some technologies demonstrate early steps towards organizational-level AI: 虽然我认为现有的人工智能系统没有接近第五级的能力,但一些技术展示了迈向组织级人工智能的早期步骤:

Hypothetical AI Corporate Management: In the future, we might see AI systems that can analyze market trends and competitor actions to formulate business strategies, manage human and AI workforce allocation and performance, and make high-level decisions on mergers, acquisitions, and corporate restructuring. 假设的人工智能企业管理:未来,我们可能会看到人工智能系统可以分析市场趋势和竞争对手的行动,以制定业务战略,管理人力和人工智能劳动力的分配和绩效,并在合并,收购和企业重组方面做出高层决策。 AI-Driven Healthcare Systems: While ethical considerations are paramount, future AI could potentially manage hospital operations, from patient scheduling to resource allocation, coordinate research efforts across multiple institutions, and analyze global health data to predict and respond to pandemics. AI驱动的医疗保健系统:虽然伦理考虑至关重要,但未来的人工智能可能会管理医院运营,从患者调度到资源分配,协调多个机构的研究工作,并分析全球健康数据以预测和应对流行病。 Autonomous Trading Systems: While limited to financial markets, some hedge funds and trading firms use AI systems that operate with significant autonomy, making real-time trading decisions based on market data and complex algorithms, managing risk and portfolio allocation autonomously, and adapting strategies based on changing market conditions. 自主交易系统:虽然仅限于金融市场,但一些对冲基金和交易公司使用的人工智能系统具有很大的自主性,可以根据市场数据和复杂算法做出实时交易决策,自主管理风险和投资组合分配,并根据不断变化的市场条件调整策略。 These examples illustrate the potential for AI to impact organizational processes, but they are a far cry from the fully autonomous, AGI-level systems envisioned for Level 5. The realization of this level would represent a paradigm shift in how we think about organizational structure and management. 这些例子说明了人工智能影响组织流程的潜力,但它们与第5级所设想的完全自主的AGI级系统相去甚远。这一层次的实现将代表我们如何看待组织结构和管理的范式转变。