人工知能

Artificial intelligence — Grokipedia Fact-checked by Grok 16 days ago Artificial intelligence Ara Eve Leo Sal 1x Acronym AI Parent Discipline computer science Coined By John McCarthy Coined Year 1955 Founding Event Dartmouth Conference Founding Year 1956 Key Pioneers Alan Turing John McCarthy Marvin Minsky Allen Newell Herbert Simon Subfields machine learning natural language processing computer vision robotics expert systems deep neural networks Major Approaches rule-based systems expert systems machine learning deep neural networks Current Paradigm machine learning, particularly deep neural networks First Program Logic Theorist First Program Year 1956 Turing Test Year 1950 Notable Milestones Dartmouth Conference (1956) Deep Blue defeating Garry Kasparov (1997) AlphaGo mastering Go (2016) large-scale generative models producing coherent text, images, and code Ai Winters cycles of optimism followed by setbacks—periods known as AI winters Deep Learning Breakthrough 2012, AlexNet winning the ImageNet competition Transformer Year 2017 Related Disciplines computer science mathematics philosophy psychology linguistics neuroscience Applications medical diagnosis from imaging autonomous vehicle navigation generating coherent text, images, and code Status Predominantly narrow AI excelling at specialized tasks, with ongoing efforts toward artificial general intelligence (AGI) Artificial intelligence (AI) is a subfield of computer science focused on developing systems that perform tasks requiring human intelligence, such as perception, reasoning, learning, and decision-making. The term was coined by John McCarthy in a 1955 proposal for the Dartmouth Conference held in 1956, which convened researchers to explore machine simulation of intelligence for solving human problems. AI research has experienced cycles of optimism and setbacks. Recent breakthroughs in machine learning —particularly deep neural networks —have produced milestones like IBM 's Deep Blue defeating chess champion Garry Kasparov in 1997, DeepMind 's AlphaGo mastering Go in 2016, and generative models creating coherent text, images, and code. Today's predominantly narrow AI excels at specialized tasks, such as medical diagnosis from imaging and autonomous vehicle navigation, while efforts toward artificial general intelligence (AGI) proceed amid debates over feasibility, timelines, and societal implications. Fundamentals Defining Artificial Intelligence Artificial Intelligence (AI) is a field of computer science focused on creating machines capable of tasks that typically require human intelligence, such as reasoning, learning from experience, pattern recognition, and decision-making under uncertainty. It encompasses approaches including rule-based systems , expert systems , robotics , natural language processing (NLP), and computer vision . [1] The term artificial intelligence was introduced by John McCarthy in the 1955 proposal for the 1956 Dartmouth Summer Research Project, and was subsequently popularized through the 1956 meeting at the Dartmouth Conference . In the proposal, it was defined as "the science and engineering of making intelligent machines," specifically aiming to explore whether "every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it." This foundational definition emphasized simulation of human-like cognitive processes through computational means. [2] However, the concept remains contested due to the lack of a universally agreed-upon measure of intelligence , with philosophical debates centering on whether intelligence entails understanding, intentionality, or merely behavioral mimicry. [3] For instance, Alan Turing's 1950 imitation game, now known as the Turing Test , operationalizes intelligence as the ability of a machine to exhibit behavior indistinguishable from a human in conversation, though critics argue it assesses deception rather than genuine cognition. [4] [5] [6] Distinctions within AI definitions categorize systems by scope: narrow artificial intelligence (ANI), or weak AI, comprises task-specific systems like those for image classification or language translation, lacking broader adaptability. [7] In contrast, artificial general intelligence (AGI), or strong AI, refers to an AI system that can understand, learn, and perform any intellectual task a human can do, across virtually all domains, with general adaptability, reasoning, and problem-solving at or beyond human level—without needing task-specific training or massive domain-specific data. [8] As of 2025, deployed AI remains narrow, excelling via statistical pattern recognition in delimited applications rather than comprehensive intelligence. [7] This divide underscores empirical hurdles in scaling task-specific prowess to general versatility, driven by data optimization over innate comprehension. [9] Classifications of artificial intelligence AI systems are commonly classified according to their capabilities (scope of intelligence) and functionality (how they process information and learn). By capabilities (scope of intelligence) This classification describes how broadly an AI can apply its intelligence. Artificial Narrow Intelligence (ANI) , also known as narrow AI or weak AI, is designed for specific tasks and lacks generalization beyond its training. It represents all current real-world AI systems. Examples include voice assistants (e.g., Siri), recommendation algorithms, image recognition, and large language models like the GPT series for text generation. See main article: Weak artificial intelligence . Artificial General Intelligence (AGI) , or strong AI, refers to hypothetical systems capable of understanding, learning, and performing any intellectual task a human can across diverse domains, with adaptability and reasoning comparable to or exceeding humans. AGI remains unrealized as of 2026. See main article: Artificial general intelligence . Artificial Superintelligence (ASI) , or super AI, is a theoretical future stage where AI surpasses human intelligence in all fields, including creativity, scientific discovery, and strategic planning, potentially leading to rapid self-improvement. ASI is purely speculative and has no current implementations. By functionality (how AI processes and learns) This framework outlines progressive levels of complexity in AI systems' interaction with the world. Reactive machines : The most basic type; no memory or learning from past experiences. They react solely to current inputs based on predefined rules. Examples: IBM Deep Blue (chess-playing program), basic spam filters. Limited memory AI : Can learn from historical data, store experiences, and improve over time. This includes most modern AI, such as self-driving cars (using past sensor data), large language models trained on vast datasets, and fraud detection systems. Theory of mind AI : Would understand human emotions, beliefs, intentions, and social dynamics to enable natural social interactions. This level remains emerging in research but not fully achieved. Self-aware AI : Hypothetical AI with consciousness, self-awareness, and subjective experiences. This is purely theoretical and far from realization. These classifications are not mutually exclusive; current AI is predominantly narrow with limited memory functionality. Emerging trends include multimodal and agentic systems, but they fall under narrow AI. Intelligence Metrics and Benchmarks AI systems are evaluated using metrics for pattern recognition, reasoning, language understanding, and problem-solving, benchmarked against human performance or specific tasks. Early methods emphasized behavioral imitation; modern ones prioritize scalable, task-specific assessments amid rapid capability advances. These metrics track progress toward general intelligence but struggle to measure causal reasoning and robustness outside training data. [7] The Turing Test , proposed by Alan Turing in 1950, checks if machines can hold indistinguishable text conversations with humans, as a behavioral benchmark. Though influential philosophically, it overlooks non-linguistic abilities like manipulation or ethics and can be gamed via mimicry without understanding. Modern large language models pass variants in controlled settings but falter on targeted weakness probes, highlighting limits in assessing cognitive depth. [8] [9] [10] Task-specific benchmarks have advanced narrow domains. In vision, the ImageNet dataset (2009) tests classification; convolutional neural networks exceeded human ~5% error by 2015, reached 88.4% top-1 accuracy by 2020 via EfficientNet and Noisy Student , and surpassed 90% later. Games demonstrate strategy: IBM's Deep Blue beat Garry Kasparov in chess (1997) using search and evaluation; DeepMind's AlphaGo defeated Lee Sedol in Go (2016) with Monte Carlo tree search and deep networks, achieving superhuman play in higher complexity. AlphaZero refined this via self-play, exceeding 3400 Elo in chess by 2017. [11] [12] Language benchmarks include GLUE (2018) and SuperGLUE (2019) for understanding tasks like sentiment and entailment; saturation above 90% led to BIG-bench (2022) with over 200 tasks probing scaling. MMLU , spanning 57 subjects with multiple-choice questions, shows 2025 leading models at 90–95%, nearing or exceeding human expert ~89–90%, though novel reasoning gaps remain. Reasoning tests like ARC yield AI ~40–50% vs. human 85%, revealing generalization shortfalls; GPQA on graduate questions hits over 90% (e.g., GPT-5.2 at 93.2%). Coding via SWE-bench sees GPT-5 at 74.9% on Verified tasks, with frontier models resolving over 50% of real GitHub issues autonomously. [13] [14] [15] [16] [17] Benchmark Focus Area Top AI Performance (circa 2025) Human Baseline Key Limitation MMLU Multitask knowledge 92–95% accuracy ~89% Saturation and contamination ARC Abstract reasoning ~50% 85% Poor generalization to novel patterns GPQA Expert Q