=09

Artificial intelligence

=09

From Wikipedia, the free encycl= opedia

=09 =09

Intelligence of machines

"AI" redirec= ts here. For other uses, see AI (disambiguation) and Artificial intelligence (disambiguat= ion).

<= /tbody>

Artific= ial intelligence (AI)
Part of a series on

Major goals
Approaches
Applications
Philosophy
History
Controversies
Glossary =
v t <= a href=3D"https://en.wikipedia.org/wiki/Special:EditPage/Template:Artificia= l_intelligence" title=3D"Special:EditPage/Template:Artificial intelligence"= >e

Artificial intelligence (AI) is the capability of computationa= l systems to perform tasks typically associated with hum= an intelligence, such as learning, reasoning, = problem-solving, perception, and decision-making. It is= a field of research in c= omputer science that develops and studies methods and software that enab= le machines to perceive their environment and use learning a= nd intelligence to take actions that maximize their chances of achie= ving defined goals.^[1]

High-profile applications of AI= include advanced web search engines= (e.g., Google Search); recommendation systems (used by YouTube, Amazon, a= nd Netf= lix); virtual assistants (e.g., Google Assi= stant, Si= ri, and Alexa); autonomou= s vehicles (e.g., Waymo); gener= ative and creative tools (e.g., = language models and AI art); and superhum= an play and analysis in strategy games (e.g., chess and Go). However= , many AI applications are not perceived as AI: "A lot of cutting edge AI h= as filtered into general applications, often without being called AI becaus= e once something becomes useful enough and common enough it's not labeled AI a= nymore."^[2]^[3]

Various subfields of AI research are centered around particular goal= s and the use of particular tools. The traditional goals of AI research inc= lude learning, reasoning, knowledge representation, planning, natural= language processing, perception, and support for robotic= s.^[a] To reach these goals, AI researchers have adapted and= integrated a wide range of techniques, including search and mathematical optimi= zation, formal logic, artificial neural networks, and methods = based on statistics, operations research, and economic= s.^[b] AI also draws upon psychology, linguistics, philosophy, neuroscien= ce, and other fields.^[4]= Some companies, such as OpenAI, Google DeepMind and = Meta,^[5] aim to cr= eate artificial general intelligen= ce (AGI) =E2=80=93 AI that can complete virtually any cognitive task at= least as well as a human.

Artificial intelligence was founded as an academic discipline in 195= 6,^[6] and the field went through multiple cycles of opt= imism throughout its history= ,^[7]= ^[8] = followed by periods of disappointment and loss of funding, known as AI winters= .^[9]^[10] Funding and interest vastly inc= reased after 2012 when graphics processing units= started being used to accelerate neural networks, and deep learni= ng outperformed previous AI techniques.^{[11] This growth accelerated further after 2017 with the transformer architecture.^[12] In the 2020s, an ongoing period of rapid progress in advanced generative AI became known =
as the =
AI boom. Generative AI's ability to create and modify content has led t=
o several unintende=
d consequences and harms. =
Ethical concerns have been raised about AI's lon=
g-term effects and potential existential risks, prompting discussions abo=
ut regulatory policies=
to ensure the safety and benefits of the technology.}

Goals

The general problem of simulating (or creating) intelligence has been br= oken into subproblems. These consist of particular traits or capabilities t= hat researchers expect an intelligent system to display. The traits describ= ed below have received the most attention and cover the scope of AI researc= h.^[a]

Reasoning and problem-solving

Early researchers developed algorithms that imitated step-by-step reason= ing that humans use when they solve puzzles or make logical [13] By the late 1980s and 1990s, methods were developed for dealing with unc= ertain or incomplete information, employing concepts from probability<= /a> and economics.^[14]

Many of these algorithms are insufficient for solving large reasoni= ng problems because they experience a "combinatorial explosion": They becom= e exponentially slower as the problems grow.^{[15] Even humans rarely use the step-by-step deduction that early AI resear=
ch could model. They solve most of their problems using fast, intuitive jud=
gments.^[16] Accurate and efficient reasoning is an unsolved problem.}

K= nowledge representation

An ontolog= y represents knowledge as a set of concepts within a domain and the relatio= nships between those concepts.

Knowledge representatio= n and knowledge engineering^[17] allow AI programs to answer questi= ons intelligently and make deductions about real-world facts. Formal knowle= dge representations are used in content-based indexing and retrieval,^[18] scene interpretation,^[19] clinical decision support,^[20] knowledge discovery (mining "interesting"= and actionable inferences from large databases),^[21]= and other areas.^[22]

A knowledge base is a body of knowledge represented in a = form that can be used by a program. An ontology is the set of objects, relations, concepts, and properties = used by a particular domain of knowledge.^{<=
span class=3D"cite-bracket">[23]=} Knowledge bases need to represent things such as objects, proper= ties, categories, and relations between objects;^[24] situations, events, states, and time;^[25] causes and effects;<= sup id=3D"cite_ref-28" class=3D"reference">[26] knowledge about = knowledge (what we know about what other people know);^[27] d= efault reasoning (things that humans assume are true until they are tol= d differently and will remain true even when other facts are changing);[28= ] and many other aspects and domains of knowledge.

Among the most difficult problems in knowledge representation are th= e breadth of commonsense knowled= ge (the set of atomic facts that the average person knows is enormous);= ^[29] and the sub-symbolic f= orm of most commonsense knowledge (much of what people know is not represen= ted as "facts" or "statements" that they could express verbally).^[16] Th= ere is also the difficulty of knowledge acquisition, the problem of obtaining knowledge for AI applications.[c]

Planning and decision-making

An "agent" is anything that perceives and takes actions in the world. A = rational agent has goals or preferences and takes actions to mak= e them happen.^[d]<= sup id=3D"cite_ref-FOOTNOTERussellNorvig2021528_36-0" class=3D"reference"><= a href=3D"https://en.wikipedia.org/wiki/Artificial_intelligence#cite_note-F= OOTNOTERussellNorvig2021528-36">[32] In automated planning, the agent has a specific goal.^[33=
] In automated decision-making, the agent has preferences=E2=80=94there= are some situations it would prefer to be in, and some situations it is tr= ying to avoid. The decision-making agent assigns a number to each situation= (called the "utility") that measures how much the agent prefers it. For each = possible action, it can calculate the "exp= ected utility": the utility of all possible outcomes of the action, wei= ghted by the probability that the outcome will occur. It can then choose th= e action with the maximum expected utility.^[34]

In c= lassical planning, the agent knows exactly what the effect of any actio= n will be.^[35] In mo= st real-world problems, however, the agent may not be certain about the sit= uation they are in (it is "unknown" or "unobservable") and it may not know = for certain what will happen after each possible action (it is not "determi= nistic"). It must choose an action by making a probabilistic guess and then= reassess the situation to see if the action worked.^[36]

In some problems, the agent's preferences may be uncertain, especial= ly if there are other agents or humans involved. These can be learned (e.g.= , with inverse= reinforcement learning), or the agent can seek information to improve = its preferences.^[37]= Information value theor= y can be used to weigh the value of exploratory or experimental actions= .^[38] The space of p= ossible future actions and situations is typically i= ntractably large, so the agents must take actions and evaluate situatio= ns while being uncertain of what the outcome will be.

A Markov decision process has a transition model that describes the probability that a part= icular action will change the state in a particular way and a reward functi= on that supplies the utility of each state and the cost of each action. A <= a href=3D"https://en.wikipedia.org/wiki/Reinforcement_learning#Policy" titl= e=3D"Reinforcement learning">policy associates a decision with each pos= sible state. The policy could be calculated (e.g., by iteration), be heuristic, or it can be learned.^[3=
9]

Game theory describes the rational behavior of multiple interac= ting agents and is used in AI programs that make decisions that involve oth= er agents.^[40]

Learning

Machine learning is the study of programs that can improv= e their performance on a given task automatically.^[41] It has been a part of AI from the begi= nning.^[e]

There are several kinds of machine learning. Unsupe= rvised learning analyzes a stream of data and finds patterns and makes = predictions without any other guidance.^[44] Supervised learning requires l= abeling the training data with the expected answers, and comes in two main = varieties: classification (where the pro= gram must learn to predict what category the input belongs in) and regression (where the program must deduce a numeric function = based on numeric input).^[45]

In reinforcement learning, the agent is r= ewarded for good responses and punished for bad ones. The agent learns to c= hoose responses that are classified as "good".^[46] Transfer learning is when t= he knowledge gained from one problem is applied to a new problem.^[=
47] Deep learning is= a type of machine learning that runs inputs through biologically inspired = artificial neural net= works for all of these types of learning.^[48]

Computational learning theory can assess learners by computational complexi= ty, by sample complexity (how much data is required), = or by other notions of optimization.^[=
49]

Natural language processing

Natural language processing (NLP) a= llows programs to read, write and communicate in human languages.^[=
50] Specific problems include= speech recognition, speech synthesis, machine translation, information = extraction, information retrieval and question answering.^=
[51]

Early work, based on Noam Chomsky's generative= grammar and semantic networks, had difficulty with word-sense disambiguation^[f] unless restricted to small domains ca= lled "micro-worlds" (due to the common sense knowledge problem= [2= 9]). Margaret = Masterman believed that it was meaning and not grammar that was the key= to understanding languages, and that thesauri and not= dictionaries should be the basis of computational language structure.

Modern deep learning techniques for NLP include word embeddin= g (representing words, typically as vectors encoding their meani= ng),^[52] tran= sformers (a deep learning architecture using an attention mechanism),^[53] and others.= ^[54] In 2019, generative pre-trained transformer= (or "GPT") language models began to generate coherent text,^[55]^[56] and by 2023, these models were able to get human-level scores on t= he bar exam, SAT test, GRE test, and many other real-w= orld applications.^{[57<=
span class=3D"cite-bracket">]}

Perception

Machine perception is the ability to use input from s= ensors (such as cameras, microphones, wireless signals, active lidar, sonar, radar= , and tactile sensors) to deduce aspects of the world. [58]

The field includes speech recognition,^=
[59] image classification,^[60] f= acial recognition, object recognit= ion,^{[61] object tracking,^[62] and rob=
otic perception.^[63]}

Social= intelligence

= Kismet, a robot head which was made in the 1990s; it is a machine that = can recognize and simulate emotions.^[64]

Affective computing is a field that comprises syste= ms that recognize, interpret, process, or simulate human f= eeling, emotion, and mood.^{<=
a href=3D"https://en.wikipedia.org/wiki/Artificial_intelligence#cite_note-7=
1">[65]} For example, some virtual assistants are pr= ogrammed to speak conversationally or even to banter humorously; it makes t= hem appear more sensitive to the emotional dynamics of human interaction, o= r to otherwise facilitate hu= man=E2=80=93computer interaction.

However, this tends to give na=C3=AFve users an unrealistic concepti= on of the intelligence of existing computer agents.^{[66] Moderate successes related to affective computing include textual sentiment analysis and, more recently, multimodal sentiment analysis, wherein AI classifies the =
effects displayed by a videotaped subject.^[67]}

Gener= al intelligence

A machine with artificial gener= al intelligence would be able to solve a wide variety of problems with = breadth and versatility similar to human intelligence.[6= 8]

Techniques

AI research uses a wide variety of techniques to accomplish the goals ab= ove.^[b]

Se= arch and optimization

AI can solve many problems by intelligently searching through many possi= ble solutions.^[69]= There are two very different kinds of search used in AI: state space search and local search.

State s= pace search

State space search searches thr= ough a tree of possible states to try to find a goal state.^[70] For example, planning algorithms search through trees of = goals and subgoals, attempting to find a path to a target goal, a process c= alled means-ends analysis.[71]

Simple exhaustive searches [72] are rarely suf= ficient for most real-world problems: the search space (the = number of places to search) quickly grows to astronomical numbers. The result is a search that is too slow or never completes.^[15=
] "Heuristics" or "rules of thumb= " can help prioritize choices that are more likely to reach a goal.^[73]

Adversarial search is use= d for game-playing programs, such as chess or Go. It sea= rches through a tree of possible moves and countermoves, looking for a w= inning position.^[74]

Local search<= /h4>

Illustration of <= a href=3D"https://en.wikipedia.org/wiki/Gradient_descent" title=3D"Gradient= descent">gradient descent for 3 different starting points; two paramet= ers (represented by the plan coordinates) are adjusted in order to minimize= the loss function (the height)

Local search uses [75]

Gradient descent is a type of local search that optim= izes a set of numerical parameters by incrementally adjusting them to minim= ize a loss function. Variants of gradient descent are commonly use= d to train neural networ= ks,^[76] through = the backpropagation algorithm.

Another type of local search is evolution= ary computation, which aims to iteratively improve a set of candidate s= olutions by "mutating" and "recombining" them, selecting only the fittest to survive each generation.[<= /span>77]

Distributed search processes can coordinate via swarm= intelligence algorithms. Two popular swarm algorithms used in search a= re particle swarm optimization (inspir= ed by bird flocking) and ant = colony optimization (inspired by ant trails).[78]

Logic

Formal l= ogic is used for reasoning a= nd knowledge representatio= n.^[79] Formal logic comes in two main forms: propositional logic<= /a> (which operates on statements that are true or false and uses logical connectives such as "and", "or", "not" and "implies")= [80] and predicate logic (which also operates on objects, pred= icates and relations and uses quantifiers such as "Ev= ery X is a Y" and "There are some Xs that a= re Ys").^[81]

Deductive reasoning in logic is the process of= proving a new statement (c= onclusion) from other statements that are given and assumed to be true = (the pr= emises).^[82] Pro= ofs can be structured as proof trees, in which nodes are labelle= d by sentences, and children nodes are connected to parent nodes by inference rules.

Given a problem and a set of premises, problem-solving reduces to se= arching for a proof tree whose root node is labelled by a solution of the p= roblem and whose leaf nodes are labelled by premis= es or axiom= s. In the case of Horn clauses, problem-solving search can be perf= ormed by reasoning forwards from the premises or [83]= In the more general case of the clausal form of first-order logic, resolution is a single, axi= om-free rule of inference, in which a problem is solved by proving a contra= diction from premises that include the negation of the problem to be solved= .^[84]

Inference in both Horn clause logic and first-order logic is undecidable, and therefore intractable. However, backward reasoning with Horn clauses, which = underpins computation in the logic programming language Prolog,= is Turing complete. Moreover, its effic= iency is competitive with computation in other symbolic = programming languages.^{<=
span class=3D"cite-bracket">[85]=}

Fuzzy logic assigns a "degree of truth" between 0 and 1. It can= therefore handle propositions that are vague and partially true.^[=
86]

Non-monotonic logics, including logic programm= ing with negation as failure, are designed to handle <= a href=3D"https://en.wikipedia.org/wiki/Default_reasoning" class=3D"mw-redi= rect" title=3D"Default reasoning">default reasoning.^{[28] Other specialized versions of logic have been developed to describe man=
y complex domains.}

Probabilistic methods for uncertain reasoning

A simple Bayesian network, with the associated conditional probability tables

Many problems in AI (including reasoning, planning, learning, perception= , and robotics) require the agent to operate with incomplete or uncertain i= nformation. AI researchers have devised a number of tools to solve these pr= oblems using methods from probability theory and economics.^[87] Precise math= ematical tools have been developed that analyze how an agent can make choic= es and plan, using decision theory, decision anal= ysis,^[88] and information value theory= .^[89] These tools in= clude models such as Markov decision processes= ,^[90] dynamic decision networks,^[91] game theory and mechanism design.^[92]

Bayesian networks^[93] are a tool that can be used for re= asoning (using the Bayesian inference algorithm),[g]^[95] learning (using = the expectat= ion=E2=80=93maximization algorithm),^[h]^=
[97] planning (usin= g decision networks)^[98] and perception (using dynamic Bayesian networks).^[91]

Probabilistic algorithms can also be used for filtering, prediction,= smoothing, and finding explanations for streams of data, thus helping perc= eption systems analyze processes that occur over time (e.g., hidden Markov models or Kalman filters).^[91]

Expectation=E2=80=93maximization clustering of Old Faithful eruption data starts from a random guess but then success= fully converges on an accurate clustering of the two physically distinct mo= des of eruption.

Classifiers and statistical learning methods

The simplest AI applications can be divided into two types: classifiers = (e.g., "if shiny then diamond"), on one hand, and controllers (e.g., "if di= amond then pick up"), on the other hand. Classifiers^{=
[99]<=
/span>} are functions that use pattern matching to = determine the closest match. They can be fine-tuned based on chosen example= s using supervised learning. Each pattern (also called= an "observation") is labeled with a certa= in predefined class. All the observations combined with their class labels = are known as a data set. When a new observation is received, that observatio= n is classified based on previous experience.^[45]

There are many kinds of classifiers in use.^[100] The decision tree is the sim= plest and most widely used symbolic machine learning algorithm.^[=
101] K-nearest neighbor algorithm was the most widely used anal= ogical AI until the mid-1990s, and Kernel meth= ods such as the support vector machine (SVM)= displaced k-nearest neighbor in the 1990s.^[102] The naive Bayes classifier is reportedly the "m= ost widely used learner"^[103] at Google, du= e in part to its scalability.^{<=
a href=3D"https://en.wikipedia.org/wiki/Artificial_intelligence#cite_note-1=
12">[104]<=
/span>} Neural networks ar= e also used as classifiers.^[10=
5]

Artificial neural networks

A neural network is an interconnected group of nodes, akin to t= he vast network of neurons in the human brain.

An artificial neural network is based on a collection of nodes also know= n as artificial neurons, which loo= sely model the neurons in a biological brain. It is trai= ned to recognise patterns; once trained, it can recognise those patterns in= fresh data. There is an input, at least one hidden layer of nodes and an o= utput. Each node applies a function and once the weight<= /a> crosses its specified threshold, the data is transmitted to the next la= yer. A network is typically called a deep neural network if it has at least= 2 hidden layers.[105]

Learning algorithms for neural networks use local search to choose the weights that will get the right out= put for each input during training. The most common training technique is t= he backpropagation algorithm.^[106] Neural networks learn to model complex relati= onships between inputs and outputs and find patterns i= n data. In theory, a neural network can learn any function.^{[107<=
span class=3D"cite-bracket">]}

In feedforward neural networks th= e signal passes in only one direction.^[108] The term perceptron typically refers to a sing= le-layer neural network.^{<=
span class=3D"cite-bracket">[109]} In contrast, deep learning uses many layers.^[110] <= a href=3D"https://en.wikipedia.org/wiki/Recurrent_neural_network" title=3D"= Recurrent neural network">Recurrent neural networks (RNNs) feed the out= put signal back into the input, which allows short-term memories of previou= s input events. Long short-term memory networks = (LSTMs) are recurrent neural networks that better preserve longterm depende= ncies and are less sensitive to the vanishin= g gradient problem.^[111]= Convolutional neural networks= (CNNs) use layers of kernels to more = efficiently process local patterns. This local processing is especially imp= ortant in image processing, where the= early CNN layers typically identify simple local patterns such as edges an= d curves, with subsequent layers detecting more complex patterns like textu= res, and eventually whole objects.^[112]

Deep learnin= g

Deep learning uses several layers of neurons between the networ= k's inputs and outputs.^[110] The multiple layers can progressiv= ely extract higher-level features from the raw input. For example, in image processing, lower layers may identify= edges, while higher layers may identify the concepts relevant to a human s= uch as digits, letters, or faces.^{[114]<=
/span>}

Deep learning has profoundly improved the performance of programs in= many important subfields of artificial intelligence, including co= mputer vision, speech recognition, natural language processing, image classification,^[115] and others. The reason that deep learning perfor= ms so well in so many applications is not known as of 2021.^[116] The sudden success of deep learning in 2012= =E2=80=932015 did not occur because of some new discovery or theoretical br= eakthrough (deep neural networks and backpropagation had been described by = many people, as far back as the 1950s)^[i] but because of two factors: the incredible increase= in computer power (including the hundred-fold increase in speed by switchi= ng to GPUs) and the availability of vast amounts of training dat= a, especially the giant curated datasets used for benchmark testing, such a= s Ima= geNet.^[j]

GPT

Generative pre-trained tr= ansformers (GPT) are large language models (LLMs= ) that generate text based on the semantic relationships between words in s= entences. Text-based GPT models are pre-trained on a large corpus of text that can be from the Internet. The pretrain= ing consists of predicting the next token (a token being usu= ally a word, subword, or punctuation). Throughout this pretraining, GPT mod= els accumulate knowledge about the world and can then generate human-like t= ext by repeatedly predicting the next token. Typically, a subsequent traini= ng phase makes the model more truthful, useful, and harmless, usually with = a technique called reinforcement learning from human feedback (RLHF). Current GPT m= odels are prone to generating falsehoods called "hallucinations". These can be reduced wi= th RLHF and quality data, but the problem has been getting worse for reason= ing systems.^{[124] Such systems are used in chatbots, which allow people to ask a question or r=
equest a task in simple text.^[125]^[126]}

Current models and services include ChatGPT, Claude, Gemini, Copilot, and Meta AI.^[=
127] = Multimodal GPT models can process different types of data (modalities) su= ch as images, videos, sound, and text.^[128]

Hard= ware and software

= Main articles: Programming languages for artificial= intelligence and Ha= rdware for artificial intelligence

In the late 2010s, graphics processing units<= /a> (GPUs) that were increasingly designed with AI-specific enhancements an= d used with specialized TensorFlow software had replaced previously used= central processing unit (CPUs) as the dominan= t means for large-scale (commercial and academic) machine learni= ng models' training.^[129] Specialized programming languages such as Prolog were used in early AI resea= rch,^[130] but general-purpose programming = languages like Python have bec= ome predominant.^[131]

The transistor density in integrated circuits has= been observed to roughly double every 18 months=E2=80=94a trend known as <= a href=3D"https://en.wikipedia.org/wiki/Moore%27s_law" title=3D"Moore's law= ">Moore's law, named after the Intel co-founder Gordon Moore, who first= identified it. Improvements in GPUs have been even faster,[132] a trend sometim= es called Huang's law,^{<=
a href=3D"https://en.wikipedia.org/wiki/Artificial_intelligence#cite_note-1=
43">[133]<=
/span>} named after Nvidia co-founder and CEO Jensen Huang.

Applications<= /h2>

= Main article: Appl= ications of artificial intelligence

AI and machine learning tec= hnology is used in most of the essential applications of the 2020s, includi= ng:

search engines (such as = Google Search)
targeting online advertisements
recommendation systems (offered by Netflix, YouTube or Amazon) driving inter= net traffic
targeted advertisi= ng (AdSense, Facebook)
virtual assistants (such as Siri or Alexa)
autonomous vehicles (includi= ng drones, ADAS and self-driving car= s)
automatic = language translation (Microsoft Translator, Google Translate)
facial recognition (Apple's FaceID or Microsoft's DeepFace and = Google's Fac= eNet)
image labeling (used by Facebook, Appl= e's Photos and TikTok).

The deployment of AI may be overseen by a chi= ef automation officer (CAO).

Health= and medicine

= Main article: Artifici= al intelligence in healthcare

It has been suggested that AI can overcome discrepancies in funding allo= cated to different fields of research.^[134]

AlphaFold 2 (2021) demonstrated the abili= ty to approximate, in hours rather than months, the 3D structu= re of a protein.^[135] In 2023, it was reported that AI-guided drug discovery helped find = a class of antibiotics capable of killing two different types of drug-resis= tant bacteria.[136] In 2024, researchers used machine learning to accelerate the search for <= a href=3D"https://en.wikipedia.org/wiki/Parkinson%27s_disease" title=3D"Par= kinson's disease">Parkinson's disease drug treatments. Their aim was to= identify compounds that block the clumping, or aggregation, of al= pha-synuclein (the protein that characterises Parkinson's disease). The= y were able to speed up the initial screening process ten-fold and reduce t= he cost by a thousand-fold.^[137]^[138]

Games

= Main article: Artifi= cial intelligence in video games

Game playing programs have been used since the 1950s = to demonstrate and test AI's most advanced techniques.^[139] Deep B= lue became the first computer chess-playing system to beat a reigning w= orld chess champion, Garry Kasparov, on 11 May 1997.^[=
140] In 2011, in a Jeopardy!<= /a> quiz show exhibition match, IBM's question answering system, Watso= n, defeated the two greatest Jeopardy! champions, Brad Rutter and Ken Jennings, by a significant margin.^[141] In March 2016, AlphaGo won 4 out of 5 = games of Go in a match with Go champion Lee Sedol, becoming the first co= mputer Go-playing system to beat a professional Go player without handicaps. Then, in 2017, it d= efeated Ke Jie, who was the best Go player in the world.^[142=
] Other programs handle imperfect-information games, such= as the pok= er-playing program Pluribus.^[143] DeepMind de= veloped increasingly generalistic reinforcement lear= ning models, such as with MuZero, which could be trained to play chess, Go, = or Atari games.[144] In 2= 019, DeepMind's AlphaStar achieved grandmaster level in StarCraft II= , a particularly challenging real-time strategy game that involves incomple= te knowledge of what happens on the map.^[145] In 2021, an AI agent competed in a PlayStation = Gran Turismo competition, winning against four of = the world's best Gran Turismo drivers using deep reinforcement learning.[146] In 2024, Google= DeepMind introduced SIMA, a type of AI capable of autonomously playing nin= e previously unseen open-world video games by obse= rving screen output, as well as executing short, specific tasks in response= to natural language instructions.^[147]

Mathematics

Large language models, such as GPT-4, Gemini, Claude, Llama o= r Mistral, are increasingly used in mathematics. These probabilistic mod= els are versatile, but can also produce wrong answers in the form of hallucinations. They= sometimes need a large database of mathematical problems to learn from, bu= t also methods such as supervised fine-tuning^[148] or trained classifiers w= ith human-annotated data to improve answers for new problems and learn from= corrections.^{[149] A February 2024 study showed that the performance of some language models=
for reasoning capabilities in solving math problems not included in their =
training data was low, even for problems with only minor deviations from tr=
ained data.^[150] O=
ne technique to improve their performance involves training the models to p=
roduce correct reasoning steps, rather than just the c=
orrect result.^{[151] The Alibaba Group developed a version of its Qwen models called Qwen2-Math, that achieved state-of-the-art performance on several math=
ematical benchmarks, including 84% accuracy on the MATH dataset of competit=
ion mathematics problems.^[152] In Janua=
ry 2025, Microsoft proposed the technique rStar-Math that leverages =
Monte Carlo tree search and step-by-step reaso=
ning, enabling a relatively small language model like Qwen-7B to sol=
ve 53% of the AIME 2024 and 90% of the MATH benchmark problems.^[15=
3]}}

Alternatively, dedicated models for mathematical problem solving wit= h higher precision for the outcome including proof of theorems have been de= veloped such as AlphaTensor, AlphaGeometry, Alpha= Proof and AlphaEvolve^[154] all from Google DeepMind,^[=
155] Llemma from Eleuth= erAI^[156] or <= i>Julius.^[157]

When natural language is used to describe mathematical problems, con= verters can transform such prompts into a formal language such as Lean to define mathematical tasks. The experimental mod= el Gemini Deep Think accepts natural language prompts directly and a= chieved gold medal results in the International Math Olympiad of 2025.^[158] =20

Some models have been developed to solve challenging problems and re= ach good results in benchmark tests, others to serve as educational tools i= n mathematics.^[159]

Topological deep learning integrate= s various topological approaches.

Finance

Finance is one of the fastest growing sectors where applied AI tools are= being deployed: from retail online banking to investment advice and insura= nce, where automated "robot advisers" have been in use for some years.^[=
160]

According to Nicolas Firzli, director of the World Pensions & Investments Forum, i= t may be too early to see the emergence of highly innovative AI-informed fi= nancial products and services. He argues that "the deployment of AI tools w= ill simply further automatise things: destroying tens of thousands of jobs = in banking, financial planning, and pension advice in the process, but I'm = not sure it will unleash a new wave of [e.g., sophisticated] pension innova= tion."^[161]

Military

= Main article: Military applications of artificial intelligence

Various countries are deploying AI military applications.^[162] The mai= n applications enhance command and control, communicat= ions, sensors, integration and interoperability.^[163] Research is= targeting intelligence collection and analysis, logistics, cyber operation= s, information operations, and semiautonomous and autono= mous vehicles.^{<=
a href=3D"https://en.wikipedia.org/wiki/Artificial_intelligence#cite_note-C=
RS-2019-172">[162]} AI technologies enable coordination of sensors an= d effectors, threat detection and identification, marking of enemy position= s, target acquisition, coordination and deconfliction of= distributed Joint F= ires between networked combat vehicles, both human-operated and autonom= ous.^[163]

AI has been used in military operations in Iraq, Syria, Israel and U= kraine.^[162]^{[164]^[165]^[166]}

Generative A= I

These paragraphs are an excerpt from Generative artificial intelligence.[ed= it]

Generative artificial int= elligence, also known as generative AI or GenAI, is a subfield of artif= icial intelligence that uses generative models to generate t= ext, images, videos, audio, software code or other forms of data.^[167] These models learn the = underlying patterns and structures of their trai= ning data and use them to generate new data^[168] in response to input, which often t= akes the form of natural language prompts.^[169]^[170]

The prevalence of generative AI tools has increased significantly si= nce the AI boom in the 2020s. This boom was made possible by improvements in <= a href=3D"https://en.wikipedia.org/wiki/Deep_learning" title=3D"Deep learni= ng">deep neural networks, particular= ly large language models (LLMs), which are based on = the chatbots such as ChatGPT, Claude, <= a href=3D"https://en.wikipedia.org/wiki/Microsoft_Copilot" title=3D"Microso= ft Copilot">Copilot, DeepSeek, Google Gemini and Grok; text-to-image models such as Stable Diffusion, Midjour= ney, and DALL-E; and text-to-video models suc= h as Veo, LTX and = Sora.^[171]= [172]^[17=
3]

Companies in a variety of sectors have used generative AI, including= those in software development, healthcare,^[174] finance,^[175] entertainment,^[176] customer service,^[177] sales and marketing,^[178] art, writi= ng,^[179] and produ= ct design.^[180]

Agents

= Main article: Agentic AI

AI agents are software entities designed to perceive their environment, = make decisions, and take actions autonomously to achieve specific goals. Th= ese agents can interact with users, their environment, or other agents. AI = agents are used in various applications, including virtual ass= istants, chatbots, autonomous vehicles, game-playing systems, a= nd industrial robotics. AI agent= s operate within the constraints of their programming, available computatio= nal resources, and hardware limitations. This means they are restricted to = performing tasks within their defined scope and have finite memory and proc= essing capabilities. In real-world applications, AI agents often face time = constraints for decision-making and action execution. Many AI agents incorp= orate learning algorithms, enabling them to improve their performance over = time through experience or training. Using machine learning, AI agents can = adapt to new situations and optimise their behaviour for their designated t= asks.^[181]^[182]^[183]

Web search

Microsoft introduced Copilot Search in February 2023 un= der the name Bing Chat, as a built-in feature for Mi= crosoft Edge and Bing mobile app. Copilot Search provides AI-generated summ= aries^[184] and ste= p-by-step reasoning based of information from web publishers, ranked in Bin= g Search.^[185] For safety, Copilot uses AI-based classifiers and filters to reduce potenti= ally harmful content.^[186]=

Google officially pushed its AI Search at its Google I/O event on 20= May 2025.^[187] It= keeps people looking at Google instead of clicking on a search result. = AI Overviews uses Gemini 2.5 to provide contextual answers to user quer= ies based on web content.^=
[188]

Sexuality

Applications of AI in this domain include AI-enabled menstruation and fe= rtility trackers that analyze user data to offer predictions,^[18=
9] AI-integrated sex toys (e.= g., teledildonics),^[190] AI-generated sexual education content,^[191] and AI agents that simulate sexua= l and romantic partners (e.g., Replika).^[192] AI is also used for the production of non-consensu= al deepfake pornography, raising significant ethical= and legal concerns.^[193]

AI technologies have also been used to attempt to identify online gender-based violence and online sexual grooming of minors.^[194]^[195]=

Other industry-specific tasks

There are also thousands of successful AI applications used to solve spe= cific problems for specific industries or institutions. In a 2017 survey, o= ne in five companies reported having incorporated "AI" in some offerings or= processes.^[196] A= few examples are energy storage, medical diagnosis, military lo= gistics, applications that predict the result of judicial decisions, foreign policy, or supply chain management.

AI applications for evacuation and disaster management are growing. A= I has been used to investigate patterns in large-scale and small-scale evac= uations using historical data from GPS, videos or social media. Furthermore= , AI can provide real-time information on the evacuation conditions.^[197]^[198]^[199]

In agriculture, AI has helped farmers to increase yield and identify= areas that need irrigation, fertilization, pesticide treatments. Agronomis= ts use AI to conduct research and development. AI has been used to predict = the ripening time for crops such as tomatoes, monitor soil moisture, operat= e agricultural robots, conduct predictive analytics,= classify livestock pig call emotions, automate greenhouses, detect disease= s and pests, and save water.

Artificial intelligence is used in astronomy to analyze increasing a= mounts of available data and applications, mainly for "classification, regr= ession, clustering, forecasting, generation, discovery, and the development= of new scientific insights." For example, it is used for discovering exopl= anets, forecasting solar activity, and distinguishing between signals and i= nstrumental effects in gravitational wave astronomy. Additionally, it could= be used for activities in space, such as space exploration, including the = analysis of data from space missions, real-time science decisions of spacec= raft, space debris avoidance, and more autonomous operation.

During the 2024 Indian elections<= /a>, US$50 million was spent on authorized AI-generated content, notably by= creating deepfakes of allied (including sometimes deceased) politicians to = better engage with voters, and by translating speeches to various local lan= guages.^[200]

Ethics

= Main article: Ethics of artifi= cial intelligence

Fueled by growth in artificial intelligence, dat= a centers' demand for power increased in the 2020s.^[226]

Goals

Reasoning and problem-solving

K= nowledge representation

Planning and decision-making

Learning

Natural language processing

Perception

Social= intelligence

Gener= al intelligence

Techniques

Se= arch and optimization

State s= pace search

Local search<= /h4>

Logic

Probabilistic methods for uncertain reasoning

Classifiers and statistical learning methods

Artificial neural networks

Deep learnin= g

GPT

Hard= ware and software

Applications<= /h2>

Health= and medicine

Games

Mathematics

Finance

Military

Generative A= I

Agents

Web search

Sexuality

Other industry-specific tasks

Ethics

Risks and h= arm

Priv= acy and copyright

D= ominance by tech giants

Power needs and environmental impacts

Misinformat= ion

Algorithmic bias and fairness

Lack = of transparency

Bad actors and weaponized AI

Technological unemployment

Existenti= al risk

Ethical machines and alignment

Open source

Frameworks

Regulation

History

Philosophy

Defining artificial intelligence

Evaluating approaches to AI

Symbolic AI and its limits

Neat vs. = scruffy

So= ft vs. hard computing

Narr= ow vs. general AI

Machine consciousness, sentience, and mind

Consciousnes= s

Computationalism and functionalism

AI w= elfare and rights

Future

Superintelligence and the singularity

Transhumanis= m

In fiction

See also

Explanat= ory notes

References

Textbooks

History of A= I

Other source= s

External li= nks