definition of problem solving in artificial intelligence

Updated : 16 August 2024 Contributors : Cole Stryker, Eda Kavlakoglu

Artificial intelligence (AI) is technology that enables computers and machines to simulate human learning, comprehension, problem solving, decision making, creativity and autonomy.

Applications and devices equipped with AI can see and identify objects. They can understand and respond to human language. They can learn from new information and experience. They can make detailed recommendations to users and experts. They can act independently, replacing the need for human intelligence or intervention (a classic example being a self-driving car). 

But in 2024, most AI researchers and practitioners—and most AI-related headlines—are focused on breakthroughs in generative AI  (gen AI), a technology that can create original text, images, video and other content. To fully understand generative AI, it’s important to first understand the technologies on which generative AI tools are built: machine learning  (ML) and deep learning .

Learn how to choose the right approach in preparing data sets and employing AI models.

A simple way to think about AI is as a series of nested or derivative concepts that have emerged over more than 70 years:  

Directly underneath AI, we have machine learning, which involves creating models by training an algorithm to make predictions or decisions based on data. It encompasses a broad range of techniques that enable computers to learn from and make inferences based on data without being explicitly programmed for specific tasks. 

There are many types of machine learning techniques or algorithms, including linear regression ,  logistic regression , decision trees , random forest , support vector machines   (SVMs) , k-nearest neighbor (KNN), clustering and more. Each of these approaches is suited to different kinds of problems and data.

But one of the most popular types of machine learning algorithm is called a neural network (or artificial neural network). Neural networks are modeled after the human brain's structure and function. A neural network consists of interconnected layers of nodes (analogous to neurons) that work together to process and analyze complex data. Neural networks are well suited to tasks that involve identifying complex patterns and relationships in large amounts of data.

The simplest form of machine learning is called supervised learning , which involves the use of labeled data sets to train algorithms to classify data or predict outcomes accurately. In supervised learning, humans pair each training example with an output label. The goal is for the model to learn the mapping between inputs and outputs in the training data, so it can predict the labels of new, unseen data.  

Register for the guide on foundation models

Deep learning is a subset of machine learning that uses multilayered neural networks, called deep neural networks, that more closely simulate the complex decision-making power of the human brain.

Deep neural networks include an input layer, at least three but usually hundreds of hidden layers, and an output layer, unlike neural networks used in classic machine learning models, which usually have only one or two hidden layers.

These multiple layers enable unsupervised learning : they can automate the extraction of features from large, unlabeled and unstructured data sets, and make their own predictions about what the data represents.

Because deep learning doesn’t require human intervention, it enables machine learning at a tremendous scale. It is well suited to natural language processing (NLP) , computer vision , and other tasks that involve the fast, accurate identification complex patterns and relationships in large amounts of data. Some form of deep learning powers most of the artificial intelligence (AI) applications in our lives today.  

Deep learning also enables:

• Semi-supervised learning , which combines supervised and unsupervised learning by using both labeled and unlabeled data to train AI models for classification and regression tasks.
• Self-supervised learning , which generates implicit labels from unstructured data, rather than relying on labeled data sets for supervisory signals.
• Reinforcement learning , which learns by trial-and-error and reward functions rather than by extracting information from hidden patterns.
• Transfer learning , in which knowledge gained through one task or data set is used to improve model performance on another related task or different data set.

Generative AI, sometimes called "gen AI" , refers to deep learning models that can create complex original content—such as long-form text, high-quality images, realistic video or audio and more—in response to a user’s prompt or request.

At a high level, generative models encode a simplified representation of their training data, and then draw from that representation to create new work that’s similar, but not identical, to the original data.

Generative models have been used for years in statistics to analyze numerical data. But over the last decade, they evolved to analyze and generate more complex data types. This evolution coincided with the emergence of three sophisticated deep learning model types:

• Variational autoencoders  or VAEs, which were introduced in 2013, and enabled models that could generate multiple variations of content in response to a prompt or instruction.
• Diffusion models, first seen in 2014, which add "noise" to images until they are unrecognizable, and then remove the noise to generate original images in response to prompts.
• Transformers (also called transformer models), which are trained on sequenced data to generate extended sequences of content (such as words in sentences, shapes in an image, frames of a video or commands in software code). Transformers are at the core of most of today’s headline-making generative AI tools, including ChatGPT and GPT-4, Copilot, BERT, Bard and Midjourney. 

In general, generative AI operates in three phases:

• Training, to create a foundation model.
• Tuning, to adapt the model to a specific application.
• Generation, evaluation and more tuning, to improve accuracy.

Generative AI begins with a "foundation model"; a deep learning model that serves as the basis for multiple different types of generative AI applications.

The most common foundation models today are large language models (LLMs) , created for text generation applications. But there are also foundation models for image, video, sound or music generation, and multimodal foundation models that support several kinds of content.

To create a foundation model, practitioners train a deep learning algorithm on huge volumes of relevant raw, unstructured, unlabeled data, such as terabytes or petabytes of data text or images or video from the internet. The training yields a neural network of billions of parameters —encoded representations of the entities, patterns and relationships in the data—that can generate content autonomously in response to prompts. This is the foundation model.

This training process is compute-intensive, time-consuming and expensive. It requires thousands of clustered graphics processing units (GPUs) and weeks of processing, all of which typically costs millions of dollars. Open source foundation model projects, such as Meta's Llama-2, enable gen AI developers to avoid this step and its costs.

Next, the model must be tuned to a specific content generation task. This can be done in various ways, including:

• Fine-tuning, which involves feeding the model application-specific labeled data—questions or prompts the application is likely to receive, and corresponding correct answers in the wanted format.
• Reinforcement learning with human feedback (RLHF), in which human users evaluate the accuracy or relevance of model outputs so that the model can improve itself. This can be as simple as having people type or talk back corrections to a chatbot or virtual assistant.

Generation, evaluation and more tuning  

Developers and users regularly assess the outputs of their generative AI apps, and further tune the model—even as often as once a week—for greater accuracy or relevance. In contrast, the foundation model itself is updated much less frequently, perhaps every year or 18 months.

Another option for improving a gen AI app's performance is retrieval augmented generation (RAG), a technique for extending the foundation model to use relevant sources outside of the training data to refine the parameters for greater accuracy or relevance.

AI offers numerous benefits across various industries and applications. Some of the most commonly cited benefits include:

• Automation of repetitive tasks.
• More and faster insight from data.
• Enhanced decision-making.
• Fewer human errors.
• 24x7 availability.
• Reduced physical risks.

Automation of repetitive tasks  

AI can automate routine, repetitive and often tedious tasks—including digital tasks such as data collection, entering and preprocessing, and physical tasks such as warehouse stock-picking and manufacturing processes. This automation frees to work on higher value, more creative work.

Enhanced decision-making  

Whether used for decision support or for fully automated decision-making, AI enables faster, more accurate predictions and reliable, data-driven decisions . Combined with automation, AI enables businesses to act on opportunities and respond to crises as they emerge, in real time and without human intervention.

Fewer human errors  

AI can reduce human errors in various ways, from guiding people through the proper steps of a process, to flagging potential errors before they occur, and fully automating processes without human intervention. This is especially important in industries such as healthcare where, for example, AI-guided surgical robotics enable consistent precision.

Machine learning algorithms can continually improve their accuracy and further reduce errors as they're exposed to more data and "learn" from experience.

Round-the-clock availability and consistency  

AI is always on, available around the clock, and delivers consistent performance every time. Tools such as AI chatbots or virtual assistants can lighten staffing demands for customer service or support. In other applications—such as materials processing or production lines—AI can help maintain consistent work quality and output levels when used to complete repetitive or tedious tasks.

Reduced physical risk  

By automating dangerous work—such as animal control, handling explosives, performing tasks in deep ocean water, high altitudes or in outer space—AI can eliminate the need to put human workers at risk of injury or worse. While they have yet to be perfected, self-driving cars and other vehicles offer the potential to reduce the risk of injury to passengers.

The real-world applications of AI are many. Here is just a small sampling of use cases across various industries to illustrate its potential:

Customer experience, service and support  

Companies can implement AI-powered chatbots and virtual assistants to handle customer inquiries, support tickets and more. These tools use natural language processing (NLP) and generative AI capabilities to understand and respond to customer questions about order status, product details and return policies.

Chatbots and virtual assistants enable always-on support, provide faster answers to frequently asked questions (FAQs), free human agents to focus on higher-level tasks, and give customers faster, more consistent service.

Fraud detection  

Machine learning and deep learning algorithms can analyze transaction patterns and flag anomalies, such as unusual spending or login locations, that indicate fraudulent transactions. This enables organizations to respond more quickly to potential fraud and limit its impact, giving themselves and customers greater peace of mind.

Personalized marketing  

Retailers, banks and other customer-facing companies can use AI to create personalized customer experiences and marketing campaigns that delight customers, improve sales and prevent churn. Based on data from customer purchase history and behaviors, deep learning algorithms can recommend products and services customers are likely to want, and even generate personalized copy and special offers for individual customers in real time.

Human resources and recruitment  

AI-driven recruitment platforms can streamline hiring by screening resumes, matching candidates with job descriptions, and even conducting preliminary interviews using video analysis. These and other tools can dramatically reduce the mountain of administrative paperwork associated with fielding a large volume of candidates. It can also reduce response times and time-to-hire, improving the experience for candidates whether they get the job or not.

Application development and modernization  

Generative AI code generation tools and automation tools can streamline repetitive coding tasks associated with application development, and accelerate the migration and modernization (reformatting and replatorming) of legacy applications at scale. These tools can speed up tasks, help ensure code consistency and reduce errors.

Predictive maintenance  

Machine learning models can analyze data from sensors, Internet of Things (IoT) devices and operational technology (OT) to forecast when maintenance will be required and predict equipment failures before they occur. AI-powered preventive maintenance helps prevent downtime and enables you to stay ahead of supply chain issues before they affect the bottom line.

Organizations are scrambling to take advantage of the latest AI technologies and capitalize on AI's many benefits. This rapid adoption is necessary, but adopting and maintaining AI workflows comes with challenges and risks. 

Data risks  

AI systems rely on data sets that might be vulnerable to data poisoning, data tampering, data bias or cyberattacks that can lead to data breaches. Organizations can mitigate these risks by protecting data integrity and implementing security and availability throughout the entire AI lifecycle, from development to training and deployment and postdeployment.

Model risks  

Threat actors can target AI models for theft, reverse engineering or unauthorized manipulation. Attackers might compromise a model’s integrity by tampering with its architecture, weights or parameters; the core components that determine a model’s behavior, accuracy and performance.

Operational risks  

Like all technologies, models are susceptible to operational risks such as model drift, bias and breakdowns in the governance structure. Left unaddressed, these risks can lead to system failures and cybersecurity vulnerabilities that threat actors can use.

Ethics and legal risks  

If organizations don’t prioritize safety and ethics when developing and deploying AI systems, they risk committing privacy violations and producing biased outcomes. For example, biased training data used for hiring decisions might reinforce gender or racial stereotypes and create AI models that favor certain demographic groups over others.  

AI ethics is a multidisciplinary field that studies how to optimize AI's beneficial impact while reducing risks and adverse outcomes. Principles of AI ethics are applied through a system of AI governance consisted of guardrails that help ensure that AI tools and systems remain safe and ethical.  

AI governance encompasses oversight mechanisms that address risks. An ethical approach to AI governance requires the involvement of a wide range of stakeholders, including developers, users, policymakers and ethicists, helping to ensure that AI-related systems are developed and used to align with society's values.

Here are common values associated with AI ethics and responsible AI :

As AI becomes more advanced, humans are challenged to comprehend and retrace how the algorithm came to a result. Explainable AI is a set of processes and methods that enables human users to interpret, comprehend and trust the results and output created by algorithms.

Although machine learning, by its very nature, is a form of statistical discrimination, the discrimination becomes objectionable when it places privileged groups at systematic advantage and certain unprivileged groups at systematic disadvantage, potentially causing varied harms. To encourage fairness, practitioners can try to minimize algorithmic bias across data collection and model design, and to build more diverse and inclusive teams.

Robust AI effectively handles exceptional conditions, such as abnormalities in input or malicious attacks, without causing unintentional harm. It is also built to withstand intentional and unintentional interference by protecting against exposed vulnerabilities.

Organizations should implement clear responsibilities and governance structures for the development, deployment and outcomes of AI systems. In addition, users should be able to see how an AI service works, evaluate its functionality, and comprehend its strengths and limitations. Increased transparency provides information for AI consumers to better understand how the AI model or service was created.

Many regulatory frameworks, including GDPR, mandate that organizations abide by certain privacy principles when processing personal information. It is crucial to be able to protect AI models that might contain personal information, control what data goes into the model in the first place, and to build adaptable systems that can adjust to changes in regulation and attitudes around AI ethics.

In order to contextualize the use of AI at various levels of complexity and sophistication, researchers have defined several types of AI that refer to its level of sophistication:

Weak AI : Also known as “narrow AI,” defines AI systems designed to perform a specific task or a set of tasks. Examples might include “smart” voice assistant apps, such as Amazon’s Alexa, Apple’s Siri, a social media chatbot or the autonomous vehicles promised by Tesla. 

Strong AI : Also known as “artificial general intelligence” (AGI) or “general AI,” possess the ability to understand, learn and apply knowledge across a wide range of tasks at a level equal to or surpassing human intelligence . This level of AI is currently theoretical and no known AI systems approach this level of sophistication. Researchers argue that if AGI is even possible, it requires major increases in computing power. Despite recent advances in AI development, self-aware AI systems of science fiction remain firmly in that realm. 

The idea of "a machine that thinks" dates back to ancient Greece. But since the advent of electronic computing (and relative to some of the topics discussed in this article) important events and milestones in the evolution of AI include the following:

1950 Alan Turing publishes Computing Machinery and Intelligence (link resides outside ibm.com). In this paper, Turing—famous for breaking the German ENIGMA code during WWII and often referred to as the "father of computer science"—asks the following question: "Can machines think?" 

From there, he offers a test, now famously known as the "Turing Test," where a human interrogator would try to distinguish between a computer and human text response. While this test has undergone much scrutiny since it was published, it remains an important part of the history of AI, and an ongoing concept within philosophy as it uses ideas around linguistics. 

1956 John McCarthy coins the term "artificial intelligence" at the first-ever AI conference at Dartmouth College. (McCarthy went on to invent the Lisp language.) Later that year, Allen Newell, J.C. Shaw and Herbert Simon create the Logic Theorist, the first-ever running AI computer program.

1967 Frank Rosenblatt builds the Mark 1 Perceptron, the first computer based on a neural network that "learned" through trial and error. Just a year later, Marvin Minsky and Seymour Papert publish a book titled Perceptrons, which becomes both the landmark work on neural networks and, at least for a while, an argument against future neural network research initiatives. 

1980 Neural networks, which use a backpropagation algorithm to train itself, became widely used in AI applications.

1995 Stuart Russell and Peter Norvig publish Artificial Intelligence: A Modern Approach (link resides outside ibm.com), which becomes one of the leading textbooks in the study of AI. In it, they delve into four potential goals or definitions of AI, which differentiates computer systems based on rationality and thinking versus acting. 

1997 IBM's Deep Blue beats then world chess champion Garry Kasparov, in a chess match (and rematch).

2004 John McCarthy writes a paper, What Is Artificial Intelligence? (link resides outside ibm.com), and proposes an often-cited definition of AI. By this time, the era of big data and cloud computing is underway, enabling organizations to manage ever-larger data estates, which will one day be used to train AI models. 

2011 IBM Watson® beats champions Ken Jennings and Brad Rutter at Jeopardy! Also, around this time, data science begins to emerge as a popular discipline.

2015 Baidu's Minwa supercomputer uses a special deep neural network called a convolutional neural network to identify and categorize images with a higher rate of accuracy than the average human. 

2016 DeepMind's AlphaGo program, powered by a deep neural network, beats Lee Sodol, the world champion Go player, in a five-game match. The victory is significant given the huge number of possible moves as the game progresses (over 14.5 trillion after just four moves). Later, Google purchased DeepMind for a reported USD 400 million.

2022 A rise in large language models  or LLMs, such as OpenAI’s ChatGPT, creates an enormous change in performance of AI and its potential to drive enterprise value. With these new generative AI practices, deep-learning models can be pretrained on large amounts of data.

2024 The latest AI trends point to a continuing AI renaissance. Multimodal models that can take multiple types of data as input are providing richer, more robust experiences. These models bring together computer vision image recognition and NLP speech recognition capabilities. Smaller models are also making strides in an age of diminishing returns with massive models with large parameter counts. 

IBM watsonx.ai AI studio is part of the IBM  watsonx ™ AI and data platform, bringing together new generative AI (gen AI) capabilities powered by  foundation models  and traditional machine learning (ML) into a powerful studio spanning the AI lifecycle. 

Put AI to work in your business with IBM’s industry-leading AI expertise and portfolio of solutions at your side.

Transform standard support into exceptional care when you give your customers instant, accurate custom care anytime, anywhere, with conversational AI.

Reinvent critical workflows and operations by adding AI to maximize experiences, real-time decision-making and business value.

AI is changing the game for cybersecurity, analyzing massive quantities of risk data to speed response times and augment under-resourced security operations.

Led by top IBM thought leaders, the curriculum is designed to help business leaders gain the knowledge needed to prioritize the AI investments that can drive growth.

Access our full catalog of over 100 online courses by purchasing an individual or multi-user digital learning subscription today, enabling you to expand your skills across a range of our products at one low price.

IBM watsonx™ Assistant is recognized as a Customers' Choice in the 2023 Gartner Peer Insights Voice of the Customer report for Enterprise Conversational AI platforms.

Discover how machine learning can predict demand and cut costs.

2024 stands to be a pivotal year for the future of AI, as researchers and enterprises seek to establish how this evolutionary leap in technology can be most practically integrated into our everyday lives.

The European Union’s AI Act entered into force on 1 August 2024. The EU AI Act is one of the first regulations of artificial intelligence to be adopted in one of the world’s largest markets. What is it going to change for businesses?

Train, validate, tune and deploy generative AI, foundation models and machine learning capabilities with IBM watsonx.ai, a next-generation enterprise studio for AI builders. Build AI applications in a fraction of the time with a fraction of the data.

Problem-Solving Agents In Artificial Intelligence

In artificial intelligence, a problem-solving agent refers to a type of intelligent agent designed to address and solve complex problems or tasks in its environment. These agents are a fundamental concept in AI and are used in various applications, from game-playing algorithms to robotics and decision-making systems. Here are some key characteristics and components of a problem-solving agent:

• Perception : Problem-solving agents typically have the ability to perceive or sense their environment. They can gather information about the current state of the world, often through sensors, cameras, or other data sources.
• Knowledge Base : These agents often possess some form of knowledge or representation of the problem domain. This knowledge can be encoded in various ways, such as rules, facts, or models, depending on the specific problem.
• Reasoning : Problem-solving agents employ reasoning mechanisms to make decisions and select actions based on their perception and knowledge. This involves processing information, making inferences, and selecting the best course of action.
• Planning : For many complex problems, problem-solving agents engage in planning. They consider different sequences of actions to achieve their goals and decide on the most suitable action plan.
• Actuation : After determining the best course of action, problem-solving agents take actions to interact with their environment. This can involve physical actions in the case of robotics or making decisions in more abstract problem-solving domains.
• Feedback : Problem-solving agents often receive feedback from their environment, which they use to adjust their actions and refine their problem-solving strategies. This feedback loop helps them adapt to changing conditions and improve their performance.
• Learning : Some problem-solving agents incorporate machine learning techniques to improve their performance over time. They can learn from experience, adapt their strategies, and become more efficient at solving similar problems in the future.

Problem-solving agents can vary greatly in complexity, from simple algorithms that solve straightforward puzzles to highly sophisticated AI systems that tackle complex, real-world problems. The design and implementation of problem-solving agents depend on the specific problem domain and the goals of the AI application.

Hello, I’m Hridhya Manoj. I’m passionate about technology and its ever-evolving landscape. With a deep love for writing and a curious mind, I enjoy translating complex concepts into understandable, engaging content. Let’s explore the world of tech together

Which Of The Following Is A Privilege In SQL Standard

Implicit Return Type Int In C

Leave a Comment Cancel reply

Save my name, email, and website in this browser for the next time I comment.

Reach Out to Us for Any Query

SkillVertex is an edtech organization that aims to provide upskilling and training to students as well as working professionals by delivering a diverse range of programs in accordance with their needs and future aspirations.

© 2024 Skill Vertex

• Part 2 Problem-solving »
• Chapter 3 Solving Problems by Searching
• Edit on GitHub

Chapter 3 Solving Problems by Searching 

When the correct action to take is not immediately obvious, an agent may need to plan ahead : to consider a sequence of actions that form a path to a goal state. Such an agent is called a problem-solving agent , and the computational process it undertakes is called search .

Problem-solving agents use atomic representations, that is, states of the world are considered as wholes, with no internal structure visible to the problem-solving algorithms. Agents that use factored or structured representations of states are called planning agents .

We distinguish between informed algorithms, in which the agent can estimate how far it is from the goal, and uninformed algorithms, where no such estimate is available.

3.1 Problem-Solving Agents 

If the agent has no additional information—that is, if the environment is unknown —then the agent can do no better than to execute one of the actions at random. For now, we assume that our agents always have access to information about the world. With that information, the agent can follow this four-phase problem-solving process:

GOAL FORMULATION : Goals organize behavior by limiting the objectives and hence the actions to be considered.

PROBLEM FORMULATION : The agent devises a description of the states and actions necessary to reach the goal—an abstract model of the relevant part of the world.

SEARCH : Before taking any action in the real world, the agent simulates sequences of actions in its model, searching until it finds a sequence of actions that reaches the goal. Such a sequence is called a solution .

EXECUTION : The agent can now execute the actions in the solution, one at a time.

It is an important property that in a fully observable, deterministic, known environment, the solution to any problem is a fixed sequence of actions . The open-loop system means that ignoring the percepts breaks the loop between agent and environment. If there is a chance that the model is incorrect, or the environment is nondeterministic, then the agent would be safer using a closed-loop approach that monitors the percepts.

In partially observable or nondeterministic environments, a solution would be a branching strategy that recommends different future actions depending on what percepts arrive.

3.1.1 Search problems and solutions 

A search problem can be defined formally as follows:

A set of possible states that the environment can be in. We call this the state space .

The initial state that the agent starts in.

A set of one or more goal states . We can account for all three of these possibilities by specifying an \(Is\-Goal\) method for a problem.

The actions available to the agent. Given a state \(s\) , \(Actions(s)\) returns a finite set of actions that can be executed in \(s\) . We say that each of these actions is applicable in \(s\) .

A transition model , which describes what each action does. \(Result(s,a)\) returns the state that results from doing action \(a\) in state \(s\) .

An action cost function , denote by \(Action\-Cost(s,a,s\pr)\) when we are programming or \(c(s,a,s\pr)\) when we are doing math, that gives the numeric cost of applying action \(a\) in state \(s\) to reach state \(s\pr\) .

A sequence of actions forms a path , and a solution is a path from the initial state to a goal state. We assume that action costs are additive; that is, the total cost of a path is the sum of the individual action costs. An optimal solution has the lowest path cost among all solutions.

The state space can be represented as a graph in which the vertices are states and the directed edges between them are actions.

3.1.2 Formulating problems 

The process of removing detail from a representation is called abstraction . The abstraction is valid if we can elaborate any abstract solution into a solution in the more detailed world. The abstraction is useful if carrying out each of the actions in the solution is easier than the original problem.

3.2 Example Problems 

A standardized problem is intended to illustrate or exercise various problem-solving methods. It can be given a concise, exact description and hence is suitable as a benchmark for researchers to compare the performance of algorithms. A real-world problem , such as robot navigation, is one whose solutions people actually use, and whose formulation is idiosyncratic, not standardized, because, for example, each robot has different sensors that produce different data.

3.2.1 Standardized problems 

A grid world problem is a two-dimensional rectangular array of square cells in which agents can move from cell to cell.

Vacuum world

Sokoban puzzle

Sliding-tile puzzle

3.2.2 Real-world problems 

Route-finding problem

Touring problems

Trveling salesperson problem (TSP)

VLSI layout problem

Robot navigation

Automatic assembly sequencing

3.3 Search Algorithms 

A search algorithm takes a search problem as input and returns a solution, or an indication of failure. We consider algorithms that superimpose a search tree over the state-space graph, forming various paths from the initial state, trying to find a path that reaches a goal state. Each node in the search tree corresponds to a state in the state space and the edges in the search tree correspond to actions. The root of the tree corresponds to the initial state of the problem.

The state space describes the (possibly infinite) set of states in the world, and the actions that allow transitions from one state to another. The search tree describes paths between these states, reaching towards the goal. The search tree may have multiple paths to (and thus multiple nodes for) any given state, but each node in the tree has a unique path back to the root (as in all trees).

The frontier separates two regions of the state-space graph: an interior region where every state has been expanded, and an exterior region of states that have not yet been reached.

3.3.1 Best-first search 

In best-first search we choose a node, \(n\) , with minimum value of some evaluation function , \(f(n)\) .

3.3.2 Search data structures 

A node in the tree is represented by a data structure with four components

\(node.State\) : the state to which the node corresponds;

\(node.Parent\) : the node in the tree that generated this node;

\(node.Action\) : the action that was applied to the parent’s state to generate this node;

\(node.Path\-Cost\) : the total cost of the path from the initial state to this node. In mathematical formulas, we use \(g(node)\) as a synonym for \(Path\-Cost\) .

Following the \(PARENT\) pointers back from a node allows us to recover the states and actions along the path to that node. Doing this from a goal node gives us the solution.

We need a data structure to store the frontier . The appropriate choice is a queue of some kind, because the operations on a frontier are:

\(Is\-Empty(frontier)\) returns true only if there are no nodes in the frontier.

\(Pop(frontier)\) removes the top node from the frontier and returns it.

\(Top(frontier)\) returns (but does not remove) the top node of the frontier.

\(Add(node, frontier)\) inserts node into its proper place in the queue.

Three kinds of queues are used in search algorithms:

A priority queue first pops the node with the minimum cost according to some evaluation function, \(f\) . It is used in best-first search.

A FIFO queue or first-in-first-out queue first pops the node that was added to the queue first; we shall see it is used in breadth-first search.

A LIFO queue or last-in-first-out queue (also known as a stack ) pops first the most recently added node; we shall see it is used in depth-first search.

3.3.3 Redundant paths 

A cycle is a special case of a redundant path .

As the saying goes, algorithms that cannot remember the past are doomed to repeat it . There are three approaches to this issue.

First, we can remember all previously reached states (as best-first search does), allowing us to detect all redundant paths, and keep only the best path to each state.

Second, we can not worry about repeating the past. We call a search algorithm a graph search if it checks for redundant paths and a tree-like search if it does not check.

Third, we can compromise and check for cycles, but not for redundant paths in general.

3.3.4 Measuring problem-solving performance 

COMPLETENESS : Is the algorithm guaranteed to find a solution when there is one, and to correctly report failure when there is not?

COST OPTIMALITY : Does it find a solution with the lowest path cost of all solutions?

TIME COMPLEXITY : How long does it take to find a solution?

SPACE COMPLEXITY : How much memory is needed to perform the search?

To be complete, a search algorithm must be systematic in the way it explores an infinite state space, making sure it can eventually reach any state that is connected to the initial state.

In theoretical computer science, the typical measure of time and space complexity is the size of the state-space graph, \(|V|+|E|\) , where \(|V|\) is the number of vertices (state nodes) of the graph and \(|E|\) is the number of edges (distinct state/action pairs). For an implicit state space, complexity can be measured in terms of \(d\) , the depth or number of actions in an optimal solution; \(m\) , the maximum number of actions in any path; and \(b\) , the branching factor or number of successors of a node that need to be considered.

3.4 Uninformed Search Strategies 

3.4.1 breadth-first search .

When all actions have the same cost, an appropriate strategy is breadth-first search , in which the root node is expanded first, then all the successors of the root node are expanded next, then their successors, and so on.

Breadth-first search always finds a solution with a minimal number of actions, because when it is generating nodes at depth \(d\) , it has already generated all the nodes at depth \(d-1\) , so if one of them were a solution, it would have been found.

All the nodes remain in memory, so both time and space complexity are \(O(b^d)\) . The memory requirements are a bigger problem for breadth-first search than the execution time . In general, exponential-complexity search problems cannot be solved by uninformed search for any but the smallest instances .

3.4.2 Dijkstra’s algorithm or uniform-cost search 

When actions have different costs, an obvious choice is to use best-first search where the evaluation function is the cost of the path from the root to the current node. This is called Dijkstra’s algorithm by the theoretical computer science community, and uniform-cost search by the AI community.

The complexity of uniform-cost search is characterized in terms of \(C^*\) , the cost of the optimal solution, and \(\epsilon\) , a lower bound on the cost of each action, with \(\epsilon>0\) . Then the algorithm’s worst-case time and space complexity is \(O(b^{1+\lfloor C^*/\epsilon\rfloor})\) , which can be much greater than \(b^d\) .

When all action costs are equal, \(b^{1+\lfloor C^*/\epsilon\rfloor}\) is just \(b^{d+1}\) , and uniform-cost search is similar to breadth-first search.

3.4.3 Depth-first search and the problem of memory 

Depth-first search always expands the deepest node in the frontier first. It could be implemented as a call to \(Best\-First\-Search\) where the evaluation function \(f\) is the negative of the depth.

For problems where a tree-like search is feasible, depth-first search has much smaller needs for memory. A depth-first tree-like search takes time proportional to the number of states, and has memory complexity of only \(O(bm)\) , where \(b\) is the branching factor and \(m\) is the maximum depth of the tree.

A variant of depth-first search called backtracking search uses even less memory.

3.4.4 Depth-limited and iterative deepening search 

To keep depth-first search from wandering down an infinite path, we can use depth-limited search , a version of depth-first search in which we supply a depth limit, \(l\) , and treat all nodes at depth \(l\) as if they had no successors. The time complexity is \(O(b^l)\) and the space complexity is \(O(bl)\)

Iterative deepening search solves the problem of picking a good value for \(l\) by trying all values: first 0, then 1, then 2, and so on—until either a solution is found, or the depth- limited search returns the failure value rather than the cutoff value.

Its memory requirements are modest: \(O(bd)\) when there is a solution, or \(O(bm)\) on finite state spaces with no solution. The time complexity is \(O(bd)\) when there is a solution, or \(O(bm)\) when there is none.

In general, iterative deepening is the preferred uninformed search method when the search state space is larger than can fit in memory and the depth of the solution is not known .

3.4.5 Bidirectional search 

An alternative approach called bidirectional search simultaneously searches forward from the initial state and backwards from the goal state(s), hoping that the two searches will meet.

3.4.6 Comparing uninformed search algorithms 

3.5 Informed (Heuristic) Search Strategies 

An informed search strategy uses domain–specific hints about the location of goals to find colutions more efficiently than an uninformed strategy. The hints come in the form of a heuristic function , denoted \(h(n)\) :

\(h(n)\) = estimated cost of the cheapest path from the state at node \(n\) to a goal state.

3.5.1 Greedy best-first search 

Greedy best-first search is a form of best-first search that expands first the node with the lowest \(h(n)\) value—the node that appears to be closest to the goal—on the grounds that this is likely to lead to a solution quickly. So the evaluation function \(f(n)=h(n)\) .

• History & Society
• Science & Tech
• Biographies
• Animals & Nature
• Geography & Travel
• Arts & Culture
• Games & Quizzes
• On This Day
• One Good Fact
• New Articles
• Lifestyles & Social Issues
• Philosophy & Religion
• Politics, Law & Government
• World History
• Health & Medicine
• Browse Biographies
• Birds, Reptiles & Other Vertebrates
• Bugs, Mollusks & Other Invertebrates
• Environment
• Fossils & Geologic Time
• Entertainment & Pop Culture
• Sports & Recreation
• Visual Arts
• Demystified
• Image Galleries
• Infographics
• Top Questions
• Britannica Kids
• Saving Earth
• Space Next 50
• Student Center
• Introduction & Top Questions
• Problem solving
• Symbolic vs. connectionist approaches
• Artificial general intelligence (AGI), applied AI, and cognitive simulation
• Machine learning
• Large language models and natural language processing
• Autonomous vehicles
• Virtual assistants
• Is artificial general intelligence (AGI) possible?

What is artificial intelligence?

artificial intelligence

Our editors will review what you’ve submitted and determine whether to revise the article.

• Harvard University - Science in the News - The History of Artificial Intelligence
• National Center for Biotechnology Information - PubMed Central - The rise of artificial intelligence in healthcare applications
• Lifewire - What is artificial intelligence?
• Computer History Museum - AI and Robotics
• Internet Encyclopedia of Philosophy - Artificial Intelligence
• artificial intelligence - Children's Encyclopedia (Ages 8-11)
• artificial intelligence (AI) - Student Encyclopedia (Ages 11 and up)
• Table Of Contents

Artificial intelligence is the ability of a computer or computer-controlled robot to perform tasks that are commonly associated with the  intellectual processes characteristic of humans , such as the ability to reason. Although there are as yet no AIs that match full human flexibility over wider domains or in tasks requiring much everyday knowledge, some AIs perform specific tasks as well as humans. Learn more.

Are artificial intelligence and machine learning the same?

No, artificial intelligence and machine learning are not the same, but they are closely related. Machine learning is the method to train a computer to learn from its inputs but without explicit programming for every circumstance. Machine learning helps a computer to achieve artificial intelligence.

Recent News

artificial intelligence (AI) , the ability of a digital computer or computer-controlled robot to perform tasks commonly associated with intelligent beings. The term is frequently applied to the project of developing systems endowed with the intellectual processes characteristic of humans, such as the ability to reason, discover meaning, generalize, or learn from past experience. Since their development in the 1940s, digital computers have been programmed to carry out very complex tasks—such as discovering proofs for mathematical theorems or playing chess —with great proficiency. Despite continuing advances in computer processing speed and memory capacity, there are as yet no programs that can match full human flexibility over wider domains or in tasks requiring much everyday knowledge. On the other hand, some programs have attained the performance levels of human experts and professionals in executing certain specific tasks, so that artificial intelligence in this limited sense is found in applications as diverse as medical diagnosis , computer search engines , voice or handwriting recognition, and chatbots .

What is intelligence?

• Investing in AI stocks
• Using AI for Money Management
• How AI is Changing Work
• Ethical Questions and AI
• AI and Regulation

All but the simplest human behavior is ascribed to intelligence, while even the most complicated insect behavior is usually not taken as an indication of intelligence. What is the difference? Consider the behavior of the digger wasp , Sphex ichneumoneus . When the female wasp returns to her burrow with food, she first deposits it on the threshold , checks for intruders inside her burrow, and only then, if the coast is clear, carries her food inside. The real nature of the wasp’s instinctual behavior is revealed if the food is moved a few inches away from the entrance to her burrow while she is inside: on emerging, she will repeat the whole procedure as often as the food is displaced. Intelligence—conspicuously absent in the case of the wasp—must include the ability to adapt to new circumstances.

(Read Ray Kurzweil’s Britannica essay on the future of “Nonbiological Man.”)

Psychologists generally characterize human intelligence not by just one trait but by the combination of many diverse abilities. Research in AI has focused chiefly on the following components of intelligence: learning, reasoning, problem solving , perception , and using language.

There are a number of different forms of learning as applied to artificial intelligence. The simplest is learning by trial and error. For example, a simple computer program for solving mate-in-one chess problems might try moves at random until mate is found. The program might then store the solution with the position so that, the next time the computer encountered the same position, it would recall the solution. This simple memorizing of individual items and procedures—known as rote learning—is relatively easy to implement on a computer. More challenging is the problem of implementing what is called generalization . Generalization involves applying past experience to analogous new situations. For example, a program that learns the past tense of regular English verbs by rote will not be able to produce the past tense of a word such as jump unless the program was previously presented with jumped , whereas a program that is able to generalize can learn the “add -ed ” rule for regular verbs ending in a consonant and so form the past tense of jump on the basis of experience with similar verbs.

(Read Yuval Noah Harari’s Britannica essay on the future of “Nonconscious Man.”)

Kelvin Smith Library

Artificial Intelligence

What is artificial intelligence (ai).

• Ethical Considerations
• AI for Students
• AI for Faculty and Staff
• AI Citation and Academic Integrity
• AI and Copyright

This is a living document for a rapidly changing field. We welcome your feedback and recommendations!

This guide includes material from:

• Duke University YouTube
• Loyola Notre Dame Library
• University of Calgary Library: Artificial Intelligence
• Florida International University Library
• UC San Diego Library
• Cleveland State University Library

Get Online Help

• KSL Ask A Librarian Information on how to get help by email, phone, & chat.

“ Artificial intelligence (AI) is the design, implementation, and use of programs, machines, and systems that exhibit human intelligence, with its most important activities being knowledge representation, reasoning, and learning. Artificial intelligence encompasses a number of important subareas, including voice recognition, image identification, natural language processing, expert systems, neural networks , planning, robotics , and intelligent agents.” ( Salem Press Encyclopedia of Science ).

“ Generative artificial intelligence is a type of artificial intelligence (AI) technology that can make content such as audio, images, text, and videos. It involves algorithms such as ChatGPT , a chatbot that can produce essays, poetry, and other content requested by a user, and DALL-E, which generates art.” ( Salem Press Encyclopedia of Science ).

Artificial intelligence is used across all industries and academic subjects. The term is used to describe everything from finding the best route on Apple and Google Maps, self-driving cars, algorithms to display a list in a certain order on a website or in a social media app, and facial recognition software to unlock a smartphone. It is part of our everyday lives, at work, in school and at home.

• What is an AI anyway?
• AI vs Machine Learning
• How AI Could Save (Not Destroy) Education
• How will AI Change the World?
• What is Artificial Intelligence?
• Generative AI in a Nutshell

Additional Resources

Introductory resources.

• A Jargon-Free Explanation of how AI Large Language Models Work Want to really understand large language models? Here’s a gentle primer.
• Demystifying Large Language Models: A Beginner’s Guide
• Large Language Model The Wikipedia entry on Large Language Models
• Generative AI - A Primer An introduction to generative artificial intelligence technology and its implications on education
• Florida International University’s AI Guide Find resources on artificial intelligence, ChatGPT, writing with AI assistance, AI academic productivity tools, plagiarism, prompt engineering, GPT misinformation and hallucinations, AI image tools, AI literacy, and discussions related to AI ethics
• The UC San Diego Guide on Generative Artificial Intelligence: Using Generative AI Tools
• AIPRM’s Ultimate Generative AI Glossary
• Glossary of AI Terms for Educators From the Center for Integrative Research in Computer and Learning Science

Advanced Resources

• Free Code Camp One of the best computer programming resources around, Free Code Camp’s extensive tutorial library includes a number of different courses on LLM use and development.
• Hugging Face Hugging Face, Inc. is a French-American company that develops computation tools for building applications using machine learning. It is most notable for its transformers library built for natural language processing applications and its platform that allows users to share machine learning models and datasets and showcase their work. Hugging Face has extensive educational resources to get you familiar with its platform and best practices for using LLMs and other tools in natural language processing.
• Full Stack LLM Bootcamp “The Full Stack brings people together to learn and share best practices across the entire lifecycle of an AI-powered product: from defining the problem and picking a GPU or foundation model to production deployment and continual learning to user experience design.” The LLM Bootcamp is an open course designed to teach you how to leverage LLMs in application development.
• Cohere LLM University Seven Modules from Cohere that provide an in depth look at how to use LLMs in a variety of projects.
• Next: AI Literacy >>
• Last Updated: Aug 28, 2024 9:31 AM
• URL: https://researchguides.case.edu/artificialintelligence

Understanding AI Problem Formulation Basics

In artificial intelligence, problem formulation is the process of identifying, analyzing, and defining the problems that need to be solved using AI techniques. It involves breaking down complex problems into smaller, more manageable components and formulating a clear problem statement. Problem formulation plays a crucial role in shaping efficient and smart solutions in the field of AI. It helps AI agents define the goals, initial state, actions, transitions, and goal test required to solve a problem. By understanding the basics of problem formulation in AI , you can gain insights into the techniques and algorithms used to solve various problems in artificial intelligence.

• 0.1 Key Takeaways
• 1 The Three Types of Problems in AI
• 2.1 Step 1: Problem Definition
• 2.2 Step 2: Problem Analysis
• 2.3 Step 3: Knowledge Representation
• 2.4 Step 4: Problem-Solving
• 2.5 Step 5: Formulating Associated Problem Components
• 3 Problem-Solving Approaches in AI
• 4 Conclusion
• 5.1 What is problem formulation in artificial intelligence?
• 5.2 What are the three types of problems in AI?
• 5.3 What are the steps for problem solving in AI?
• 5.4 What are the problem-solving approaches in AI?
• 5.5 What is the role of problem formulation and problem solving in AI?
• 6 Source Links

Key Takeaways

• Problem formulation is the process of identifying, analyzing, and defining the problems in artificial intelligence.
• It involves breaking down complex problems into smaller, more manageable components.
• Problem formulation helps AI agents define the goals, initial state, actions, transitions, and goal test required to solve a problem.
• Understanding problem formulation is essential for implementing problem-solving techniques in AI.
• Various techniques and algorithms are used in problem formulation to solve problems in artificial intelligence.

The Three Types of Problems in AI

In the domain of artificial intelligence, there are three types of problems: ignorable, recoverable, and irrecoverable. Understanding these problem types is crucial in determining the appropriate problem-solving techniques and algorithms to be applied.

Ignorable problems are those where certain solution steps can be ignored without affecting the final outcome. These problems usually involve redundant or unnecessary actions that can be omitted in the solution process.

Recoverable problems , on the other hand, are those where solution steps can be undone or reversed if needed. This type of problem allows for flexibility in the problem-solving process, as mistakes or incorrect steps can be rectified along the way.

Irrecoverable problems are the most challenging type, as the solution steps cannot be undone once executed. This means that careful consideration and analysis must be undertaken before taking any action, as there is no turning back.

Examples of problem formulation in artificial intelligence can include tasks such as:

• Pathfinding: Finding the optimal route from one point to another in a given environment.
• Constraint Satisfaction: Satisfying a set of constraints or conditions while searching for a solution.
• Optimization: Maximizing or minimizing a specific objective within a given set of constraints.

By understanding the different types of problems in AI and their corresponding examples, you can approach problem formulation and problem-solving with a clear strategy and direction, optimizing your chances of finding effective solutions.

Steps for Problem Solving in AI

Problem-solving in AI is a multi-step process that allows you to tackle complex problems using various techniques and algorithms. By understanding and following these steps, you can effectively solve problems in the field of artificial intelligence.

Step 1: Problem Definition

Problem definition is the first crucial step in problem-solving. It involves clearly specifying the inputs and acceptable system solutions for the given problem. By defining the problem accurately, you provide a solid foundation for finding the right solution.

Step 2: Problem Analysis

Once the problem is defined, the next step is to analyze it thoroughly. This involves examining the problem from different angles, identifying any patterns or underlying factors, and gaining a deeper understanding of its complexity. Problem analysis helps you uncover valuable insights that can guide your problem-solving approach.

Step 3: Knowledge Representation

Knowledge representation involves collecting detailed information about the problem and exploring possible techniques and algorithms for solving it. By understanding the available resources and methodologies, you can choose the most effective approach to tackle the problem at hand.

Step 4: Problem-Solving

Once you have analyzed the problem and gathered the necessary knowledge, it’s time to apply problem-solving techniques. This step involves selecting the best techniques and algorithms based on the problem’s characteristics and constraints. By using the right tools, you increase the chances of finding an optimal solution.

Step 5: Formulating Associated Problem Components

To achieve the desired goal, it’s crucial to formulate the associated problem components. This includes defining the initial state, actions, transitions, goal test, and path costing required to solve the problem effectively. By carefully formulating these components, you create a structured framework that aids in problem-solving.

Understanding these steps is essential for implementing problem-solving techniques in AI. By following a systematic approach and leveraging the power of AI algorithms, you can overcome complex challenges and find innovative solutions.

Problem-Solving Approaches in AI

In the field of artificial intelligence, there are various approaches to problem-solving. These approaches utilize different algorithms and techniques to tackle complex problems and find effective solutions. Three common problem-solving approaches in AI include heuristic algorithms , searching algorithms , and genetic algorithms .

Heuristic algorithms are used to experiment and test different procedures in order to understand the problem and generate a solution. While they may not always provide the optimal solution, heuristic algorithms offer effective short-term methods for achieving goals. By leveraging prior knowledge and experience, these algorithms can guide problem-solving processes and provide valuable insights.

Searching algorithms are fundamental techniques used by rational agents or problem-solving agents to find the most appropriate solutions. These algorithms involve creating and exploring a search space to identify the desired solution. By systematically traversing the search space, searching algorithms can efficiently navigate through complex problem domains and identify potential solutions. They play a crucial role in solving problems such as pathfinding, constraint satisfaction, and optimization tasks.

Genetic algorithms are inspired by evolutionary theory and natural selection. These algorithms employ a population-based approach and simulate the natural process of evolution to solve problems. By generating and evolving populations of potential solutions, genetic algorithms mimic genetic variation, selection, and reproduction to find optimal or near-optimal solutions. Genetic algorithms are particularly effective in solving complex problems with multiple variables and constraints.

Understanding these problem-solving approaches is essential in selecting the most suitable technique for a given problem in artificial intelligence. Whether utilizing heuristic algorithms , searching algorithms , or genetic algorithms, each approach offers unique benefits and trade-offs. By leveraging these approaches, AI practitioners can develop intelligent systems and applications capable of solving complex problems effectively and efficiently.

Note: The table above summarizes the main characteristics of each problem-solving approach in AI.

Problem formulation is a critical component of problem-solving in artificial intelligence. By identifying, analyzing, and defining the problems that need to be solved using AI techniques, you can lay the foundation for efficient and smart solutions. Through problem formulation, AI agents can define clear goals, initial states, actions, transitions, and goal tests required to achieve the desired outcome.

To successfully solve problems in AI, it is essential to follow the steps for problem-solving in AI. These steps include problem definition, problem analysis, knowledge representation, problem-solving, and formulation of associated problem components. By systematically going through these steps, you can gain a deep understanding of the problem and select the most suitable techniques to solve it.

Various problem-solving approaches, such as heuristic algorithms, searching algorithms, and genetic algorithms, can be applied in AI. Heuristic algorithms allow for experimentation and testing, offering effective short-term methods for achieving goals. Searching algorithms explore a search space to find the most appropriate solutions, while genetic algorithms generate and evolve potential solutions based on fitness criteria.

In conclusion, problem formulation and problem solving in AI play a vital role in shaping the field of artificial intelligence. By applying these techniques and approaches, intelligent systems and applications can be developed to tackle complex challenges and improve our lives.

What is problem formulation in artificial intelligence?

Problem formulation in artificial intelligence is the process of identifying, analyzing, and defining the problems that need to be solved using AI techniques. It involves breaking down complex problems into smaller components and formulating a clear problem statement.

What are the three types of problems in AI?

The three types of problems in AI are ignorable, recoverable, and irrecoverable. Ignorable problems are those where certain solution steps can be ignored without affecting the final outcome. Recoverable problems are those where solution steps can be undone if needed. Irrecoverable problems are those where solution steps cannot be undone.

What are the steps for problem solving in AI?

The steps for problem solving in AI include problem definition, problem analysis, knowledge representation, problem-solving, and formulation of associated problem components. Problem definition involves specifying the inputs and acceptable system solutions, while problem analysis entails analyzing the problem thoroughly. Knowledge representation involves collecting detailed information about the problem, and problem-solving is the selection of the best techniques to solve the problem. Formulating the associated problem components includes defining the initial state, actions, transitions, goal test, and path costing.

What are the problem-solving approaches in AI?

The problem-solving approaches in AI include heuristic algorithms, searching algorithms, and genetic algorithms. Heuristic algorithms experiment and test procedures to generate a solution, searching algorithms involve creating and exploring a search space, and genetic algorithms generate and evolve populations of potential solutions based on fitness criteria.

What is the role of problem formulation and problem solving in AI?

Problem formulation and problem solving in AI play a vital role in shaping efficient and smart solutions. Problem formulation helps AI agents define the goals, initial state, actions, transitions, and goal test required to solve a problem, while problem solving techniques and approaches enable the development of intelligent systems and applications in the field of artificial intelligence.

Source Links

• https://hbr.org/2023/06/ai-prompt-engineering-isnt-the-future
• https://www.geeksforgeeks.org/problem-solving-in-artificial-intelligence/
• https://www.javatpoint.com/problem-solving-techniques-in-ai

Mark Walters

With years of experience in the tech industry, Mark is not just a writer but a storyteller who brings the world of technology to life. His passion for demystifying the intricacies of the digital realm sets Twefy.com apart as a platform where accessibility meets expertise.

Leave a Comment Cancel reply

Save my name, email, and website in this browser for the next time I comment.

At Twefy.com, we are passionate about technology and its ever-evolving landscape. We strive to be your go-to source for the latest tech news, insightful analyses, and practical guides to help you confidently navigate the digital world.

Whether you’re a seasoned tech enthusiast or just dipping your toes into the vast sea of innovation, Twefy.com empowers and inspires you.

Privacy Policy

Terms and Conditions

Latest Blogs

Unlocking Linchpin Technology: Your Key to Innovation

March 20, 2024

Ethernet Traits: Speed & Reliability Uncovered

How Cities Grew: Tech Advance Behind Urban Expansion

© Twefy.com (2024)

Terms and Condition

Privacy policy

Artificial Intelligence

Control System

• Interview Q

Intelligent Agent

Problem-solving, adversarial search, knowledge represent, uncertain knowledge r., subsets of ai, artificial intelligence mcq, related tutorials.

• Send your Feedback to [email protected]

Help Others, Please Share

Learn Latest Tutorials

Transact-SQL

Reinforcement Learning

R Programming

React Native

Python Design Patterns

Python Pillow

Python Turtle

Preparation

Verbal Ability

Interview Questions

Company Questions

Trending Technologies

Cloud Computing

Data Science

Machine Learning

B.Tech / MCA

Data Structures

Operating System

Computer Network

Compiler Design

Computer Organization

Discrete Mathematics

Ethical Hacking

Computer Graphics

Software Engineering

Web Technology

Cyber Security

C Programming

Data Mining

Data Warehouse

MIT Technology Review

• Newsletters

What is AI?

Everyone thinks they know but no one can agree. And that’s a problem.

• Will Douglas Heaven archive page

Internet nastiness, name-calling, and other not-so-petty, world-altering disagreements

AI is sexy, AI is cool. AI is entrenching inequality, upending the job market, and wrecking education. AI is a theme-park ride, AI is a magic trick. AI is our final invention, AI is a moral obligation. AI is the buzzword of the decade, AI is marketing jargon from 1955. AI is humanlike, AI is alien. AI is super-smart and as dumb as dirt. The AI boom will boost the economy, the AI bubble is about to burst. AI will increase abundance and empower humanity to maximally flourish in the universe. AI will kill us all.

What the hell is everybody talking about?

Artificial intelligence is the hottest technology of our time. But what is it? It sounds like a stupid question, but it’s one that’s never been more urgent. Here’s the short answer: AI is a catchall term for a set of technologies that make computers do things that are thought to require intelligence when done by people. Think of recognizing faces, understanding speech, driving cars, writing sentences, answering questions, creating pictures. But even that definition contains multitudes.

And that right there is the problem. What does it mean for machines to understand speech or write a sentence? What kinds of tasks could we ask such machines to do? And how much should we trust the machines to do them?

As this technology moves from prototype to product faster and faster, these have become questions for all of us. But (spoilers!) I don’t have the answers. I can’t even tell you what AI is. The people making it don’t know what AI is either. Not really. “These are the kinds of questions that are important enough that everyone feels like they can have an opinion,” says Chris Olah, chief scientist at the San Francisco–based AI lab Anthropic. “I also think you can argue about this as much as you want and there’s no evidence that’s going to contradict you right now.”

But if you’re willing to buckle up and come for a ride, I can tell you why nobody really knows, why everybody seems to disagree, and why you’re right to care about it.

Let’s start with an offhand joke.

Back in 2022, partway through the first episode of Mystery AI Hype Theater 3000 , a party-pooping podcast in which the irascible cohosts Alex Hanna and Emily Bender have a lot of fun sticking “the sharpest needles’’ into some of Silicon Valley’s most inflated sacred cows, they make a ridiculous suggestion. They’re hate-reading aloud from a 12,500-word Medium post by a Google VP of engineering, Blaise Agüera y Arcas, titled “ Can machines learn how to behave? ” Agüera y Arcas makes a case that AI can understand concepts in a way that’s somehow analogous to the way humans understand concepts—concepts such as moral values. In short, perhaps machines can be taught to behave. 

Hanna and Bender are having none of it. They decide to replace the term “AI’’ with “mathy math”—you know, just lots and lots of math.

The irreverent phrase is meant to collapse what they see as bombast and anthropomorphism in the sentences being quoted. Pretty soon Hanna, a sociologist and director of research at the Distributed AI Research Institute, and Bender, a computational linguist at the University of Washington (and internet-famous critic of tech industry hype), open a gulf between what Agüera y Arcas wants to say and how they choose to hear it.

“How should AIs, their creators, and their users be held morally accountable?” asks Agüera y Arcas.

How should mathy math be held morally accountable? asks Bender.

“There’s a category error here,” she says. Hanna and Bender don’t just reject what Agüera y Arcas says; they claim it makes no sense. “Can we please stop it with the ‘an AI’ or ‘the AIs’ as if they are, like, individuals in the world?” Bender says.

It might sound as if they’re talking about different things, but they’re not. Both sides are talking about large language models, the technology behind the current AI boom. It’s just that the way we talk about AI is more polarized than ever. In May, OpenAI CEO Sam Altman teased the latest update to GPT-4 , his company’s flagship model, by tweeting , “Feels like magic to me.”

There’s a lot of road between math and magic.

AI has acolytes, with a faith-like belief in the technology’s current power and inevitable future improvement. Artificial general intelligence is in sight, they say; superintelligence is coming behind it. And it has heretics, who pooh-pooh such claims as mystical mumbo-jumbo.

The buzzy popular narrative is shaped by a pantheon of big-name players, from Big Tech marketers in chief like Sundar Pichai and Satya Nadella to edgelords of industry like Elon Musk and Altman to celebrity computer scientists like Geoffrey Hinton . Sometimes these boosters and doomers are one and the same, telling us that the technology is so good it’s bad .

As AI hype has ballooned, a vocal anti-hype lobby has risen in opposition, ready to smack down its ambitious, often wild claims. Pulling in this direction are a raft of researchers, including Hanna and Bender, and also outspoken industry critics like influential computer scientist and former Googler Timnit Gebru and NYU cognitive scientist Gary Marcus. All have a chorus of followers bickering in their replies.

In short, AI has come to mean all things to all people, splitting the field into fandoms. It can feel as if different camps are talking past one another, not always in good faith.

Maybe you find all this silly or tiresome. But given the power and complexity of these technologies—which are already used to determine how much we pay for insurance, how we look up information, how we do our jobs, etc. etc. etc.—it’s about time we at least agreed on what it is we’re even talking about.

Yet in all the conversations I’ve had with people at the cutting edge of this technology, no one has given a straight answer about exactly what it is they’re building. (A quick side note: This piece focuses on the AI debate in the US and Europe, largely because many of the best-funded, most cutting-edge AI labs are there. But of course there’s important research happening elsewhere, too, in countries with their own varying perspectives on AI, particularly China.) Partly, it’s the pace of development. But the science is also wide open. Today’s large language models can do amazing things . The field just can’t find common ground on what’s really going on under the hood .

These models are trained to complete sentences. They appear to be able to do a lot more—from solving high school math problems to writing computer code to passing law exams to composing poems. When a person does these things, we take it as a sign of intelligence. What about when a computer does it? Is the appearance of intelligence enough?

These questions go to the heart of what we mean by “artificial intelligence,” a term people have actually been arguing about for decades. But the discourse around AI has become more acrimonious with the rise of large language models that can mimic the way we talk and write with thrilling/chilling (delete as applicable) realism.

We have built machines with humanlike behavior but haven’t shrugged off the habit of imagining a humanlike mind behind them. This leads to over-egged evaluations of what AI can do; it hardens gut reactions into dogmatic positions, and it plays into the wider culture wars between techno-optimists and techno-skeptics.

Add to this stew of uncertainty a truckload of cultural baggage, from the science fiction that I’d bet many in the industry were raised on, to far more malign ideologies that influence the way we think about the future. Given this heady mix, arguments about AI are no longer simply academic (and perhaps never were). AI inflames people’s passions and makes grownups call each other names.

“It’s not in an intellectually healthy place right now,” Marcus says of the debate. For years Marcus has pointed out the flaws and limitations of deep learning, the tech that launched AI into the mainstream, powering everything from LLMs to image recognition to self-driving cars. His 2001 book The Algebraic Mind argued that neural networks, the foundation on which deep learning is built, are incapable of reasoning by themselves. (We’ll skip over it for now, but I’ll come back to it later and we’ll see just how much a word like “reasoning” matters in a sentence like this.)

Marcus says that he has tried to engage Hinton—who last year went public with existential fears about the technology he helped invent—in a proper debate about how good large language models really are. “He just won’t do it,” says Marcus. “He calls me a twit.” (Having talked to Hinton about Marcus in the past, I can confirm that. “ChatGPT clearly understands neural networks better than he does,” Hinton told me last year.) Marcus also drew ire when he wrote an essay titled “Deep learning is hitting a wall.” Altman responded to it with a tweet : “Give me the confidence of a mediocre deep learning skeptic.”

At the same time, banging his drum has made Marcus a one-man brand and earned him an invitation to sit next to Altman and give testimony last year before the US Senate’s AI oversight committee.

And that’s why all these fights matter more than your average internet nastiness. Sure, there are big egos and vast sums of money at stake. But more than that, these disputes matter when industry leaders and opinionated scientists are summoned by heads of state and lawmakers to explain what this technology is and what it can do (and how scared we should be). They matter when this technology is being built into software we use every day, from search engines to word-processing apps to assistants on your phone. AI is not going away. But if we don’t know what we’re being sold, who’s the dupe?

“It is hard to think of another technology in history about which such a debate could be had—a debate about whether it is everywhere, or nowhere at all,” Stephen Cave and Kanta Dihal write in Imagining AI , a 2023 collection of essays about how different cultural beliefs shape people’s views of artificial intelligence. “That it can be held about AI is a testament to its mythic quality.”

Above all else, AI is an idea—an ideal—shaped by worldviews and sci-fi tropes as much as by math and computer science. Figuring out what we are talking about when we talk about AI will clarify many things. We won’t agree on them, but common ground on what AI is would be a great place to start talking about what AI should be .

What is everyone really fighting about, anyway?

In late 2022, soon after OpenAI released ChatGPT , a new meme started circulating online that captured the weirdness of this technology better than anything else. In most versions , a Lovecraftian monster called the Shoggoth, all tentacles and eyeballs, holds up a bland smiley-face emoji as if to disguise its true nature. ChatGPT presents as humanlike and accessible in its conversational wordplay, but behind that façade lie unfathomable complexities—and horrors. (“It was a terrible, indescribable thing vaster than any subway train—a shapeless congeries of protoplasmic bubbles,” H.P. Lovecraft wrote of the Shoggoth in his 1936 novella At the Mountains of Madness .)  

For years one of the best-known touchstones for AI in pop culture was The Terminator , says Dihal. But by putting ChatGPT online for free, OpenAI gave millions of people firsthand experience of something different. “AI has always been a sort of really vague concept that can expand endlessly to encompass all kinds of ideas,” she says. But ChatGPT made those ideas tangible: “Suddenly, everybody has a concrete thing to refer to.” What is AI? For millions of people the answer was now: ChatGPT.

The AI industry is selling that smiley face hard. Consider how The Daily Show recently skewered the hype, as expressed by industry leaders. Silicon Valley’s VC in chief, Marc Andreessen: “This has the potential to make life much better … I think it’s honestly a layup.” Altman: “I hate to sound like a utopic tech bro here, but the increase in quality of life that AI can deliver is extraordinary.” Pichai: “AI is the most profound technology that humanity is working on. More profound than fire.”

Jon Stewart: “Yeah, suck a dick, fire!”

But as the meme points out, ChatGPT is a friendly mask. Behind it is a monster called GPT-4, a large language model built from a vast neural network that has ingested more words than most of us could read in a thousand lifetimes. During training, which can last months and cost tens of millions of dollars, such models are given the task of filling in blanks in sentences taken from millions of books and a significant fraction of the internet. They do this task over and over again. In a sense, they are trained to be supercharged autocomplete machines. The result is a model that has turned much of the world’s written information into a statistical representation of which words are most likely to follow other words, captured across billions and billions of numerical values.

It’s math—a hell of a lot of math. Nobody disputes that. But is it just that, or does this complex math encode algorithms capable of something akin to human reasoning or the formation of concepts?

Many of the people who answer yes to that question believe we’re close to unlocking something called artificial general intelligence , or AGI, a hypothetical future technology that can do a wide range of tasks as well as humans can. A few of them have even set their sights on what they call superintelligence , sci-fi technology that can do things far better than humans. This cohort believes AGI will drastically change the world—but to what end? That’s yet another point of tension. It could fix all the world’s problems—or bring about its doom. 

kinda mad how the so called godfathers of AI managed to convince seemingly smart people within AI field & many regulators to buy into the absurd idea that a sophisticated curve fitting (to a dataset) machine can have the urge to exterminate humans — Abeba Birhane (@Abebab) June 30, 2024

Today AGI appears in the mission statements of the world’s top AI labs. But the term was invented in 2007 as a niche attempt to inject some pizzazz into a field that was then best known for applications that read handwriting on bank deposit slips or recommended your next book to buy. The idea was to reclaim the original vision of an artificial intelligence that could do humanlike things (more on that soon).

It was really an aspiration more than anything else, Google DeepMind cofounder Shane Legg, who coined the term, told me last year: “I didn’t have an especially clear definition.”

AGI became the most controversial idea in AI . Some talked it up as the next big thing: AGI was AI but, you know, much better . Others claimed the term was so vague that it was meaningless.

“AGI used to be a dirty word,” Ilya Sutskever told me, before he resigned as chief scientist at OpenAI.

But large language models, and ChatGPT in particular, changed everything. AGI went from dirty word to marketing dream.

Which brings us to what I think is one of the most illustrative disputes of the moment—one that sets up the sides of the argument and the stakes in play. 

Seeing magic in the machine

A few months before the public launch of OpenAI’s large language model GPT-4 in March 2023, the company shared a prerelease version with Microsoft, which wanted to use the new model to revamp its search engine Bing.

At the time, Sebastian Bubeck was studying the limitations of LLMs and was somewhat skeptical of their abilities. In particular, Bubeck—the vice president of generative AI research at Microsoft Research in Redmond, Washington—had been trying and failing to get the technology to solve middle school math problems. Things like: x – y = 0; what are x and y ? “My belief was that reasoning was a bottleneck, an obstacle,” he says. “I thought that you would have to do something really fundamentally different to get over that obstacle.”

Then he got his hands on GPT-4. The first thing he did was try those math problems. “The model nailed it,” he says. “Sitting here in 2024, of course GPT-4 can solve linear equations. But back then, this was crazy. GPT-3 cannot do that.”

But Bubeck’s real road-to-Damascus moment came when he pushed it to do something new.

The thing about middle school math problems is that they are all over the internet, and GPT-4 may simply have memorized them. “How do you study a model that may have seen everything that human beings have written?” asks Bubeck. His answer was to test GPT-4 on a range of problems that he and his colleagues believed to be novel.

Playing around with Ronen Eldan, a mathematician at Microsoft Research, Bubeck asked GPT-4 to give, in verse, a mathematical proof that there are an infinite number of primes.

Here’s a snippet of GPT-4’s response: “If we take the smallest number in S that is not in P / And call it p, we can add it to our set, don’t you see? / But this process can be repeated indefinitely. / Thus, our set P must also be infinite, you’ll agree.”

Cute, right? But Bubeck and Eldan thought it was much more than that. “We were in this office,” says Bubeck, waving at the room behind him via Zoom. “Both of us fell from our chairs. We couldn’t believe what we were seeing. It was just so creative and so, like, you know, different.” 

The Microsoft team also got GPT-4 to generate the code to add a horn to a cartoon picture of a unicorn drawn in Latex, a word processing program. Bubeck thinks this shows that the model could read the existing Latex code, understand what it depicted, and identify where the horn should go.

“There are many examples, but a few of them are smoking guns of reasoning,” he says—reasoning being a crucial building block of human intelligence.

Bubeck, Eldan, and a team of other Microsoft researchers described their findings in a paper that they called “ Spark s of artificial general intelligence ”: “We believe that GPT-4’s intelligence signals a true paradigm shift in the field of computer science and beyond.” When Bubeck shared the paper online, he tweeted : “time to face it, the sparks of #AGI have been ignited.”

The Sparks paper quickly became infamous—and a touchstone for AI boosters. Agüera y Arcas and Peter Norvig, a former director of research at Google and coauthor of Artificial Intelligence: A Modern Approach , perhaps the most popular AI textbook in the world, cowrote an article called “ Artificial General Intelligence Is Already Here .” Published in Noema , a magazine backed by an LA think tank called the Berggruen Institute, their argument uses the Sparks paper as a jumping-off point: “Artificial General Intelligence (AGI) means many different things to different people, but the most important parts of it have already been achieved by the current generation of advanced AI large language models,” they wrote. “Decades from now, they will be recognized as the first true examples of AGI.”

Since then, the hype has continued to balloon. Leopold Aschenbrenner, who at the time was a researcher at OpenAI focusing on superintelligence, told me last year: “AI progress in the last few years has been just extraordinarily rapid. We’ve been crushing all the benchmarks, and that progress is continuing unabated. But it won’t stop there. We’re going to have superhuman models, models that are much smarter than us.” (He was fired from OpenAI in April because, he claims, he raised security concerns about the tech he was building and “ ruffled some feathers .” He has since set up a Silicon Valley investment fund.)

In June, Aschenbrenner put out a 165-page manifesto arguing that AI will outpace college graduates by “2025/2026” and that “we will have superintelligence, in the true sense of the word” by the end of the decade. But others in the industry scoff at such claims. When Aschenbrenner tweeted a chart to show how fast he thought AI would continue to improve given how fast it had improved in last few years, the tech investor Christian Keil replied that by the same logic, his baby son, who had doubled in size since he was born, would weigh 7.5 trillion tons by the time he was 10.

It’s no surprise that “sparks of AGI” has also become a byword for over-the-top buzz. “I think they got carried away,” says Marcus, speaking about the Microsoft team. “They got excited, like ‘Hey, we found something! This is amazing!’ They didn’t vet it with the scientific community.” Bender refers to the Sparks paper as a “fan fiction novella.”

Not only was it provocative to claim that GPT-4’s behavior showed signs of AGI, but Microsoft, which uses GPT-4 in its own products, has a clear interest in promoting the capabilities of the technology. “This document is marketing fluff masquerading as research,” one tech COO posted on LinkedIn.

Some also felt the paper’s methodology was flawed. Its evidence is hard to verify because it comes from interactions with a version of GPT-4 that was not made available outside OpenAI and Microsoft. The public version has guardrails that restrict the model’s capabilities, admits Bubeck. This made it impossible for other researchers to re-create his experiments.

One group tried to re-create the unicorn example with a coding language called Processing, which GPT-4 can also use to generate images . They found that the public version of GPT-4 could produce a passable unicorn but not flip or rotate that image by 90 degrees. It may seem like a small difference, but such things really matter when you’re claiming that the ability to draw a unicorn is a sign of AGI.

The key thing about the examples in the Sparks paper, including the unicorn, is that Bubeck and his colleagues believe they are genuine examples of creative reasoning. This means the team had to be certain that examples of these tasks, or ones very like them, were not included anywhere in the vast data sets that OpenAI amassed to train its model. Otherwise, the results could be interpreted instead as instances where GPT-4 reproduced patterns it had already seen.

Bubeck insists that they set the model only tasks that would not be found on the internet. Drawing a cartoon unicorn in Latex was surely one such task. But the internet is a big place. Other researchers soon pointed out that there are indeed online forums dedicated to drawing animals in Latex . “Just fyi we knew about this,” Bubeck replied on X. “Every single query of the Sparks paper was thoroughly looked for on the internet.”

(This didn’t stop the name-calling: “I’m asking you to stop being a charlatan,” Ben Recht, a computer scientist at the University of California, Berkeley, tweeted back before accusing Bubeck of “being caught flat-out lying.”)

Bubeck insists the work was done in good faith, but he and his coauthors admit in the paper itself that their approach was not rigorous—notebook observations rather than foolproof experiments. 

Still, he has no regrets: “The paper has been out for more than a year and I have yet to see anyone give me a convincing argument that the unicorn, for example, is not a real example of reasoning.”

That’s not to say he can give me a straight answer to the big question—though his response reveals what kind of answer he’d like to give. “What is AI?” Bubeck repeats back to me. “I want to be clear with you. The question can be simple, but the answer can be complex.”

“There are many simple questions out there to which we still don’t know the answer. And some of those simple questions are the most profound ones,” he says. “I’m putting this on the same footing as, you know, What is the origin of life? What is the origin of the universe? Where did we come from? Big, big questions like this.”

Seeing only math in the machine

Before Bender became one of the chief antagonists of AI’s boosters, she made her mark on the AI world as a coauthor on two influential papers. (Both peer-reviewed, she likes to point out—unlike the Sparks paper and many of the others that get much of the attention.) The first, written with Alexander Koller, a fellow computational linguist at Saarland University in Germany, and published in 2020, was called “ Climbing towards NLU ” (NLU is natural-language understanding).

“The start of all this for me was arguing with other people in computational linguistics whether or not language models understand anything,” she says. (Understanding, like reasoning, is typically taken to be a basic ingredient of human intelligence.)

Bender and Koller argue that a model trained exclusively on text will only ever learn the form of a language, not its meaning. Meaning, they argue, consists of two parts: the words (which could be marks or sounds) plus the reason those words were uttered. People use language for many reasons, such as sharing information, telling jokes, flirting, warning somebody to back off, and so on. Stripped of that context, the text used to train LLMs like GPT-4 lets them mimic the patterns of language well enough for many sentences generated by the LLM to look exactly like sentences written by a human. But there’s no meaning behind them, no spark . It’s a remarkable statistical trick, but completely mindless.

They illustrate their point with a thought experiment. Imagine two English-speaking people stranded on neighboring deserted islands. There is an underwater cable that lets them send text messages to each other. Now imagine that an octopus, which knows nothing about English but is a whiz at statistical pattern matching, wraps its suckers around the cable and starts listening in to the messages. The octopus gets really good at guessing what words follow other words. So good that when it breaks the cable and starts replying to messages from one of the islanders, she believes that she is still chatting with her neighbor. (In case you missed it, the octopus in this story is a chatbot.)

The person talking to the octopus would stay fooled for a reasonable amount of time, but could that last? Does the octopus understand what comes down the wire? 

Imagine that the islander now says she has built a coconut catapult and asks the octopus to build one too and tell her what it thinks. The octopus cannot do this. Without knowing what the words in the messages refer to in the world, it cannot follow the islander’s instructions. Perhaps it guesses a reply: “Okay, cool idea!” The islander will probably take this to mean that the person she is speaking to understands her message. But if so, she is seeing meaning where there is none. Finally, imagine that the islander gets attacked by a bear and sends calls for help down the line. What is the octopus to do with these words?

Bender and Koller believe that this is how large language models learn and why they are limited. “The thought experiment shows why this path is not going to lead us to a machine that understands anything,” says Bender. “The deal with the octopus is that we have given it its training data, the conversations between those two people, and that’s it. But then here’s something that comes out of the blue and it won’t be able to deal with it because it hasn’t understood.”

The other paper Bender is known for, “ On the Dangers of Stochastic Parrots ,” highlights a series of harms that she and her coauthors believe the companies making large language models are ignoring. These include the huge computational costs of making the models and their environmental impact; the racist, sexist, and other abusive language the models entrench; and the dangers of building a system that could fool people by “haphazardly stitching together sequences of linguistic forms … according to probabilistic information about how they combine, but without any reference to meaning: a stochastic parrot.”

Google senior management wasn’t happy with the paper, and the resulting conflict led two of Bender’s coauthors, Timnit Gebru and Margaret Mitchell, to be forced out of the company, where they had led the AI Ethics team. It also made “stochastic parrot” a popular put-down for large language models—and landed Bender right in the middle of the name-calling merry-go-round.

The bottom line for Bender and for many like-minded researchers is that the field has been taken in by smoke and mirrors: “I think that they are led to imagine autonomous thinking entities that can make decisions for themselves and ultimately be the kind of thing that could actually be accountable for those decisions.”

Always the linguist, Bender is now at the point where she won’t even use the term AI “without scare quotes,” she tells me. Ultimately, for her, it’s a Big Tech buzzword that distracts from the many associated harms. “I’ve got skin in the game now,” she says. “I care about these issues, and the hype is getting in the way.”

Extraordinary evidence?

Agüera y Arcas calls people like Bender “AI denialists”—the implication being that they won’t ever accept what he takes for granted. Bender’s position is that extraordinary claims require extraordinary evidence, which we do not have.

But there are people looking for it, and until they find something clear-cut—sparks or stochastic parrots or something in between—they’d prefer to sit out the fight. Call this the wait-and-see camp.

As Ellie Pavlick, who studies neural networks at Brown University, tells me: “It’s offensive to some people to suggest that human intelligence could be re-created through these kinds of mechanisms.”

She adds, “People have strong-held beliefs about this issue—it almost feels religious. On the other hand, there’s people who have a little bit of a God complex. So it’s also offensive to them to suggest that they just can’t do it.”

Pavlick is ultimately agnostic. She’s a scientist, she insists, and will follow wherever the science leads. She rolls her eyes at the wilder claims, but she believes there’s something exciting going on. “That’s where I would disagree with Bender and Koller,” she tells me. “I think there’s actually some sparks—maybe not of AGI, but like, there’s some things in there that we didn’t expect to find.”

The problem is finding agreement on what those exciting things are and why they’re exciting. With so much hype, it’s easy to be cynical.

Researchers like Bubeck seem a lot more cool-headed when you hear them out. He thinks the infighting misses the nuance in his work. “I don’t see any problem in holding simultaneous views,” he says. “There is stochastic parroting; there is reasoning—it’s a spectrum. It’s very complex. We don’t have all the answers.”

“We need a completely new vocabulary to describe what’s going on,” he says. “One reason why people push back when I talk about reasoning in large language models is because it’s not the same reasoning as in human beings. But I think there is no way we can not call it reasoning. It is reasoning.”

Anthropic’s Olah plays it safe when pushed on what we’re seeing in LLMs, though his company, one of the hottest AI labs in the world right now, built Claude 3, an LLM that has received just as much hyperbolic praise as GPT-4 (if not more) since its release earlier this year.

“I feel like a lot of these conversations about the capabilities of these models are very tribal,” he says. “People have preexisting opinions, and it’s not very informed by evidence on any side. Then it just becomes kind of vibes-based, and I think vibes-based arguments on the internet tend to go in a bad direction.”

Olah tells me he has hunches of his own. “My subjective impression is that these things are tracking pretty sophisticated ideas,” he says. “We don’t have a comprehensive story of how very large models work, but I think it’s hard to reconcile what we’re seeing with the extreme ‘stochastic parrots’ picture.”

That’s as far as he’ll go: “I don’t want to go too much beyond what can be really strongly inferred from the evidence that we have.”

Last month, Anthropic released results from a study in which researchers gave Claude 3 the neural network equivalent of an MRI. By monitoring which bits of the model turned on and off as they ran it, they identified specific patterns of neurons that activated when the model was shown specific inputs.

Anthropic also reported patterns that it says correlate with inputs that attempt to describe or show abstract concepts. “We see features related to deception and honesty, to sycophancy, to security vulnerabilities, to bias,” says Olah. “We find features related to power seeking and manipulation and betrayal.”

ASK IT FOR A RECIPE pic.twitter.com/0ZM3uGRJi9 — heron (@iamaheron_) May 23, 2024

These results give one of the clearest looks yet at what’s inside a large language model. It’s a tantalizing glimpse at what look like elusive humanlike traits. But what does it really tell us? As Olah admits, they do not know what the model does with these patterns. “It’s a relatively limited picture, and the analysis is pretty hard,” he says.

Even if Olah won’t spell out exactly what he thinks goes on inside a large language model like Claude 3, it’s clear why the question matters to him. Anthropic is known for its work on AI safety—making sure that powerful future models will behave in ways we want them to and not in ways we don’t (known as “alignment” in industry jargon). Figuring out how today’s models work is not only a necessary first step if you want to control future ones; it also tells you how much you need to worry about doomer scenarios in the first place. “If you don’t think that models are going to be very capable,” says Olah, “then they’re probably not going to be very dangerous.”

Why we all can’t get along

In a 2014 interview with the BBC that looked back on her career, the influential cognitive scientist Margaret Boden, now 87, was asked if she thought there were any limits that would prevent computers (or “tin cans,” as she called them) from doing what humans can do.

“I certainly don’t think there’s anything in principle,” she said. “Because to deny that is to say that [human thinking] happens by magic, and I don’t believe that it happens by magic.”

But, she cautioned, powerful computers won’t be enough to get us there: the AI field will also need “powerful ideas”—new theories of how thinking happens, new algorithms that might reproduce it. “But these things are very, very difficult and I see no reason to assume that we will one of these days be able to answer all of those questions. Maybe we will; maybe we won’t.” 

Boden was reflecting on the early days of the current boom, but this will-we-or-won’t-we teetering speaks to decades in which she and her peers grappled with the same hard questions that researchers struggle with today. AI began as an ambitious aspiration 70-odd years ago and we are still disagreeing about what is and isn’t achievable, and how we’ll even know if we have achieved it. Most—if not all—of these disputes come down to this: We don’t have a good grasp on what intelligence is or how to recognize it. The field is full of hunches, but no one can say for sure.

We’ve been stuck on this point ever since people started taking the idea of AI seriously. Or even before that, when the stories we consumed started planting the idea of humanlike machines deep in our collective imagination. The long history of these disputes means that today’s fights often reinforce rifts that have been around since the beginning, making it even more difficult for people to find common ground.

To understand how we got here, we need to understand where we’ve been. So let’s dive into AI’s origin story—one that also played up the hype in a bid for cash.

A brief history of AI spin

The computer scientist John McCarthy is credited with coming up with the term “artificial intelligence” in 1955 when writing a funding application for a summer research program at Dartmouth College in New Hampshire.

The plan was for McCarthy and a small group of fellow researchers, a who’s-who of postwar US mathematicians and computer scientists—or “John McCarthy and the boys,” as Harry Law, a researcher who studies the history of AI at the University of Cambridge and ethics and policy at Google DeepMind, puts it—to get together for two months (not a typo) and make some serious headway on this new research challenge they’d set themselves.

“The study is to proceed on the basis of the conjecture that 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,” McCarthy and his coauthors wrote. “An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves.”

That list of things they wanted to make machines do—what Bender calls “the starry-eyed dream”—hasn’t changed much. Using language, forming concepts, and solving problems are defining goals for AI today. The hubris hasn’t changed much either: “We think that a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer,” they wrote. That summer, of course, has stretched to seven decades. And the extent to which these problems are in fact now solved is something that people still shout about on the internet. 

But what’s often left out of this canonical history is that artificial intelligence almost wasn’t called “artificial intelligence” at all.

More than one of McCarthy’s colleagues hated the term he had come up with. “The word ‘artificial’ makes you think there’s something kind of phony about this,” Arthur Samuel, a Dartmouth participant and creator of the first checkers-playing computer, is quoted as saying in historian Pamela McCorduck’s 2004 book Machines Who Think . The mathematician Claude Shannon, a coauthor of the Dartmouth proposal who is sometimes billed as “the father of the information age,” preferred the term “automata studies.” Herbert Simon and Allen Newell, two other AI pioneers, continued to call their own work “complex information processing” for years afterwards.

In fact, “artificial intelligence” was just one of several labels that might have captured the hodgepodge of ideas that the Dartmouth group was drawing on. The historian Jonnie Penn has identified possible alternatives that were in play at the time, including “engineering psychology,” “applied epistemology,” “neural cybernetics,” “non-numerical computing,” “neuraldynamics,” “advanced automatic programming,” and “hypothetical automata.” This list of names reveals how diverse the inspiration for their new field was, pulling from biology, neuroscience, statistics, and more. Marvin Minsky, another Dartmouth participant, has described AI as a “suitcase word” because it can hold so many divergent interpretations.

But McCarthy wanted a name that captured the ambitious scope of his vision. Calling this new field “artificial intelligence” grabbed people’s attention—and money. Don’t forget: AI is sexy, AI is cool.

In addition to terminology, the Dartmouth proposal codified a split between rival approaches to artificial intelligence that has divided the field ever since—a divide Law calls the “core tension in AI.”

McCarthy and his colleagues wanted to describe in computer code “every aspect of learning or any other feature of intelligence” so that machines could mimic them. In other words, if they could just figure out how thinking worked—the rules of reasoning—and write down the recipe, they could program computers to follow it. This laid the foundation of what came to be known as rule-based or symbolic AI (sometimes referred to now as GOFAI, “good old-fashioned AI”). But coming up with hard-coded rules that captured the processes of problem-solving for actual, nontrivial problems proved too hard.

The other path favored neural networks, computer programs that would try to learn those rules by themselves in the form of statistical patterns. The Dartmouth proposal mentions it almost as an aside (referring variously to “neuron nets” and “nerve nets”). Though the idea seemed less promising at first, some researchers nevertheless continued to work on versions of neural networks alongside symbolic AI. But it would take decades—plus vast amounts of computing power and much of the data on the internet—before they really took off. Fast-forward to today and this approach underpins the entire AI boom.

The big takeaway here is that, just like today’s researchers, AI’s innovators fought about foundational concepts and got caught up in their own promotional spin. Even team GOFAI was plagued by squabbles. Aaron Sloman, a philosopher and fellow AI pioneer now in his late 80s, recalls how “old friends” Minsky and McCarthy “disagreed strongly” when he got to know them in the ’70s: “Minsky thought McCarthy’s claims about logic could not work, and McCarthy thought Minsky’s mechanisms could not do what could be done using logic. I got on well with both of them, but I was saying, ‘Neither of you have got it right.’” (Sloman still thinks no one can account for the way human reasoning uses intuition as much as logic, but that’s yet another tangent!)

As the fortunes of the technology waxed and waned, the term “AI” went in and out of fashion. In the early ’70s, both research tracks were effectively put on ice after the UK government published a report arguing that the AI dream had gone nowhere and wasn’t worth funding. All that hype, effectively, had led to nothing. Research projects were shuttered, and computer scientists scrubbed the words “artificial intelligence” from their grant proposals.

When I was finishing a computer science PhD in 2008, only one person in the department was working on neural networks. Bender has a similar recollection: “When I was in college, a running joke was that AI is anything that we haven’t figured out how to do with computers yet. Like, as soon as you figure out how to do it, it wasn’t magic anymore, so it wasn’t AI.”

But that magic—the grand vision laid out in the Dartmouth proposal—remained alive and, as we can now see, laid the foundations for the AGI dream.

Good and bad behavior

In 1950, five years before McCarthy started talking about artificial intelligence, Alan Turing had published a paper that asked: Can machines think? To address that question, the famous mathematician proposed a hypothetical test, which he called the imitation game. The setup imagines a human and a computer behind a screen and a second human who types questions to each. If the questioner cannot tell which answers come from the human and which come from the computer, Turing claimed, the computer may as well be said to think.

What Turing saw—unlike McCarthy’s crew—was that thinking is a really difficult thing to describe. The Turing test was a way to sidestep that problem. “He basically said: Instead of focusing on the nature of intelligence itself, I’m going to look for its manifestation in the world. I’m going to look for its shadow ,” says Law.

In 1952, BBC Radio convened a panel to explore Turing’s ideas further. Turing was joined in the studio by two of his Manchester University colleagues—professor of mathematics Maxwell Newman and professor of neurosurgery Geoffrey Jefferson—and Richard Braithwaite, a philosopher of science, ethics, and religion at the University of Cambridge.

Braithwaite kicked things off: “Thinking is ordinarily regarded as so much the specialty of man, and perhaps of other higher animals, the question may seem too absurd to be discussed. But of course, it all depends on what is to be included in ‘thinking.’”

The panelists circled Turing’s question but never quite pinned it down.

When they tried to define what thinking involved, what its mechanisms were, the goalposts moved. “As soon as one can see the cause and effect working themselves out in the brain, one regards it as not being thinking but a sort of unimaginative donkey work,” said Turing.

Here was the problem: When one panelist proposed some behavior that might be taken as evidence of thought—reacting to a new idea with outrage, say—another would point out that a computer could be made to do it.

As Newman said, it would be easy enough to program a computer to print “I don’t like this new program.” But he admitted that this would be a trick.

Exactly, Jefferson said: He wanted a computer that would print “I don’t like this new program” because it didn’t like the new program. In other words, for Jefferson, behavior was not enough. It was the process leading to the behavior that mattered.

But Turing disagreed. As he had noted, uncovering a specific process—the donkey work, to use his phrase—did not pinpoint what thinking was either. So what was left?

“From this point of view, one might be tempted to define thinking as consisting of those mental processes that we don’t understand,” said Turing. “If this is right, then to make a thinking machine is to make one which does interesting things without our really understanding quite how it is done.”

It is strange to hear people grapple with these ideas for the first time. “The debate is prescient,” says Tomer Ullman, a cognitive scientist at Harvard University. “Some of the points are still alive—perhaps even more so. What they seem to be going round and round on is that the Turing test is first and foremost a behaviorist test.”

For Turing, intelligence was hard to define but easy to recognize. He proposed that the appearance of intelligence was enough—and said nothing about how that behavior should come about.

And yet most people, when pushed, will have a gut instinct about what is and isn’t intelligent. There are dumb ways and clever ways to come across as intelligent. In 1981, Ned Block, a philosopher at New York University, showed that Turing’s proposal fell short of those gut instincts. Because it said nothing of what caused the behavior, the Turing test can be beaten through trickery (as Newman had noted in the BBC broadcast).

“Could the issue of whether a machine in fact thinks or is intelligent depend on how gullible human interrogators tend to be?” asked Block. (Or as computer scientist Mark Reidl has remarked : “The Turing test is not for AI to pass but for humans to fail.”)

Imagine, Block said, a vast look-up table in which human programmers had entered all possible answers to all possible questions. Type a question into this machine, and it would look up a matching answer in its database and send it back. Block argued that anyone using this machine would judge its behavior to be intelligent: “But actually, the machine has the intelligence of a toaster,” he wrote. “All the intelligence it exhibits is that of its programmers.”

Block concluded that whether behavior is intelligent behavior is a matter of how it is produced, not how it appears. Block’s toasters, which became known as Blockheads, are one of the strongest counterexamples to the assumptions behind Turing’s proposal.

Looking under the hood

The Turing test is not meant to be a practical metric, but its implications are deeply ingrained in the way we think about artificial intelligence today. This has become particularly relevant as LLMs have exploded in the past several years. These models get ranked by their outward behaviors, specifically how well they do on a range of tests. When OpenAI announced GPT-4, it published an impressive-looking scorecard that detailed the model’s performance on multiple high school and professional exams. Almost nobody talks about how these models get those results.

That’s because we don’t know. Today’s large language models are too complex for anybody to say exactly how their behavior is produced. Researchers outside the small handful of companies making those models don’t know what’s in their training data; none of the model makers have shared details. That makes it hard to say what is and isn’t a kind of memorization—a stochastic parroting. But even researchers on the inside, like Olah, don’t know what’s really going on when faced with a bridge-obsessed bot.

This leaves the question wide open: Yes, large language models are built on math—but are they doing something intelligent with it?

And the arguments begin again.

“Most people are trying to armchair through it,” says Brown University’s Pavlick, meaning that they are arguing about theories without looking at what’s really happening. “Some people are like, ‘I think it’s this way,’ and some people are like, ‘Well, I don’t.’ We’re kind of stuck and everyone’s unsatisfied.”

Bender thinks that this sense of mystery plays into the mythmaking. (“Magicians do not explain their tricks,” she says.) Without a proper appreciation of where the LLM’s words come from, we fall back on familiar assumptions about humans, since that is our only real point of reference. When we talk to another person, we try to make sense of what that person is trying to tell us. “That process necessarily entails imagining a life behind the words,” says Bender. That’s how language works.

“The parlor trick of ChatGPT is so impressive that when we see these words coming out of it, we do the same thing instinctively,” she says. “It’s very good at mimicking the form of language. The problem is that we are not at all good at encountering the form of language and not imagining the rest of it.”

For some researchers, it doesn’t really matter if we can’t understand the how . Bubeck used to study large language models to try to figure out how they worked, but GPT-4 changed the way he thought about them. “It seems like these questions are not so relevant anymore,” he says. “The model is so big, so complex, that we can’t hope to open it up and understand what’s really happening.”

But Pavlick, like Olah, is trying to do just that. Her team has found that models seem to encode abstract relationships between objects, such as that between a country and its capital. Studying one large language model, Pavlick and her colleagues found that it used the same encoding to map France to Paris and Poland to Warsaw. That almost sounds smart, I tell her. “No, it’s literally a lookup table,” she says.

But what struck Pavlick was that, unlike a Blockhead, the model had learned this lookup table on its own. In other words, the LLM figured out itself that Paris is to France as Warsaw is to Poland. But what does this show? Is encoding its own lookup table instead of using a hard-coded one a sign of intelligence? Where do you draw the line?

“Basically, the problem is that behavior is the only thing we know how to measure reliably,” says Pavlick. “Anything else requires a theoretical commitment, and people don’t like having to make a theoretical commitment because it’s so loaded.”

Not all people. A lot of influential scientists are just fine with theoretical commitment. Hinton, for example, insists that neural networks are all you need to re-create humanlike intelligence. “Deep learning is going to be able to do everything,” he told MIT Technology Review in 2020 . 

It’s a commitment that Hinton seems to have held onto from the start. Sloman, who recalls the two of them arguing when Hinton was a graduate student in his lab, remembers being unable to persuade him that neural networks cannot learn certain crucial abstract concepts that humans and some other animals seem to have an intuitive grasp of, such as whether something is impossible. We can just see when something’s ruled out, Sloman says. “Despite Hinton’s outstanding intelligence, he never seemed to understand that point. I don’t know why, but there are large numbers of researchers in neural networks who share that failing.”

And then there’s Marcus, whose view of neural networks is the exact opposite of Hinton’s. His case draws on what he says scientists have discovered about brains.

Brains, Marcus points out, are not blank slates that learn fully from scratch—they come ready-made with innate structures and processes that guide learning. It’s how babies can learn things that the best neural networks still can’t, he argues.

“Neural network people have this hammer, and now everything is a nail,” says Marcus. “They want to do all of it with learning, which many cognitive scientists would find unrealistic and silly. You’re not going to learn everything from scratch.”

Not that Marcus—a cognitive scientist—is any less sure of himself. “If one really looked at who’s predicted the current situation well, I think I would have to be at the top of anybody’s list,” he tells me from the back of an Uber on his way to catch a flight to a speaking gig in Europe. “I know that doesn’t sound very modest, but I do have this perspective that turns out to be very important if what you’re trying to study is artificial intelligence.”

Given his well-publicized attacks on the field, it might surprise you that Marcus still believes AGI is on the horizon. It’s just that he thinks today’s fixation on neural networks is a mistake. “We probably need a breakthrough or two or four,” he says. “You and I might not live that long, I’m sorry to say. But I think it’ll happen this century. Maybe we’ve got a shot at it.”

The power of a technicolor dream

Over Dor Skuler’s shoulder on the Zoom call from his home in Ramat Gan, Israel, a little lamp-like robot is winking on and off while we talk about it. “You can see ElliQ behind me here,” he says. Skuler’s company, Intuition Robotics, develops these devices for older people, and the design—part Amazon Alexa, part R2-D2—must make it very clear that ElliQ is a computer. If any of his customers show signs of being confused about that, Intuition Robotics takes the device back, says Skuler.

ElliQ has no face, no humanlike shape at all. Ask it about sports, and it will crack a joke about having no hand-eye coordination because it has no hands and no eyes. “For the life of me, I don’t understand why the industry is trying to fulfill the Turing test,” Skuler says. “Why is it in the best interest of humanity for us to develop technology whose goal is to dupe us?”

Instead, Skuler’s firm is betting that people can form relationships with machines that present as machines. “Just like we have the ability to build a real relationship with a dog,” he says. “Dogs provide a lot of joy for people. They provide companionship. People love their dog—but they never confuse it to be a human.”

ElliQ’s users, many in their 80s and 90s, refer to the robot as an entity or a presence—sometimes a roommate. “They’re able to create a space for this in-between relationship, something between a device or a computer and something that’s alive,” says Skuler.

But no matter how hard ElliQ’s designers try to control the way people view the device, they are competing with decades of pop culture that have shaped our expectations. Why are we so fixated on AI that’s humanlike? “Because it’s hard for us to imagine something else,” says Skuler (who indeed refers to ElliQ as “she” throughout our conversation). “And because so many people in the tech industry are fans of science fiction. They try to make their dream come true.”

How many developers grew up today thinking that building a smart machine was seriously the coolest thing—if not the most important thing—that they could possibly do?

It was not long ago that OpenAI launched its new voice-controlled version of ChatGPT with a voice that sounded like Scarlett Johansson, after which many people—including Altman—flagged the connection to Spike Jonze’s 2013 movie Her .

Science fiction co-invents what AI is understood to be. As Cave and Dihal write in Imagining AI : “AI was a cultural phenomenon long before it was a technological one.”

Stories and myths about remaking humans as machines have been around for centuries. People have been dreaming of artificial humans for probably as long as they have dreamed of flight, says Dihal. She notes that Daedalus, the figure in Greek mythology famous for building a pair of wings for himself and his son, Icarus, also built what was effectively a giant bronze robot called Talos that threw rocks at passing pirates.

The word robot comes from robota , a term for “forced labor” coined by the Czech playwright Karel Čapek in his 1920 play Rossum’s Universal Robots . The “laws of robotics” outlined in Isaac Asimov’s science fiction, forbidding machines from harming humans, are inverted by movies like The Terminator , which is an iconic reference point for popular fears about real-world technology. The 2014 film Ex Machina is a dramatic riff on the Turing test. Last year’s blockbuster The Creator imagines a future world in which AI has been outlawed because it set off a nuclear bomb, an event that some doomers consider at least an outside possibility.

Cave and Dihal relate how another movie, 2014’s Transcendence , in which an AI expert played by Johnny Depp gets his mind uploaded to a computer, served a narrative pushed by ur-doomers Stephen Hawking, fellow physicist Max Tegmark, and AI researcher Stuart Russell. In an article published in the Huffington Post on the movie’s opening weekend, the trio wrote: “As the Hollywood blockbuster Transcendence debuts this weekend with … clashing visions for the future of humanity, it’s tempting to dismiss the notion of highly intelligent machines as mere science fiction. But this would be a mistake, and potentially our worst mistake ever.”

Right around the same time, Tegmark founded the Future of Life Institute, with a remit to study and promote AI safety. Depp’s costar in the movie, Morgan Freeman, was on the institute’s board, and Elon Musk, who had a cameo in the film, donated $10 million in its first year. For Cave and Dihal, Transcendence is a perfect example of the multiple entanglements between popular culture, academic research, industrial production, and “the billionaire-funded fight to shape the future.”

On the London leg of his world tour last year, Altman was asked what he’d meant when he tweeted : “AI is the tech the world has always wanted.” Standing at the back of the room that day, behind an audience of hundreds, I listened to him offer his own kind of origin story: “I was, like, a very nervous kid. I read a lot of sci-fi. I spent a lot of Friday nights home, playing on the computer. But I was always really interested in AI and I thought it’d be very cool.” He went to college, got rich, and watched as neural networks became better and better. “This can be tremendously good but also could be really bad. What are we going to do about that?” he recalled thinking in 2015. “I ended up starting OpenAI.”

Why you should care that a bunch of nerds are fighting about AI

Okay, you get it: No one can agree on what AI is. But what everyone does seem to agree on is that the current debate around AI has moved far beyond the academic and the scientific. There are political and moral components in play—which doesn’t help with everyone thinking everyone else is wrong.

Untangling this is hard. It can be difficult to see what’s going on when some of those moral views take in the entire future of humanity and anchor them in a technology that nobody can quite define.

But we can't just throw our hands up and walk away. Because no matter what this technology is, it’s coming, and unless you live under a rock, you’ll use it in one form or another. And the form that technology takes—and the problems it both solves and creates—will be shaped by the thinking and the motivations of people like the ones you just read about. In particular, by the people with the most power, the most cash, and the biggest megaphones.

Which leads me to the TESCREALists. Wait, come back! I realize it’s unfair to introduce yet another new concept so late in the game. But to understand how the people in power may mold the technologies they build, and how they explain them to the world’s regulators and lawmakers, you need to really understand their mindset.

Gebru, who founded the Distributed AI Research Institute after leaving Google, and Émile Torres, a philosopher and historian at Case Western Reserve University, have traced the influence of several techno-utopian belief systems on Silicon Valley. The pair argue that to understand what’s going on with AI right now—both why companies such as Google DeepMind and OpenAI are in a race to build AGI and why doomers like Tegmark and Hinton warn of a coming catastrophe—the field must be seen through the lens of what Torres has dubbed the TESCREAL framework .

The clunky acronym (pronounced tes-cree-all ) replaces an even clunkier list of labels: transhumanism , extropianism , singularitarianism , cosmism , rationalism , effective altruism , and longtermism . A lot has been written (and will be written) about each of these worldviews, so I’ll spare you here. (There are rabbit holes within rabbit holes for anyone wanting to dive deeper. Pick your forum and pack your spelunking gear.)

This constellation of overlapping ideologies is attractive to a certain kind of galaxy-brain mindset common in the Western tech world. Some anticipate human immortality; others predict humanity’s colonization of the stars. The common tenet is that an all-powerful technology—AGI or superintelligence, choose your team—is not only within reach but inevitable. You can see this in the do-or-die attitude that’s ubiquitous inside cutting-edge labs like OpenAI: If we don’t make AGI, someone else will.

What’s more, TESCREALists believe that AGI could not only fix the world’s problems but level up humanity. “The development and proliferation of AI—far from a risk that we should fear—is a moral obligation that we have to ourselves, to our children and to our future,” Andreessen wrote in a much-dissected manifesto last year. I have been told many times over that AGI is the way to make the world a better place—by Demis Hassabis , CEO and cofounder of Google DeepMind; by Mustafa Suleyman , CEO of the newly minted Microsoft AI and another cofounder of DeepMind; by Sutskever , Altman , and more.

But as Andreessen noted, it’s a yin-yang mindset. The flip side of techno-utopia is techno-hell. If you believe that you are building a technology so powerful that it will solve all the world’s problems, you probably also believe there’s a non-zero chance it will all go very wrong. When asked at the World Government Summit in February what keeps him up at night, Altman replied: “It’s all the sci-fi stuff.”

It’s a tension that Hinton has been talking up for the last year. It’s what companies like Anthropic claim to address. It’s what Sutskever is focusing on in his new lab , and what he wanted a special in-house team at OpenAI to focus on last year before disagreements over the way the company balanced risk and reward led most members of that team to leave.

Sure, doomerism is part of the spin. (“Claiming that you have created something that is super-intelligent is good for sales figures,” says Dihal. “It’s like, ‘Please, someone stop me from being so good and so powerful.’”) But boom or doom, exactly what (and whose) problems are these guys supposedly solving? Are we really expected to trust what they build and what they tell our leaders?

Gebru and Torres (and others) are adamant: No, we should not. They are highly critical of these ideologies and how they may influence the development of future technology, especially AI. Fundamentally, they link several of these worldviews—with their common focus on “improving” humanity—to the racist eugenics movements of the 20th century.

One danger, they argue, is that a shift of resources toward the kind of technological innovations that these ideologies demand, from building AGI to extending life spans to colonizing other planets, will ultimately benefit people who are Western and white at the cost of billions of people who aren’t. If your sight is set on fantastical futures, it’s easy to overlook the present-day costs of innovation, such as labor exploitation, the entrenchment of racist and sexist bias, and environmental damage.  

“Are we trying to build a tool that’s useful to us in some way?” asks Bender, reflecting on the casualties of this race to AGI. If so, who’s it for, how do we test it, how well does it work? “But if what we’re building it for is just so that we can say that we’ve done it, that’s not a goal that I can get behind. That’s not a goal that’s worth billions of dollars.”

Bender says that seeing the connections between the TESCREAL ideologies is what made her realize there was something more to these debates. “Tangling with those people was—” she stops. “Okay, there’s more here than just academic ideas. There’s a moral code tied up in it as well.”

Of course, laid out like this without nuance, it doesn’t sound as if we—as a society, as individuals—are getting the best deal. It also all sounds rather silly. When Gebru described parts of the TESCREAL bundle in a talk last year, her audience laughed. It’s also true that few people would identify themselves as card-carrying students of these schools of thought, at least in their extremes.

But if we don’t understand how those building this tech approach it, how can we decide what deals we want to make? What apps we decide to use, what chatbots we want to give personal information to, what data centers we support in our neighborhoods, what politicians we want to vote for?

It used to be like this: There was a problem in the world, and we built something to fix it. Here, everything is backward: The goal seems to be to build a machine that can do everything, and to skip the slow, hard work that goes into figuring out what the problem is before building the solution.

And as Gebru said in that same talk, “A machine that solves all problems: if that’s not magic, what is it?”

Semantics, semantics … semantics?

When asked outright what AI is, a lot of people dodge the question. Not Suleyman. In April, the CEO of Microsoft AI stood on the TED stage and told the audience what he’d told his six-year-old nephew in response to that question. The best answer he could give, Suleyman explained, was that AI was “a new kind of digital species”—a technology so universal, so powerful, that calling it a tool no longer captured what it could do for us.

“On our current trajectory, we are heading toward the emergence of something we are all struggling to describe, and yet we cannot control what we don’t understand,” he said. “And so the metaphors, the mental models, the names—these all matter if we are to get the most out of AI whilst limiting its potential downsides.”

Language matters! I hope that’s clear from the twists and turns and tantrums we’ve been through to get to this point. But I also hope you’re asking: Whose language? And whose downsides? Suleyman is an industry leader at a technology giant that stands to make billions from its AI products. Describing the technology behind those products as a new kind of species conjures something wholly unprecedented, something with agency and capabilities that we have never seen before. That makes my spidey sense tingle. You?

I can’t tell you if there’s magic here (ironically or not). And I can’t tell you how math can realize what Bubeck and many others see in this technology (no one can yet). You’ll have to make up your own mind. But I can pull back the curtain on my own point of view.

Writing about GPT-3 back in 2020, I said that the greatest trick AI ever pulled was convincing the world it exists. I still think that: We are hardwired to see intelligence in things that behave in certain ways, whether it’s there or not. In the last few years, the tech industry has found reasons of its own to convince us that AI exists, too. This makes me skeptical of many of the claims made for this technology.

With large language models—via their smiley-face masks—we are confronted by something we’ve never had to think about before. “It’s taking this hypothetical thing and making it really concrete,” says Pavlick. “I’ve never had to think about whether a piece of language required intelligence to generate because I’ve just never dealt with language that didn’t.”

AI is many things. But I don’t think it’s humanlike. I don’t think it’s the solution to all (or even most) of our problems. It isn’t ChatGPT or Gemini or Copilot. It isn’t neural networks. It’s an idea, a vision, a kind of wish fulfillment. And ideas get shaped by other ideas, by morals, by quasi-religious convictions, by worldviews, by politics, and by gut instinct. “Artificial intelligence” is a helpful shorthand to describe a raft of different technologies. But AI is not one thing; it never has been, no matter how often the branding gets seared into the outside of the box. 

“The truth is these words”—intelligence, reasoning, understanding, and more—“were defined before there was a need to be really precise about it,” says Pavlick. “I don’t really like when the question becomes ‘Does the model understand—yes or no?’ because, well, I don’t know. Words get redefined and concepts evolve all the time.”

Artificial intelligence

What are AI agents? 

The next big thing is AI tools that can do more complex tasks. Here’s how they will work.

• Melissa Heikkilä archive page

Google DeepMind’s new AI systems can now solve complex math problems

AlphaProof and AlphaGeometry 2 are steps toward building systems that can reason, which could unlock exciting new capabilities.

• Rhiannon Williams archive page

Robot-packed meals are coming to the frozen-food aisle

Found everywhere from airplanes to grocery stores, prepared meals are usually packed by hand. AI-powered robotics is changing that.

• James O'Donnell archive page

OpenAI has released a new ChatGPT bot that you can talk to

The voice-enabled chatbot will be available to a small group of people today, and to all ChatGPT Plus users in the fall. 

Stay connected

Get the latest updates from mit technology review.

Discover special offers, top stories, upcoming events, and more.

Thank you for submitting your email!

It looks like something went wrong.

We’re having trouble saving your preferences. Try refreshing this page and updating them one more time. If you continue to get this message, reach out to us at [email protected] with a list of newsletters you’d like to receive.

What is AI (artificial intelligence)?

Humans and machines: a match made in productivity  heaven. Our species wouldn’t have gotten very far without our mechanized workhorses. From the wheel that revolutionized agriculture to the screw that held together increasingly complex construction projects to the robot-enabled assembly lines of today, machines have made life as we know it possible. And yet, despite their seemingly endless utility, humans have long feared machines—more specifically, the possibility that machines might someday acquire human intelligence  and strike out on their own.

Get to know and directly engage with senior McKinsey experts on AI

Sven Blumberg is a senior partner in McKinsey’s Düsseldorf office; Michael Chui is a partner at the McKinsey Global Institute and is based in the Bay Area office, where Lareina Yee is a senior partner; Kia Javanmardian is a senior partner in the Chicago office, where Alex Singla , the global leader of QuantumBlack, AI by McKinsey, is also a senior partner; Kate Smaje and Alex Sukharevsky are senior partners in the London office.

But we tend to view the possibility of sentient machines with fascination as well as fear. This curiosity has helped turn science fiction into actual science. Twentieth-century theoreticians, like computer scientist and mathematician Alan Turing, envisioned a future where machines could perform functions faster than humans. The work of Turing and others soon made this a reality. Personal calculators became widely available in the 1970s, and by 2016, the US census showed that 89 percent of American households had a computer. Machines— smart machines at that—are now just an ordinary part of our lives and culture.

Those smart machines are also getting faster and more complex. Some computers have now crossed the exascale threshold, meaning they can perform as many calculations in a single second as an individual could in 31,688,765,000 years . And beyond computation, which machines have long been faster at than we have, computers and other devices are now acquiring skills and perception that were once unique to humans and a few other species.

About QuantumBlack, AI by McKinsey

QuantumBlack, McKinsey’s AI arm, helps companies transform using the power of technology, technical expertise, and industry experts. With thousands of practitioners at QuantumBlack (data engineers, data scientists, product managers, designers, and software engineers) and McKinsey (industry and domain experts), we are working to solve the world’s most important AI challenges. QuantumBlack Labs is our center of technology development and client innovation, which has been driving cutting-edge advancements and developments in AI through locations across the globe.

AI is a machine’s ability to perform the cognitive functions we associate with human minds, such as perceiving, reasoning, learning, interacting with the environment, problem-solving, and even exercising creativity. You’ve probably interacted with AI even if you don’t realize it—voice assistants like Siri and Alexa are founded on AI technology, as are some customer service chatbots that pop up to help you navigate websites.

Applied AI —simply, artificial intelligence applied to real-world problems—has serious implications for the business world. By using artificial intelligence, companies have the potential to make business more efficient and profitable. But ultimately, the value of AI isn’t in the systems themselves. Rather, it’s in how companies use these systems to assist humans—and their ability to explain to shareholders and the public what these systems do—in a way that builds trust and confidence.

For more about AI, its history, its future, and how to apply it in business, read on.

Learn more about QuantumBlack, AI by McKinsey .

Looking for direct answers to other complex questions?

What is machine learning.

Machine learning is a form of artificial intelligence that can adapt to a wide range of inputs, including large sets of historical data, synthesized data, or human inputs. (Some machine learning algorithms are specialized in training themselves to detect patterns; this is called deep learning . See Exhibit 1.) These algorithms can detect patterns and learn how to make predictions and recommendations by processing data, rather than by receiving explicit programming instruction. Some algorithms can also adapt in response to new data and experiences to improve over time.

The volume and complexity of data that is now being generated, too vast for humans to process and apply efficiently, has increased the potential of machine learning, as well as the need for it. In the years since its widespread deployment, which began in the 1970s, machine learning has had an impact on a number of industries, including achievements in medical-imaging analysis  and high-resolution weather forecasting.

The volume and complexity of data that is now being generated, too vast for humans to process and apply efficiently, has increased the potential of machine learning, as well as the need for it.

What is deep learning?

Deep learning is a more advanced version of machine learning that is particularly adept at processing a wider range of data resources (text as well as unstructured data including images), requires even less human intervention, and can often produce more accurate results than traditional machine learning. Deep learning uses neural networks—based on the ways neurons interact in the human brain —to ingest data and process it through multiple neuron layers that recognize increasingly complex features of the data. For example, an early layer might recognize something as being in a specific shape; building on this knowledge, a later layer might be able to identify the shape as a stop sign. Similar to machine learning, deep learning uses iteration to self-correct and improve its prediction capabilities. For example, once it “learns” what a stop sign looks like, it can recognize a stop sign in a new image.

What is generative AI?

Case study: vistra and the martin lake power plant.

Vistra is a large power producer in the United States, operating plants in 12 states with a capacity to power nearly 20 million homes. Vistra has committed to achieving net-zero emissions by 2050. In support of this goal, as well as to improve overall efficiency, QuantumBlack, AI by McKinsey worked with Vistra to build and deploy an AI-powered heat rate optimizer (HRO) at one of its plants.

“Heat rate” is a measure of the thermal efficiency of the plant; in other words, it’s the amount of fuel required to produce each unit of electricity. To reach the optimal heat rate, plant operators continuously monitor and tune hundreds of variables, such as steam temperatures, pressures, oxygen levels, and fan speeds.

Vistra and a McKinsey team, including data scientists and machine learning engineers, built a multilayered neural network model. The model combed through two years’ worth of data at the plant and learned which combination of factors would attain the most efficient heat rate at any point in time. When the models were accurate to 99 percent or higher and run through a rigorous set of real-world tests, the team converted them into an AI-powered engine that generates recommendations every 30 minutes for operators to improve the plant’s heat rate efficiency. One seasoned operations manager at the company’s plant in Odessa, Texas, said, “There are things that took me 20 years to learn about these power plants. This model learned them in an afternoon.”

Overall, the AI-powered HRO helped Vistra achieve the following:

• approximately 1.6 million metric tons of carbon abated annually
• 67 power generators optimized
• $60 million saved in about a year

Read more about the Vistra story here .

Generative AI (gen AI) is an AI model that generates content in response to a prompt. It’s clear that generative AI tools like ChatGPT and DALL-E (a tool for AI-generated art) have the potential to change how a range of jobs  are performed. Much is still unknown about gen AI’s potential, but there are some questions we can answer—like how gen AI models are built, what kinds of problems they are best suited to solve, and how they fit into the broader category of AI and machine learning.

For more on generative AI and how it stands to affect business and society, check out our Explainer “ What is generative AI? ”

What is the history of AI?

The term “artificial intelligence” was coined in 1956  by computer scientist John McCarthy for a workshop at Dartmouth. But he wasn’t the first to write about the concepts we now describe as AI. Alan Turing introduced the concept of the “ imitation game ” in a 1950 paper. That’s the test of a machine’s ability to exhibit intelligent behavior, now known as the “Turing test.” He believed researchers should focus on areas that don’t require too much sensing and action, things like games and language translation. Research communities dedicated to concepts like computer vision, natural language understanding, and neural networks are, in many cases, several decades old.

MIT physicist Rodney Brooks shared details on the four previous stages of AI:

Symbolic AI (1956). Symbolic AI is also known as classical AI, or even GOFAI (good old-fashioned AI). The key concept here is the use of symbols and logical reasoning to solve problems. For example, we know a German shepherd is a dog , which is a mammal; all mammals are warm-blooded; therefore, a German shepherd should be warm-blooded.

The main problem with symbolic AI is that humans still need to manually encode their knowledge of the world into the symbolic AI system, rather than allowing it to observe and encode relationships on its own. As a result, symbolic AI systems struggle with situations involving real-world complexity. They also lack the ability to learn from large amounts of data.

Symbolic AI was the dominant paradigm of AI research until the late 1980s.

Neural networks (1954, 1969, 1986, 2012). Neural networks are the technology behind the recent explosive growth of gen AI. Loosely modeling the ways neurons interact in the human brain , neural networks ingest data and process it through multiple iterations that learn increasingly complex features of the data. The neural network can then make determinations about the data, learn whether a determination is correct, and use what it has learned to make determinations about new data. For example, once it “learns” what an object looks like, it can recognize the object in a new image.

Neural networks were first proposed in 1943 in an academic paper by neurophysiologist Warren McCulloch and logician Walter Pitts. Decades later, in 1969, two MIT researchers mathematically demonstrated that neural networks could perform only very basic tasks. In 1986, there was another reversal, when computer scientist and cognitive psychologist Geoffrey Hinton and colleagues solved the neural network problem presented by the MIT researchers. In the 1990s, computer scientist Yann LeCun made major advancements in neural networks’ use in computer vision, while Jürgen Schmidhuber advanced the application of recurrent neural networks as used in language processing.

In 2012, Hinton and two of his students highlighted the power of deep learning. They applied Hinton’s algorithm to neural networks with many more layers than was typical, sparking a new focus on deep neural networks. These have been the main AI approaches of recent years.

Traditional robotics (1968). During the first few decades of AI, researchers built robots to advance research. Some robots were mobile, moving around on wheels, while others were fixed, with articulated arms. Robots used the earliest attempts at computer vision to identify and navigate through their environments or to understand the geometry of objects and maneuver them. This could include moving around blocks of various shapes and colors. Most of these robots, just like the ones that have been used in factories for decades, rely on highly controlled environments with thoroughly scripted behaviors that they perform repeatedly. They have not contributed significantly to the advancement of AI itself.

But traditional robotics did have significant impact in one area, through a process called “simultaneous localization and mapping” (SLAM). SLAM algorithms helped contribute to self-driving cars and are used in consumer products like vacuum cleaning robots and quadcopter drones. Today, this work has evolved into behavior-based robotics, also referred to as haptic technology because it responds to human touch.

• Behavior-based robotics (1985). In the real world, there aren’t always clear instructions for navigation, decision making, or problem-solving. Insects, researchers observed, navigate very well (and are evolutionarily very successful) with few neurons. Behavior-based robotics researchers took inspiration from this, looking for ways robots could solve problems with partial knowledge and conflicting instructions. These behavior-based robots are embedded with neural networks.

Learn more about  QuantumBlack, AI by McKinsey .

What is artificial general intelligence?

The term “artificial general intelligence” (AGI) was coined to describe AI systems that possess capabilities comparable to those of a human . In theory, AGI could someday replicate human-like cognitive abilities including reasoning, problem-solving, perception, learning, and language comprehension. But let’s not get ahead of ourselves: the key word here is “someday.” Most researchers and academics believe we are decades away from realizing AGI; some even predict we won’t see AGI this century, or ever. Rodney Brooks, an MIT roboticist and cofounder of iRobot, doesn’t believe AGI will arrive until the year 2300 .

The timing of AGI’s emergence may be uncertain. But when it does emerge—and it likely will—it’s going to be a very big deal, in every aspect of our lives. Executives should begin working to understand the path to machines achieving human-level intelligence now and making the transition to a more automated world.

For more on AGI, including the four previous attempts at AGI, read our Explainer .

What is narrow AI?

Narrow AI is the application of AI techniques to a specific and well-defined problem, such as chatbots like ChatGPT, algorithms that spot fraud in credit card transactions, and natural-language-processing engines that quickly process thousands of legal documents. Most current AI applications fall into the category of narrow AI. AGI is, by contrast, AI that’s intelligent enough to perform a broad range of tasks.

How is the use of AI expanding?

AI is a big story for all kinds of businesses, but some companies are clearly moving ahead of the pack . Our state of AI in 2022 survey showed that adoption of AI models has more than doubled since 2017—and investment has increased apace. What’s more, the specific areas in which companies see value from AI have evolved, from manufacturing and risk to the following:

• marketing and sales
• product and service development
• strategy and corporate finance

One group of companies is pulling ahead of its competitors. Leaders of these organizations consistently make larger investments in AI, level up their practices to scale faster, and hire and upskill the best AI talent. More specifically, they link AI strategy to business outcomes and “ industrialize ” AI operations by designing modular data architecture that can quickly accommodate new applications.

What are the limitations of AI models? How can these potentially be overcome?

We have yet to see the longtail effect of gen AI models. This means there are some inherent risks involved in using them—both known and unknown.

The outputs gen AI models produce may often sound extremely convincing. This is by design. But sometimes the information they generate is just plain wrong. Worse, sometimes it’s biased (because it’s built on the gender, racial, and other biases of the internet and society more generally).

It can also be manipulated to enable unethical or criminal activity. Since gen AI models burst onto the scene, organizations have become aware of users trying to “jailbreak” the models—that means trying to get them to break their own rules and deliver biased, harmful, misleading, or even illegal content. Gen AI organizations are responding to this threat in two ways: for one thing, they’re collecting feedback from users on inappropriate content. They’re also combing through their databases, identifying prompts that led to inappropriate content, and training the model against these types of generations.

But awareness and even action don’t guarantee that harmful content won’t slip the dragnet. Organizations that rely on gen AI models should be aware of the reputational and legal risks involved in unintentionally publishing biased, offensive, or copyrighted content.

These risks can be mitigated, however, in a few ways. “Whenever you use a model,” says McKinsey partner Marie El Hoyek, “you need to be able to counter biases  and instruct it not to use inappropriate or flawed sources, or things you don’t trust.” How? For one thing, it’s crucial to carefully select the initial data used to train these models to avoid including toxic or biased content. Next, rather than employing an off-the-shelf gen AI model, organizations could consider using smaller, specialized models. Organizations with more resources could also customize a general model based on their own data to fit their needs and minimize biases.

It’s also important to keep a human in the loop (that is, to make sure a real human checks the output of a gen AI model before it is published or used) and avoid using gen AI models for critical decisions, such as those involving significant resources or human welfare.

It can’t be emphasized enough that this is a new field. The landscape of risks and opportunities is likely to continue to change rapidly in the coming years. As gen AI becomes increasingly incorporated into business, society, and our personal lives, we can also expect a new regulatory climate to take shape. As organizations experiment—and create value—with these tools, leaders will do well to keep a finger on the pulse of regulation and risk.

What is the AI Bill of Rights?

The Blueprint for an AI Bill of Rights, prepared by the US government in 2022, provides a framework for how government, technology companies, and citizens can collectively ensure more accountable AI. As AI has become more ubiquitous, concerns have surfaced  about a potential lack of transparency surrounding the functioning of gen AI systems, the data used to train them, issues of bias and fairness, potential intellectual property infringements, privacy violations, and more. The Blueprint comprises five principles that the White House says should “guide the design, use, and deployment of automated systems to protect [users] in the age of artificial intelligence.” They are as follows:

• The right to safe and effective systems. Systems should undergo predeployment testing, risk identification and mitigation, and ongoing monitoring to demonstrate that they are adhering to their intended use.
• Protections against discrimination by algorithms. Algorithmic discrimination is when automated systems contribute to unjustified different treatment of people based on their race, color, ethnicity, sex, religion, age, and more.
• Protections against abusive data practices, via built-in safeguards. Users should also have agency over how their data is used.
• The right to know that an automated system is being used, and a clear explanation of how and why it contributes to outcomes that affect the user.
• The right to opt out, and access to a human who can quickly consider and fix problems.

At present, more than 60 countries or blocs have national strategies governing the responsible use of AI (Exhibit 2). These include Brazil, China, the European Union, Singapore, South Korea, and the United States. The approaches taken vary from guidelines-based approaches, such as the Blueprint for an AI Bill of Rights in the United States, to comprehensive AI regulations that align with existing data protection and cybersecurity regulations, such as the EU’s AI Act, due in 2024.

There are also collaborative efforts between countries to set out standards for AI use. The US–EU Trade and Technology Council is working toward greater alignment between Europe and the United States. The Global Partnership on Artificial Intelligence, formed in 2020, has 29 members including Brazil, Canada, Japan, the United States, and several European countries.

Even though AI regulations are still being developed, organizations should act now to avoid legal, reputational, organizational, and financial risks. In an environment of public concern, a misstep could be costly. Here are four no-regrets, preemptive actions organizations can implement today:

• Transparency. Create an inventory of models, classifying them in accordance with regulation, and record all usage across the organization that is clear to those inside and outside the organization.
• Governance. Implement a governance structure for AI and gen AI that ensures sufficient oversight, authority, and accountability both within the organization and with third parties and regulators.
• Data management. Proper data management includes awareness of data sources, data classification, data quality and lineage, intellectual property, and privacy management.
• Model management. Organizations should establish principles and guardrails for AI development and use them to ensure all AI models uphold fairness and bias controls.
• Cybersecurity and technology management. Establish strong cybersecurity and technology to ensure a secure environment where unauthorized access or misuse is prevented.
• Individual rights. Make users aware when they are interacting with an AI system, and provide clear instructions for use.

How can organizations scale up their AI efforts from ad hoc projects to full integration?

Most organizations are dipping a toe into the AI pool—not cannonballing. Slow progress toward widespread adoption is likely due to cultural and organizational barriers. But leaders who effectively break down these barriers will be best placed to capture the opportunities of the AI era. And—crucially—companies that can’t take full advantage of AI are already being sidelined by those that can, in industries like auto manufacturing and financial services.

To scale up AI, organizations can make three major shifts :

• Move from siloed work to interdisciplinary collaboration. AI projects shouldn’t be limited to discrete pockets of organizations. Rather, AI has the biggest impact when it’s employed by cross-functional teams with a mix of skills and perspectives, enabling AI to address broad business priorities.
• Empower frontline data-based decision making . AI has the potential to enable faster, better decisions at all levels of an organization. But for this to work, people at all levels need to trust the algorithms’ suggestions and feel empowered to make decisions. (Equally, people should be able to override the algorithm or make suggestions for improvement when necessary.)
• Adopt and bolster an agile mindset. The agile test-and-learn mindset will help reframe mistakes as sources of discovery, allaying the fear of failure and speeding up development.

Learn more about QuantumBlack, AI by McKinsey , and check out AI-related job opportunities if you’re interested in working at McKinsey.

Articles referenced:

• “ As gen AI advances, regulators—and risk functions—rush to keep pace ,” December 21, 2023, Andreas Kremer, Angela Luget , Daniel Mikkelsen , Henning Soller , Malin Strandell-Jansson, and Sheila Zingg
• “ What is generative AI? ,” January 19, 2023
• “ Tech highlights from 2022—in eight charts ,” December 22, 2022
• “ Generative AI is here: How tools like ChatGPT could change your business ,” December 20, 2022, Michael Chui , Roger Roberts , and Lareina Yee  
• “ The state of AI in 2022—and a half decade in review ,” December 6, 2022, Michael Chui , Bryce Hall , Helen Mayhew , Alex Singla , and Alex Sukharevsky  
• “ Why businesses need explainable AI—and how to deliver it ,” September 29, 2022, Liz Grennan , Andreas Kremer, Alex Singla , and Peter Zipparo
• “ Why digital trust truly matters ,” September 12, 2022, Jim Boehm , Liz Grennan , Alex Singla , and Kate Smaje
• “ McKinsey Technology Trends Outlook 2023 ,” July 20, 2023, Michael Chui , Mena Issler, Roger Roberts , and Lareina Yee  
• “ An AI power play: Fueling the next wave of innovation in the energy sector ,” May 12, 2022, Barry Boswell, Sean Buckley, Ben Elliott, Matias Melero , and Micah Smith  
• “ Scaling AI like a tech native: The CEO’s role ,” October 13, 2021, Jacomo Corbo, David Harvey, Nicolas Hohn, Kia Javanmardian , and Nayur Khan
• “ What the draft European Union AI regulations mean for business ,” August 10, 2021, Misha Benjamin, Kevin Buehler , Rachel Dooley, and Peter Zipparo
• “ Winning with AI is a state of mind ,” April 30, 2021, Thomas Meakin , Jeremy Palmer, Valentina Sartori , and Jamie Vickers
• “ Breaking through data-architecture gridlock to scale AI ,” January 26, 2021, Sven Blumberg , Jorge Machado , Henning Soller , and Asin Tavakoli  
• “ An executive’s guide to AI ,” November 17, 2020, Michael Chui , Brian McCarthy, and Vishnu Kamalnath
• “ Executive’s guide to developing AI at scale ,” October 28, 2020, Nayur Khan , Brian McCarthy, and Adi Pradhan
• “ An executive primer on artificial general intelligence ,” April 29, 2020, Federico Berruti , Pieter Nel, and Rob Whiteman
• “ The analytics academy: Bridging the gap between human and artificial intelligence ,” McKinsey Quarterly , September 25, 2019, Solly Brown, Darshit Gandhi, Louise Herring , and Ankur Puri  

This article was updated in April 2024; it was originally published in April 2023.

Want to know more about AI?

Related articles.

Ten unsung digital and AI ideas shaping business

Driving innovation with generative AI

As gen AI advances, regulators—and risk functions—rush to keep pace

• Data Science
• Data Analysis
• Data Visualization
• Machine Learning
• Deep Learning
• Computer Vision
• Artificial Intelligence
• AI ML DS Interview Series
• AI ML DS Projects series
• Data Engineering
• Web Scrapping

Characteristics of Artificial Intelligence Problems

Problems in Artificial Intelligence (AI) come in different forms, each with its own set of challenges and potential for innovation. From image recognition to natural language processing, AI problems exhibit distinct characteristics that shape the strategies and techniques used to tackle them effectively. In this article, we delve into the fundamental characteristics of AI problems, providing light on what makes them so fascinating and formidable.

Table of Content

Key Terminologies in Artificial Intelligence Problems

Addressing the challenges of ai problems, examples of ai applications and challenges across domains, characteristics of artificial intelligence problems – faqs.

Before exploring the characteristics, let’s clarify some essential AI concepts:

• Problem-solving: Problem-solving is a process that is a solution provided to a complex problem or task. When dealing with AI, problem-solving involves creating algorithms and methods of artificial intelligence that will empower machines to imitate humans’ capabilities of logical and reasonable thinking in certain situations.
• Search Space: Searching space refers to the area where an agent involved in the problem-solving process can examine all the possible states or settings with the hope of discovering a solution. It covers a gamut of options that the agent might select for arriving at the same destination.
• State: An entity represents some unique and specific arrangement of elements in a problem-solving situation. States can be assigned to different locations, challenges, or dangers that the problem-solving agent faces while looking for a solution to the problem within the search space.
• Search Algorithm: A search algorithm describes any process or method targeted for examining and exploring the given problem space to find a solution. Algorithm decision-making has diverging levels of complexity and effectiveness. They are studied to help in the discovery of the most suitable results.
• Heuristic: Heuristic is a thumb rule or guiding principle that is used to make intelligent decisions or solve the problems that are encountered during the process. Applying heuristics in AI is prevalent in prioritizing search paths or evaluating probable solutions based on their likelihood of finishing successfully.
• Optimization: The problem of optimization implies finding the best solution for process selection among the set of feasible alternatives submitted to some previously set objectives or criteria. AI optimization approaches are employed to deal optimally with complex issues through performance and efficiency improvement.

By understanding these key terminologies, we can better grasp the characteristics of AI problems and the techniques used to address them. These concepts form the foundation of AI problem-solving and provide the framework for developing innovative solutions to real-world challenges.

Let’s explore the core characteristics that differentiate AI problems:

• Learning and adaptation: AI systems should be capable of learning from data or experiences and adapting their behaviour accordingly. This enables them to improve performance over time and handle new situations more effectively.
• Complexity: AI problems often involve dealing with complex systems or large amounts of data. AI systems must be able to handle this complexity efficiently to produce meaningful results.
• Uncertainty: AI systems frequently operate in environments where outcomes are uncertain or incomplete information is available. They must be equipped to make decisions or predictions under such conditions.
• Dynamism: Environments in which AI systems operate can change over time. These changes may occur unpredictably or according to specific rules, requiring AI systems to continually adjust their strategies or models.
• Interactivity : Many AI applications involve interaction with users or other agents. Effective AI systems should be able to perceive, interpret, and respond to these interactions in a meaningful way.
• Context dependence: The behavior or performance of AI systems may depend on the context in which they operate. Understanding and appropriately responding to different contexts is essential for achieving desired outcomes.
• Multi-disciplinary: AI problems often require knowledge and techniques from multiple disciplines, including computer science, mathematics, statistics, psychology, and more. Integrating insights from these diverse fields is necessary for developing effective AI solutions.
• Goal-oriented Design: AI systems are typically designed to achieve specific objectives or goals. Designing AI systems with clear objectives in mind helps guide the development process and ensures that the resulting systems are focused on achieving meaningful outcomes.

These characteristics collectively shape the challenges and opportunities involved in developing and deploying AI systems across various domains and applications.

The characteristics of AI problems present unique challenges that require innovative approaches to solution development. Some of the key aspects to consider in tackling these challenges include:

• Complexity and Uncertainty: AI difficulties are sometimes characterized by highly variable domains that are difficult to predict exactly. Hence, AI algorithms should be installed with the skill of dealing with unclear circumstances and should make decisions that are based on imperfect data or noisy information.
• Algorithmic Efficiency: Among the key challenges of this approach are the enormous search spaces, computational resources, and the efficiency of the algorithms in terms of problem-solving. Strategies like caching, pruning, and parallelization are among the most widely used implementations for better algorithmic speed.
• Domain Knowledge Integration: Such numerous AI problems involve the ability to capture the rules and reasoning of the real world to model and solve the questions correctly. The AI machines that have been trained with expertise from relevant domains improve the accuracy and effectiveness of the applications in the real world.
• Scalability and Adaptability: AI solutions should be able to process large datasets and complex cases at the same time, and they should also be versatile by responding to shifts in conditions and requirements. Strategies such as machine learning and reinforcement learning allow systems to do more than just perform according to the given tasks at hand; they empower systems to learn and progress over time.
• Ethical and Social Implications: AI technologies elicit ethical and social limitations concerning problems of bias, justice, privacy, and responsible office. Taking these implications into account, along with ethical frameworks, compliance frameworks, and stakeholder engagement, is essential. This approach will help position cryptocurrencies as a secure and trustworthy investment.
• Interpretability and Explainability: To achieve interpretability and explainability of AI algorithms for the sake of understanding and confidence among users and stakeholders, these algorithms should be knowable and comprehensible enough. Examples like chatbots producing natural-like conversation could better clarify the working scheme of AI technology.
• Robustness and Resilience: AI machinery should perform against its being hacked or affected by adversarial attacks, inaccuracies (errors), and environmental changes. Robustness testing, the construction of mechanisms for error handling, and the building up of redundancy must be taken seriously by AI systems to ensure their reliability and stability.
• Human-AI Collaboration: Successful human-AI entente is the key component to making the most of our advantages as well as artificial intelligence skills. Achieving AI solutions that are capable of supporting human skills and more importantly, preferences will reduce human efforts correspondingly and bring the best performance.

By addressing these challenges through innovative methodologies and interdisciplinary collaboration, we can harness the full potential of AI to solve complex problems and drive societal progress.

1. Robotics

Problem: A delivery robot navigating a busy warehouse to locate and retrieve a specific item.

Characteristics:

• Complexity: Industrial storage is networked, in the middle of things, with obstacles, and other robots and people moving unpredictably. This robot must process the visual scene, plan the route effectively, and detect and avoid possible collisions.
• Dynamism: A combination of outside factors leads to change, which is a constant inside the warehouse. Unpredictable system failures or spontaneous tasks can make the robot change its means and decision-making at the moment of need.
• Uncertainty: Sensor data (such as images obtained from a camera) might be noisy, incomplete, and unstable. The robot could be handling decisions based on fragmented or formless pieces of information.

2. Natural Language Processing (NLP)

Problem: A sentiment analysis system in NLP classifying customer reviews as positive, negative, or neutral.

• Subjectivity: Human language is nuanced. Sarcasm, irony, and figurative expressions can be difficult for machines to accurately interpret.
• Need for Context: Understanding sentiment may depend on cultural references, product-specific knowledge, or even the reviewer’s prior interactions with the company.
• Ambiguity: A single word or phrase could have multiple meanings, affecting the overall sentiment of the text.

3. Computer Vision

Problem: A medical image recognition system in Computer Vision designed to detect tumors in X-rays or MRI scans.

• Complexity: Medical images are highly detailed and can exhibit subtle variations. The system needs to distinguish between healthy tissue and potential abnormalities.
• Uncertainty: Images may contain noise or artifacts. The presence of a tumor might not be immediately obvious, requiring the system to handle ambiguity.
• Ethical Considerations: False positives or false negatives have serious consequences for patient health. Accuracy, transparency, and minimizing bias are crucial.

The premises of AI-based problems – complexity, uncertainty, subjectivity, and more, – bring an unavoidable difficulty to the table. These features must be known for building appropriate AI because this is necessary. Through the use of machine learning, probabilistic reasoning, and knowledge representation which are referred to as the tools in AI development alongside the ethical considerations, these designers and scientists can face such complexities well and give shape to AI in a way that will be beneficial to society.

Q. What are the core characteristics that differentiate AI problems?

The core characteristics of AI problems include complexity, uncertainty and ambiguity, lack of clear problem definition, non-linearity, dynamism, subjectivity, interactivity, context sensitivity, and ethical considerations.

Q. Can you explain the concept of problem-solving in AI?

Problem-solving in AI involves creating algorithms and methods that enable machines to imitate human capabilities of logical and reasonable thinking in certain situations.

Q. What is meant by the term “search space” in AI?

Search space refers to the area where an agent involved in the problem-solving process can examine all the possible states or settings with the hope of discovering a solution.

Q. How do AI algorithms address challenges such as complexity and uncertainty?

AI algorithms are designed to handle unclear circumstances and make decisions based on imperfect data or noisy information.

Q. What are some examples of AI applications and the challenges they face?

Examples include robotics (e.g., delivery robots navigating busy warehouses), natural language processing (e.g., sentiment analysis of customer reviews), and computer vision (e.g., medical image recognition for detecting tumors).

Q. What role do ethical considerations play in AI development?

Ethical considerations are crucial in AI development to address issues such as bias, justice, privacy, and responsibility, ensuring that AI technologies are deployed responsibly and ethically.

Please Login to comment...

Similar reads.

• How to Delete Discord Servers: Step by Step Guide
• Google increases YouTube Premium price in India: Check our the latest plans
• California Lawmakers Pass Bill to Limit AI Replicas
• Best 10 IPTV Service Providers in Germany
• Content Improvement League 2024: From Good To A Great Article

Improve your Coding Skills with Practice

What kind of Experience do you want to share?

Can AI Match Human Ingenuity in Creative Problem-Solving?

When ChatGPT and other large language models began entering the mainstream two years ago, it quickly became apparent the technology could excel at certain business functions, yet it was less clear how well artificial intelligence could handle more creative tasks.

Sure, generative AI can summarize the content of an article, identify patterns in data, and produce derivative work—say, a song in the style of Taylor Swift or a poem in the mood of Langston Hughes—but can the technology develop truly innovative ideas?

Specifically, Harvard Business School Assistant Professor Jacqueline Ng Lane was determined to find out “how AI handled open-ended problems that haven’t been solved yet—the kind where you need diverse expertise and perspectives to make progress.”

In a working paper published in the journal Organization Science , Lane and colleagues compare ChatGPT’s creative potential to crowdsourced innovations produced by people. Ultimately, the researchers found that both humans and AI have their strengths—people contribute more novel suggestions while AI creates more practical solutions—yet some of the most promising ideas are the ones people and machines develop together.

Lane cowrote the paper with Léonard Bouissioux, assistant professor at the University of Washington’s Foster School of Business; Miaomiao Zhang, an HBS doctoral student, Karim Lakhani, the Dorothy & Michael Hintze Professor of Business Administration at HBS; and Vladimir Jacimovic, CEO and founder of ContinuumLab.ai and executive fellow at HBS.

Crowdsourcing people for ‘moonshots’

Any innovation process usually starts with brainstorming, says Lane, whose research has long looked at how creative ideas are produced.

“You start with defining the problem, then you generate ideas, then you evaluate them and choose which ones to implement.”

“It’s like a funnel,” she says. “You start with defining the problem, then you generate ideas, then you evaluate them and choose which ones to implement.”

Research has shown that crowdsourcing can be an effective way to generate initial ideas. However, the approach can be time-consuming and expensive. Creative teams typically offer incentives to respondents for their ideas. Then teams often must wait for input and then comb through ideas to come up with the most promising leads.

An off-the-shelf large language model such as ChatGPT, however, is free or low cost for end users, and can generate an infinite number of ideas quickly, Lane says. But are the ideas any good?

To find out, Lane and her fellow researchers asked people to come up with business ideas for the sustainable circular economy, in which products are reused or recycled to make new products. They disseminated a request on an online platform, offering $10 for participating and $1,000 for the best idea. Here’s part of their request:

We would like you to submit your circular economy idea, which can be a unique new idea or an existent idea that is used in the industry.

Here is an example: Car sharing in order to reduce the carbon footprint associated with driving. …

Submit your real-life use cases on how companies can implement the circular economy in their businesses. New ideas are also welcome, even if they are “moonshots.”

Seeking creative ideas from ChatGPT

The researchers asked for ideas that would involve “sharing, leasing, reusing, repairing, refurbishing [or] recycling existing materials and products as long as possible.” Suggestions would be scored for uniqueness, environmental benefits, profit potential, and feasibility.

Some 125 people replied with contributions, offering insights from a variety of industries and professional backgrounds. One, for example, proposed a dynamic pricing algorithm for supermarkets to cut down on food waste, while another suggested a mobile app that could store receipts to reduce paper waste.

At the same time, the research team employed prompt engineering techniques to craft a variety of AI prompts. Using these carefully designed prompts, they generated several hundred additional solutions through ChatGPT. The team strategically modified their prompts to:

• Challenge the model to create more ideas.
• Mimic the perspective of someone from a particular industry, job title, and place—a persona.
• Remind the model to provide ideas that reflect the scoring criteria.

The team then recruited some 300 evaluators well-versed in the circular economy to evaluate a randomized selection of the ideas based on the scoring criteria.

People are creative, but AI ideas are more feasible

The evaluators judged the human solutions as more novel, employing more unique “out of the box” thinking. However, they found the AI-generated ideas to be more valuable and feasible.

For example, one participant from Africa proposed creating interlocking bricks using foundry dust and waste plastic, creating a new construction material and cutting down on air pollution at the same time. “The evaluators said, ‘Wow, this is really innovative, but it would never work,’” Lane says.

“We were surprised at how powerful these technologies were.”

One ChatGPT response, meanwhile, created an idea to convert food waste into biogas, a renewable energy source that could be used for electricity and fertilizer. Not the most novel idea, the researchers noted, but one that could be implemented and might show a clear financial return.

“We were surprised at how powerful these technologies were,” Lane says, “especially in these early stages in the creative process.”

How to reach the best solutions

The “best” ideas, Lane says, may come from those in which humans and AI collaborate, with people engineering prompts and continually working with AI to develop more original ideas.

“We consistently achieved higher quality results when AI would come up with an idea and then we had an instruction that said: Make sure before you create your next idea, it’s different from all the ones before it,” Lane explains.

Additional prompts increased the novelty of the ideas, generating everything from waste-eating African flies to beverage containers tracked by smart chips that instantly pay consumers for recycling them.

Based on the findings, the researchers suggest business leaders keep a few points in mind when implementing AI to develop creative solutions:

• Knowing how to ask the right questions is important. Organizations might want to invest in cultivating an “AI-literate” workforce that can understand the capabilities and limitations of AI to generate the most successful ideas.
• Organizations should resist the temptation to rely excessively on AI. That could “dumb down” the overall level of creative output over time, leading to more incremental improvements than radical breakthroughs, the team says.
• People should view generative AI models as collaborative tools. In a sequential approach, humans could brainstorm solutions, then submit them to AI to refine them and increase their value and feasibility. Alternatively, humans could work more iteratively with AI, constantly shaping and improving the ideas it provides.

The most productive way to use generative AI, the research suggests, is to combine the novelty that people excel at with the practicality of the machine. Says Lane, “We still need to put our minds toward being forward-looking and envisioning new things as we are guiding the outputs of AI to create the best solutions.”

Rapport: The Hidden Advantage That Women Managers Bring to Teams

• 11 Jun 2024
• In Practice

The Harvard Business School Faculty Summer Reader 2024

How transparency sped innovation in a $13 billion wireless sector.

• 24 Jan 2024

Why Boeing’s Problems with the 737 MAX Began More Than 25 Years Ago

• 27 Jun 2016

These Management Practices, Like Certain Technologies, Boost Company Performance

• Problems and Challenges
• Innovation and Invention
• Disruptive Innovation
• Search Technology
• Information Technology
• Information

Sign up for our weekly newsletter

Problem-Solving Methods in Artificial Intelligence

Information & contributors, bibliometrics & citations, view options.

• Haigh T (2024) How the AI Boom Went Bust Communications of the ACM 10.1145/3634901 67 :2 (22-26) Online publication date: 25-Jan-2024 https://dl.acm.org/doi/10.1145/3634901
• Zheng L Xing Y Yu L Zhang J (2023) Uncovering the Dark Side of Artificial Intelligence in Electronic Markets Journal of Organizational and End User Computing 10.4018/JOEUC.327278 35 :1 (1-25) Online publication date: 1-Aug-2023 https://dl.acm.org/doi/10.4018/JOEUC.327278
• Sleeman D Gilhooly K (2023) Groups of experts often differ in their decisions AI Magazine 10.1002/aaai.12135 44 :4 (555-567) Online publication date: 8-Dec-2023 https://dl.acm.org/doi/10.1002/aaai.12135
• Show More Cited By

Index Terms

Computing methodologies

Artificial intelligence

Control methods

Search methodologies

General and reference

Document types

Reference works

Theory of computation

Design and analysis of algorithms

Algorithm design techniques

Dynamic programming

Recommendations

Methods and standards for research on explainable artificial intelligence: lessons from intelligent tutoring systems.

The DARPA Explainable Artificial Intelligence (AI) (XAI) Program focused on generating explanations for AI programs that use machine learning techniques. This article highlights progress during the DARPA Program (2017‐2021) relative to research ...

Lessons learned in the work on intelligent tutoring systems that apply to system design in Explainable AI. image image

Artificial intelligence (AI) is the Science and Engineering domain concerned with the theory and practice of developing systems that exhibit the characteristics we associate with intelligence in human behavior. Starting with a brief history of ...

Artificial Intelligence: Theories, Models and Applications 7th Hellenic Conference on AI, SETN 2012, Lamia, Greece, May 28-31, 2012, Proceedings

Information, published in.

McGraw-Hill Pub. Co.

Publication History

Contributors, other metrics, bibliometrics, article metrics.

• 414 Total Citations View Citations
• 0 Total Downloads
• Downloads (Last 12 months) 0
• Downloads (Last 6 weeks) 0
• Segev D Shaposhnik Y (2022) A Polynomial-Time Approximation Scheme for Sequential Batch Testing of Series Systems Operations Research 10.1287/opre.2019.1967 70 :2 (1153-1165) Online publication date: 1-Mar-2022 https://dl.acm.org/doi/10.1287/opre.2019.1967
• Pearl J (2022) Asymptotic Properties of Minimax Trees and Game-Searching Procedures Probabilistic and Causal Inference 10.1145/3501714.3501723 (61-90) Online publication date: 28-Feb-2022 https://dl.acm.org/doi/10.1145/3501714.3501723
• Ardizzoni S Consolini L Laurini M Locatelli M (2021) Efficient solution algorithms for the Bounded Acceleration Shortest Path problem 2021 60th IEEE Conference on Decision and Control (CDC) 10.1109/CDC45484.2021.9683191 (5729-5734) Online publication date: 14-Dec-2021 https://dl.acm.org/doi/10.1109/CDC45484.2021.9683191
• Garrett C Lozano-Pérez T Kaelbling L (2020) Sampling-based methods for factored task and motion planning International Journal of Robotics Research 10.1177/0278364918802962 37 :13-14 (1796-1825) Online publication date: 17-Jun-2020 https://dl.acm.org/doi/10.1177/0278364918802962
• Dehesa J Vidler A Lutteroth C Padget J Bernhaupt R Mueller F Verweij D Andres J McGrenere J Cockburn A Avellino I Goguey A Bjørn P Zhao S Samson B Kocielnik R (2020) Touché: Data-Driven Interactive Sword Fighting in Virtual Reality Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems 10.1145/3313831.3376714 (1-14) Online publication date: 21-Apr-2020 https://dl.acm.org/doi/10.1145/3313831.3376714
• Alves M de Oliveira Oliveira M Silva J dos Santos Souza U (2020) Succinct Monotone Circuit Certification: Planarity and Parameterized Complexity Computing and Combinatorics 10.1007/978-3-030-58150-3_40 (496-507) Online publication date: 29-Aug-2020 https://dl.acm.org/doi/10.1007/978-3-030-58150-3_40
• Barto A (2019) Reinforcement Learning AI Magazine 10.1609/aimag.v40i1.2844 40 :1 (3-15) Online publication date: 1-Mar-2019 https://dl.acm.org/doi/10.1609/aimag.v40i1.2844

View options

Login options.

Check if you have access through your login credentials or your institution to get full access on this article.

Full Access

Share this publication link.

Copying failed.

Share on social media

Affiliations, export citations.

• Please download or close your previous search result export first before starting a new bulk export. Preview is not available. By clicking download, a status dialog will open to start the export process. The process may take a few minutes but once it finishes a file will be downloadable from your browser. You may continue to browse the DL while the export process is in progress. Download
• Download citation
• Copy citation

We are preparing your search results for download ...

We will inform you here when the file is ready.

Your file of search results citations is now ready.

Your search export query has expired. Please try again.

share this!

August 31, 2024

This article has been reviewed according to Science X's editorial process and policies . Editors have highlighted the following attributes while ensuring the content's credibility:

fact-checked

peer-reviewed publication

trusted source

Research team proposes solution to AI's continual learning problem

by Scott Lilwall, Alberta Machine Intelligence Institute

A team of Alberta Machine Intelligence Institute (Amii) researchers has revealed more about a mysterious problem in machine learning—a discovery that might be a major step towards building advanced AI that can function effectively in the real world.

The paper, titled "Loss of Plasticity in Deep Continual Learning," is published in Nature . It was authored by Shibhansh Dohare, J. Fernando Hernandez-Garcia, Qingfeng Lan, Parash Rahman, as well as Amii Fellows & Canada CIFAR AI Chairs A. Rupam Mahmood and Richard S. Sutton.

In their paper, the team explores a vexing problem that has long been suspected in deep learning models but has not received much attention: for some reason, many deep learning agents engaged in continual learning lose the ability to learn and have their performance degrade drastically.

"We have established that there is definitely a problem with current deep learning," said Mahmood. "When you need to adapt continually, we have shown that deep learning eventually just stops working. So effectively you can't keep learning."

He points out that not only does the AI agent lose the ability to learn new things, but it also fails to relearn what it learned in the past after it is forgotten. The researchers dubbed this phenomenon "loss of plasticity," borrowing a term from neuroscience where plasticity refers to the brain's ability to adapt its structure and form new neural connections.

The state of current deep learning

The researchers say that loss of plasticity is a major challenge to developing AI that can effectively handle the complexity of the world and would need to be solved to develop human-level artificial intelligence.

Many existing models aren't designed for continual learning. Sutton references ChatGPT as an example; it doesn't learn continually. Instead, its creators train the model for a certain amount of time. When training is over, the model is then deployed without further learning.

Even with this approach, merging new and old data into a model's memory can be difficult. Most of the time, it is more effective to just start from scratch, erasing the memory, and training the model on everything again. For large models like ChatGPT, that process can take a lot of time and cost millions of dollars each time.

It also limits the kind of things a model can do. For fast-moving environments that are constantly changing, like financial markets for instance, Sutton says continual learning is a necessity.

Hidden in plain sight

The first step to addressing loss of plasticity, according to the team, was to show that it happens and it matters. The problem is one that was "hiding in plain sight"—there were hints suggesting that loss of plasticity could be a widespread problem in deep learning, but very little research had been done to actually investigate it.

Rahman says he first became interested in exploring the problem because he kept seeing hints of the issue—and that intrigued him.

"I'd be reading through a paper, and you'd see something in the appendices about how performance dropped off. And then you'd see it in another paper a while later," he said.

The research team designed several experiments to search for loss of plasticity in deep learning systems. In supervised learning, they trained networks in sequences of classification tasks. For example, a network would learn to differentiate between cats and dogs in the first task, then between beavers and geese on the second task, and so on for many tasks. They hypothesized that as the networks lost their ability to learn, their ability to differentiate would decrease in each subsequent task.

And that's exactly what happened.

"We used several different data sets to test, to show that it could be widespread. It really shows that it isn't happening in a little corner of deep learning, " Sutton said.

Dealing with the dead

With the problem established, the researchers then had to ask: could it be solved? Was loss of plasticity an inherent issue for continual deep-learning networks, or was there a way to allow them to keep learning?

They found some hope in a method based on modifying one of the fundamental algorithms that make neural networks work: backpropagation.

Neural networks are built to echo the structure of the human brain: They contain units that can pass information and make connections with other units, just like neurons. Individual units can pass information along to other layers of units, which do the same. All of this contributes to the network's overall output.

However, when adapting the connection strength or "weights" of the network with backpropagation, a lot of the time these units will calculate outputs that don't actually contribute to learning. They also won't learn new outputs, so they will become dead weight to the network and stop contributing to the learning process.

Over long-term continual learning, as many as 90% of a network's units might become dead, Mahmood notes. And when enough stops contributing, the model loses plasticity.

So, the team came up with a modified method that they call "continual backpropagation."

Dohare says that it differs from backpropagation in a key way: While backpropagation randomly initializes the units only at the very beginning, continual backpropagation does so continually. Once in a while, during learning, it selects some of the useless units, like the dead ones, and reinitializes them with random weights. By using continual backpropagation, they find that models can continually learn much longer, sometimes seemingly indefinitely.

Sutton says that other researchers might come up with better solutions to tackle loss of plasticity, but their continual backprop approach at least shows the problem can be solved, and this tricky problem isn't inherent to deep networks.

He's hopeful that the team's work will bring more attention to loss of plasticity and encourage other researchers to examine the issue.

"We established this problem in a way that people sort of have to acknowledge it. The field is gradually getting more willing to acknowledge that deep learning, despite its successes, has fundamental issues that need addressing," he said. "So, we're hoping this will open up this question a little bit."

Explore further

Feedback to editors

AI may not steal many jobs after all. It may just make workers more efficient

3 hours ago

A quantum neural network can see optical illusions like humans do. Could it be the future of AI?

Aug 31, 2024

Pilot plant demonstrates iron-based hydrogen storage feasibility

Exploring the fundamental reasoning abilities of LLMs

Virtual and augmented reality can temporarily change the way people perceive distances, finds study

Aug 30, 2024

Researchers develop ultra-high efficiency perovskite LEDs by strengthening lattice

Transparency is often lacking in datasets used to train large language models, study finds

Morphing facial technology sheds light on the boundaries of self-recognition

Silicon chip propels 6G communications forward

Aug 29, 2024

Scalable graphene technology could significantly enhance battery safety and performance

Related stories.

New method allows AI to learn indefinitely

Aug 22, 2024

Mesoscale neural plasticity helps in AI learning

Oct 22, 2021

Framework improves 'continual learning' for artificial intelligence

May 15, 2019

Brain-inspired learning algorithm realizes metaplasticity in artificial and spiking neural networks

Sep 1, 2023

A new perspective on Petri net learning

Nov 8, 2023

Researchers develop new strategies to teach computers to learn like humans do

Aug 30, 2022

Recommended for you

Let us know if there is a problem with our content.

Use this form if you have come across a typo, inaccuracy or would like to send an edit request for the content on this page. For general inquiries, please use our contact form . For general feedback, use the public comments section below (please adhere to guidelines ).

Please select the most appropriate category to facilitate processing of your request

Thank you for taking time to provide your feedback to the editors.

Your feedback is important to us. However, we do not guarantee individual replies due to the high volume of messages.

E-mail the story

Your email address is used only to let the recipient know who sent the email. Neither your address nor the recipient's address will be used for any other purpose. The information you enter will appear in your e-mail message and is not retained by Tech Xplore in any form.

Your Privacy

This site uses cookies to assist with navigation, analyse your use of our services, collect data for ads personalisation and provide content from third parties. By using our site, you acknowledge that you have read and understand our Privacy Policy and Terms of Use .

E-mail newsletter

As artificial intelligence rapidly advances, experts debate level of threat to humanity

Paul Solman Paul Solman

Ryan Connelly Holmes Ryan Connelly Holmes

Leave your feedback

• Copy URL https://www.pbs.org/newshour/show/as-artificial-intelligence-rapidly-advances-experts-debate-level-of-threat-to-humanity

The development of artificial intelligence is speeding up so quickly that it was addressed briefly at both Republican and Democratic conventions. Science fiction has long theorized about the ways in which machines might one day usurp their human overlords. As the capabilities of modern AI grow, Paul Solman looks at the existential threats some experts fear and that some see as hyperbole.

Read the Full Transcript

Notice: Transcripts are machine and human generated and lightly edited for accuracy. They may contain errors.

Geoff Bennett:

The development of artificial intelligence is speeding up so quickly that it was addressed briefly at both political conventions, including the Democratic gathering this week.

Of course, science fiction writers and movies have long theorized about the ways in which machines might one day usurp their human overlords.

As the capabilities of modern artificial intelligence grow, Paul Solman looks at the existential threats some experts fear and that some see as hyperbole.

Eliezer Yudkowsky, Founder, Machine Intelligence Research Institute:

From my perspective, there's inevitable doom at the end of this, where, if you keep on making A.I. smarter and smarter, they will kill you.

Paul Solman:

Kill you, me and everyone, predicts Eliezer Yudkowsky, tech pundit and founder back in the year 2000 of a nonprofit now called the Machine Intelligence Research Institute to explore the uses of friendly A.I.; 24 years later, do you think everybody's going to die in my lifetime, in your lifetime?

Eliezer Yudkowsky:

I would wildly guess my lifetime and even your lifetime.

Now, we have heard it before, as when the so-called Godfather of A.I., Geoffrey Hinton, warned Geoff Bennett last spring.

Geoffrey Hinton, Artificial Intelligence Pioneer:

The machines taking over is a threat for everybody. It's a threat for the Chinese and for the Americans and for the Europeans, just like a global nuclear war was.

And more than a century ago, the Czech play "R.U.R.," Rossum's Universal Robots, from which the word robot comes, dramatized the warning.

And since 1921 — that's more than 100 years ago — people have been imagining that the robots will become sentient and destroy us.

Jerry Kaplan, Author, "Generative Artificial Intelligence: What Everyone Needs to Know": That's right.

A.I. expert Stanford's Jerry Kaplan at Silicon Valley's Computer History Museum.

Jerry Kaplan:

That's created a whole mythology, which, of course, has played out in endless science fiction treatments.

Like the Terminator series.

Michael Biehn, Actor:

A new order of intelligence decided our fate in a microsecond, extermination.

Judgment Day forecast for 1997. But, hey, that's Hollywood. And look on the bright side, no rebel robots or even hoverboards or flying cars yet.

On the other hand, robots will be everywhere soon enough, as mass production drives down their cost. So will they soon turn against us?

I got news for you. There's no they there. They don't want anything. They don't need anything. We design and build these things to our own specifications. Now, that's not to say we can't build some very dangerous machines and some very dangerous tools.

Kaplan thinks what humans do with A.I. is much scarier than A.I. on its own, create super viruses, mega drones, God knows what else.

But whodunit aside, the big question still is, will A.I. bring doomsday?

A.I. Reid Hoffman avatar: I'd rate the existential threat of A.I. around a three or four out of 10.

That's the avatar of LinkedIn founder Reid Hoffman, to which we fed the question, 1 being no threat, 10 extinction. What does the real Reid Hoffman say?

Reid Hoffman, Creator, LinkedIn Corporation:

I'm going to go for two on that answer.

I'm going to tell you that your avatar said 3 to 4.

Reid Hoffman:

All right. Well, my avatar is trying to approximate something that is what kind of a bulk of educated, reasonable people would want to hear.

But, hey, even if you told me there was only a one in 10 chance of a ticking time bomb in my room, I would be out of there.

Yudkowsky started warning of rogue A.I. back in 2005. And since?

Things have gone a bit worse than hoped for. And, to be clear, the sting at the end of this is A.I. gets smarter than us, is poorly controlled, and probably humanity becomes collateral damage to its own expansion.

What's the doomsday scenario here?

It is smarter than humanity. From its perspective, it now wants to get independence of humanity. It doesn't want to be running on computers that require electricity that humans have to generate.

There are many more A.I. companies now than just OpenAI. And some of the people there are explicitly like, yes, sure like maybe it'll wipe out humanity, but I don't care about that.

And of all the world's problems, tribalism, climate change, et cetera, you think A.I. is a bigger problem?

I think climate change is unlikely to kill literally everyone, down to the last man, woman and child. At all of the top three companies, I expect the leaders would prefer not to wipe out humanity, but, as they see it, it's not up to them. They got to stay ahead of their competitors.

So what does A.I. itself think?

I asked Amica, a robot hooked up to ChatGPT. Destruction, 1 to 10? The response, a 3.

Amica, ChatGPT A.I.:

The potential for harmony and enhancement exists, but vigilance is paramount. Would you roll the dice on that?

I'm afraid we're all going to be rolling the dice on that. And I'm worried that in your answer, you're saying there's something like a 30 percent chance that A.I., artificial intelligence, will destroy humanity. Shouldn't I be scared?

Your concern is valid, Paul. The 30 percent isn't a direct prediction, but a cautious nod to the uncertainties and ethical dilemmas of handling powerful technology.

But I'm afraid that human beings may not be capable of protecting themselves from the risks of A.I.

That dread is profound. It's daunting, but not without hope. Humanity's resilience has often defied expectations.

Perhaps it's no surprise that the actual human who created ChatGPT, Sam Altman, thinks the same.

Sam Altman, CEO, OpenAI:

I believe that A.I. will be very net good, tremendously net good, but I think, like with any other tool, it'll be misused. Like, you can do great things with a hammer and you can, like, kill people with a hammer. I don't think that absolves us, or you all, or society from trying to mitigate the bad as much as we can and maximize the good.

And Reid Hoffman thinks we can maximize the good.

We have a portfolio risk. We have climate change as a possibility. We have pandemic as a possibility. We have nuclear war as a possibility. We have asteroids as a possibility. We have human world war as a possibility. We have all of these existential risks.

And you go, OK, A.I., is it also an additional existential risk? And the answer is, yes, potentially. But you look at its portfolio and say, what improves our overall portfolio? What reduces existential risk for humanity? And A.I. is one of the things that adds a lot in the positive column.

So, if you think, how do we prevent future natural or manmade pandemic, A.I. is the only way that I think can do that. And also, like, it might even help us with climate change things. So you go, OK, in the net portfolio, our existential risk may go down with A.I.

For the sake of us all, grownups, children, grandchildren, let's hope he's right.

For the "PBS News Hour" in Silicon Valley, Paul Solman.

Listen to this Segment

Watch the Full Episode

Paul Solman has been a correspondent for the PBS News Hour since 1985, mainly covering business and economics.

Support Provided By: Learn more

More Ways to Watch

Educate your inbox.

Subscribe to Here’s the Deal, our politics newsletter for analysis you won’t find anywhere else.

Thank you. Please check your inbox to confirm.

Ensemble of physics-informed neural networks for solving plane elasticity problems with examples

• Original Paper
• Published: 29 August 2024

Cite this article

• Aliki D. Mouratidou   ORCID: orcid.org/0000-0002-8382-1263 1 ,
• Georgios A. Drosopoulos 2 , 3 &
• Georgios E. Stavroulakis 1  

Two-dimensional (plane) elasticity equations in solid mechanics are solved numerically with the use of an ensemble of physics-informed neural networks (PINNs). The system of equations consists of the kinematic definitions, i.e. the strain–displacement relations, the equilibrium equations connecting a stress tensor with external loading forces and the isotropic constitutive relations for stress and strain tensors. Different boundary conditions for the strain tensor and displacements are considered. The proposed computational approach is based on principles of artificial intelligence and uses a developed open-source machine learning platform, scientific software Tensorflow, written in Python and Keras library, an application programming interface, intended for a deep learning. A deep learning is performed through training the physics-informed neural network model in order to fit the plain elasticity equations and given boundary conditions at collocation points. The numerical technique is tested on an example, where the exact solution is given. Two examples with plane stress problems are calculated with the proposed multi-PINN model. The numerical solution is compared with results obtained after using commercial finite element software. The numerical results have shown that an application of a multi-network approach is more beneficial in comparison with using a single PINN with many outputs. The derived results confirmed the efficiency of the introduced methodology. The proposed technique can be extended and applied to the structures with nonlinear material properties.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Subscribe and save.

• Get 10 units per month
• Download Article/Chapter or eBook
• 1 Unit = 1 Article or 1 Chapter
• Cancel anytime

Price includes VAT (Russian Federation)

Instant access to the full article PDF.

Rent this article via DeepDyve

Institutional subscriptions

Similar content being viewed by others

Comparison of neural FEM and neural operator methods for applications in solid mechanics

Deep learning operator network for plastic deformation with variable loads and material properties

Teaching solid mechanics to artificial intelligence—a fast solver for heterogeneous materials

Explore related subjects.

• Artificial Intelligence

The entire software package is available upon request.

Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean, J., Devin, M., Ghemawat, S., Irving, G., Isard, M., Kudlur, M., Levenberg, J., Monga, R., Moore, S., Murray, D.G., Steiner, B., Tucker, P., Vasudevan, V., Warden, P., Wicke, M., Yu, Y., Zheng, X.: Tensorflow: a system for large-scale machine learning. In: 12th USENIX Symposium on Operating Systems Design and Implementation (OSDI 16), USENIX Association, Savannah, GA, pp. 265-283. URL https://www.usenix.org/conference/osdi16/technical-sessions/presentation/abadi. (2016)

Baydin, A., Pearlmutter, B.A., Radul, A.A., Siskind, J.M.: Automatic differentiation in machine learning: a survey. J. Mach. Learn. Res. 18 , 1 (2018)

MathSciNet   Google Scholar  

Bazmara, M., Mianroodi, M., Silani, M.: Application of physics-informed neural networks for nonlinear buckling analysis of beams. Acta. Mech. Sin. 39 , 422438 (2023)

Article   MathSciNet   Google Scholar  

Cai, S., Mao, Z., Wang, Z., Yin, M., Karniadakis, G.E.: Physics-informed neural networks (PINNs) for fluid mechanics: a review. Acta Mech. Sin. 37 , 1–12 (2022)

Chadha, C., He, J., Abueidda, D., Koric, S., Guleryuz, E., Jasiuk, I.: Improving the accuracy of the deep energy method. Acta Mech. 234 , 5975–5998 (2023)

Chen, Y., Lu, L., Karniadakis, G.E., Dal Negro, L.: Physics-informed neural networks for inverse problems in nano-optics and metamaterials. Opt. Express 28 (8), 11618–11633 (2020)

Article   Google Scholar  

Chollet, F.: Deep learning with python, Manning Publications Company, URL https://books.google.ca/books?id= Yo3CAQAACAAJ (2017)

Drosopoulos, G.A., Stavroulakis, G.E.: Non-linear mechanics for composite. In: Heterogeneous Structures, CRC Press, Taylor and Francis (2022)

Drosopoulos, G.A., Stavroulakis, G.E.: Data-driven computational homogenization using neural networks: FE2-NN application on damaged masonry. ACM J. Comput. Cult. Herit. 14 , 1–19 (2020)

Google Scholar  

Faroughi, S., Darvishi, A., Rezaei, S.: On the order of derivation in the training of physics-informed neural networks: case studies for non-uniform beam structures. Acta Mech. 234 , 5673–5695 (2023)

Faroughi, S.A., Pawar, N., Fernandes, C., Raissi, M., Das, S., Kalantari, N.K., Mahjour, S.K.: Physics-guided, physics-informed, and physics-encoded neural networks. Sci. Comput. (2023). https://doi.org/10.48550/arXiv.2211.07377

Fletcher, R.: Practical methods of optimization, 2nd edn. John Wiley & Sons, New York (1987)

Güne, A., Baydin, G., Pearlmutter, B.A., Siskind, J.M.: Automatic differentiation in machine learning: a survey, J. Mach. Learn. Res. 18 , 1-43, (2018) URL http://www.jmlr.org/papers/volume18/17-468/17-468.pdf

Guo, M., Haghighat, E.: An energy-based error bound of physics-informed neural network solutions in elasticity. J. Eng. Mech. (2020). https://doi.org/10.48550/arXiv.2010.09088

Haghighat, E., Juanes, R.: Sciann: a keras/tensorflow wrapper for scientific computations and physics-informed deep learning using artificial neural networks. Comp. Meth. Appl. Mech. Eng. 373 , 113552 (2021)

Haghighat, E., Raissi, M., Moure, A., Gomez, H., and Juanes, R.: A deep learning framework for solution and discovery in solid mechanics. https://arxiv.org/abs/2003.02751 (2020)

Harandi, A., Moeineddin, A., Kaliske, M., Reese, S., Rezaei, S.: Mixed formulation of physics-informed neural networks for thermo-mechanically coupled systems and heterogeneous domains. Int J Numer. Methods Eng. 125 , e7388 (2024)

Kadeethum, T., Jorgensen, T., Nick, H.: Physics-informed neural networks for solving nonlinear diffusivity and Biot’s equations. PLoS ONE 15 , e0232683 (2020)

Karniadakis, G.E., Kevrekidis, I.G., Lu, L., et al.: Physics-informed machine learning. Nat. Rev. Phys. 3 , 422–440 (2021)

Katsikis, D., Muradova, A.D., Stavroulakis, G.S.: A gentle introduction to physics-informed neural networks, with applications in static rod and beam problems. Jr. Adv. Appl. & Comput. Math. 9 , 103–128 (2022)

Kingma, D.P., Ba, J.: Adam: a method for stochastic optimization, computer science, mathematics. In: The International Conference on Learning Representations (ICLR) (2015)

Kortesis, S., Panagiotopoulos, P.D.: Neural networks for computing in structural analysis: methods and prospects of applications. Int. Jr. Numeric. Meth. Eng. 36 , 2305–2318 (1993)

Kovachki, N., Lanthaler, S., Mishra, S.: On universal approximation and error bounds for Fourier neural operator. J. Mach. Learn. Res. 22 , 1–76 (2021)

Lagaris, E., Likas, A., Fotiadis, D.I.: Artificial neural networks for solving ordinary and partial differential equations. IEEE Trans. Neural Netw. 9 , 987–1000 (1998)

Li, Z., Kovachki, N., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A., Anandkumar, A.: Fourier neural operator for parametric partial differential equations, arXiv preprint arXiv:2010.08895 (2020)

Liu, D.C., Nocedal, J.: On the limited memory BFGS method for large scale optimization. Math. Program. 45 , 503–528 (1989)

Lu, L., Jin, P., Pang, G., Zhang, Z., Karniadakis, G.E.: Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators. Nat. Mach. Intell. 3 , 218–229 (2021). https://doi.org/10.1038/s42256-021-00302-5

Meade, A.J., Fernandez, A.A.: The numerical solution of linear ordinary differential equations by feed-forward neural networks. Math. Comput. Model. 19 , 1–25 (1994). https://doi.org/10.1016/0895-7177(94)90095-7

Muradova, A.D., Stavroulakis, G.E.: Physics-informed neural networks for elastic plate problems with bending and Winkler-type contact effects. J. Serb. Soc. Comput. Mech. 15 , 45–54 (2021)

Muradova, A.D, Stavroulakis, G.E.: Physics-informed Neural Networks for the solution of unilateral contact problems. In: Book of Proceedings, 13th International Congress on Mechanics HSTAM, pp. 451-459. (2022). https://hstam2022.eap.gr/book-of-proceedings/

Muradova, A.D., Stavroulakis, G.E.: The projective-iterative method and neural network estimation for buckling of elastic plates in nonlinear theory. Commun. Nonlin. Sci. Num. Sim. 12 , 1068–1088 (2007)

Niu, S., Zhang, E., Bazilevs, Y., Srivastava, V.: Modeling finite-strain plasticity using physics-informed neural network and assessment of the network performance. J. Mech. Phys. Solids 172 , 105177 (2023)

Raissi, M., Perdikaris, P., Karniadakis, G.E.: Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. J. Comput. Phys. 378 , 686–707 (2019)

Rezaei, S., Harandi, A., Moeineddin, A., Xu, B.-X., Reese, S.: A mixed formulation for physics-informed neural networks as a potential solver for engineering problems in heterogeneous domains: comparison with finite element method. Comput. Meth. Appl. Mech. Eng. 401 (Part B), 115616 (2022)

Ruder, S.: An overview of gradient descent optimization algorithms (2017) https://arxiv.org/abs/1609.04747

Shin, Y., Darbon, J., Karniadakis, G.E.: On the convergence of physics informed neural networks for linear second-order elliptic and parabolic type PDEs. Commun. Comput. Phys. 28 , 2042–2074 (2020)

Stavroulakis, G.E.: Inverse and crack identification problems in engineering mechanics. Springer, New York (2000)

Stavroulakis, G.E., Avdelas, A., Abdalla, K.M., Panagiotopoulos, P.D.: A neural network approach to the modelling, calculation and identification of semi-rigid connections in steel structures. J. Construct. Steel Res. 44 , 91–105 (1997)

Stavroulakis, G., Bolzon, G., Waszczyszyn, Z., Ziemianski, L.: Inverse analysis. In: Karihaloo, B., Ritchie, R.O., Milne, I. (eds) Comprehensive structural integrity, numerical and computational methods, vol. 3, Chap 13, pp. 685–718. Elsevier, Amsterdam (2003)

Tartakovsky, A. M., Marrero, C. O., Perdikaris, P., Tartakovsky, G. D., and Barajas- Solano D.: Learning parameters and constitutive relationships with physics informed deep neural networks. arXiv preprint arXiv:1808.03398 (2018)

Theocaris, P.S., Panagiotopoulos, P.D.: Plasticity including the Bauschinger effect, studied by a neural network approach. Acta Mech. 113 , 63–75 (1995). https://doi.org/10.1007/BF01212634

Waszczyszyn, Z., Ziemiański, L.: Neural networks in the identification analysis of structural mechanics problems. In: Mróz, Z., Stavroulakis, G.E. (eds.) Parameter identification of materials and structures, CISM International Centre for Mechanical Sciences (Courses and Lectures), p. 469. Springer, Vienna (2005)

Xiao, Y., Wei, Z., Wang, Z.: A limited memory BFGS-type method for large-scale unconstrained optimization. Comput. Math. Appl. 56 , 1001–1009 (2008)

Yagawa, G., Oishi, A.: Computational mechanics with neural networks. Springer, New York (2021)

Book   Google Scholar  

Download references

The work of A.D.M. and G.E.S. has been supported by the Project Safe-Aorta, which was implemented in the framework of the Action “Flagship actions in interdisciplinary scientific fields with a special focus on the productive fabric”, through the National Recovery and Resilience Fund Greece 2.0 and funded by the European Union-NextGenerationEU (Project ID:TAEDR-0535983)

Author information

Authors and affiliations.

School of Production Engineering and Management, Institute of Computational Mechanics and Optimization, Technical University of Crete, Kounoupidiana, 73100, Chania, Crete, Greece

Aliki D. Mouratidou & Georgios E. Stavroulakis

Discipline of Civil Engineering, School of Engineering and Computing, University of Central Lancashire, Preston campus, Preston, PR1 2HE, UK

Georgios A. Drosopoulos

Discipline of Civil Engineering, School of Engineering, University of KwaZulu-Natal, Durban campus, Durban, 4041, South Africa

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Aliki D. Mouratidou .

Ethics declarations

Conflict of interest.

The authors declare that they have no conflict of interest.

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Reprints and permissions

About this article

Mouratidou, A.D., Drosopoulos, G.A. & Stavroulakis, G.E. Ensemble of physics-informed neural networks for solving plane elasticity problems with examples. Acta Mech (2024). https://doi.org/10.1007/s00707-024-04053-3

Download citation

Received : 26 February 2024

Revised : 27 May 2024

Accepted : 31 July 2024

Published : 29 August 2024

DOI : https://doi.org/10.1007/s00707-024-04053-3

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Find a journal
• Publish with us
• Track your research