Understanding AI: From Concepts to Capabilities

Artificial intelligence (AI) is a transformative technology that empowers computers and machines to mimic and execute cognitive functions traditionally associated with human minds. These functions include learning, comprehending, problem-solving, decision-making, creativity, and even operating with a degree of autonomy. AI-driven applications and devices possess the remarkable ability to perceive and identify objects, understand and respond to human language, learn from new information and experiences, and provide detailed recommendations. A classic example of AI in action is a self-driving car, which operates independently without direct human intervention.

In 2024, a significant portion of the AI landscape, particularly in research and industry headlines, revolves around breakthroughs in generative AI (gen AI). This advanced form of AI is capable of creating original content such as text, images, video, and more in response to user prompts. To truly grasp generative AI, it’s essential to first understand its foundational technologies: machine learning (ML) and deep learning.

The Building Blocks of AI: Machine Learning and Deep Learning

Think of AI as a series of interconnected concepts that have evolved over more than 70 years. At its core, underneath the broader umbrella of AI, lies machine learning.

Machine Learning (ML)

Machine learning involves developing models by training algorithms to make predictions or decisions based on data. It encompasses a wide array of techniques that allow computers to learn from and make inferences based on data without being explicitly programmed for every specific task. Various machine learning algorithms exist, including linear regression, logistic regression, decision trees, and neural networks. Each is suited to different types of problems and data sets.

One of the most prominent types of machine learning algorithms is the neural network (or artificial neural network). Inspired by the human brain’s structure and function, neural networks consist of interconnected layers of nodes that collaboratively process and analyze complex data. They are particularly effective at identifying intricate patterns and relationships within large datasets.

The simplest form of machine learning is supervised learning, where algorithms are trained on labeled datasets to accurately classify data or predict outcomes. In this method, each training example is paired with an output label by humans, enabling the model to learn the mapping between inputs and outputs and predict labels for new, unseen data.

Deep Learning

Deep learning is a specialized subset of machine learning that utilizes multi-layered neural networks, known as deep neural networks. These networks more closely simulate the complex decision-making power of the human brain. Unlike classic machine learning models, which typically have one or two hidden layers, deep neural networks include an input layer, at least three, and often hundreds, of hidden layers, and an output layer.

These numerous layers facilitate unsupervised learning, allowing the networks to automatically extract features from large, unlabeled, and unstructured datasets and make their own predictions about what the data represents. Because deep learning minimizes the need for human intervention, it enables machine learning at an immense scale. It excels in tasks such as natural language processing (NLP) and computer vision, where fast and accurate identification of complex patterns in vast amounts of data is crucial. Most of the AI applications we interact with daily are powered by some form of deep learning.

Deep learning also underpins:

Transfer learning: Knowledge gained from one task or dataset is applied to improve performance on a related task or different dataset.

Semi-supervised learning: Combines labeled and unlabeled data for training AI models in classification and regression.

Self-supervised learning: Generates implicit labels from unstructured data, reducing reliance on explicit labeling.

Reinforcement learning: Models learn through trial-and-error and reward functions rather than explicit pattern extraction.

Generative AI: Creating Original Content

Generative AI (gen AI) refers to deep learning models capable of creating complex, original content like long-form text, high-quality images, realistic video, or audio in response to a user’s prompt. Fundamentally, generative models encode a simplified representation of their training data and then draw upon this representation to produce new work that is similar to, but not identical to, the original.

While generative models have been used in statistics for years to analyze numerical data, their evolution over the past decade has enabled them to analyze and generate more complex data types. This advancement coincided with the emergence of three sophisticated deep learning model types:

• Variational Autoencoders (VAEs): Introduced in 2013, these models can generate multiple variations of content from a single prompt.
• Diffusion Models: First appearing in 2014, these models generate original images by iteratively adding and then removing “noise” from images.
• Transformers (or Transformer Models): Trained on sequenced data, transformers generate extended sequences of content (like words in sentences or frames in a video). These models are at the heart of many prominent generative AI tools today, including ChatGPT, GPT-4, Copilot, BERT, Bard, and Midjourney.

How Generative AI Works

Generative AI typically operates in three phases: training, tuning, and generation/evaluation/more tuning.

Generation, Evaluation, and More Tuning: Developers and users continuously assess the outputs of their generative AI applications, tuning the model for greater accuracy or relevance, sometimes even weekly. The foundation model itself is updated less frequently, perhaps annually or every 18 months. Another technique for improving performance is retrieval augmented generation (RAG), which extends the foundation model to use external, relevant sources beyond its training data for more accurate or relevant outputs.

Training: This phase involves creating a foundation model. These are deep learning models that serve as a base for various generative AI applications. The most common are Large Language Models (LLMs) for text generation, but foundation models also exist for image, video, sound, or multimodal content. To build a foundation model, a deep learning algorithm is trained on vast amounts of raw, unstructured, and unlabeled data (e.g., terabytes of text or images from the internet). This intensive process yields a neural network with billions of parameters, allowing it to generate content autonomously in response to prompts. This training is computationally intensive, time-consuming, and expensive, often costing millions of dollars and requiring thousands of clustered GPUs for weeks. Open-source foundation models, like Meta’s Llama-2, help mitigate these costs for developers.

Tuning: After training, the model is refined for a specific content generation task. This can involve:

Fine-tuning: Feeding the model application-specific labeled data, likely questions, and corresponding correct answers.

Reinforcement Learning with Human Feedback (RLHF): Human users evaluate the accuracy or relevance of model outputs, allowing the model to improve itself (e.g., correcting a chatbot’s response).

AI Agents and Agentic AI: The Next Evolution

An AI agent is an autonomous AI program designed to perform tasks and achieve goals on behalf of a user or another system without human intervention. These agents can design their own workflows and utilize available tools (other applications or services). Agentic AI refers to a system of multiple AI agents whose efforts are coordinated to accomplish more complex tasks or greater goals than any single agent could achieve.

Unlike chatbots and other AI models that operate within predefined constraints and often require human intervention, AI agents and agentic AI exhibit autonomy, goal-driven behavior, and adaptability to changing circumstances. This concept highlights their “agency,” or their capacity for independent, purposeful action.

One way to understand agents is as a natural progression from generative AI. While gen AI models focus on creating content based on learned patterns, agents leverage that content to interact with each other and other tools to make decisions, solve problems, and complete tasks. For instance, a gen AI app might tell you the best time to climb Mt. Everest, but an AI agent could then use an online travel service to book your flight and hotel.

Curabitur varius eros et lacus rutrum consequat. Mauris sollicitudin enim condimentum, luctus enim justo non, molestie nisl.

The Benefits of AI

AI offers numerous advantages across various industries and applications. Key benefits include:

• Automation of repetitive tasks: AI automates routine digital tasks like data collection and preprocessing, as well as physical tasks like warehouse stock-picking, freeing humans for higher-value, more creative work.
• Enhanced decision-making: AI facilitates faster, more accurate predictions and reliable, data-driven decisions. Combined with automation, it allows businesses to act on opportunities and crises in real-time.
• Fewer human errors: AI reduces errors by guiding processes, flagging potential mistakes, and fully automating tasks, especially critical in fields like healthcare where AI-guided surgical robotics ensure precision. Machine learning algorithms continuously improve accuracy with more data and experience.
• 24/7 availability and consistency: AI is always on, providing consistent performance. AI chatbots and virtual assistants can reduce staffing demands for customer service, while in manufacturing, AI maintains consistent work quality and output.
• Reduced physical risks: By automating dangerous work (e.g., handling explosives, tasks in extreme environments), AI eliminates the need to put human workers at risk. Self-driving vehicles, though still evolving, aim to reduce injury risks for passengers.

Real-World Applications of AI

AI’s real-world applications are extensive. Here’s a brief look at some common use cases across industries:

• Customer experience, service, and support: AI-powered chatbots and virtual assistants handle inquiries using NLP and generative AI to understand and respond to customer questions, providing 24/7 support and freeing human agents for complex issues.
• Fraud detection: Machine learning and deep learning algorithms analyze transaction patterns to flag anomalies (e.g., unusual spending or login locations) indicative of fraud, enabling quicker responses and limiting impact.
• Personalized marketing: Retailers and banks use AI to create personalized customer experiences and marketing campaigns. Deep learning algorithms recommend products and services based on purchase history and behavior, even generating personalized copy and offers.
• Human resources and recruitment: AI-driven platforms streamline hiring by screening resumes, matching candidates to jobs, and conducting preliminary interviews via video analysis, reducing administrative burden and improving response times.
• Application development and modernization: Generative AI code generation and automation tools streamline repetitive coding tasks, accelerating the migration and modernization of legacy applications, ensuring code consistency and reducing errors.
• Predictive maintenance: Machine learning models analyze data from sensors and IoT devices to forecast maintenance needs and predict equipment failures, preventing downtime and addressing supply chain issues proactively.

Challenges and Risks in AI Adoption

While organizations are eager to embrace AI’s benefits, rapid adoption comes with inherent challenges and risks:

• Data risks: AI systems rely on datasets vulnerable to data poisoning, tampering, bias, or cyberattacks leading to breaches. Mitigating these requires protecting data integrity and implementing security throughout the entire AI lifecycle.
• Model risks: AI models can be targeted for theft, reverse engineering, or unauthorized manipulation. Attackers might compromise model integrity by tampering with architecture, weights, or parameters.
• Operational risks: Models are susceptible to operational issues like model drift, bias, and governance breakdowns, which can lead to system failures and cybersecurity vulnerabilities.
• Ethics and legal risks: Without prioritizing safety and ethics, organizations risk privacy violations and biased outcomes. For example, biased training data in hiring can reinforce stereotypes, creating models that favor certain demographic groups.

AI Ethics and Governance: Ensuring Responsible AI

AI ethics is a multidisciplinary field focused on maximizing AI’s beneficial impact while minimizing risks. Principles of AI ethics are put into practice through AI governance, a system of safeguards ensuring AI tools and systems remain safe and ethical. Effective AI governance requires input from a diverse range of stakeholders, including developers, users, policymakers, and ethicists, to align AI systems with societal values.

Common values associated with AI ethics and responsible AI include:

• Explainability and interpretability: As AI becomes more advanced, understanding its decision-making process is challenging. Explainable AI provides processes and methods to help humans interpret, comprehend, and trust algorithmic results.
• Fairness and inclusion: While machine learning inherently involves statistical discrimination, it becomes problematic when it systematically disadvantages certain groups. Promoting fairness involves minimizing algorithmic bias in data collection and model design, and building diverse teams.
• Robustness and security: Robust AI effectively handles unusual conditions or malicious attacks without causing harm and is built to resist intentional and unintentional interference by protecting against vulnerabilities.
• Accountability and transparency: Organizations should establish clear responsibilities and governance for AI system development, deployment, and outcomes. Users should understand how an AI service works, its functionality, and its limitations.
• Privacy and compliance: Many regulatory frameworks (e.g., GDPR) mandate privacy principles for personal information processing. It’s crucial to protect AI models containing personal data, control data input, and build adaptable systems for changing regulations and ethical views.

Weak AI vs. Strong AI: Levels of Sophistication

Researchers categorize AI by its level of complexity and sophistication:

• Weak AI (Narrow AI): These AI systems are designed for specific tasks or a limited set of tasks. Examples include voice assistants like Alexa or Siri, social media chatbots, or autonomous vehicles. Most current AI applications fall into this category.
• Strong AI (Artificial General Intelligence – AGI): Also known as “general AI,” AGI would possess the ability to understand, learn, and apply knowledge across a wide range of tasks at or surpassing human intelligence. This level of AI is currently theoretical, and no known AI systems approach this sophistication. Researchers believe AGI, if possible, requires significant increases in computing power, and some predict it may not emerge for decades, or even centuries.

The Evolution of AI: A Historical Perspective

The concept of a “thinking machine” dates back to ancient Greece. However, significant milestones in AI’s evolution since the advent of electronic computing include:

• 1950: Alan Turing publishes “Computing Machinery and Intelligence,” posing the question, “Can machines think?” and introducing the “Turing Test.”
• 1956: John McCarthy coins the term “artificial intelligence” at the first AI conference at Dartmouth College. Allen Newell, J.C. Shaw, and Herbert Simon create the Logic Theorist, the first running AI computer program.
• 1967: Frank Rosenblatt builds the Mark 1 Perceptron, a neural network-based computer that learned through trial and error.
• 1980: Neural networks, using backpropagation algorithms for self-training, become widely used in AI applications.
• 1995: Stuart Russell and Peter Norvig publish “Artificial Intelligence: A Modern Approach,” a leading AI textbook defining four potential goals of AI.
• 1997: IBM’s Deep Blue defeats then-world chess champion Garry Kasparov.
• 2004: John McCarthy proposes an often-cited definition of AI. The era of big data and cloud computing begins, enabling the large datasets needed for AI model training.
• 2011: IBM Watson® beats Jeopardy! champions Ken Jennings and Brad Rutter. Data science emerges as a popular discipline.
• 2015: Baidu’s Minwa supercomputer, using a convolutional neural network, identifies and categorizes images with higher accuracy than the average human.
• 2016: DeepMind’s AlphaGo program, powered by a deep neural network, defeats world Go champion Lee Sodol, a significant victory due to the game’s vast number of possible moves.
• 2022: The rise of Large Language Models (LLMs) like OpenAI’s ChatGPT dramatically enhances AI performance and its potential for enterprise value. These generative AI practices enable deep-learning models to be pretrained on massive amounts of data.
• 2024: Current AI trends indicate a continuing renaissance, with multimodal models (combining computer vision and NLP) offering richer experiences, and smaller models making strides.

Expanding Use of AI in Business

AI is profoundly impacting businesses, with some companies significantly outpacing others in adoption. A 2022 survey indicated that AI model adoption has more than doubled since 2017, with corresponding increases in investment. The areas where companies derive value from AI have evolved, now extending beyond manufacturing and risk to include:

• Marketing and sales
• Product and service development
• Strategy and corporate finance

Leading organizations consistently invest more in AI, scale practices faster, and recruit and upskill top AI talent. They link AI strategy to business outcomes and “industrialize” AI operations by designing modular data architectures that can quickly accommodate new applications.

Limitations of AI Models and Overcoming Them

Despite their capabilities, generative AI models have inherent risks and limitations:

• Convincing but incorrect outputs: Gen AI outputs can sound highly convincing even when the information is wrong or biased (due to biases present in internet data and society).
• Potential for misuse: Models can be manipulated for unethical or criminal activities (e.g., “jailbreaking” to generate inappropriate content). Gen AI organizations are responding by collecting user feedback on inappropriate content and training models against such generations.
• Reputational and legal risks: Organizations relying on gen AI risk unintentionally publishing biased, offensive, or copyrighted content.

These risks can be mitigated by:

• Careful data selection: Meticulously choosing initial training data to avoid toxic or biased content.
• Specialized or customized models: Using smaller, specialized models or customizing general models with proprietary data to minimize biases.
• Human in the loop: Ensuring a human reviews gen AI outputs before publication or use, especially for critical decisions.

As generative AI integrates further into business and daily life, the landscape of risks and opportunities will continue to evolve rapidly, leading to new regulatory environments. Organizations should proactively implement “no-regrets” actions to avoid legal, reputational, and financial risks.

The AI Bill of Rights and Global AI Governance

The Blueprint for an AI Bill of Rights, published by the US government in 2022, provides a framework for accountable AI, addressing concerns about transparency, bias, intellectual property, and privacy. It outlines five core principles:

1. The right to safe and effective systems: Systems should undergo pre-deployment testing, risk identification, mitigation, and ongoing monitoring.
2. Protections against discrimination by algorithms: Automated systems should not contribute to unjustified differential treatment based on race, color, ethnicity, sex, religion, age, etc.
3. Protections against abusive data practices: Built-in safeguards and user agency over data usage.
4. The right to know that an automated system is being used: Clear explanations of how and why it affects the user.
5. The right to opt out: Access to a human for problem resolution.

Currently, over 60 countries or blocs, including Brazil, China, the EU, Singapore, South Korea, and the US, have national strategies for responsible AI use. Approaches range from guidelines (like the US Blueprint) to comprehensive regulations (like the EU’s AI Act, due in 2024). Collaborative international efforts, such as the US–EU Trade and Technology Council and the Global Partnership on Artificial Intelligence (with 29 members including Brazil, Canada, Japan, and several European countries), are also working towards setting AI standards.

Even with evolving regulations, organizations should take preemptive action to manage risks:

Individual rights: Inform users when they are interacting with an AI system and provide clear usage instructions.

Transparency: Inventory and classify AI models, recording all usage clearly for internal and external stakeholders.

Governance: Implement a strong AI governance structure with oversight, authority, and accountability for both internal operations and third parties/regulators.

Data, model, and technology management:

Data management: Be aware of data sources, classification, quality, lineage, intellectual property, and privacy.

Model management: Establish principles and guardrails for AI development to ensure fairness and bias controls.

Cybersecurity and technology management: Ensure a secure environment to prevent unauthorized access or misuse.

Scaling AI: From Projects to Full Integration

Many organizations are still cautiously approaching AI. Slow progress towards widespread adoption often stems from cultural and organizational barriers. However, leaders who effectively overcome these barriers will be best positioned to seize AI’s opportunities. Companies failing to fully leverage AI are already being sidelined by competitors in industries like auto manufacturing and financial services.

To scale AI initiatives, organizations should make three major shifts:

1. Move from siloed work to interdisciplinary collaboration: AI’s biggest impact comes when cross-functional teams with diverse skills and perspectives address broad business priorities.
2. Empower frontline data-based decision-making: AI can enable faster, better decisions at all organizational levels. This requires trust in algorithms’ suggestions and empowering individuals to make decisions (while also allowing them to override or suggest improvements).
3. Adopt and bolster an agile mindset: An agile test-and-learn approach reframes mistakes as learning opportunities, alleviating fear of failure and accelerating development.