AI - not just generative
Unless you are involved in an AI project at work in some way or form or have a passion for the technology that led you down the rabbit hole, you might only think of AI as the generative type, best know for creating images, audio or text such as poems, summaries or the like.
However there are many different categories in the rapidly developing world of AI technology. Let’s look at these categories, the models they use, their typical real-world application and a few examples:
Creates new content like text, images, or music using patterns from training data.
Helps humans understand why an AI made a certain decision.
Classifies inputs into categories, like identifying spam or recognizing faces.
Learns actions by trial and error using rewards and penalties.
Enables machines to interpret and understand visual data like images or video.
Allows machines to understand and generate human language.
Forecasts future outcomes based on historical data.
Suggests content or products based on user preferences and behavior.
Uses predefined rules to make decisions like a human expert.
Performs tasks independently in real-world environments.
Trains models collaboratively without sharing raw data.
Automates the process of building, training, and optimizing machine learning models.
Uses logic and rules instead of learning from data.
Runs AI locally on devices without needing a cloud connection.
Mimics human thinking by combining learning and logic.
Advanced AI systems, that are specifically designed to perform complex, multi-step problem-solving tasks.
GenAI is designed to create new content, such as text, images, music, or even video. It learns from large datasets and uses that knowledge to produce outputs that resemble what it has seen. For example, it can write articles, generate realistic images, or compose music by mimicking patterns it has learned.
Models: GPT (Generative Pretrained Transformer), GANs (Generative Adversarial Networks), Variational Autoencoders (VAEs).
Applications: Content generation (text, images, music), synthetic data creation, personalized marketing materials, chatbots.
Examples: OpenAI ChatGPT, DALL·E, Midjourney, Jasper.ai, Synthesia, Runway ML.
XAI focuses on making the decisions made by AI systems understandable to humans. Instead of giving a result without context, XAI shows why and how an AI made a particular decision. This is especially important in fields like healthcare or finance where trust and transparency are critical.
Models: SHAP (SHapley Additive exPlanations), LIME (Local Interpretable Model-agnostic Explanations).
Applications: Regulatory compliance, healthcare diagnostics, credit scoring, fraud detection.
Examples: IBM AI Explainability 360, Fiddler AI, Google What-If Tool, Microsoft InterpretML.
This type of AI is trained to distinguish between different types or categories of data. It does this by learning to draw boundaries between different groups, like telling cats apart from dogs in photos. It's focused on classification and decision-making.
Models: Support Vector Machines (SVM), Logistic Regression, Convolutional Neural Networks (CNNs).
Applications: Spam filtering, facial recognition, medical diagnostics, credit approval.
Examples: Google Cloud Vision, Amazon Rekognition, Apple Face ID, Email spam classifiers.
Reinforcement learning teaches AI through a trial-and-error approach. It interacts with an environment and receives rewards or penalties based on its actions. Over time, it learns which actions lead to the best outcomes.
Models: Q-Learning, Deep Q-Networks (DQN), Proximal Policy Optimization (PPO).
Applications: Game playing (AlphaGo), robotics, real-time bidding, dynamic pricing.
Examples: DeepMind AlphaGo & AlphaZero, OpenAI Five, Unity ML-Agents, Tesla Autopilot.
This field of AI helps machines to see and understand images and videos. It mimics human vision by recognizing objects, people, or patterns in visual data, allowing AI to analyze and respond to visual inputs.
Models: CNNs, YOLO (You Only Look Once), ResNet, VGG.
Applications: Autonomous vehicles, security cameras, medical imaging, augmented reality.
Examples: Google Lens, Amazon Rekognition, Zebra Medical Vision, Meta's AR Glasses.
NLP enables computers to understand, interpret, and generate human language. It allows AI to read, listen, and respond in a way that mimics human conversation and comprehension.
Models: Transformer-based models (GPT, BERT), RNNs, LSTMs.
Applications: Virtual assistants, chatbots, translation, sentiment analysis, information extraction.
Examples: OpenAI GPT-4, Google BERT, Grammarly, Google Translate, IBM Watson NLP.
Description: Predictive AI analyzes past data to make informed guesses about future events. For instance, it can forecast sales, detect potential failures in machinery, or predict customer behavior.
Models: Regression models, Time-Series Forecasting (ARIMA, Prophet), Gradient Boosting Machines (XGBoost).
Applications: Demand forecasting, financial market predictions, predictive maintenance, risk management.
Examples: Salesforce Einstein, IBM SPSS, Amazon Forecast, Microsoft Azure ML.
These AI systems suggest products, content, or services based on user preferences and past behavior. They help personalize experiences, like recommending movies on Netflix or products on Amazon.
Models: Collaborative Filtering, Matrix Factorization, Deep Learning recommendation models.
Applications: E-commerce, streaming platforms, personalized content recommendations.
Examples: Netflix recommendation engine, Amazon product suggestions, Spotify Discover Weekly, YouTube Up Next.
Expert systems use a set of pre-programmed rules to simulate human expert decision-making. They work like digital consultants, offering advice based on logical reasoning.
Models: Rule-based logic, decision trees, knowledge-based inference engines.
Applications: Medical diagnosis, financial advisory systems, customer service automation.
Examples: MYCIN (medical diagnosis), IBM Watson for Oncology, Pegasystems Decision Hub.
Self-operating AI systems that can perform tasks on their own, often in real time and in changing environments. They combine sensing, thinking, and acting without human input.
Models: Sensor fusion models, real-time decision-making algorithms, RL-based navigation.
Applications: Self-driving cars, drones, automated industrial machinery.
Examples: Waymo autonomous vehicles, Tesla FSD, Boston Dynamics robots, Skydio drones.
Federated learning allows AI to learn from data across many devices or organizations without the data ever leaving its original location. This protects privacy while still training effective models.
Models: Federated Averaging, secure aggregation techniques.
Applications: Privacy-sensitive healthcare AI, mobile device learning, collaborative financial services.
Examples: Google Gboard (mobile keyboard), NVIDIA Clara for healthcare, Flower framework, OpenMined.
AutoML automates the process of building, training, and optimizing machine learning models. It enables users—especially non-data scientists—to develop high-performing models without needing to manually code or tune parameters.
Models: Meta-learning frameworks, Neural Architecture Search (NAS), Tree-based AutoML (e.g., TPOT, H2O AutoML), Google AutoML.
Applications: Customer churn prediction, fraud detection, demand forecasting, healthcare diagnostics.
Examples: Google Vertex AI (AutoML), Microsoft Azure AutoML, H2O.ai, Amazon SageMaker Autopilot, Aible.
Symbolic AI uses predefined rules and logic to simulate intelligent behavior. It's like teaching a computer step-by-step instructions to solve problems or understand symbols and language.
Models: Logic programming, Prolog, semantic networks.
Applications: Knowledge management systems, complex rule-based decisions, theorem proving.
Examples: Cyc knowledge base, Wolfram Alpha, CLIPS, expert rules in SAP systems.
Edge AI brings intelligence to devices right where data is generated (like phones or sensors), instead of sending data to the cloud. This makes decisions faster and keeps sensitive data local.
Models: Lightweight CNNs (MobileNet, EfficientNet), TinyML models.
Applications: IoT devices, smart appliances, real-time analytics, wearables.
Examples: Apple Neural Engine, Google Coral, AWS Panorama, NVIDIA Jetson.
Cognitive AI tries to replicate human thinking processes, such as reasoning, problem-solving, and learning from experience. It blends machine learning with symbolic reasoning to create systems that can adapt and think more like people.
Models: Neural-symbolic integration, cognitive architectures (e.g., SOAR, ACT-R).
Applications: Virtual assistants with human-like interactions, advanced decision support, adaptive learning systems.
Examples: IBM Watson Assistant, Microsoft Cognitive Services, Expert System Cogito, OpenCog.
Description: Reasoning models are designed to emulate logical, human-like problem-solving. They can break complex tasks into step-by-step solutions, apply inference rules, and maintain contextual understanding across multiple steps. Unlike traditional pattern-matching models, they simulate structured thinking processes.
Models: Chain-of-thought prompting, ReAct (Reasoning + Acting), Tree of Thoughts, Meta-reasoning frameworks, and LLMs with structured output.
Applications: Multi-step decision making, data-to-insight transformation, business agent orchestration, scientific and financial reasoning.
Examples: ChatGPT with tool use, Aible Agents, Google's PaLM 2 reasoning tasks, OpenAI function calling, Microsoft AutoAgents.
AI Agents are not really a category of their own. They are autonomous systems that:
Perceive their environment (via data or sensors),
Reason about that data (via logic or learning),
Take actions toward a goal (often in a loop),
Learn and adapt over time.
As such, they often combine multiple AI technologies to achieve this:
| AI Category | Role in AI Agents |
|---|---|
| Generative AI | Allows agents to create responses, content, or even code to fulfill tasks. |
| Natural Language Processing | Allows agents to understand instructions and interact using human language. |
| Reasoning Models | Provides step-by-step logic and planning capabilities to solve complex tasks. |
| Reinforcement Learning | Helps agents learn optimal actions through trial and error in interactive settings. |
| Computer Vision | Supports agents in understanding visual input when required (e.g., from images). |
| Predictive AI | Allows agents to forecast outcomes and make proactive decisions. |
| Expert Systems / Symbolic AI | Provide structured reasoning and logic when rule-based decisions are needed. |
| Autonomous AI | Describes the agent's ability to operate independently across tasks or systems. |
| AutoML | Helps agents select, train, and optimize models on the fly, without human input. |
| Explainable AI | Enhances agent transparency by clarifying how decisions are made. |
| Cognitive AI | Empowers agents with more human-like reasoning and memory across steps. |
Many AI platforms today draw from multiple AI categories and often incorporate AI agents, further expanding their capabilities.
AI is definitely not just your friendly Chatbot that can answer questions and paint you a picture. And the list above is by no means complete. The more we put AI to work, the more categories will emerge. It’s an exciting, fascinating, and fast-moving field—and I’m looking forward to seeing even more clever and beneficial applications of this versatile technology.
