Core Concepts of AI

Image generated with DALL-E by OpenAI

What is AI?

Artificial intelligence is a set of methods that produce useful behavior from software systems. Common tasks include classification, prediction, ranking, search, planning, text generation, image generation, speech processing, and anomaly detection. Most modern AI systems learn from data during training rather than relying on hand written rules.

Types of AI: narrow AI and AGI

Most AI in use today is narrow AI. It performs a specific task or a bounded set of tasks, such as speech recognition, recommendations, fraud detection, image classification, or game play. Narrow systems often perform well when inputs match training conditions and evaluation data. Performance drops when conditions change or when the system is pushed outside its training scope.

Artificial general intelligence refers to a system that can learn and perform across many tasks at a human level with reliable transfer to new tasks and settings. AGI is not a deployed engineering reality today. Definitions vary, but common requirements include broad capability, stable performance under new conditions, and the ability to learn new tasks without task specific retraining.

Machine learning

Machine learning is a set of techniques where a model learns patterns from data and uses those patterns to produce outputs. Training typically means selecting a model family, defining an objective, and adjusting model parameters to reduce error on training data while checking generalization on held out data.

Common categories include supervised learning, unsupervised learning, self supervised learning, and reinforcement learning.

Supervised learning uses labeled examples where inputs have target outputs. It is widely used for tasks such as spam detection, medical image classification, speech to text, and quality inspection.

Unsupervised learning uses unlabeled data to find structure. Common uses include clustering, anomaly detection, and representation learning.

Self supervised learning trains on raw data by creating training targets from the data itself, such as predicting missing or next parts of a sequence. Many large language models and vision models use this approach because it scales with large amounts of unlabeled data.

Reinforcement learning trains an agent through interaction with an environment using rewards. It is used in control problems, robotics research, and some game settings. Many deployed systems combine reinforcement learning with supervised or self supervised training rather than relying on reinforcement learning alone.

Deep learning and neural networks

Deep learning is a subset of machine learning that uses neural networks with multiple layers. These models can learn useful representations from large datasets and perform well on perception and generation tasks.

A neural network consists of layers of parameters that map inputs to outputs. Training adjusts these parameters to reduce error on an objective, commonly through gradient based optimization.

Common architectures include convolutional neural networks, recurrent neural networks, and transformers.

Convolutional neural networks are widely used for images and other grid like data.

Recurrent neural networks were used for sequence data, though they are less common in current large scale systems.

Transformers are widely used for language and are now common in vision, audio, and multimodal systems. Many modern language models are transformer based.

Natural language processing

Natural language processing focuses on systems that process text and speech. Common tasks include translation, summarization, classification, extraction, retrieval, and question answering.

Modern NLP often uses pretrained transformer models trained on large corpora, then adapted to specific tasks or used through prompting. These systems can produce fluent output and follow many instructions, but they can also produce incorrect statements, misread context, or follow malicious instructions when placed in tool using settings.

Computer vision

Computer vision focuses on systems that interpret images and video. Tasks include classification, object detection, segmentation, tracking, and image generation.

Deep learning models often learn features directly from pixels. Performance can drop under changes in lighting, viewpoint, image quality, or data distribution. For real deployments, evaluation often needs tests for drift, edge cases, and failure modes, not only accuracy on a single benchmark.

Robotics

Robotics combines software with sensors and actuators to act in the physical world. AI methods support perception, planning, and control.

Many robots run models trained offline, then execute policies with safety checks and control logic. Robotics work often exposes constraints that are less visible in purely digital systems, including uncertainty in sensing, partial observability, safety requirements, and the cost of real world errors.

Where this fits

AI systems are defined by their models, data, compute budgets, evaluation practices, and deployment conditions. These elements set limits on reliability and use. When systems read untrusted input or act through tools, security and misuse become part of the engineering work.