AI terms.
Plain English first.

AI includes systems built from explicit rules as well as systems trained on data. Calling something AI does not mean it is conscious, understands like a person, or can handle every task.

Distinguish the task, the method, and the evaluation. A rule-based planner and a learned classifier can both be AI systems; neither implies artificial general intelligence.

Training adjusts a model using data. Testing checks how well it works on examples it did not train on. Learning during training does not mean the system automatically learns from every later conversation.

Supervised learning uses target labels; unsupervised learning looks for structure; reinforcement learning uses rewards. Evaluate generalization, data leakage, distribution shift, and the cost of different errors.

Think of a series of adjustable filters. Each layer transforms the information passed to the next. The brain inspired the name, but a neural network is not a digital copy of a human brain.

Units apply learned weights and usually nonlinear transformations. Training commonly uses backpropagation to compute gradients and an optimizer to update parameters. Neural networks can be shallow or deep.

Early layers can capture simpler features; later layers can combine them into more useful representations. Deep learning is a type of neural-network learning, not a container that includes all neural networks.

Depth refers to successive learned transformations, not human-like depth of thought. Performance depends on architecture, data, optimization, and evaluation; more layers alone do not guarantee better results.

An attention mechanism weighs relationships between pieces of information. Transformers can process text, images, audio, and other data. They are not limited to chatbots.

Attention relates query, key, and value representations. Architectures may use encoder blocks, decoder blocks, or both, with positional information and feed-forward layers. Attention weights alone are not a reliable explanation of a model’s reasoning.

The output can be useful without being accurate. A convincing sentence or realistic image is not evidence that the underlying claim or event is real. Generative AI is a capability, not one particular model design.

Generative approaches include autoregressive models, diffusion models, variational autoencoders, and generative adversarial networks. Transformer-based systems are one important family, not the entire category.

Many familiar chat models generate text by predicting the next token, a small piece of text, repeatedly. Their answers can reflect useful learned patterns and still contain invented details.

Separate pretraining, later instruction or preference training, and inference, when the model is used. Context windows limit what fits in a request. Retrieval and tools can supply outside information without changing model weights. Some models also accept images or audio.

Generative means it produces output. Pre-trained means its initial learning happened before your request. Transformer names its underlying architecture. GPT is not a synonym for every AI assistant.

The GPT lineage uses autoregressive transformer language modeling. Later training and product tools affect behavior. A custom “GPT” inside ChatGPT is a configured assistant, which is a different use of the term from an underlying model family.

The application is the experience you use; the model is part of the technology powering it. Available tools and features depend on the service and settings.

Evaluate the complete system: model selection, instructions, conversation context, retrieval, tools, and permissions. A product name does not identify one fixed model or guarantee live web access.

Claude is a separate provider’s system. It is not an OpenAI GPT product, even though it can perform many similar tasks.

Distinguish the Claude application from a particular Claude model and an API integration. Document handling, tool access, retention settings, and limits depend on the product configuration.

Gemini is Google’s system, not an OpenAI GPT product. Some experiences can work with several kinds of input, such as text and images.

Distinguish Gemini models from the Gemini application and integrations in other Google products. Multimodal capability, account permissions, and enabled tools are separate considerations.