Transformers
A neural architecture that uses attention to combine information across a sequence.
- Tokens
- Embeddings + position
- Attention + feed-forward
- Output
Key ideas
- Token
- Attention
- Position
- Layer
Learn next: Attention · Embeddings
Concepts
Understand the ideas behind modern AI.
Start with an overview, then sign in to explore the research and build it yourself.
10 concepts
Attention
Compute a weighted combination of information according to its relevance to a query.
- Query + keys
- Similarity scores
- Normalized weights
- Weighted values
Key ideas
- Query
- Key
- Value
- Weight
Learn next: Transformers
Embeddings
Represent items as numerical vectors so a model can compare and use them.
- Text or item
- Embedding model
- Vector
- Similarity comparison
Key ideas
- Vector
- Dimension
- Similarity
- Representation
Learn next: Vector Databases · Retrieval-Augmented Generation
Vector Databases
Store vectors and retrieve similar items, often alongside metadata filters.
- Documents
- Embeddings
- Vector index
- Query + filters
- Matches
Key ideas
- Index
- Nearest neighbor
- Metadata
- Recall
Learn next: Retrieval-Augmented Generation
Retrieval-Augmented Generation
Retrieve relevant external context before asking a model to generate an answer.
- Question
- Retrieve passages
- Context + model
- Answer + references
Key ideas
- Retriever
- Context
- Grounding
- Generation
Learn next: Vector Databases · Model Evaluation
AI Agents
Systems that select actions, use tools, observe results, and continue toward a goal.
- Goal
- Select action
- Tool
- Observe
- Continue or stop
Key ideas
- Goal
- Tool
- Observation
- Policy
- Stop condition
Learn next: Tool Calling · Agent Memory · Model Evaluation
Tool Calling
A model requests a structured operation; the application validates and executes it.
- Model request
- Validate arguments
- Authorize + execute
- Return result
Key ideas
- Schema
- Arguments
- Authorization
- Result
Learn next: AI Agents · Guardrails
Agent Memory
Preserve selected information so an agent can use it in later steps or sessions.
- Observation
- Select useful facts
- Store
- Retrieve into context
Key ideas
- Working context
- Persistence
- Retrieval
- Retention
Learn next: AI Agents · Guardrails
Model Evaluation
Measure whether an AI system performs its intended task under defined conditions.
- Task + criteria
- Test cases
- Run system
- Score + inspect
- Improve
Key ideas
- Test set
- Metric
- Baseline
- Failure case
Learn next: Guardrails · Retrieval-Augmented Generation
Guardrails
Application controls that constrain inputs, actions, and outputs of an AI system.
- Input checks
- Scoped tools
- Output checks
- Review or release
Key ideas
- Policy
- Validation
- Least privilege
- Human review
Learn next: Tool Calling · Model Evaluation