Describes upper bound of time/space complexity of an algorithm.

Bubble sort is O(n²).

Used in system design to estimate scalability.

Object-Oriented Programming: paradigm using classes and objects.

class Dog: pass

Used in production codebases for maintainability.

Ability to present the same interface for different data types.

def area(shape): return shape.area()

Interfaces in APIs, pluggable modules.

Solving by breaking into overlapping subproblems and storing results.

Fibonacci with memoization.

Optimization, resource allocation.

Binary Search Tree: tree where left < root < right.

Insert/search in O(log n) average.

Databases, sorted data structures.

Data structure mapping keys to values with O(1) average lookup.

Learning from labeled input-output pairs.

Learning patterns from unlabeled data.

Model fits training data too closely and fails to generalize.

Algorithm for computing gradients in neural networks.

Convolutional Neural Network: for grid-like data (e.g. images).

Recurrent Neural Network: for sequences.

Mechanism to weigh relevance of different inputs.

Architecture based on self-attention, no recurrence. Enables parallel processing of sequences.

BERT, GPT, T5 architectures.

ChatGPT, translation, summarization.

Bidirectional Encoder Representations from Transformers.

Generative Pre-trained Transformer: autoregressive language model.

Retrieval Augmented Generation: augmenting LLMs with retrieved docs.

LLM generating plausible but incorrect or nonsensical content.

Database optimized for similarity search on embeddings. Stores vectors and supports nearest-neighbor queries.

Pinecone, Weaviate, pgvector.

RAG systems, semantic search.

Reinforcement Learning from Human Feedback. Fine-tuning LLMs using human preferences.

ChatGPT alignment.

Alignment of generative models.

AI that uses tools and multi-step reasoning to accomplish goals.

ReAct, Tool Use agents.

Copilots, autonomous assistants.

Reasoning + Acting: pattern for agent decision-making.

Practices for deploying and maintaining ML systems.

Change in input distribution over time affecting model performance.

Responsible, Accountable, Consulted, Informed: responsibility matrix.

Technique for array/string subarray problems with O(n) time.

Technique using two indices to traverse array/string efficiently.

Least Recently Used cache: evicts oldest unused item when full.

Tree for storing strings; supports prefix search efficiently.

Caching computed results to avoid redundant work in recursion.

Service Level Objective: target for reliability (e.g. 99.9% uptime).

Service Level Agreement: contract with customers on service guarantees.