Concepts · Definitions · Insights · Interview Prep
Transformers, LLMs, RAG, Agents, and the generative AI stack
Classical ML, deep learning, training patterns, and foundations
Data engineering, pipelines, and data-centric AI
Linear algebra, calculus, probability for ML
The share of predictions a model gets right, usually written as a percentage. It's the simplest way to summarize how well a classifier is doing.
Analogy: It's a report-card percentage — 90% accuracy means 9 out of 10 calls were correct.
Related: loss, softmax
The little non-linear function each neuron applies to its input, letting a network learn curved, complex patterns instead of just straight lines. Without it, stacking layers would be pointless.
Analogy: The bend in the wire — without a kink, ten straight segments are still just one straight line.
Related: neural-network, softmax, gradient
Read the full conceptA mechanism that lets a model weigh which parts of the input matter most for each piece it's processing. Instead of treating every word equally, it focuses on the ones that are actually relevant.
Analogy: Like reading with a highlighter — you weight the important words and skim the rest.
Related: softmax, embedding
Read the full conceptThe number of training examples a model looks at before it updates its weights once. Bigger batches give steadier updates but need more memory; smaller batches update more often and add helpful noise.
Analogy: It's how many flashcards you review before pausing to update your notes.
Related: epoch, gradient, loss
The maximum amount of text — measured in tokens — a language model can consider at once, covering both your prompt and its reply. Go past it and the earliest tokens fall out of view.
Analogy: It's the model's desk space — only so much paper fits on top at one time.
Related: token, prompt, attention
Read the full conceptA way to test a model by splitting the data into several folds, training on most and testing on the rest, then rotating so every fold gets a turn as the test set. You get a sturdier estimate than a single split.
Analogy: Grading with several different pop quizzes instead of betting everything on one big exam.
Related: overfitting, underfitting, accuracy
Amazon Bedrock, Azure OpenAI, and Google Vertex AI all sell the same promise: call a frontier model over an API and skip the GPUs. Here's how they actually work, and how to pick one.
Every big model launch arrives wrapped in benchmark charts and superlatives. Here's the short checklist that tells you whether a release actually matters for what you're building.
Quick answer
K-means groups points into k clusters by alternating two steps until they stop changing: assign each point to its nearest centroid, then move each centroid to the mean of the points assigned to it. It's simple and fast, but you have to pick k, it only finds a local optimum so initialization matters (use k-means++), and it assumes roughly round, similarly-sized clusters.
Quick answer
Softmax turns a vector of raw scores into a probability distribution: exponentiate each value, then divide by the sum of the exponentials. The one thing you must get right is numerical stability — subtract the max value before exponentiating so you never overflow on large inputs. The result is unchanged mathematically but safe in floating point.
Quick answer
In a deep network, backpropagation multiplies many small local derivatives together, so the gradient shrinks exponentially as it flows back toward the early layers — those layers barely update and the network stops learning. You fix it with ReLU-style activations, residual (skip) connections, good weight initialization, and normalization, all of which keep the gradient from collapsing.
Explore our full library of concepts, definitions, and tutorials. Let's build something amazing.