Concepts
The vocabulary of AI engineering, one animated page at a time.
Start hereLarge Language Models are built on neural networks — vast webs of connected nodes that learn patterns from text. Scroll to see how it all works, explained simply.
Read this one first: every other concept builds on it →
How models run (9)
Words become model-ready pieces
Route each token to a few specialists
Reuse past attention work
Draft quickly, verify in batches
Keep the GPU busy
One fixed budget, shared by everything
The same distribution, read two ways
Fewer bits per number, same shape
Put the unchanging part first
Retrieval (9)
Split documents by meaning
Rank documents by keyword relevance
Combine keyword and semantic retrieval
Turn conversation into a search query
Reorder candidates by relevance
Meaning, written down as coordinates
Close enough, much faster
Did the needed evidence appear in the first k?
Recall@k cannot see the order
Agents & tools (8)
Make the response match a schema
One standard interface for tools and context
Pick the right model for each request
Reason, act, observe, then decide again
The model asks; your code does it
Split what splits, and nothing else
Three different things wearing one name
Spend tokens thinking before answering
Running in production (7)
Save workflow state
Stop repeatedly calling a failing service
See the full request path
A retry must not do it twice
Say no early instead of failing late
Find out on 5%, not on everyone
Nothing changed, and it got worse
Evaluation & safety (6)
Use a model to grade another model's output
The cases you refuse to break
Fluent, confident, and not in the source
Checks on the way in and on the way out
Untrusted text starts giving the model orders
The user is the adversary this time