SECAI Core · phase 2 of 15 · noun
Transformer
A neural network architecture that excels at processing and analyzing sequences of data by focusing on the connections between different pieces of information in a sequence
The Explain card
- Plain English
- A transformer is a neural network architecture built for sequences. It uses attention, which lets every piece of the input weigh its connection to every other piece, rather than reading strictly in order.
- Example
- Nearly every modern large language model is a transformer. When it reads "the key was in the safe, and it was rusty", attention links "it" back to "key".
- Why it matters
- Attention is why prompts can be influenced from anywhere in the context window. Instructions buried deep in a pasted document can carry as much weight as the user's actual request.
- Hook
- Every word looks at every other word, and so can an injected one.
Word knowledge
How the word is built, where it came from, and what it sits beside in memory.
In a sentence
In a prompt-injection evaluation, testers insert an instruction into a retrieved PDF and inspect the attention weights the transformer assigns to those tokens.
How it is built
- trans- prefix across
- form root shape
- -er suffix one that does
Where it came from
- Origin
- Latin transformare, from trans (across) plus formare (to form), via Old French transformer; English agent noun in -er
- Entered the language
- c. 1600
- What changed
- The verb transform entered English in the late 1300s for a change of shape. Around 1600 the agent noun named a person who made that change. Electrical engineers applied it in the 1880s to the coil that steps voltage. In 2017 Vaswani and colleagues applied the same agent noun to their sequence model.
How it is spelled
- Pattern
- trans- keeps its s before a consonant
- Pattern
- final -rm does not double before -er
- Pattern
- agent -er attaches to the verb stem with no extra letter
- Breaks the pattern
- Latin and German keep -ator (transformator); English built transformer on the verb with native -er
Spelled like
- transfer
- transmit
- former
- performer
Broken into chunks
-
trans
- transfer
- transmit
- transparent
-
form
- formation
- reform
- conform
-
er
- encoder
- decoder
- performer
What it sits beside
Same subject
Same shape
- performer
- informer
- reformer
Neural network architectures
Latin trans- compounds
- transfer
- transmit
- transform
Agent nouns in -er
- encoder
- decoder
- tokenizer
Where it sits in the deck
Phase 2: How Models Are Built: Architectures and Learning Mechanics
With the paradigms named, zoom in on the architectural building blocks that turn data into trained artifacts.