STEP 01
Prompt from User
The user supplies text or another input to the interface.
Related HAAW lessons
↑ Back to navigationHow AI's Actually Work — HAAW
From User Prompt to Screen
19 learning checkpoints in the text-generation pipeline. Select a term in the navigation to jump to its explanation; each card links to the existing HAAW lessons that explore the idea in greater depth.
This is a teaching sequence, not a claim that every term is a distinct software component. In particular, “Translator” and “Embedded ID” are mnemonic labels explained in their cards.
← Full HAAW indexSTEP 02
Translator
The tokenizer converts incoming text into pieces the model can process. “Translator” is a learning shorthand, not a separate language translator.
Related HAAW lessons
↑ Back to navigationSTEP 03
Token ID
Each token is assigned an integer identifier from the model’s vocabulary.
Related HAAW lessons
↑ Back to navigationSTEP 04
Token Vector
The model looks up a learned numerical representation for each token ID.
Related HAAW lessons
↑ Back to navigationSTEP 05
Embedded ID
Conceptual bridge: the token ID serves as the lookup key for an embedding. This is not normally a second, distinct ID.
Related HAAW lessons
↑ Back to navigationSTEP 06
Embedded Vector
The embedding vector represents a token numerically, with position information incorporated by the model architecture.
Related HAAW lessons
↑ Back to navigationSTEP 07
Transformer
A stack of learned layers processes the vectors using attention and other transformations.
Related HAAW lessons
↑ Back to navigationSTEP 08
Attention
Tokens exchange context-dependent information, allowing the model to use relevant earlier tokens.
Related HAAW lessons
↑ Back to navigationSTEP 09
Feed-Forward Network
Within transformer layers, learned nonlinear transformations further process each token’s state.
Related HAAW lessons
↑ Back to navigationSTEP 10
Hidden State
The final layers produce context-sensitive numerical representations used to predict the next token.
Related HAAW lessons
↑ Back to navigationSTEP 11
Logits
The model projects its final state into a score for each possible next token.
Related HAAW lessons
↑ Back to navigationSTEP 12
Probability Distribution
The logits are transformed into a distribution over candidate tokens, subject to decoding settings.
Related HAAW lessons
↑ Back to navigationSTEP 13
Next Token Selection
A decoding rule chooses the next token, such as greedy selection or sampling.
Related HAAW lessons
↑ Back to navigationSTEP 14
Token ID
The chosen next token is represented by its vocabulary ID and added to the sequence.
Related HAAW lessons
↑ Back to navigationSTEP 15
Translator
The tokenizer’s decoding side maps generated token IDs back into text fragments.
Related HAAW lessons
↑ Back to navigationSTEP 16
Output Token
A decoded token fragment becomes part of the emerging response.
Related HAAW lessons
↑ Back to navigationSTEP 17
Generation Loop
The model repeats next-token prediction until a stopping condition is reached.
Related HAAW lessons
↑ Back to navigationSTEP 18
Response
The interface assembles or streams generated fragments into a readable answer.
Related HAAW lessons
↑ Back to navigationSTEP 19
Output on User Screen
The user sees the rendered answer in the application interface.
Related HAAW lessons
↑ Back to navigation