How AI Translation Works
AI translation turns text into numbers, learns patterns across languages, and generates translated text by predicting one word at a time.
A complete interactive classroom, not just a preview.
Start when you are ready to enter this Stage's 11 scenes and explore, respond, and learn as you go.
When you tap Translate, what actually happens inside the app between your sentence and the translated one?
- three-eras
- Three generations of machine translation: rules, statistics, and neural networks.
- tokens-embeddings
- How text is split into tokens and converted into numerical vectors.
- encoder-decoder
- The encoder reads source text into meaning; the decoder writes target text.
- attention
- Attention lets the decoder focus on relevant source words at each step.
- translation-quality
- Why AI translation still makes mistakes: ambiguity, idioms, and low-resource languages.
AI translation just swaps each word in one language for a word in another.
Show that modern systems translate meaning as a whole, not word-by-word.
Neural translation works like a giant bilingual dictionary.
Reveal that the model learns patterns from data, not stored word pairs.
Long sentences are translated all at once in a single step.
Demonstrate that translation is generated token by token, with attention guiding each choice.
- building a production translation model
- low-level math of transformers
- speech translation pipelines
- Learner explains why word-by-word swap fails on idioms.
- Learner traces a sentence through tokenization, encoder, attention, and decoder.
- Learner identifies why a translation output is wrong or awkward.
- Apply the encoder-decoder + attention mental model to understand how AI chatbots and summarizers work.
Curious learners with no AI or programming background, comfortable with everyday technology.
- 01Why "Translate" Is Harder Than It LooksslideOrientationObserve
Set up the puzzle: everyday translation hides surprisingly complex machinery.
- What users see vs what the system does
- Why simple word swap fails
- What this lesson will reveal
- 02Try the Three Eras of TranslationinteractiveModel buildingApply
A simulation where the learner feeds the same sentence into a rule-based, statistical, and neural translator to feel the difference.
- Rule-based: dictionary + grammar rules
- Statistical: pick the most likely phrase match
- Neural: generate holistically
- 03From Words to NumbersslideOrientationObserve
Bridge from eras to the modern neural pipeline by introducing tokenization and embeddings.
- Text becomes tokens
- Tokens become vectors
- Similar meanings cluster in vector space
- 04See Tokens and EmbeddingsinteractiveModel buildingObserve
A simulation that tokenizes a sentence and shows its embeddings projected onto a 2D map, where related words cluster.
- Tokens highlighted on the sentence
- Embedding positions plotted
- Clusters reveal meaning
- 05Encoder-Decoder: Read, Then WriteinteractiveModel buildingConstruct
A diagram-style simulation walking through how the source sentence is encoded and then decoded step by step.
- Encoder compresses meaning
- Decoder generates target tokens
- Each step depends on the encoded meaning
- 06Spot the MisalignmentinteractiveMisconception repairPredict
Predict: learners see a naive word-swap translation of an idiom and must mark which word pairs are misleading.
- Idioms defeat word-for-word mapping
- Meaning is not 1:1 across languages
- Context matters more than vocabulary
- 07Attention: Looking Back at the SourceslideOrientationObserve
Explain attention as the decoder's way of re-reading the source sentence while writing.
- Decoder generates token by token
- Attention weights highlight relevant source words
- Weights shift with each output step
- 08Watch Attention MoveinteractiveModel buildingObserve
An interactive visualization showing attention weights lighting up between source words and each generated target word.
- See which source word is being focused on
- Notice how focus shifts per step
- Heavier lines = stronger attention
- 09Where AI Translation Still Trips UpslideOrientationExplain
Survey quality challenges: ambiguity, idioms, names, low-resource languages, and hallucination.
- Multiple valid translations exist
- Rare languages have less training data
- Models can confidently produce wrong output
- 10Translation Quality ChallengeinteractiveApplicationChoose
Learners compare three AI outputs side by side, choose the best one, and justify their pick by spotting errors.
- Read critically, not passively
- Detect dropped meaning, awkward phrasing, hallucinations
- Justify with translation concepts
- 11The Full Pipeline in One LineslideSynthesisExplain
Recap the end-to-end journey: sentence → tokens → embeddings → encoder → attention → decoder → translated tokens.
- All stages connect into one pipeline
- Same blueprint powers chatbots and summaries
- Translation is prediction, not lookup
Discussion threads for a Stage aren't available yet.
