AI AdventureAICreate with AIDay 5
🔪 Create with AI · Day 5

Story Surgery

Some AI doesn't make pictures — it operates on text. Same idea, brand-new form: shrink it, simplify it, translate it, make it fancy. One model, many surgeries. Text goes in the door, different text walks out.

↓ scrub in
🔀 Warmup · Same Words, New Shape

Reshape your text — instantly

Type anything. These buttons are simple rules (no AI yet) that transform your words: SHOUT, whisper, reverse, pig-latin. The content stays; only the form changes. That's the whole idea of text-to-text — warm up on it here.

output

These are just rules a kid could write. The real model below does the same move — same content, new form — but it actually understands the words, so it can summarize and translate, not just flip letters.

🏷️ The pro wordsA model that takes text in and gives new text out is doing text-to-text generation. The shape of model that reads a whole input, then writes a whole output, is called sequence-to-sequence (seq2seq) — built from an encoder (reads) plus a decoder (writes). The famous one is T5. Ours is FLAN-T5: a T5 that was instruction-tuned (FLAN) so it follows commands like "summarize:" or "translate to French:".
🔬 Real Model Lab · The Operating Table

One real model, five surgeries

This is FLAN-T5 — a real text-to-text model — running 100% on your own machine. Load it once, then put any paragraph on the table and press an operation. It really summarizes, simplifies, translates, and rephrases. No server, no fakery.

not loaded

First load downloads the model from Hugging Face, then it's cached for offline use. On a 4GB laptop give it a minute. ☕

🚀 Beyond the Basics

Text-to-text is everywhere

Once you can turn text into new text, a hundred tools fall out of the same idea. You already use these:

⌨️

Autocomplete

Your keyboard and email predict the next words — text in, more text out. Same seq2seq family, tuned to finish your sentence.

Rewrite tools

"Make it shorter / friendlier / more formal" buttons in docs and email are literally this model with a fixed instruction baked in.

🌍

Translation

Translate apps are the original seq2seq success story: read one language, write another. That's the T in T5's family tree.

🏷️ Pro names to look upReal tasks in this family: summarization, machine translation, paraphrasing, grammar correction, and style transfer for text. The trick that lets ONE model do them all is instruction tuning — training on thousands of "do X to this text" examples (that's the FLAN in FLAN-T5).
🎯 Boss · Surgeon's Challenge

Pick the right operation

A patient comes in with a goal. Choose the operation that gets them there. A real prompt engineer matches the tool to the job in a heartbeat — get 3 right to win the badge!

Correct: 0 / 3

🎉 Day 5 Complete!

You met text-to-text: one model that transforms writing into new writing — summarize, simplify, translate, restyle. Same meaning, new form. You even ran a real FLAN-T5 seq2seq model on your own machine and found where it breaks.

Day 6: Next Adventure →
Tomorrow: we keep building your generative toolkit — more real models, more making. See you there! 🚀
AI Adventure · Create with AI · Day 5 — made for young inventors 🚀