Skip to main content
DATUM
Enter

Practice the call.

One hand a day, twelve lessons, and toys you can move.

How Play works

No account, and nothing here costs anything to get wrong. The Table is new every day. Lessons show how AI works by letting you move it.

Today's Table

Five calls, the same hand for everyone today.Today leans on: Guard what leaves.The next hand arrives at midnight UTC.

Your lesson path

0 / 12
  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
  6. 6
  7. 7
  8. 8
  9. 9
  10. 10
  11. 11
  12. 12
Start with lesson 1: The context window

Pick any lesson. Each one lets you try the idea for yourself.

Opening your Table history…

This week: Guard what leavesLook at what is about to go out, and whose it is.

This week's thread

A lesson, a shift and a project that lean on Guard what leaves. Take any, or none.

Twelve lessons

Start with work.

Using AI at work0 of 7 explored

  1. The context windowAI can only see a set amount of text at once. In a long chat, the oldest parts quietly fall out of view.
  2. RetrievalBefore answering, the system looks up the most relevant bits of your documents and hands them to the AI along with your question.
  3. Prompt injectionText inside an email, web page or file can trick an AI into treating it as an order, as if you had written it.
  4. EvalsBefore you swap one prompt or tool for another, test both on the same set of real examples and score them the same way.
  5. Picking the next tokenAI writes one chunk at a time by guessing what comes next, with a dial for how safe or adventurous each guess is.
  6. TokensAI does not read letters or whole words. It reads chunks, and the chunks it knows came from what it read most.
  7. Precision and recallAny system that flags things has to choose between catching everything and only flagging what is truly a match.

How AI goes wrong3 core failure modes

  1. Sure but wrongInvented facts and citations sound just as smooth as real ones unless you ask for quotes from your pasted source.
  2. What it cannot seeIf a fact lives in this morning's spreadsheet or your internal intranet, the model cannot see it until you put it in the prompt.
  3. It agrees with youLeading questions get flattering answers. Ask for the strongest case against your draft before you send it.

How it is built0 of 5 explored

  1. EmbeddingsAI turns words into lists of numbers, so words that mean similar things end up pointing the same way.
  2. AttentionWhen reading a sentence, each word looks back at the other words and decides which ones matter most for its meaning.
  3. Underfitting and overfittingA model can be too simple to see the pattern, or so eager that it memorizes every quirk and mistake as if it mattered.
  4. Hold some data backTo find out how good a model really is, test it on examples it never saw while it was learning.
  5. Gradient descentTraining is many small adjustments, each one making the model a little less wrong, with a setting for how big each step is.
Write the brief

Two sentences, four checks

You have anonymised shift notes for six beds and need a short handover note for the incoming night charge nurse.

Words you will meet

Shifts

A working day, a few minutes.

Seven working daysWhat you skip at nine comes back at three.

The Lab

Move an AI and watch.

All Lab rooms