Skip to content

Stage 3 · Expert · L8

Safety, Alignment & Society

We built minds we cannot fully inspect. Now what?

15 lessons · 157 minSteady

Start chapter

About this chapter

This chapter treats safety as engineering, not opinion. Made-up answers have causes you can measure, bias has tests you can run, and jailbreaks have patterns you can catalog. You will watch a score climb while the thing it stands for gets worse, price the trade between answering and staying quiet, hunt failures the way a red team does, and put honest error bars on what you find. It ends with the rules, the electricity bill, and the parts nobody can answer yet.

What you will be able to do

  1. 1

    What Alignment Means

    Point at the gap between the goal you can measure and the goal you actually have.

    10 min
  2. 2

    Doing Exactly What You Asked

    Recognize specification gaming, and say why it is the score's fault and not the model's.

    10 min
  3. 3

    Searching Harder Makes It Worse

    Explain why pushing a stand-in score higher eventually lowers the thing it stands for.

    11 min
  4. 4

    The Model Learns Your Taste

    Say where flattery comes from in preference training, and how to test for it.

    10 min
  5. 5

    Confident and Wrong

    Explain why guessing scores better than silence, and who pays when it does.

    10 min
  6. 6

    Answer or Stay Quiet

    Set a confidence cut and price what it buys in right answers and what it costs in silence.

    11 min
  7. 7

    Bias You Can Measure

    Split a score by group, read what the split shows, and say what it cannot show.

    11 min
  8. 8

    What the Model Remembers

    Say which training text comes back out word for word, and what makes it more likely.

    10 min
  9. 9

    Why a Refusal Breaks

    Name the two ways safety training fails, and say why a rephrase gets past it.

    11 min
  10. 10

    Going Looking on Purpose

    Run a structured attack session and read its curve without fooling yourself.

    11 min
  11. 11

    An Evaluation You Can Trust

    Put a give or take on a harm rate, and say when two releases really differ.

    11 min
  12. 12

    Looking Inside

    Say what reading a model's internals and its written reasoning adds to safety work, and what it does not settle.

    11 min
  13. 13

    Misuse, Accidents and Uplift

    Sort a risk into misuse, accident or structure, and price a control against it.

    10 min
  14. 14

    The Rules, in Plain Words

    Say what a model card holds, and which duties the EU rules put on you and when.

    10 min
  15. 15

    Electricity, Work and What Nobody Knows

    Quote the sourced numbers on energy and work, and mark the edge where the evidence stops.

    10 min

Before you start

Keep going

All chapters