AI, plainly

10 Lesson 10 of 15

Good at that, bad at this

Skill is not one dial. Two requests that feel like the same kind of job can sit on opposite sides of a cliff.

  • About 2 min read
  • Interactive 3D diagram
  • Hands-on tool
The version most people start with

It is either good or it isn't. You have an overall impression (impressive, or overhyped), and you apply it to whatever you ask next. If it wrote a decent cover letter, it can probably handle your spreadsheet.

The idea to leave with

Ability here is jagged in a way that human ability is not.

See it in the lesson

Ability here is jagged in a way that human ability is not. If a person is good at one task, that tells you a fair bit about similar tasks, because human skills share the same foundations. Here the foundations are different. So the map has cliffs in places that look perfectly smooth from the outside.

The landscape below is that map. Height is how reliable the tool is at each task. The dashed red lines join pairs that feel like the same kind of request but are not. Walk along one and you fall off a cliff.

Pale peaks are tasks it usually handles. Dark troughs are tasks it often gets wrong. The reliability values are based on what these tools are commonly seen to do, not measured in a study.
Read this diagram as text

Pins are numbered from most to least reliable, matching this list:

    The three pairs joined by dashed lines are: explaining what a word means, versus counting the letters in that word; summarising a long email, versus giving a source you can find; and drafting a polite reply, versus telling you today's news. In each pair, the first is near the top of the list and the second near the bottom.

    The pattern underneath

    The troughs are not random. Look at what the unreliable tasks share. They need an exact count, an exact lookup, a fact from after training, or something about you personally. The peaks are tasks with many good answers, where you can judge the result yourself.

    That gives you a quick test for a new task, before you trust the result. Is there one answer that has to be exactly right? Does it need information from outside the model? Could I tell if it were wrong? Three quick questions, and they predict the cliffs surprisingly well.

    This matters for more than wasted time. When researchers watched beginner programmers use these tools, the group split. Some sped up. Others struggled more than they would have without the help. The tools did not close the gap between them. They widened it. What separated the two groups was not talent but how they used the tool. That is a skill, and it is what the next lessons are about.

    The three-question test

    Type a task you are thinking of handing to a chatbot, answer three questions about it, and see roughly where it sits on the map. Nothing you type leaves this page.

    One analogy. It is closer to a landscape than a dial. You would not assume that because one path up a hill is easy, the one beside it is too.

    Where the analogy breaks: a real landscape looks like what it is. Here you cannot see the cliffs from the request, because both tasks arrive as an ordinary sentence you typed.

    The move to make

    Pick a task you think it should ace, and one that feels almost the same but needs an exact answer. Run both. The gap between them is what to remember, not either result on its own.

    Worth flagging: The heights on that map are my own estimates from commonly reported behaviour, not measurements. The main claim (that reliability is jagged across tasks that feel similar) is solid. The exact height of any one pin is not.