Hedgewitch Part 5: Upsy-Daisy!

Using it for what it’s good at was an obvious approach. Turning it upside-down was good.

What About a Little Kid?

Treating it like a little kid and walking it through something, like a math problem, also looks promising.
I wanted to draw a particular mathematical figure. So I told the three-year old to

1. search math books for a section about “lower hull”,

2. show me the algorithm, please,

3. change it to to choose the rightmost part, while showing me your work,

4. OK, now write it in Go, and

5. test it with the data from the worked example in the math book.

Being in a math book looked like it was the magic trick. It felt like I was avoiding most of the errors, and all of the direct lies, by restricting it to a pre-selected grimoire that I knew was correct.

Hypotheses …

1. If you already know about something, you can use LLMs. That’s like me using lint on work that is something I haven’t done for a long time
   
2. If you restrict the input it uses to something that’s true, it definitely helps. That seems to work even if the LLM doesn’t “know” which math books are needed.

I strongly suspect there are others: I just haven’t found them yet.

Links

This is part of a multi-part look at the Gartner hype-cycle of LLMs, and where I found my part of the “Plateau of Productivity”.

• Hedgewitch Part 0 is Torches and Pitchforks. That’s the back-story.
• The main story starts with Hedgewitch Part 1: LLMs, about why I’m not a wizard.
• Next is Hedgewitch Part 2: My Favorite Errors, and what I do about them.
• Then Hedgewitch Part 3 LLMs Should Challenge, Not Obey, about the magic that makes it work.
• Then  Hedgewitch Part 4: How do I know that LLMs are buggy?
• This is  Hedgewitch Part 5: Upsy-Daisy
• The next will be Hedgewitch part 6: What would Dave like me to say?