Focus on decisions, not tasks

This fits the way I like to use LLMs: I always ask them for options, then I decide myself which of those options makes the most sense.

Essentially I'm using them as weird magical documentation that can spit out (incomplete but still useful) available options to guide my decision making at any turn.

I like to think of it as the apprentices working for famous artists like Leonardo. The master would draw the outline/sketch, and then the students would fill in the blanks under supervision. Sometimes, the master would steal ideas from the students.

Would you be willing to give an example of this?

Smells like reinforcement learning in real life. Master sets ups the task environment, collects samples from students, picks the best and maybe even augments them. Students watch the master and learn... and the cycle continues.

And then the master becomes a grandmaster (unless entropy explosion occurs).

You should actually ask for 3 examples so you can select one.

Not exactly the same… But recently I wanted to pick a library in Python or Julia for simulating differential equations using a GPU. So I asked ChatGPT which libraries exist for this (JAX, CuPy, etc.), asked it to generate code to solve e.g. the 2D heat equation on a 1000x1000 grid for 100 time steps using each of those frameworks. Then I stepped in and verified that each code appeared to do the same thing, and proceeded to benchmark their performance on my hardware. Afterwards I had an informed choice of which framework to use for my project, even though ChatGPT gave me the benchmark code instead of the answer.

Here's one: if you want it to rewrite something or show you a better way to say something you've already written, just ask for different options. I usually find that mixing what I wanted to write with parts of its suggestions gives me a great result.

Yea, and it's what ChatGPT does sometimes even if you don't ask it for multiple options. It shows two options and asks you to select the best one. Essentially using customers to help with training.

Similarly I find Bing copilot very handy for exploring problem spaces and their solution spaces. Then i dig/ask deeper and deeper until i understand the solution. It is a quick way to brainstorm without diverting too much, and hence converging, since that tech knows a lot details about a lot of things.

Here's one: I asked Claude this:

"Options for JavaScript to turn a JPEG into a vector SVG"

Result: https://gist.github.com/simonw/d2e724c357786371d7cc4b5b5bb87...

I ended up building this: https://tools.simonwillison.net/svg-render

More details here: https://simonwillison.net/2024/Oct/6/svg-to-jpg-png/

> I ended up building this:

I'm confused about how you ended up building the opposite of your initial prompt? The initial prompt seemed like puzzle worth solving.