[IsaacL]

I've been able to give it arbitrary blocks of code and have it explain how they work.

(Asking this makes GPT more effective when I ask it make further changes. One reason I do this is when I start a new session with ChatGPT discussing code it helped me write previously, especially if I've gone away and done a big refactoring myself.)

A very simple example is that I asked it to write some Ruby functions that would generate random creature descriptions (e.g., "a ferocious ice dragon", "a mysterious jungle griffin"). It did this by generating three arrays (adjectives, locations, creature types) and randomly selecting from them to build the output string. I then asked it to explain how many different descriptions it could generate, and it explained that multiplying the length of the three arrays would give the number of outputs. (125 for the first iteration, 5x5x5).

I then asked it how it would increase the number of possible outputs to 1000, and it did so by increasing each of the three arrays to length 10. I then asked it how it would generate millions of possible outputs, and it added extra arrays to make the creature descriptions more complicated, increasing the number of permutations of strings.

This is not the most sophisticated example, but it shows what GPT can do when it can combine "knowledge" of different areas.

If it's able to combine the solutions to known problems in a straightforward way, it can accomplish a lot. Beyond a certain point it needs guidance from the user, but if used as a tool to fill in the gaps in your own knowledge, its enormously powerful. I it more as an "intelligence-augmenter" than a "human-replacer".

See my comment here where I went into more detail on how I work with GPT: https://news.ycombinator.com/item?id=35197613