Skip to content

gottfired/advent-of-code-2021-copilot-edition

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

24 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Advent of code 2021 github copilot edition

None of the code was written by me. I only write comments and then use autocomplete.

Exceptions: day6/pt2 and day8/pt2, where I gave up and implemented manually.

Screencast of day 7

https://youtu.be/XvoCyVFQOm8

Lessons learned

  • Use types. Days 1-4 I used plain JS, but without type information copilot often didn't have enough information to use data structures in a correct way. I had to explicitely e.g. say "data is a 2d matrix of numbers"
  • Use natural language to describe your functions. This works better than describing in pseudo code. E.g. don't say "read a file of lines which contains vectors separated by -> where coordinates are comma separated". Instead say, read file with lines that look like this x,y -> x,y.
    • For example on day4 I used a pseudo code approach. This became very tedious, where writing the code myself would've been way faster and less error prone
    • On day5 I used types and natural language. The result was a lot better.
  • day7 part 2. Had to put the cost algorithm into a separate function, otherwise codepilot insisted on reusing the same cost function as for part 1.
  • In general already start off by separating your code into functions. It's easier for copilot to reference those steps later on. Also refactoring with copilot later on doesn't work and I'm not allowed to touch the code which is a problem. Sometimes I had the case where an algorithm was autocompleted correctly and as I wanted to describe the same thing as a separate function using the same wording, copilot refused to write it correctly again. I have the suspicion that copilot, if a code is once generated doesn't repeat itself thinking that the code before might have been wrong and thus tries to find a more correct solution if provided with the same comments.
  • Starting from day11 avoiding manual course correction get very tedious. I decided to allow minimal correctiony by hand, because the code is often 99% correct, but reformulating the comments to get it 100% correct takes too much time. Unfortunately 1% wrong e.g. with recursive functions often leads to a stack overflow :)

Highlights so far

  • day2 I could copy and paste the instructions from AOC and it worked
  • day5 I could say // function part2 is the same as part1 but also with diagonal lines. Thats pretty amazing!
  • day9 Recursive flood fill algorithm without me needing to touch the code. Also whole day went very smooth with minimal rephrasing.

Lowlights so far

  • day6 part 2. Copilot didn't understand what I want. Had to hand code it.
  • day8 part 2. I have no idea how I should've explained the solution to copilot, so again manual implementation.
  • day11. I was too dumb to read the instructions correctly (thoght the octopus flash at 9, not > 9). This lead to wrong autocompleted code, which I had to fix up manually. Also copilot refused to increase the neighbours after flashing for whatever reason. Had to +1 by adding a separate // increase neighbour comment. Also copilot used lodash fill for allocating the 2d flashed array which was wrong. Had to tell copilot to init the array not using lodash.

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published