- All tasks were solved by ChatGPT with minor fixes from human where the fix was obvious.
- No refactoring or code restyling was made apart from renaming some methods.
- No code/performance optimizations were made manually, but in some cases such optimizations were requested from ChatGPT.
- In the majority of cases the first working implementation was accepted without trying to make the solution more elegant.
- The primary goal was to understand how quickly ChatGPT is able to solve relatively simple coding tasks with complex descriptions.
- In the majority of cases ChatGPT was unable to solve a task at first try. Usually it required human programmer to test the code, understand the problem and either fix it manually or request ChatGPT to fix it.
- ChatGPT is quite bad at fixing the defects, especially the complicated ones. It tends to go in circles proposing the same solutions again and again without trying to think outside this loop.
- ChatGPT is fundamentally unable to test the code it suggests even if requested explicitly. This inability leads to providing the erroneous code again and again.
- ChatGPT is unable to even check the proposed code for compilability. In Kotlin some standard core classes share names with Java core classes but have different methods, e.g. ArrayDeque. ChatGPT is unable to actually distinguish them and tries to use them interchangeably causing compilation errors.
- 2022, day 5: Hours were spent trying to get correct crates stacks input parsing from ChatGPT without any luck. It was always trying to parse the lines of crates not columns. And even if explained explicitly that the crates are stacked in columns the parsing was incorrect. Had to give up and write the parsing code manually.
- 2022, day 7: It was very difficult to get the correctly working code out of ChatGPT. Also, interestingly, the new Turbo model produced cleaner and more elegant code but the Legacy model version was closer to the correct one, so in the end I've used it instead of trying to fix the Turbo one. Overall, Legacy model is slower but more accurate.
- 2022, day 8: Similarly to the previous day it was quite challenging to get something working. Required some additional effort to fix the defects. Also, had to re-write the requirements for part 2 to make them more structured and simple for ChatGPT to parse. Again, the Legacy model proved to be more accurate, but then Turbo model was used for some refactoring work.