It is interesting to consider that such tools can impact not only education but also this kind of competitions.
I have not experimented much. Some *.txt files in the repo show output for simulating pi using Monte Carlo and estimating Eulerâs number using a Taylor series. ChatGPT can generate codes that donât compile, give correct results, or give wrong results. The last case is the most problematic.
1 Like
Thanks, I did not see. It seems those small problems need typically 4 attempts which is reasonable.
Yes, TDD would be necessary for such cases.
No wonder someone wanted to bite the bullet and write a program to automate the tedious process to make AI work after several tries. However, quite often AI will only give you a code snippet and you are supposed to fill the details (which is ok, but I wonder how a python program will cope with that.)
In addition, AI may never find a correct solution. Iâm sure I am not the only one who saw it failing several times in a row, then falling down the rabbit hole (the infinite loop I mentioned earlier - exact same answer, presented as a new one.)
Programs like that prove we are just playing some form of the âimitation game reversedâ here. We know in advance the one we are âtalkingâ with is not a human, itâs not intelligent no matter how we stress the term. And we know it probably wonât make it by itself. The problem is then shifted from âusing AI to accelerate some actual workâ to âfind a good way to hold AIâs hand step by step so it can finally walkâ.
As a search engine, AI can be quite useful. As far translation is concerned, a specialized AI is actually quite decent (but again, not to be trusted, you canât just copy/paste its translation.) However as far programming is concerned, it is far away from being even that.
Have you ever tried a chess program that has âcharactersâ available? It gives you the option to choose the playing style of your opponent. You can pick an offensive opponent, a defensive opponent, a queen hunter, a Sicilian Defense advocate, whatever. While playing, it gives you the impression you are dealing with an intelligent opponent who also has some character. But itâs all fake, itâs just pre-programmed to play a specific way. All it does is comparisons, and it does it quickly. No intelligence involved.
Same thing here. We have an advanced parser able to (usually) âunderstandâ what we ask and compose a readable answer. But itâs all fake, it understood nothing because it has no intelligence to understand anything, and the answer is not to be trusted.
That is the good news, it wonât decide by itself (as it has no self) to register to the Advent of Code 2023 and ruin your pleasure!
Finally we can play the imitation game, and we like it although we are sadly playing alone. We have not yet found an E.T., so maybe we are looking for an A.I.
1 Like
Because some of you were curious what the capabilities of ChatGPT4 are.
I asked the GPT to write tests and the main program for task one of day one:
https://chat.openai.com/share/269d5bfb-9924-4b32-a7f7-f517dee93bee
BTW, in Python it can run its own code and directly verify it.
1 Like
It is impressive that it could work well on that subject that sounds rather original (therefore probably not present in the learning process like could be something about prime numbers for example).
And it does not seem perturbed (and not even âinterestedâ) by the Elves story. All information useless seems to have been automatically filtered.
Advent of Code participants now look like Kasparov⌠for this 1st problem at least.
Iâve used AI to generate an image of what I will look like around day 20⌠seems pretty accurate!!
10 Likes
@jacobwilliams Great image, except it got the Fortran logo wrong!
1 Like
That was day 19 for me last year. And I still donât know dynamic programming 
1 Like
The answer is most definitely yes for me. It took less than 2 minutes to get a working solution in Python. I couldnât have written it in such a short time frame, and I am good at Python. The Fortran solution took maybe 3 minutes, as I had to workaround some simple bugs. I would not be able to get it done in Fortran in 3 minutes either. It would probably take me 10 to 20 minutes. So for me, for this task, I got at least 4x speedup (possibly more).
2 Likes
In some ways chatGPT is kind of like reviewing PRs for me: I am still responsible for the final outcome, but it does most of the real work, I only check the solution and focus on figuring out proper ways to fix it if the approach is incorrect. For Day 1 it was a big help. I am sure for some of the more complicated problems it will fail miserably and I might need to actually do them myself.
Definitely âinspiredâ by this logo found in Bing:
Be careful with copyright! 
Although the fortran-tech.com site seems not online anymore.
I think using such a tool to generate your own company logo would be dangerousâŚ
1 Like
Either the task was really simple, or you are using a much better AI than the one I have access to (Iâll definitely never pay for such a thing.)
But imagine it wasnât @certik but a student who is just learning Fortran. They would need way more time and effort to write the program by themselves than your 10-20 minutes. But they will learn something in the process. If the AI can do it for them, they will learn absolutely nothing (unless of course they opt not to use it, which I doubt.) And they need to learn a lot of things.
It makes perfect sense for you to use the AI in this case. You wonât learn anything if you write the program yourself, because for you itâs just a simple task, you did something similar countless of times already. So it makes sense to just say âthis is boring, write the program for meâ. What about the student I mentioned though?
1 Like
As She said:
Its province is to assist us in making available what we are already acquainted with.
To be acquainted with or not to be acquainted with, that is the questionâŚ
How to teach students to use such meta tools, as they donât even master the basics? At which level? How to prevent them to use such tools to avoid learning the basics? (classroom teaching is needed more than ever but the tendency to reduce public spending is reducing face-to-face)
If you find a way to do that for more than a tiny minority of them, let me know. Because I see no way to achieve this. I think itâs just human nature. If they have something that will save them a few steps, they will use it without a second thought. Even though they do need to take those steps themselves, without help.
2 Likes
Thatâs a different (but important) question, how to teach students. We should probably start a separate thread for that.
Chess engines are better than humans for the past 20 years. If you want to learn chess, you have to learn how to play yourself, not let chess engines play for you. But you can use chess engines to help you get better. Isnât this similar? If you want to learn programming, do it yourself. But you can use AI to help you get better. And if students donât want to learn programming, then maybe they should do something else?
2 Likes
I see your point, I might have something to say but my verbosity will make in just another long sheet of text, which wonât even be the first today. And like you said this is stuff for another thread.
Let me just say that I realized I liked Mathematics more than Physics (which, technically, was what I was studying) when I had that big moment of âwow, I actually proved that theorem concerning cubic splines all by myselfâ. I could have an excuse to use an AI instead (after all, I was studying Physics, right? why should I care about proving a theorem, I just need to use it). I could very well use that AI, if I had it handy, losing that âbig momentâ forever, and never realizing Numerical Analysis is actually my thing, not Physics.
2 Likes
day 2 completed: GitHub - dacarnazzola/Advent-of-Code-2023: advent of code 2023: https://adventofcode.com/2023
If they keep on with this string parsing for every question, I will probably end up making some little string parsing module with nice utility functions rather than rewriting custom parsing for each day.
This one runs in 1-2ms, interesting 1/10th the input still takes 1/3 the time as day 1 for me. Itâs probably a little less efficient to build up the character(len=5) :: string rather than just keeping a first and last index, but doing that approach would require more toggling logic. In the beginning I was close to creating a âsplit on delimiterâ function to return a derived-type array of strings, since then itâs an easy pattern of every other element being number or color.
1 Like
I donât worry about speed unless it becomes a real problem. Typically my Fortran solutions for these puzzles are fast enough. I enjoy stretching myself and seeing which âmodernâ Fortran features I can make use of. The early days are fairly easy, but they get harder during the month and often require knowledge of some advanced mathematics thing I donât know, so I end up going to the discussion board for a hint.
I also donât strive to get them done as early in the day as possible for max points.
Iâm visiting my mom for the next week, so donât know how many of these I will be able to do, but I am taking my laptop. I use Intel Fortran for this, and now use IFX exclusively - it will be interesting to see if I encounter any bugs.
2 Likes