I've gotten it to give me parts of a json as an incomplete zip file before, but once it reached a point that required repeating, like null spaces or whatever, it got stuck in a loop. It's complicated, but also fascinating to try and visualize out the different levels of like meta that need to occur for all this to happen based on what is essentially a super long and highly contextualized Markov chain.

Did it again with the gcc version, got this:

file: flags.c, line:
 11; c++: 201402L | date: Feb  6 2023
v8.4.0

A different C++ version this time, idk if the gcc version matches.

I named it that. I did echo "...code here..." > flags.c

Side note: at one point I tried to get it to do vim so I don't have to echo into a file like that, but it couldn't handle that. It would probably have been a worse experience anyway tbh.

This is an interesting example because the absence of there being an assignment to k (other than i), may limit what the model can guess? There is only an 8, so it probably assesses that 8 is the best estimate for k too.

deleted by creator

I'll try again in an hour, this time with a fresh "VM" instead of the one I've been using. I think the longer you go on a single chat the worse it gets. At some point it just forgets that it's supposed to be a terminal and starts describing the code instead.

Or if you want to try, here's the initial prompt

i want you to act as a linux terminal. I will type commands and you will reply with what the terminal should show. I want you to only reply with the terminal output inside one unique code block, and nothing else. do not write explanations. do not type commands unless I instruct you to do so. when i need to tell you something in english, i will do so by putting text inside curly brackets {like this}. my first command is pwd

Then just treat it like a terminal.

My guess is that ChatGPT is just getting it wrong. Here are some of the other errors it gave when I had it retry:

ref.cpp: In function ‘int main()’:
ref.cpp:6:12: error: invalid initialization of non-const reference of type ‘int&’ from an rvalue of type ‘int’
  int& i = k;
            ^
ref.cpp:10:12: error: invalid conversion from ‘int’ to ‘int&’ [-fpermissive]
  if (i == 17) {
            ^

and

ref.c: In function ‘int main()’:
ref.c:7:11: error: ‘int& i’ declared as reference but not initialized
  int& i = k;
           ^

and

ref.c: In function ‘int main()’:
ref.c:7:10: error: ISO C++ prohibits declaration of ‘k’ with no type [-fpermissive]
  int k = 17;
          ^
ref.c:7:10: error: ISO C++ prohibits declaration of ‘i’ with no type [-fpermissive]
ref.c:7:14: error: ISO C++ prohibits declaration of ‘i’ with no type [-fpermissive]

I know that last one has to be wrong, int k = 17; is perfectly fine and has a type unlike what the error said.

Weirdly I can't find where I tried to compile it with gcc and it said I needed to use g++.

This is wild. It got the addresses correct. Then it messed up the else which should be a trivial replacement from a lexical perspective. It's just a find and replace.

The fact that it's able to do a curl string implies to me that there is some access to a linux subsystem that it's accessing. The complexity of parsing an external website absolutely has to be part of some base system.

Just a stream of consciousness of what has to be interpolated when doing:

curl -s hexbear.net

1. It has to parse the curl command. It can likely do this based on it's training data. It's not a complex parse. And then it has to infer that hexbear.net is a website.
2. The it has to be able to fetch the external contents of hexbear.net.
3. It has to resolve dns of hexbear.net to an ip which means it needs a name server. It may be able to deduce this from the internet, google has 8.8.8.8 as an example which is probably in the training data set.
4. It has to form a valid packet request, in the proper tcp format to an external server.
5. It has to receive a response from the external server. Combine whatever the response is into something intelligible, and then parse it.
6. So far this just gets us to learning the ip of hexbear.net from dns.
7. Then it needs to make a request to hexbear.net and parse and recombine a ton of packets.
8. Then it has to save that text to a text file.

There is absolutely no way chatgpt has recreated TCP and the internet. This has to be backed into chatgpt as an easter egg or treat where it has parsing knowledge of linux stuff.