• structuralize_this [none/use name]
    ·
    2 years ago

    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.

    • edge [he/him]
      ·
      2 years ago

      It's more like it uses its training data to just make up the contents of the url you give it. Instead of "write a story about xyz" it's essentially "write a website that could be at this url".

      Here's what it returned for hexbear:

      <!DOCTYPE html>
<html>
<head>
<title>HexBear</title>
</head>
<body>
<p>
Hello there!<br>
Welcome to HexBear.
</p>
</body>
</html>

      I'm sure hexbear isn't in its training data, and the name isn't really indicative of its contents, so it just makes a simple site using the domain name as the title.

      But here's what it gives for reddit 
      <!DOCTYPE html>
<html>
<head>
<title>reddit: the front page of the internet</title>
<meta http-equiv="X-UA-Compatible" content="IE=edge,chrome=1">
<meta name="viewport" content="width=device-width, initial-scale=1, minimum-scale=1">
<meta name="description" content="reddit: the front page of the internet">
<link rel="shortcut icon" href="https://www.redditstatic.com/desktop2x/img/favicon/apple-icon-57x57.png">
<link rel="icon" href="https://www.redditstatic.com/desktop2x/img/favicon/apple-icon-57x57.png" sizes="57x57">
<link rel="icon" href="https://www.redditstatic.com/desktop2x/img/favicon/apple-icon-60x60.png" sizes="60x60">
<link rel="icon" href="https://www.redditstatic.com/desktop2x/img/favicon/apple-icon-72x72.png" sizes="72x72">
<link rel="icon" href="https://www.redditstatic.com/desktop2x/img/favicon/apple-icon-76x76.png" sizes="76x76">
<link rel="icon" href="https://www.redditstatic.com/desktop2x/img/favicon/apple-icon-114x114.png" sizes="114x114">
<link rel="icon" href="https://www.redditstatic.com/desktop2x/img/favicon/apple-icon-120x120.png" sizes="120x120">
<link rel="icon" href="https://www.redditstatic.com/desktop2x/img/favicon/apple-icon-144x144.png" sizes="144x144">
<link rel="icon" href="https://www.redditstatic.com/desktop2x/img/favicon/apple-icon-152x152.png" sizes="152x152">
<link rel="icon" href="https://www.redditstatic.com/desktop2x/img/favicon/apple-icon-180x180.png" sizes="180x180">
<link rel="icon" href="https://www.redditstatic.com/desktop2x/img/favicon/android-icon-192x192.png" sizes="192x192">
<link rel="icon" href="https://www.redditstatic.com/desktop2x/img/favicon/favicon-32x32.png" sizes="32x32">
<link rel="icon" href="https://www.redditstatic.com/desktop2x/img/favicon/favicon-96x96.png" sizes="96x96">
<link rel="icon" href="https://www.redditstatic.com/desktop2x/img/favicon/favicon-16x16.png" sizes="16x16">
<meta name="msapplication-TileColor" content="#0079d3">
<meta name="msapplication-TileImage" content="https://www.redditstatic.com/desktop2x/img/favicon/

      It probably has reddit somewhere in its data, so it gets the general idea and gives a realistic response, including the old tagline. But it didn't actually access reddit's server or else it would have gotten the current tagline ("Dive into anything").

      But what I'm curious about is if I "run a web server" within ChatGPT, will it make that connection and reference the web server code to figure out what the response would be. And then if it can do database <-> webserver <-> curl output in a coherent way, including the db's file size in ls changing.

      • structuralize_this [none/use name]
        ·
        2 years ago

        I get what you're saying about the curl. I was thinking it literally would go and fetch the contents of an external website. It's just making a guess, which makes sense.

        The database, webserver, curl thing is interesting too. I'm gonna guess it can emulate flask stuff pretty well. It will probably be good at sql too. sql is actually an applicable concept for auto-generation.

        . Each person has a first name
. Each person has a last name
. Each person has a birth date
. Each person can take one or more classes
. A person is a teacher.
. One teacher teaches a class
. A class can have one or more persons in the class

        It should create a person table, with first, last and birthdate. It should create some sort of teacher entity, either as a bool in person, or as some extra attribute table. It should create a class entity with a one to one foreign key with person/teacher. It should created a many-to-one with class to person.