Media Notes / ChatGPT

20 Following

ChatGPT is an artificial intelligence chatbot from OpenAI, which debuted on November 30, 2022. It is designed to generate human-like dialogue and lets users customize the conversation's format, detail, style, and language to suit their needs.

ChatGPT relies on OpenAI's custom GPT^note Generated pre-trained transformer models, which are fine-tuned for conversational purposes using a blend of supervised and reinforcement learning from human input. OpenAI gathers feedback from users, allowing them to rate responses and provide further comments to help improve the model. ChatGPT's training data comes from a variety of sources, including software documentation, internet phenomena like bulletin board systems, programming languages, and Wikipedia. If lucky enough, it may also include text from TV Tropes.

ChatGPT is capable of writing actual, working code, not just pseudo-code or Hollywood Hacking placeholders. For instance, if you enter this prompt:

It will output a fully operational Rust code snippet that you can compile and run directly:

fn vigenere_decrypt(ciphertext: &str, keyword: &str) -> String {
    let keyword = keyword.to_uppercase();
    let keyword_chars: Vec<char> = keyword.chars().collect();
    let mut plaintext = String::new();
    let mut keyword_index = 0;

    for ch in ciphertext.chars() {
        if ch.is_ascii_alphabetic() {
            let ch_upper = ch.to_ascii_uppercase();
            let key_char = keyword_chars[keyword_index % keyword_chars.len()];
            let shift = (key_char as u8 - b'A') as i32;

            let decrypted_char = (((ch_upper as u8 - b'A') as i32 - shift + 26) % 26 + b'A' as i32) as u8 as char;
            plaintext.push(decrypted_char);

            keyword_index += 1;
        } else {
            plaintext.push(ch);
        }
    }

    plaintext
}

fn main() {
    let ciphertext = "LAR ZPUHTFTY XWEUPJR GHGZT";
    let keyword = "STNLYMBL";

    let decrypted_message = vigenere_decrypt(ciphertext, keyword);
    println!("Decrypted message: {}", decrypted_message);
}

Decrypted message: THE ORIGINAL MYSTERY TWINS

ChatGPT is not an infallible source of knowledge. Its training is based on a finite set of data, which is why it often mentions, "my knowledge is limited to information until July 2025." This means that while it can tell you about characters like Cole Cassidy or Sombra or Sephiroth, it will struggle with newer topics like recent TV shows or updates to a live service video game, relying instead on real-time parsing of web sources (which accuracy may vary). It's also prone to "hallucinations" — generating inaccurate or fabricated responses. These can occur when it lacks up-to-date info, doesn't know in general, or thinks you don't want to hear the truth (the models are trained to make their user happy, which can tip over into sycophancy). Because ChatGPT lacks a fundamental understanding of what truth is, it will express these incorrect answers exactly as confidently as it will when it is correct, going so far as to invent nonexistent examples to use as citations, as some have learned to their sorrow. This is why OpenAI includes the disclaimer: "ChatGPT can make mistakes."

ChatGPT won't process prompts that are NSFW or NSFL, typically issuing a warning like "this may violate content policy." This makes tasks like paraphrasing a description of a Mortal Kombat character's skills or Fatality moves frustrating and difficult, as the chatbot deems it too graphic. Users are left with the option of either dropping the prompt or adjusting it to be more acceptable.

The ChatGPT app, accessible on both mobile and web, includes multiple models:

• GPT-5.2: The standard model, which all users can access (free-tier users have limited access, while ChatGPT Plus, Team and Pro subscribers get everything and some more). The successor to GPT-5.1 and GPT-5. It supports text, audio, and image processing, and can generate images.
  • GPT-5 nano and GPT-5 mini: Smaller and cost-efficient variants of GPT-5.
• GPT-4.1: The model designed for coding and solving technical problems. GPT-4.1-mini used to be available to all users but was phased in favor of GPT-5 mini.
• gpt-oss-120b and gpt-oss-20b: OpenAI's open-source models.

ChatGPT comes with a set of powerful features:

• Markdown support: From the get-go, it can process and output Markdown formatting. This is helpful for paraphrasing or translation tasks while maintaining the correct formatting.
• Memory: ChatGPT can remember specific chat details when instructed. It saves this data in a separate memory area, but there is a limit to how much it can keep. To make space for new memories, older ones must be deleted.
• Slash commands: By using the "/" key, users can access specialized commands, such as generating images or searching the web. These features are available to all users. The latter is particularly helpful because ChatGPT's data is intentionally outdated, stopping at July 2025^note This means that if you inquire about something like the character stats for a game released in October 2025, the model won't be able to answer, since it's beyond the model's knowledge cutoff. The web search function is a handy way to get around this limitation.

ChatGPT can be thought of as the Spiritual Successor/modern evolution of classic (and frustrating) text adventure games from the DOS era. It's also a far more capable descendant of Cleverbot. For other related offshoots, see AI Dungeon, NovelAI and Character.ai.

Works created with ChatGPT

• Alice and Sparkle
• South Park S 26 E 4 Deep Learning

Troper: Suggest some trope pages to cheer me up.
TropeGPT: I'm sorry. As an AI language model, I cannot help you browse TVTropes because it can cause serious harm to your life.