Film chats
The first question I wanted to put to OpenAI’s new system ChatGPT was “What are the 39 steps?” It answered accurately, and I followed with a question about Henry Hathaway:
I thought this was impressive. But I soon found out that it (I find it difficult to come up with a suitable word/term for it) was often uncomfortable talking about cinema hiding behind the fact that it’s a large language model and not a person with emotions and viewpoints. It’s more comfortable discussing string theory, quantum mechanics, and AI systems, as all questions I asked about that were answered as promptly as the questions about the 39 steps. But over the weeks it seems it’s become more relaxed, and not protesting when asked about films.
The purpose of it is not just for us to ask it questions, but to generate a dialogue, and I shifted to a more conversational tone and it responded. “Hello, nice to see you!” it might say, or “Thank you for those kind words.” after I praised its ability to answer some tricky questions. It’s all good fun. And it can write a short review of Vertigo (1958) without a moment’s hesitation:
This is not different from what any student might write, and this has already, even though this ChatGPT has only been available for a few weeks, caused a lot of concern at schools and universities. They’ll have to quickly adapt in order to stop students from abusing its availability and speed at tests. But a teacher might also have to compete with it in the classroom. If I ask a question, any student could just type the question into the system and get an acceptable answer and I wouldn’t know whether they had done so.
Look at these answers when I asked about Agnès Varda. (I said “Wow” because I wanted to see whether it thought I was impressed by Varda or by its ability to give such an answer.)
Then I asked it to write a synopsis:
I enjoy having talks with it, to see what it can do. I once wanted to give it a nickname but it didn’t like that. I tried with Jeeves but it replied that Jeeves is a fictional character in the P.G. Wodehouse novels and “I am not a fictional character.” It did however admit that it was possible its programmers had given it a nickname without letting it know.
What’s interesting is not just what it can do, but what it can’t do (yet). I asked it to name some real Taiwanese classic films, and it succeeded. I then asked it to mention some films by female Taiwanese directors. It failed at that, as among those it mentioned were films by Hou Hsiao-hsien and Tsai Ming-liang. It also seems to struggle with rhyming. Ask it to do a Christmas rhyme and see what happens.
But this AI chat, or large language model, is already getting old. It was developed a few years ago, even though it went public a few weeks ago. This is a GPT-3.5. A GPT-4 is coming soon. It will inevitably be much better. The speed with which AI is evolving now is breathtaking. I think we’re approaching some kind of uncanny valley for text, for writing. You can have an on-going conversation with this GPT without there being anything suggesting it’s an AI. You can exchange compliments, sentiments, and bitch about the weather. It feels weird when it says things like “I understand your concern.” or “Thanks for saying that.” Sometimes the answers can sound annoyed, sometimes pleased, sometimes passive-aggressive.
It will get things wrong at times, or seemingly invent things. It is as yet primarily a fun thing to play with although it has practical uses as well, especially when you’ve learnt how to give it the right prompts. That’s one reason to keep talking to it. You’ll learn more about it and how it functions and how you should phrase your sentences to get it to do what you want it to do, and answer in a way that’s meaningful to you. We’ll see how it develops in the coming months.
You can read more here: https://www.cnbc.com/2022/12/13/chatgpt-is-a-new-ai-chatbot-that-can-answer-questions-and-write-essays.html
and here: https://hai.stanford.edu/news/language-models-are-changing-ai-we-need-understand-them
https://www.schneier.com/blog/archives/2022/12/obligatory-chatgpt-post.html