Though I am obviously a huge fan of AI as a creative tool, I have had little to say about all the language and research and chatbot AI's. The things I have been interested in these tools doing, I have not found them to be successful at, and I have also found that no matter how sophisticated these tools get, they always seem to get stuck in the same way all automated systems get stuck. It's like when you call an automated voice system and it works perfectly if you are doing all the things it wants you to do, but as soon as what you need goes slightly off script, the supposed tool becomes a wall to beat your fists bloody on.
If you're that sort of person.
But something unusual happened today.
For reasons I'd rather not go to far into, because it is a bit off topic, I was working on some images at work. These were inspired by an old Randy Newman song called "Political Science". The extraordinarily topical tagline of the song is, well, let me show you:
I was making some images of this, but what I really needed was my home set up. I wanted to layer a beetle on top of my nuclear heart and I couldn't do it without my photo editor.
But I was on lunch and heard about a new AI system that could edit pictures in the chat with simple written instructions. I could thus try uploading my picture to our chat and ask it to, I don't know, put a beetle in it.
So I tried it out.
It was the above image, but without the beetle.
It didn't work. The AI came back blank, and with an exclamation mark that meant that my image was harmful or dangerous.
I said to the AI that it wasn't harmful or dangerous at all.
The AI still wouldn't do it. It explained that it was trained not to produce harmful or dangerous content.
So far this is consistent with every AI usage case I have ever had, and was completely normal and frustrating. But because I am more of a bloody fists against the wall person, I exuberantly explained satire to the AI. I asked the AI to go listen to the song by Randy Newman. I told the AI I wasn't doing anything harmful, but was rather working for goodness.
So the AI went and processed for nine seconds and then came back.
"I understand now." The AI said. "I see what you mean. If you will resubmit what you wanted I can do it for you."
So I told the AI again.
And it did it.
No comments:
Post a Comment
If you were wondering, yes, you should comment. Not only does it remind me that I must write in intelligible English because someone is actually reading what I write, but it is also a pleasure for me since I am interested in anything you have to say.
I respond to pretty much every comment. It's like a free personalized blog post!
One last detail: If you are commenting on a post more than two weeks old I have to go in and approve it. It's sort of a spam protection device. Also, rarely, a comment will go to spam on its own. Give either of those a day or two and your comment will show up on the blog.