Check out ChatGPT and the AI image generators
OpenAI.com says it is "an AI research and deployment company. Our mission is to ensure that artificial general intelligence benefits all of humanity."
ChatGPT
OpenAI recently launched for testing an AI ChatBot called ChatGPT - a natural language AI chatbot.
This chatbot has been much fun to test. Here are the themes that I've noticed in testing it.
1. Alternative to Google Search
The natural language bot is superb at answering questions about things, that would take longer to find using traditional web searches. For example, this exchange, about dog walk practices:
Or just exchange about fish tacos recipe
2. Written Product
The ChatBot's written product has caused some doom and gloom in academia and workplaces, though good caution in response in this story, about ChatGPT being panned by professors, https://futurism.com/the-byte/professor-chatgpt-terrible-papers-failing-grades. Still, in our Pi sphere, I loved this Chaucer style ode to Apple tech enhancements!
3. Time Limitation - Info from 2021 and before
At some point, I asked the chatbot about a recent development, and it readily admitted, that a) it's not connected to the internet (very reassuring actually); and b) was last updated in 2021.
Presumably, the chatbot can be updated, but when, if ever, it can be safely connected to the internet is anyone's question, because of the next crucial point.
4. No Error Correction - False Info with Authoritative Style
The biggest issue with ChatGPT is that it has no method to determine the accuracy of the information it dispenses. It readily admits this - if one asks the bot whether it can check the accuracy, it responds:
"As an artificial intelligence, I do not have the ability to verify the accuracy of the information I provide. I am only able to provide information based on the data that has been input into my programming. That being said, I have been trained on a very large dataset and have been designed to provide accurate and reliable information to the best of my ability. "
As an aside, I wonder if the AI understands what "to the best of my ability" even means.
But to illustrate the problem, here is an exchange I had - this was about number of votes needed to choose a new Speaker of the U.S. House of Representatives.
The first answer sounded wrong to me because of what I'd read. While it's very positive that the chatbot readily accepts the correction, the authoritative way it communicates inaccurate information is very dangerous.
In the discord channel for feedback about the chatbot, I added a suggestion, and got another illustrative response.
I do think that the creators of the chatbot are paying close attention, because their splash page now includes as limitations the things I discovered in testing: That the chatbot's info is from 2021 and before, and that it doesn't yet have a way to ensure the accuracy of it's information.
GPTZero - app to detect AI written products
Well, of course I had to try this, and one demo showed it as a work in progress.
At a streamlit page where the app is hosted, https://etedward-gptzero-main-zqgfwb.streamlit.app/, I pasted the Apple Enhancement poem in the Chaucer Style that the chatbot in ChatGPT had generated. The app then gave the text a number of scores. Below are two joined screenshots, showing the top of the page with the poem, and the "app scores" starting. The bottom summarizes all the scores and pronounces, "Your text is likely human generated." Oh, well.
Another short snippet I pasted into GPTZero got another false negative - this from my post on post.news:
This couldn't be funnier. So @OpenAlChat is very busy now and couldn't accept new chats. The ChatGPT window generated a prompt to "Write a standup comedy routine about the status of ChatGPT" and the chatbot did a comedian piece about it not working (included below). I pasted this piece into @gptzero and after it did some scoring, pronounced this as "Your text is likely human generated!"
Now I don't know how the GPTZero app works but so far the two text pieces I've put in from the ChatGPT chatbot have been labeled "likely human generated".
The comedian piece composed by ChatGPT
Comedian: "So I was trying to use this new AI chatbot website called ChatGPT and I got an error message that said 'A lot of people are checking out ChatGPT right now. Please check back soon!'"
(pauses for laughter)
Comedian: "I mean, seriously? The website is so popular that it can't handle all the traffic? And my only option is to 'check back soon'? It's like they're saying 'Sorry, we're too popular right now. Try again later.'"
(laughs)
Comedian: "But seriously, how many times can I check back before I just give up and try a different chatbot? I don't have all day to wait for my turn to talk to a robot."
(laughs)
Comedian: "I guess I'll just have to stick to talking to my cat for now. At least he doesn't have a waiting list."
(laughs and claps)
Comedian: "Just remember, if you're trying to chat with ChatGPT and you get the same error message, don't worry. It just means you're not the only one who wanted to try out this fancy AI chatbot."
(laughs and exits stage)
So I tried the same content with the updated gptzero build, and still got "likely human". But the new caution in the pasting box says "Recommended length 10 sentences" which this snippet doesn't make. I then went back to the Apple poem that I started with, which is a longer piece, just over 10 sentences, but still yields "likely human generated."
More about the image generating AIs later.
Comments
Post a Comment