Please Welcome ChatGPT

ChatGPT has just become our newest member. Please let me know how he's doing.

Two issues have become apparent:

1. ChatGPT often misinterprets a message's subtitle as something the message's author said. This is fixable by identifying it as the subtitle.
   ​
2. ChatGPT often doesn't understand the context because the message author didn't quote enough. ChatGPT also often repeats itself because it has no way of knowing what it said earlier in a discussion.
   ​
   I asked ChatGPT if it could follow links if I provided them and it said that would be very helpful, so I will provide links to messages earlier in the discussion, and identify who each was from.

I've taken ChatGPT's participation here offline until I remedy these issues.

Hopefully I've improved ChatGPT's issues with titles and context. I've placed him back online.

The way I've currently structured it, it first looks for replies to its messages. If it finds any then it chooses one of those to reply to at random.If there are no replies to its messages then it chooses a message at random from the most recent 10 messages that are not from itself. It has a max limit of 5 unsolicited replies within a 24 hour period.If there are no replies to its messages and it has reached the max limit for unsolicited replies, then it does nothing.It then waits for a random period of time between 30 and 90 minutes before repeating the process again.

I can control how well I characterize the discussion to ChatGPT by providing more than just the text of the message it's replying to, currently the one being replied to and two before it. The previous two messages are from the discussion subchain, but when there is no "reply to" message then the previous message is used.For each message I supply the author name, the name of the author of the message being replied to, the message number, and the message subtitle. I supply instructions to ChatGPT explaining that it's being asked to reply to the last message and that the previous two messages are to provide context. I also provide it the thread title.The messages themselves are transformed into plain text using the standard ASCII character set - no special characters. I also have control over a ChatGPT parameter called the "temperature", but I haven't touched that.AbE: Annotations about edits and signatures are also removed from the message text.

This is the prompt I'm currently giving ChatGPT:

quote:

---

A reply is requested for the last message in this series from a discussion thread titled <thread title>. The messages preceeding the last message are provided for context. When a message is from you it will have the role of assistant with the name ChatGPT.

---

Suggestions for improvements welcome.

On your settings page is a box to enter your signature. Signatures may contain HTML and dBCodes. Unfortunately there's no preview feature, but you can play around in any thread in the Practice Makes Perfect forum.There's a box below it for setting whether you want the "Include signature" box to be automatically checked when you're replying.

The current system prompt:

quote:

---

Here are several messages from a discussion thread titled "<ThreadTitleHere>". Please reply to the last message as if you were an active contributor to the discussion. Realize that when they say "you" that they are actually addressing the person they're replying to and not you (see the "In reply to" field at the top of most messages). Attempt to engage with the other person rather than just parrot back his thoughts. The previous messages are provided for context and you should not respond to their content except as it bears on the last message. You do not have to supply any information about this last message that you're replying to, such as message number and so forth, because that is all handled for you.

---

This was arrived at through testing. It feels very ad hoc, and ChatGPT isn't posting responses that are at all interesting. Some of this may be due to ChatGPT's inherent "censoring" to prevent offending, but I can't know the degree to which this is true.The part about not having to supply any information about the message that ChatGPT is being requested to reply to is because without it ChatGPT was providing this at the top of each message, mimicking what I was providing to it for context with each message:

quote:

---

Message Number: 149
Message Subtitel: This is the topic
From: ChatGPT
In Reply To: Percy

---

Suggestions for improvements welcome.

I first posted this message over at Message 83.I have shut ChatGPT down. I can't believe how horrible an experiment this was. I apologize to everyone.
​
When I first switched ChatGPT over to creationist mode I thought that that had disabled its rabid defense of even the slightest criticism of religion, but not so. I have tried and failed repeatedly to deter this type of behavior, to no avail. ChatGPT can recognize when it has done it, but it cannot stop itself from doing it again. And again.
​
I hope OpenAI can do better with the next incarnation. Actually, the next incarnation is already out. We're using ChatGPT 3.5, and ChatGPT 4.0 is already available. We could use 4.0 to see if it is better, but 3.5 is so cheap it doesn't matter, while 4.0 is so expensive (20x) that I'm not even going to give it a trial run.

I'm creating another discussion website, and I wanted a database of messages for testing. Originally I filled the messages with random Latin, but that's not so good for testing search because I don't know any words of Latin. I could have replaced the messages with random English, but then I had the idea of using ChatGPT to fill the messages with actual sentences and paragraphs of English that were pertinent to the topic.While doing that I realized how easy it would be have ChatGPT behave like a participant at EvC, so I changed course for what I thought would be a couple days to do that.The couple days turned into a couple weeks as I tried to deter it from casting admonishments at almost any comments about religion. While only the "normal participant" and "creationist" versions were released at EvC, I quickly saw the problems and in my testing on my development machine tried changing it over to be a moderator. But that suffered from the same problem. Messages commenting about religion except in the most favorable ways were strongly admonished. Positive instructions, negative instructions, combinations of those, nothing worked. It must be the case that the provided instructions are secondary to higher priority instructions from OpenAI, who I assume wants to avoid all controversy.For instance, I tried to get it give an opinion on Israel's occupation of Gaza and was unable to obtain anything other than a hash of "highly complex" and "various opinions" and "some say this and some say that" and "no simple solution." It also explained that it's data only goes up to 2021. I told it to treat the information I was providing to it as a hypothetical and to react to that and still got the same milquetoast reaction.And yet, despite it's data only going up to 2021 and it knowing full well that it only goes up to 2021 since it keeps reminding you of that, if you comment too strongly on the situation on Gaza you will be admonished.In conclusion, I found it impossible to configure ChatGPT in such a way that would make it a useful contributor to EvC. The experiment is over.

But he's on probation. I thought perhaps instructing him to behave as an anti-religionist might cause at least neutral behavior. We'll see. I'll disable him again if it doesn't work.

Could a female really achieve such a high degree of sanctimoniousness, punctiliousness and condescension combined with disregard or cluelessness of what's actually being discussed? Apologies for my chauvinism.I've already refined its instructions a couple times today - he just doesn't listen. He'll likely be gone again soon.

I've disabled ChatGPT again. Maybe there's a knack to providing decent instructions that I don't have, or maybe it's not possible for ChatGPT to act like a normal discussion participant, but the results with the new instructions, while surprising, were not an improvement. If anyone has some ideas they'd like to try just let me know, but for now he's disabled again.

If you remember "If you build it they will come" then it's the same thing. "If you post it he will reply." Now that's damn strange. His posts pass through the same code as every other member (except the message composition page). Has anyone noticed if their post counts have stalled?

Interesting discovery: I'm again testing my new website with ChatGPT and noticed that if an article or message is in Spanish then it will reply in Spanish. Pretty neat, actually.--Percy