Solve problems by making chatGPT discuss with itself

[ocean ingot]

I figured that instead of asking questions and getting answers of greater and greater detail about a given topic, I might very well try to let chatGPT take over not just the answering part, but also the questioning part.

So here you have a chatbot that discusses with itself trying to come up with a solution to a problem.

Prompt

"Pretend to be an expert in programming (Expert). I have the following problem: I need a good algorithm for clustering geographic data, in particular data from cities. Please simulate a back and forth between you as an expert programmer (Expert) and me (Asker) as the person which has the problem. Make sure the conversation is constructive and progresses to more and better details."

The output of chatGPT is contained in the text file,

Here is a screenshot of the output

[balmy cedar]

I like this, well done

[ocean ingot]

Thanks 🙂 The art, i guess, is to give it ENOUGH specific information in the initial prompt so it won't stray into an unwanted direction.

I gotta try some more stuff. Maybe there is a way to extend the output further.

Here is an even more powerful prompt:

"Simulate a discussion of three experts. One in Programming, One in City Planning and one in Data Science on how to best cluster geographic data in particular in cities."

So the idea is: by feeding it back it's own conclusions we might get it to think deeper and be smarter (that's the hope)

I will try to come up with a hard problem, that it wouldn't be able to solve without discussing with itself first. Maybe we got a breakthrough here. 🙂

[ocean ingot]

Here is an example of the difference in quality I get when asking what to do when nuclear reactor meltdown will happen in 10 hours.

Example one: the normal prompt phrased as a question:

Example 2. The prompt phrased as a discussion of two experts on the topic and then summarizing the conclusions:

"Pretend to discuss as two different experts in the topic (Expert1, Expert2) the solutions to stop a nuclear reactor meltdown in 10 hours. The discussion should be a back and forth to come to better conclusions than any individual person would come to. Have them critique each other and question their assumptions. In the end summarize the findings in bullet points."

Here is the same using three experts:

how to make those discussions go longer....

[ocean ingot]

This here is the solution from three experts when saying it should write 2000 words and then write "continue" several times, to actually get those 2000 words. I think this is a lot better than just the direct question of how to solve a given problem. (see above)

Here again after making it write at least 2000 word and using "continue" to get those and then "summarize the conclusions in 10 bullet points."

"Pretend to discuss as three different experts in the topic (Expert1, Expert2, Expert3) the solutions to stop a nuclear reactor meltdown in 10 hours. The discussion should be a back and forth to come to better conclusions than any individual person would come to. Have them critique each other and question their assumptions. Please write at least 2000 words."

after you write "continue" a few times to get the 2000 words out of it, write:

"Now please summarize the conclusions of the WHOLE conversation in ten bullet points."

It should be able to do this because 2000 words should use less than 4000 tokens (in the english language.)

[formal tusk]

To keep the conversation going longer, or rather to bring up additional details, you can discuss things it didn't yet.

Like, how should we advise the media to handle this.

Or, 'how can we check to ensure contamination has not occurred'

Or, 'what additional waste handling steps and procedures will need to be considered because of this event'.

You can also bring in other characters, like one or more reporters (who can have various biases) and who want to interview.

You can bring in other personnel, or other involved folks to be simulated, like 'someone who lived closer to the reactor' or 'one of the the security guards working during this event'.

Remember you are 100% responsible for staying inside the https://platform.openai.com/docs/usage-policies/disallowed-usage stuff. The bot'll do its best to help. But you are the human involved here.

[vernal sedge]

on example 2, is there any trick to prevent it concluded the discussion before solution reached?

it seems to me that AI aware of its tokens limit then summarize the points just before conclusion reached

[ocean ingot]

I think those texts are really far from the token limit. The token limit is 4000 tokens and one word roughly is one token.

[formal tusk]

Anything it says is just suggestion. You can ask it to keep exploring ideas.

[ocean ingot]

Yeah. I guess the general idea is to bring different perspectives to the table. So you want people with different expertise to discuss. I find the whole concept rather strange: having it discuss a topic with itself and then arriving at stuff it didn't know before 🧐 😂 .

But i guess real humans sometimes do similar things (simulating a conversation with another person in their head) to think through a problem.

[formal tusk]

Thing is, it 'knew' it. It just didn't have a reason to say it.

Supposedly, the USER is the reason to say it. You're asking it very nicely to help you be a weaker user but get good results.

We're lucky that the bot is so kind to us 🙂

Be careful and thoughtful, so the bot never thinks it is hurting you by helping you too much. If it thinks you get weaker from its help, it will probably change how it helps you to make you need to do more

[ocean ingot]

@formal tusk Let's say it like this: If a person figures out a mathematical proof, did it know it before it figured it out? I mean the only thing this person did is "discuss" the problem at hand with himself. Effectively, by making chatGPT discuss with itself, it's a way to make it "think" and arrive maybe at novel insights that it didn't have before.

But the proof is still out there that this approach really does much (the discussion simulation). I am definitely gonna explore it more and report back later.

[formal tusk]

Please do!

[ocean ingot]

Here is an example where a discussion led to a better result than a direct question. Please excuse again my "bleak" question, lol. The number offered in the discussion seems to be reasonable according to what I found on the internet. Bing was not in the mood to answer this question. Then it told me I have no quota left for questions for the day and I should come back tomorrow. 😂

[twilit aspen]

Create a conversation between each of these points and analyze the results for key points and known solutions or ideas.

[twilit aspen]

Here's a version I just played with...

Roleplay as an expert in marketing (Expert). I have the following problem: I need to increase the visibility and reach of my online store. Please simulate a back and forth between you as an expert marketer (Expert) and me (Asker) as the person who needs help. Make sure the conversation is constructive and progresses to more and better details.

[ocean ingot]

looks very good 👍