oak sequoia
Hi guys! I would love your help to develop a small educational project.
The idea is to completely dismantle the current educational system, using tools like OpenIA (ChatGPT...), to create a research paper using the scientific method completely synthesized by an AI, with references, experimental part...
Generating a 40 pages paper.
Can you think of anything?
blazing minnow
so... you want to make research papers minus the research?
oak sequoia
so... you want to make research papers minus the research?
The idea is to generate the entire research paper, solely with AI, yes.
It is only the experimental part of a much larger work. The concept is to demonstrate that it can or cannot be done.
As a model I had thought of using gpt4-32k (I have access to it).
The problem is that I don't know how to handle the request and prompts system to get a meaningful response.
Above all, because of the issue of context
blazing minnow
yep, your main problem on that is simply the fact that the you can't fit that much text into 32K tokens
it will require a smart chunking of the requests and some prompt engeneering for the sections of the ""paper"" it wil be writing
oak sequoia
yep, your main problem on that is simply the fact that the you can't fit that mu...
Yes, that's the problem 😦
I don't quite know how to handle it
Obviously this will only serve to demonstrate that an AI is not yet capable of replacing a good researcher, the idea is to prove it somehow.
A few months ago I managed to generate something with gpt3.5-turbo, and the results without having a context and only with the inputs from an input index were not too bad, but the sentences always ended with "finally", "in summary", and they were all exactly the same length. So I think yes, it will be a big challenge.
blazing minnow
when I have this sort of problem I usually use the AI to summarize the big chunks of text, so it knows the bullet points of what it already generated, that way, it can coherently reference previous text or avoid repeating itself when generating
oak sequoia
when I have this sort of problem I usually use the AI to summarize the big chunk...
And how can I divide the requests?
oak sequoia
@blazing minnow
blazing minnow
just make individual requests for diferent goals, take a look at the docs: https://platform.openai.com/docs/api-reference/chat
oak sequoia
just make individual requests for diferent goals, take a look at the docs: https...
And the prompt?
oak sequoia
@blazing minnow
blazing minnow
hello
stoic tundra
You need to stack it with the information it needs from the real sources. That keeps it from fabulating.
oak sequoia
You need to stack it with the information it needs from the real sources. That ...
Well... that may not be a problem, the problem is how do you structure the requests to get a valid response?
stoic tundra
It's a mirror. Give it the kind of thing you want with the code converter, say: completely understand the structure of this document. Then take your time, think about it, and create a text file that replicates the structure but with x topic? The way to figure it out is to endlessly bang your head against it with lateral approaches. It's demigod smart, but only if you ask it in a way that indicates the style of response you want.
Ask it directly what you want and say "What am I missing here?" Give it your full context.
oak sequoia
Code converter? I can't afford the PRO
Or use the API frequently for testing (due to high costs).
stoic tundra
Then you can feed it a table of contents and some chunks. The PRO is gonna be necessary to get it really smart, but there are hacks around. The problem is the context issue. I just needs to know what you want
3.5 isn't bright enough for writing a paper you'd want to sign your name to.
oak sequoia
3.5 isn't bright enough for writing a paper you'd want to sign your name to.
The article will evidently not be signed, published or anything else.
What I have done is to do some work on how LLMs work and check whether or not they can generate a scientific research paper.
stoic tundra
Keep poking at it. Each of the models shows its own weaknesses through asking it to do what you think are impossible things. In GPT4, just defining how much output you want with each one helps.
oak sequoia
The API can be accessed by other methods, yes, I know. But I want to be "legal".
But how do I test it, without access to the API? At most I can do a generation. No more
stoic tundra
THink it all through, then DM me your ideas. I'll do a couple of rounds of input. You'd need to tell me your full context for wanting the paper and what it can do. I can dump it into Plus, which is very close to the API, just in a box with a leash on.
Is this a school paper or for what? But yeah, we can take it off this forum if you're interested. Otherwise, ask in the 3 forums, if that's what you're using.
oak sequoia
THink it all through, then DM me your ideas. I'll do a couple of rounds of inp...
Of course thanks
I explain everything to you privately