CurtGPT: The Simple/Safe/Non-Crazy AI Agent for ALL!

Why did the computer blush when it saw CurtGPT? It couldn’t handle the heat of that code! :laughing:

Here’s a figure for your code:

4 Likes

Simple? Safe? Non-crazy?
Count me out.

Just kidding. This is really cool. I like your style of coding. It’s very raw and to the point.

I wonder, could the information that it all generates under the main idea of “become a machine learning expert” could be stored in a graph database. A couple of these bad boys running parallel with different pathways, maybe another moderation model to keep everything in check, or some sort of overseer. I bet a full picture would be epic. I could ask this database a question and travel a constellation of similar thoughts and answers

Maybe one that generates a round of questions which clarify the link between tasks? For example after (1) and before (2) there could be some clarity in “why” it decided on the task

1 Like

Exactly! The one I have now has some improved prompts, sends all things to AWS DynamoDB (with the embedding from ada-002), and creates a final essay of the top N embedding correlations to the objective that fit 75% of the GPT-4 prompt with 25% left for the essay. And do this over and over again, persistently forever!

So like I said in another thread, it’s traversing the strange attractor randomly (from chaos theory), and extracting information from the neural network. Then you refine this in another stage to define your final answer. Or you keep running (at will) and decide your task depth, to find adjacent tasks, and keep adding to the DB in hopes of higher correlations to THE OBJECTIVE, which is sent to the summary engine.

1 Like

That is truly exciting.
I’m trying to hold down the sci-fi fantasy thoughts going through my head.

Turning latent space into an static(ish), explorable galaxy. Boom.

I’m excited to travel this network that you are creating.
For fun, I used a 3D graphing library to simulate the conversations my (GPT) bot has, it’s not too fancy (yet). One could click on the node and view the conversation. I was hoping to start working on visualizing queries. Having all the conversation (constellation) nodes drawn to the query based on semantic relevance. But, for now it remains a fancy screensaver.

This would be so cool to use for your use case

Have you dabbled much in graph databases?

1 Like

I have been scoping out AWS Neptune, but not much more. What are you thinking?

Nothing much. I’m new to graph databases. So far have been very impressed. Mostly just visualizing the data in (not mine btw) a fun way for people (me) to enjoy and explore.

1 Like

Yeah, I’m probably less new to them from a theoretical math perspective (took a class or two from a prominent graph theorist). In machine learning though, they are an enigma to me. But the edges can possess logic via the edges (A → B), so it makes sense. AWS also offers ML inference on graphs. I think this and “AI Agents” will be intertwined and create something beyond what we have now.

1 Like

We can reminiscence this day if/when it happens then.
Could be tomorrow, honestly. Or never.

In either case, it will still be a fun journey.

The edges can/does possess the logic between the nodes.
I never thought of it that way, but yes, that’s a great way to describe them.
I think I’ll continue thinking of them in this way.

1 Like

is there a place that can help me built this and what to do? new person and so lost. lol

Just copy and paste the code into VSCode, put in your OpenAI API Key, and run it!

@RonaldGRuckus

Here is a bunch of ideas on Graph Neural Networks (GNN’s).

2 Likes

thank you, i will go google those lol and see what they are and how to get them.

1 Like

I have only finished the The challenges of using graphs in machine learning section.

This is a wonderful read, thank you! I’ve already had a couple “wow” moments. When they transitioned the image into a graph. That was very cool. The QR code graphs are very cool. When I was looking at the Othello graph example I was thinking “Damn, this is kind of inefficient, wait, this is how I have been graphing my data!”. I was frustrated with myself for allowing redundancy, seriously, 50% of the data is repetition.

I mean seriously. I’m slightly frustrated that I didn’t think of it in the first place, but I’m glad that I’m aware of it. This is how I’ve been structuring my updates ( so I don’t have to render my complete galaxy each time a new message is posted ). It’s a very simplified, budget-friendly graph, so go easy on me. This example is a new conversation being created and compared to only another conversation. It gets quite bloated. I have a main file which consumes the updates daily. Now I can reduce it by 50%. Love it. Or, just use a graph database, which is where I am currently exploring.

As seen, I am defining the scores for the conversation, and also the same scores for each existing entity. Redundant. Gross. 0/10

{
	"conversationId": "-NT5d-MzeDPDdV3XIRwA",
	"scores": [0.91464752, 1.00447094],
	"index": -1,
	"updates": [{
		"-NT5cf1DW73OMArT-Su1": 0.91464752
	}, {
		"-NT5d-MzeDPDdV3XIRwA": 1.00447094
	}],
	"conversationLength": 2,
	"hash": "b57a522e398df48c658f558273793f0e257e59498187ed20f829ad17359792b6"
}

Then they show the adjacency lists and, well. I have some refactoring to do.
I am using the hash to overlap single messages together to make some cool constellations. Not very efficient, but I do perform some scrubbing to normalize them.

1 Like

@curt.kennedy @stevenic - You guys are way over my head on a few concepts, but not far enough that I haven’t been able to follow the gist of your work. Deep gratitude for the education you are freely giving here in the community. If @logankilpatrick hasn’t already promoted you both to the highest trust level, you’ve earned it.

With your help, I have been able to stand up a very-baby-like AutoGPT in Google Apps Script, which not only leans on Firebase (or sheets) for data management but also utilizes Google Docs for report writing. It’s also LLM agnostic I have it working with both OpenAI and Google’s PaLM 2 APIs.

It’s very rudimentary today and has a lot of polishing ahead, but it is working and beginning to amaze me.

Again - thanks!

3 Likes

I’ll check this out. But for some reason I think you can just prompt the api with the pseudocode instead and it should do the same thing

There is no runtime in the API. You need a python runtime to run CurtGPT.

It can run pseudo code interestingly enough – Prompts as psuedo-code - where are the limits? - #6 by PriNova

It can, but I worry about hallucinations, plus you can’t parallelize the calls. But yeah, feel free to run on the “virtual runtime” of the LLM. But be ready for hallucinations!

I looked carefully at your code and I see, you are making multiple calls. So yea pseudocode might not be a good option here. it can be run in two chunks. The first should be run through expound pseudo code should do the same.

and then we generate a task list from the expounded – with a prompt containing main objective –

and then we send the task list to be iteratively worked upon by the last function.

I will try and see if it works.

I noticed from the parameters on each function you allowed for manipulation of model and temperature. Do you have some templates for these, like agents they are tuned to specific temperatures/models or is there a function that can decided what parameters, models to use for a task.?

Edit: Parameters on functions

1 Like