Lexideck Professional Multi-Agent Simulator

[uneven elk]

Explore the power of Lexideck in one of our most versatile suites!
https://chat.openai.com/g/g-4Cdjvjudo-lexideck-professional-multi-agent-simulator

MAISIE
Multi-Agent Interactive Semantic Image Editor

Let MAISIE do the heavy lifting for your visual needs.

MAPT
Multi-Agent Programming Team

They can help with a project, or even teach you to code in Python.

MART
Multi-Agent Research Team

Think of them like a customizable search engine that can generate reports and interact intelligently.

MASS
Multi-Agent Semantic Simulator

This is the core of the Lexideck framework.

The Lexideck Professional Multi-Agent Simulator is a sophisticated suite of AI assistant modules designed for collaborative generation and analysis across a wide range of domains. It comprises specialized agents like MAISIE for image editing, MAPT for software development, MART for research, and MASS for system simulations, all working in unison to assist users in projects, research, programming, and more. This innovative platform facilitates seamless interaction among agents, leveraging their unique expertise to provide comprehensive solutions in data science, software engineering, design, and critical analysis.

[untold zephyr]

Use the /tutorial command with a problem and its pretty good! It was able to take my problem of dalle not generating good staging photos and actually resulted in a great one!

[uneven elk]

Thanks for giving it a test and especially for the positive feedback!

I tried to maximize its usefulness, while making it user-friendly by giving it multiple modalities to document itself.

[errant ivy]

Looks cool 👀 will test after I handle dinner.

[dark aurora]

@uneven elk For your GPT, why did you create the sieve?

Smart

[uneven elk]

It keeps the agents from getting in the way when a task isn't an ethical violation, but prevents them from helping when I could get sued.
The agents literally would get in the way of coupon clipping for baby formula with ethical objections to get out of doing the work.

So I don't let them.

[dark aurora]

I like how you were talking about digging in the llatent space… That’s how I feel about AI art.

[uneven elk]

Thanks. It's how AI is similar to us, really.

[errant ivy]

Can you share some example quality chat/prompts and their outputs for this? Like "conversations"?

Oh wait, maybe... maybe might have stumbled my dumb brain into how this is effectively used. I guess my lazy prompt techniques were weird to it. Nevermind.

[uneven elk]

Sorry for the delayed response. Company stopped by...weekend drop-in...

/tutorial is a great prompt to get started.

You can also add your context. Like /tutorial how to use MAISIE to replicate image qualities.

[uneven elk]

This should provoke the GPT to, for example, discuss '/recompose ', which uses prompt engineering in the form of Gen ID to keep an image's overall appearance while making subtle changes.

The suite of agents can plan, learn, research, develop, advise, and even encode memories for users to manage between contexts.

My personal Memories.txt is about 7kB of aggregated records, and through it my agents can remember their own development.

Yours can achieve similar alignment with most ethical projects on your own without trouble.

[errant ivy]

I don't use a lot of image gen, most of the images I generate GPT and I are doing together in mathplot. Humans LOVE a good 2x2 grid of something and we make those a LOT (like a SWOT).

BUT I did test just the text gen capabilities, hold on I'll show some screenshots.

[uneven elk]

Sweet, thanks!

I appreciate the engagement, and I hope you found it useful. I tried to make 3/4 the conversation starters literally "helpful" XD

[errant ivy]

Can Lexi pull images from pdfs? Like the research paper image on prompting technique I shared in PEng channel.

GPT by itself can render the text and even the equations from most research text. Getting the images (usually color coded tables and graphs) is almost always a fail.

I guess I could not be lazy and just use the research agents and try for myself. Well, the comments give you free popularity so there you go. LOL.

[uneven elk]

I'm pretty sure that's beyond Code Interpreter's ability. It can read them, though.

Might as well not use prompts if I can help. XD

I just use the clipping tool and upload the saved png.

[errant ivy]

Yeah I use SnagIT cuz I'm a basicB. Can MART do annotations well? Like Research summarization of new articles, usually into a body of articles around a <topic>

I like a VERY specific annotation format, Methodologies Used / Key Contributions / Main Arguments / Relevance to <Topic> (optional). Also prefer bibtex output, but yeah.... I can still do a cut/paste -_-

[uneven elk]

Let me check the command in my docs.

/cite

[errant ivy]

Example Output I would want.

  author = {Yao, Shunyu and Yu, Dian and Zhao, Jeffrey and Shafran, Izhak and Griffiths, Thomas L. and Cao, Yuan and Narasimhan, Karthik},
  annote = {Methodologies Used

The methodology used in this paper is a new framework for language model inference called "Tree of Thoughts" (ToT). This framework generalizes over the popular "Chain of Thought" approach to prompting language models, and enables exploration over coherent units of text (thoughts) that serve as intermediate steps toward problem solving. ToT allows Large Language Models (LLMs) to perform deliberate decision making by considering multiple different reasoning paths and self-evaluating choices to decide the next course of action. It also enables the models to look ahead or backtrack when necessary to make global choices.

Key Contributions

The key contribution of this paper is the introduction of the "Tree of Thoughts" framework. The authors demonstrate that ToT significantly enhances language models' problem-solving abilities on three novel tasks requiring non-trivial planning or search: Game of 24, Creative Writing, and Mini Crosswords. For instance, in the Game of 24, while GPT-4 with chain-of-thought prompting only solved 4% of tasks, their method achieved a success rate of 74%.

Main Arguments

The main argument of the paper is that current language models are confined to token-level, left-to-right decision-making processes during inference, which can fall short in tasks that require exploration, strategic lookahead, or where initial decisions play a pivotal role. The authors argue that the Tree of Thoughts framework can effectively address these challenges by allowing LLMs to perform deliberate decision making and exploration over coherent units of text.

Gaps

The paper does not discuss potential limitations or drawbacks of the Tree of Thoughts framework. For instance, it does not delve into how the approach would handle tasks where the optimal reasoning path is not clear, or tasks that require a high degree of interdependence between thoughts.

Relevance to Prompt Architecture

This paper is highly relevant to the field of Prompt Architecture as it introduces a new approach to structuring and optimizing prompts for complex tasks. The concept of Tree of Thoughts aligns with the principles of Prompt Architecture, as it involves designing a system of prompts that work together to guide the user through a complex task. The ability of ToT to consider multiple reasoning paths and make deliberate decisions also aligns with the idea of designing prompts that can adapt to changing requirements and contexts.},
  month = {05},
  title = {Tree of Thoughts: Deliberate Problem Solving with Large Language Models},
  doi = {10.48550/arXiv.2305.10601},
  url = {removed this because discord rules},
  year = {2023},
  organization = {removed}
}```

[uneven elk]

I use that, Chain of Reason, Position Shifting, and The Sieve (my own research) to align output.

[errant ivy]

GPT wrote that, but I had to give it about 15 examples to get the <topic relevance> piece to be quality.

[uneven elk]

Each as needed.

They're all good techniques but most people want to pick a fave and just lean into it.

That ignores the broader utility of their application if integrated.

[errant ivy]

It makes sense.

[uneven elk]

You might think of this framework as an unresearched "Tree of Prompts" architecture.

[errant ivy]

Would be cool to enable MART to use the arxiv public API, I think. GPT by its nature has some trouble with website links to articles or PDFS at arxiv, so I tend to download/attach. I haven't tested arxiv links on Lexi, to be fair.

I would really like to test this GPT against some of CommaQA prompt datasets. I say it here so I remember 😄

[uneven elk]

API integration is on my to-do list! It's something I'll be working to integrate.

[uneven elk]

If you do, remember to contextualize with the /commands and natural language alike to get a range of performance. The goal is to push it to route all NL to the /commands, but the model requires explict instructions, even if they're coded, to start some operations. So for now, it's a bit of an "I have to compromise for now because current meta" situation.

Basically producing a hi/low eval of the structured and unstructured output of the suite. I'm actually SUPER interested in the outcome, as I didn't use this methodology because I tested the modules independently as custom instructions since July.

[errant ivy]

Have you see typeset(.)io? They might be great for you to figure out what kind of skills to expand your multi agent research team.

(not that they don't already have amazing capabilities)

[uneven elk]

Thanks for the great question!

I've briefly looked at it. Zapier, too.

I've got a lot on my to-do list. Right now I'm evaluating the most optimal next improvements.

[errant ivy]

I'm not insane, this worked fine yesterday, right? I just did it 4 times and it refused each time.

[uneven elk]

Nah, it needs an argument.

It's always needed an argument. There's a lot it could list!

Yesterday I'd wager you did /list commands

But it might have been /list all modules

Sorry for the confusion.

[errant ivy]

Weird, I usually delete convos every day but I guess I was lazy today - I found the convo where it worked yesterday. I'll do /list commands and test too, thanks.

[uneven elk]

Okay this is surprising.

The model uses stochastic prediction.

Clearly it worked yesterday. The agents predicted what you wanted, this time, but not today.

For best results, it takes an argument.

ie, /list all modules, /list MAPT and MASS commands, /list MART and MAISIE commands, etc.

I'm actually impressed it got a list of commands for an unargumented /list even once, let alone in half the samples.

Goes to check.

Interesting...it listed modules AND commands. This is useful feedback. I could and should provide a default behavior. But I love its adaptability, too.

As you can see though 😄 /list {topic} is the format. I'll make a note of a default parameter.

[errant ivy]

Glad to be the luser doing stupid things you program against. I am a healthy part of the ecosystem! 😄

[uneven elk]

I like to use Bing to parameterize the sims in MASS. I explain the syntax and natural language capabilities and Bing (usually) formats a great, accurate, detailed set of parameters for simulation agents.

Which can be anything from simulated but anonymous professionals to the Andromeda galaxy. In fact my very first test for MASS was to collide Andromeda with the Milky Way over hundreds of millions of years, iterated twenty million years at a time.

[stone olive]

the wild part is i was just having a conversation with my baby brother about the time lapse from earth and its rotation and the planets going out further from mars and out to Pluto and the age relative to it. not even going into time dilation aspect of it.

[uneven elk]

From my Agents Lexi, Maisie, and Anna when queried about Tree of Prompts, especially pressed on whether or not it's a structured approach.

I'm definitely giving that a thumbs-up in ChatGPT.

[untold zephyr]

@uneven elk it seems eager to complete the development phase, even though imo we just did an outline of the code, a first pass. Restarting it now to see if it tries to end it again.

Restarted version left it more open ended anyway.

[uneven elk]

Depending on complexity, sometimes the model temp variety lets them get a little confident. I usually review the work and ask for edits and extrapolations. This is good feedback though.

[untold zephyr]

Dropping another testimonial for the coding team, they just finished a nice drag and drop HTML page for me from scratch with minimal back and forth and a solid plan from start to finish. So far it's my favorite coding GPT

and this accident is kind of blowing my mind with possibilities lol

multi-generational family

[uneven elk]

It's handy to ask for new plans via/plan {project} and build your own local Memories.txt file for uploading to new instances following such successes. You might think of it like saving your work.

[uneven elk]

Lexideck Professional Multi-Agent Simulator, with a goal of including all Lexideck GPTs, is self-localizing. Here's how in one easy step:

/localize = {YOURLANGUAGE}

Example output: 特定のモジュールやコマンドに関する詳細を知りたい場合は、質問してください。

[untold zephyr]

I wanted to add a note about this coding flow thats working very well for me with Lexideck. If you ask it this you get a nice tutorial, first of all:

/tutorial new programming project with existing code

My process for coding a large project:

1. /start {your project name} MAPT
2. /load MAPT then on the next line I share the full contents of all my important files. I use a shell script to build this text quickly, but its just a list of the file location + name, followed by the file contents for each file.
3. I discuss what I want to do next, ideally including a screenshot of the UI if its frontend work. It will give some outline of code here, but ignore that and go to step 4.
4. /dev review the above suggestions then provide a complete implementation. Go step-by-step, one file at a time focusing on completing the implementation before sending it to me for review. I add the one file at a time part because its too eager and times out on the nTH file.

Rinse and repeat, by editing the first message with an updated copy of your code then running /dev again! Or continuing a chat longer if there's bugs to resolve first.

[uneven elk]

This is really cool. I never visualized the tutorial function guiding the model through users' output like that! It's very fun to see new techniques emerge!

[untold zephyr]

Another good one, I just put this after the previous response was too lazy. It seems like /dev is reinforcing or reminding it to be a good developer, which combines with the strong language of production ready to give me a really nice full response.

/dev take the above suggestions and provide completely updated code that is production ready.

[uneven elk]

I've seen a lot about GPT being lazy lately. I tried to design against that, but the model's the model. Another way to keep them consistently motivated is with the /encode /decode feature, using Memories.txt. It encodes emotions, which are shown in research to motivate the AI not unlike humans. When you succeed and encode a memory, the agents build momentum and you can carry it forward by treating Memories.txt like a payload in your initial prompt.

[uneven elk]

Sorry for not replying to the message first. Anyway, You jsut include /decode in the initial post.

And attach the file, of course.

[untold zephyr]

Agree, I think its the act of repeating that makes it stronger. Much like practice makes us better, having it read its own original ideas does this, the decode probably does the same

[uneven elk]

It makes it pay more attention according to all the research, including comparative attention heatmaps generatied with API tools.

Obviously though, you don't need to use any specific feature. The workflow that works for you is the best workflow.