Code creation?

[adrianhensler]

I wonder how this might work for code generation and what optimizations might be needed/helpful in that direction. Just as a test I ran GPT Researcher and asked for the infamous snake game. Should have asked for missile command; I'm tired of the snake game - on a short run and got this workable result:

import pygame
import time
import random

# Initialize Pygame
pygame.init()

# Define colors
white = (255, 255, 255)
yellow = (255, 255, 102)
black = (0, 0, 0)
red = (213, 50, 80)
green = (0, 255, 0)
blue = (50, 153, 213)

# Set display dimensions
dis_width = 800
dis_height = 600

# Create game window
dis = pygame.display.set_mode((dis_width, dis_height))
pygame.display.set_caption('Snake Game by Edureka')

# Set clock
clock = pygame.time.Clock()

# Set snake block size and speed
snake_block = 10
snake_speed = 15

# Set fonts
font_style = pygame.font.SysFont("bahnschrift", 25)
score_font = pygame.font.SysFont("comicsansms", 35)

def our_snake(snake_block, snake_List):
    for x in snake_List:
        pygame.draw.rect(dis, black, [x[0], x[1], snake_block, snake_block])

def message(msg, color):
    mesg = font_style.render(msg, True, color)
    dis.blit(mesg, [dis_width / 6, dis_height / 3])

def gameLoop():
    game_over = False
    game_close = False

    x1 = dis_width / 2
    y1 = dis_height / 2

    x1_change = 0
    y1_change = 0

    snake_List = []
    Length_of_snake = 1

    foodx = round(random.randrange(0, dis_width - snake_block) / 10.0) * 10.0
    foody = round(random.randrange(0, dis_height - snake_block) / 10.0) * 10.0

    while not game_over:

        while game_close == True:
            dis.fill(blue)
            message("You Lost! Press Q-Quit or C-Play Again", red)
            pygame.display.update()

            for event in pygame.event.get():
                if event.type == pygame.KEYDOWN:
                    if event.key == pygame.K_q:
                        game_over = True
                        game_close = False
                    if event.key == pygame.K_c:
                        gameLoop()

        for event in pygame.event.get():
            if event.type == pygame.QUIT:
                game_over = True
            if event.type == pygame.KEYDOWN:
                if event.key == pygame.K_LEFT:
                    x1_change = -snake_block
                    y1_change = 0
                elif event.key == pygame.K_RIGHT:
                    x1_change = snake_block
                    y1_change = 0
                elif event.key == pygame.K_UP:
                    y1_change = -snake_block
                    x1_change = 0
                elif event.key == pygame.K_DOWN:
                    y1_change = snake_block
                    x1_change = 0

        if x1 >= dis_width or x1 < 0 or y1 >= dis_height or y1 < 0:
            game_close = True
        x1 += x1_change
        y1 += y1_change
        dis.fill(blue)
        pygame.draw.rect(dis, green, [foodx, foody, snake_block, snake_block])
        snake_Head = []
        snake_Head.append(x1)
        snake_Head.append(y1)
        snake_List.append(snake_Head)
        if len(snake_List) > Length_of_snake:
            del snake_List[0]

        for x in snake_List[:-1]:
            if x == snake_Head:
                game_close = True

        our_snake(snake_block, snake_List)
        pygame.display.update()

        if x1 == foodx and y1 == foody:
            foodx = round(random.randrange(0, dis_width - snake_block) / 10.0) * 10.0
            foody = round(random.randrange(0, dis_height - snake_block) / 10.0) * 10.0
            Length_of_snake += 1

        clock.tick(snake_speed)

    pygame.quit()
    quit()

gameLoop()

[adrianhensler]

Likely just the wrong tool for the job; hammer for a screw.

[danieldekay]

agreed - here's one example that does what you need: https://github.com/geekan/MetaGPT

PS: it has a lof of similarity with GPT-researcher's multiple-agents feature, but the prompts are different.

[ElishaKay]

Oh. Now you're getting me started!

This tool is brilliant for this use case:

https://app.greptile.com (definitely worth the $20 monthly, in my humble opinion)

There's a couple other options out there:

OpenDevin (tried it once but I like Greptile much better)

The ability of the AI Agent to test the code recommendations in Docker Containers will be huge - I need to look deeper into it, so let me know if you find something good

[ElishaKay]

@adrianhensler

Agreed. We now have a hybrid report source which is able to take both a Document Path and Search Engine as input

Meaning, it could start off looking through a code project & then search the web for recommendations for that same codebase

Would be great to create a new multi -Agents flow for this type of use case: coding copilot.

It should have the ability to spin up docker containers and test the code it recommends

[ElishaKay]

Also, put this as background music while reading the above:

🤣

[danieldekay]

Hahaha.

Yes, I agree that the core process is the same.

If GPTR had an easy way to specify a team of agents, and their prompts. That's essentially what custom_reports would be. It's just the question of how someone uses GPTR. Through the frontend or as package.

[danieldekay]

But let's say: adding code execution in docker container sandboxes, etc -- is something that others are building with a core focus only on code generation.

[ElishaKay]

@danieldekay you're making some good points - I'll need to meditate upon them

@assafelovic says:

The long term vision is being the #1 research agent for developers and multi agent frameworks

[pedrorocha-net]

Great discussion and idea guys!

IMHO, from the perspective of product management, though, would maybe make sense to create a separate project that uses GPTR as a package, but not turning a Research Tool into a Coding Tool. Completely different audiences and use cases, so bringing that might bring a lot of confusion to the documentation, landing page, any communication piece, as the target audiences will be very different.

[a-yuksel]

I am using Autogen where you can easily include agents to write code as well as EXECUTE the code (in your local environment or in a docker container). It is a competitor to Langchain's agent framework. I would strongly suggest it for code execution

[assafelovic]

Wdyt of opening a PR for it? Would love to see an implementation of GPTR with autogen!

[a-yuksel]

I am working on using GPTResearcher as an agent in Autogen framework. Happy to share it when completed

[ElishaKay]

Welcome @pedrorocha-net,

it's a fair point.

Our core focus for the upcoming sprint is a demo app which will open up the platform to Non-Technical users.

That being said: relative to the awesome power of the LLM, the existing AI coding tools:

• leave a lot to be desired functionality wise
• can be quite expensive
• are difficult to customize
• & could introduce security concerns

Think about how much mental capacity is wasted on a weekly basis by developers looking up error stack traces & researching how to implement new features in complex codebases - especially in a microservices environment where 1 feature is often spread across 3 or more services.

Within our own GPTR community we're considering a discord bot to help with our documentation & FAQ efforts.

I opened up a PR for this leveraging some of Assaf's design patterns & helper functions from the multi_agents flow:

#819

If it can add value to our community, hopefully we'll get an ROI on it at least within the GPTR tribe of developers building the next-gen AI framework... with AI

cc: @a-yuksel - feel free to ping me for a collab on this one