How we implement this about devil HOT 9 OPEN

@benjefferies, @DuNuNuBatman, @garrettg123, @senovr, @camel2243, @jakob1379, @spinute, @km-tr
hi folks, sorry for bothering, but i thought you might be interested too as we all collaborated on opencommit

from devil.

let me keep it here https://github.com/Skyvern-AI/skyvern, mb useful for automating browsing

from devil.

@malpou, @jaroslaw-weber, @sebastienfi, @avighnac, @jlopezlira, @zavodnyuk, @evandeininger, @joshuahamlet, @matt-degraffenreid, @nader-zouaoui

hi guys, i've taken the liberty to invite you here in case you want to help build this piece of software :) This will be a great fun i promise and good foundation to other tech we can collaborate on

if so—please share your knowledge and vision on how to build this. I'm thinking about same CLI approach we do for https://github.com/di-sukharev/opencommit

from devil.

if you recall any libraries we can use to speedrun the 0.0.1 version and see how it behaves—please share <3

from devil.

i also played with GPT solving a test in a REPL mode here https://github.com/di-sukharev/AI-TDD, so it kinda feels that a Devin like agent can be built in few weeks (assuming we use CLI to chat and just log all messages to stdout)

from devil.

lets be honest Devin is shit, but when it gets to gpt6, then it should be capable of solving complex tasks such as Code and deploy instagram clone.

The main idea behind this project is never write a single line of code again or fuck with configuring zoo of 3rd party libraries, but make software engineering fun as building lego we used to build as kids—you help the Tool (i called it Devinl, open-source is a meme anyway) to build the app, it performs the heavy lifting and manual labour (aka coding) while you are answering clarification questions, sharing links, explain how you see the implementation.

we can now build a gpt4 expensive and slow version of such a tool, then put it on gpt5 rails and see how it behaves

i think we can build it today by smart hacks like splitting one Devin into 10 subDevins or putting a Supervisor to look at every subDevin response and tell him where and how to improve. and even if we ending up with 10 SubDevins and 10 Supervisors for each of the SubDevin is slow and expensive, we solve this problems later. first is making it consistently solve problems we give him

from devil.

Dude. your funny. i wish we could grab a coffee in berlin or smthn

from devil.

thanks man lol, this makes me think about the Devil again, not sure this is actually possibly with gpt4 even as a proof of concept. Devin is dead because it's not capable of anything harder than ugly todo app, bc when you show a simple console error (or any issue you struggle with) to gpt—it makes it even worse, it doesnt help (this is my latest experience), it wont be able to solve anything more complex than generate and deploy an empty html page, and we would end up scripting a yet one more no-code tool

i should probably look into gpt-engineer deeper or auto-gpt or alternatives, still thinking

from devil.

btw, if anyone wants to play with "how to automate coding", you may look at https://github.com/di-sukharev/AI-TDD — LLM generates code in a loop until it solves your test, it's already kinda built, but still requires fixing bugs here and there :)

i want to experiment with this further in coming days, bc it seems the only working way how to explicitly explain LLM what you want it to code, otherwise prompting with free text works on a very very simple cases only which is not enough for me.

from devil.