Introducing Devid, AI Software Engineer You Can Actually Use
Вставка
- Опубліковано 4 чер 2024
- Meet Devid, an actually usable open-source implementation of Devin.👨🏼💻
🧪Agency Swarm Lab (Repo from the video):
github.com/VRSEN/agency-swarm...
🤖 Hire AI Agents for Business:
agents.vrsen.ai/
🐙 Agency Swarm:
github.com/VRSEN/agency-swarm
🔋 OAI-Widget (My SaaS)👇
www.oai-widget.com/
🚀 Custom AI Projects:
www.vrsen.ai/
🤝 Consulting/coaching:
calendly.com/vrsen/ai-project...
🐦 Twitter:
/ __vrsen__
💬 Discord:
/ discord
Intro - 00:00
Devin Demo Overview - 2:34
Installing Docker - 05:16
Importing Agents - 9:06
Adjusting Agents - 13:44
Benchmarking LLaMA - 19:20
Fine Tuning Your Agency - 23:45
About: Discover Devid, the groundbreaking AI Software Enginee that runs with minimal supervision. Devid surpasses traditional coding assistance by autonomously tackling software engineering tasks from planning to execution, utilizing tools like command lines, code editors, and web browsers for API documentation. Witness Devin's unique capability to plan, code, and even self-learn API integration, challenging the limits of AI-assisted coding interfaces. However, our critique highlights potential improvements in autonomy and efficiency, offering insights into creating a truly independent AI software engineer. Join us as we explore the future of AI in software development. - Наука та технологія
VRSEN is the most exciting project in this space
Watch me automate myself out of the job
@@vrsen God
Agreed. I've looked at a number of different Agent packages. VRSEN's work is the only project I've seen which contains code that I felt like I had a real chance of actually understanding and using, and doesn't look exclusively like a pitch for investment money.
@@petrus4 bingo!! That time stamp callout was classic 🤣
@@vrsen epic answer :)))
I think other UA-camrs are bored and post all these other codes, thank you for actually improving and building for production and not all these snake games everyone loves 😂
Thanks 😎
UA-cam is only good because of people like you, doing what you do and sharing it. Thanks🙏
The quality and quantity of your kick ass videos are a demonstration that Agency Swarm works… how on earth would you have time for all that otherwise? Thanks for these high quality videos! ❤👏
dude you are the original, so happy to have such experienced expert as part of the community, lately seems a lot of bullshit is happening with those AI models, so happy to have you again for opening our eyes :)
This is very impressive. Original ideas, excellent execution.
Cant wait for the addition of being able to use a local llm. Open ai gets enough money lol.
I would really love to see a pre-defined vendor and model selector like selecting to use Groq or ollama and the respective models of them
Thanks so much for that amazing walkthrough, can't wait to try Devid!
Very cool! Looking forward to rewatching, following the steps, and trying it out for myself. Great!!!
This is a no BS channel. Great content every single time 🙏🏿🙏🏿🙏🏿
People are sleeping on this channel and from what I can tell you are the real deal. Devid is awesome!
Thinking between gpt pilot and this which to adopt. Love that open source is taking over and there is so much choice now. Will test your solution over the weekend!
I can't tell you how much I appreciate this, thank you very much!
Brilliant! Sysreq? And can it be taught a different language? Fit example, if I give Devid a url to the instruction page(s) for hyperapp js (a strange one, I know) will it learn the language's methods?
You're not getting enough respect my man. Love watching your videos
Great abstractions and even better explanations!
Have you evaluated using some of the most recent open source models to power your agentic framework? Mixtral 8x7B or Mixtral 8x22B, DBRX or similar models? Thanks for sharing your project with the world!
Implement RAG functionality as an independent module then add it as a Tool. In that way, microservices architecture can be implemented.
Arsen can you give us the AI enthusiasts the luxury of having you in an interview with David Ondrejf or his podcast?
Sure if he’s down
Great content was wondering if we can replace open ai key with other providers or giving a endpoint from ollama. Is this possible?
This is so sick. Thanks!
Such a cool project! Love it ❤️❤️❤️
thanks man for you fantastic work, from now i will follow you ...thanks again .sharing is caring
Brazil here. Fantástico. Gratidão eterna por essa ferramenta maravilhosa. Desejo todo sucesso do mundo meu amigo.
Conseguiu fazer rodar? Abs
Deu boa pra rodar man?
One of the best channel for ai learning, so much value in your videos thanks for the quality content.
Question : why arent you using Crew Ai as your agents framework ? Their tools are quite good and oss ?
He did address the limitations of CrewAI, as well as AutoGen, in the video about his framework. It was only a couple of videos ago if you want to look for it.
@@brianWreaves Thank Brian, i'll definnitely look fort it
Thanks for your contribution to the knowledge of this earth!
What code editor are you using?
Slavic roots say hello ! I shake your hand firmly ! your research is great ! we are waiting for new videos !
I see the file thing but does this mean chat with docs like csv files or just files it provides you with?
Can never find one that can actually use the spreadsheet you give them and fix the errors by trying the sample document you gave them? It always thinks it complete and you have to repeat the errors making it no better then gpt
Since vison is available on the assistants API now when do you think you'll have an update? There are some interesting approaches here. I only wish the gradio UI wasn't required to interact with the agent. solid work!
I love it. This is why I subscribe! You pointed out that Devin lacks the "fully" agentic ability -- with their continued prompts and lengthy time to build, per time stamp! Nice! But what was the issue with OpenAI API for retrieval? The laziness sucks. I get it. What about using Claude instead?
As always, amazing, thanks.
additionally having the RAG capability on curated documents and examples would have been wonderful!
Very cool. Thanks for posting!
Any plans to add multiple LLM support? Including local LLM and Claude/Gemini specifically.
Btw, looks like the latest model and api update from OpenAI allows vision, I haven’t tested yet. Cheers!
It seems heavily dependent on the Assistants API and its tools for file retrieval. But if I had to guess, his abstractions here probably leave room for future implementation of RAG without OpenAI dependencies.
@@avi7278 True, we could use LLAMAIndex or Langchain for RAG and then use it with them with Devid as Tools. Using Python's automated Prefect library would be magically
@@generationgap416 i got my system to read files interpret whats there relevant to a prompt and then continue writing in a completely new session till finished! This is fine as is as long as it works similar to his system on an open llm its done
any plans to support local llms using something like lm-studio with their api? they now support embeddings endpoint so can handle retrieval and everything all locally
@vrsen Please let us know your plans for local llm
I wonder what the biggest project is that has been written entirely by agents. So far I've only seen small projects like a snake clone or game of life like in this video. But I'm really curious how these agents scale to bigger projects, stuff that you would normally build with a team over the course of a year or more. Anyone know of someone who's working on this?
Hey, I was able to get it to work, my question is what is the difference between Agency-Swarm-Lab, and your other Agency-Swarm ? Also, where can I learn more to build my own ai agents swarm?
yeah buddy! addicted to slanina and ai agents. subscribed
Could it handle dataset of node.js and react.js application?
i wish i understood what you was doing cause i installed docker but idk how to get the agent swarm lab to run on it and the program you using don't look like either github or docker so now im really confused.
The thumbnail cover image is gangster tho 😎🤣 I see what you did there 😂
At the end some idea came up. You update the manifesto in the end to adjust. But if you add an extra agent, which will update the manifesto so the next time it will auto adjust. Must be possible right?
Seems pretty cool, if an error occurs because of not detailed enough prompting it will add the fix to to prompt so it teaches itself.
So everytime you run it, it will improve itself. Feels so cool 😊
what about genesis i put agency-swarm genesis it say may be want open ai key or something because it give me error
As Andrew Ng said LDB+Reflexion brought gpt3.5 from 48% to 95%, what kind are you implementing here?
I hope you'll continue developing this even for local LLM, llama3 will be available at the end of May, and in conjunction with this, it would be like almost a real engineer.
Nice I will use this to perfect my coding skills.
Hey man, pretty cool channel. Just subscribed!
Has anyone used it with frameworks like Refine JS? It would be great to know that it can stick to a certain framework and have content awareness of the entire protect built with it.
What have you made with this agents stuff
You got one subscriber sir!
Can you use Wizard LLM instead?
Hey VRSEN, massive fan of your content. Really appreciationg what you do.
When trying to replicate the project on my local machine i got the problem, that the agency.py file can't find the agency_swarm file to import Agency from it. Since i've seen the open issues with no comments and no status on them i thought i'll ask here to find out how i could fix it. I tried pip install agency-swarm from my agency folder but to no avail.
Hope you keep up the good work and again thank you for the golden content and projects you provide us with.
Awesome again. Now if we can only get this to work without relying on open ai nonsense it'd be great. Would love to see the ability to use local llms
which editor do you use?
IntelliJ Idea?
Where is the Set-up tutorial so I can get this on my Local Host?
did you solve that? I have trouble to run this on my local host.
OUTSTANDING(!)
This better work.
I just watched that dude (Internet of Bugs) debunk that Devin program.🙄😆
Scott Wu needs to get jumped in the parking LOL.
He up here acting like he's a genius
How does it perform in SWE-Bench?
@vrsen did you consider adding a web GUI or any GUI to Agency Swarm Lab it could sky rocket the project
Phenomenal
congrats on continued work to give control to devs on implementing these projects that dont rely on langchain/lama - both of which which are a pain in production tbh
you my friend will be a co author, thank you so much
Indian engineer two weeks from now: “Hey everyone, today we are releasing Devender”
Korean engineer next month: “Annyeong everyone, today we have released Daeyoon”
Can we get a groq mistral/mixtral api to run this?
Datastax recently released open source assistants api implementation, I will be testing it next week
@@vrsen Yes please, I love agency SWARM, but I can't base an architecture on Openai's Assistants'Agents, Openai specially is lazy at coding, but nobody wants to do that for more reasons, their quality is lowering, their prices are high and nobody knows what they do with the data, did I mentioned it's hyper-woke?.
Got it
Installed Devika and it sucks. Is this better?
For sure. You can customize it in all ways
While it might not be necessary to view the browser window It could be very useful when trouble shooting.
Does it work in Linux?
Add support google germini pro 1.5 for increased context of the whole project files
nice job
AWESOME
Keep it going, a lot of improvements can be done. Manual steps are not an option, as you understand.
can we integrate groq? instead of openai
Did you try?
@@generationgap416 yep and it’s working, it’s similar to openai endpoint
Claude streaming for Sonnet & Opus seems to be something that can be used, I found it is mire intelligent than OpenAI
Calling out git, shots fired
if someone is able to make this thing works with llama locally without relying on any API thats going to be a game changer .
VRSEN is goat ❤️. If you happen to launch a company please I’m not rich but I want to buy some shares 😂
Silicon Valley spends millions to create Devin, Vren creates something 10x better with his bare hands. I'm here for this
Devin, Devika, now Devid 😢 I think it's time we find this folks wives. Watch out for Devisha
cold sniper killer.
nice. just sub. will be trying it out asap
Designing products like this, stop relying on OpenAI API, start using open-source models. I appreciate you open sourcing it though so it can be easily updated to use open source models.
No, Claude 3 API instead, you should use the best model available.
@@TheRealUsername Correction, use the best open source model. Stop relying on these big corporations. They harvest your data. I would never use OpenAi/Anthropic/Groq api in a production environment as your putting that companies data at risk.
Is it like
Is it free?Is it open source?Is it linked to specific countries, such as? ديفينوب
Красава!
Devid dude I love it. Totally agree, way to stick it to them. Devin is such a d ..... Bag.
Bro came with receipts
OHHHHHHHHH yes
Why do you share such gold all the time :O
Same question here haha
I wouldn't want to do anything else in the world instead. I appreciate you guys being on this journey with me.
Dibs on Devine.
I got dibs on Devine.
Hi,
Please create a vscode plugin.
And, it would be great if it can scan / Index folders and files and optimize databases.
👍
Thanks.
do it again in LLAMA 3
Devin is cooked here.
This is just a crewai fork lol 😂
This is a spoon
follow the white rabbit
Cheekii bureekie
I must be dumb -- what would you actually use this for?
Your accent 😍
Thanks for showing what you are doing. No need to bash Devin though. It causes a weird vibe and does not generate any value for your viewers, at least for me.
Hi VRSEN ❤ I'm also an ai developer, let me know if you like to collab! All the best!
Наш слоняра!
You need to make this much easier to install, you are making a code assistant for many people who don't understand code. the process is confusing and tedious
Dam, ok this is cool and I can't Even code. Well now I can pretend I can lol .
4:45 You mention how long devin takes. Devin doesn't take 2 hours to complete the task. It only took about 30 minutes to set up the training and complete the initial 30 epochs. Then the user checks back later and the training is on epoch 404.
I love that you're doing all this open source for the community, but if your going criticize others work please try to be accurate.