Links: ChatGPT Server Commander GitHub Repo: github.com/wonderwhy-er/ChatGPTServerCommander WebSim IDE Plugin: ?plugin=@xech/ide New WebSim adventure: websim.ai/c/t8etqIvkNTppQQ2Uz Chapters: 00:00 - Intro 00:47 - Jump-in joke test with RunwayML 01:00 - Last video recap: WebSim limitations 02:44 - Plan for today 03:07 - Review of WebSim Text Adventure app and what we will be changing with ChatGPT 05:48 - Download WebSim code, using Voice In with ChatGPT Server Commander 06:07 - Creating folders, moving files, renaming file, and opening Finder window in the new folder 08:12 - Asking ChatGPT to explain what the app is and how it works 10:48 - Asking ChatGPT to extract, run, and test the initial story generation feature in isolation 13:36 - Brainstorming with ChatGPT on how to improve story generation 19:43 - Asking ChatGPT to edit and run the code 21:19 - ChatGPT makes a syntax error, its detected and reported, ChatGPT iterating on it by itself 24:09 - Testing and iterating on new, sometimes hilarious, story setup generation: almost 40k unique combinations 35:10 - Integrating and reviewing the new story generator back into WebSim 37:51 - Uploading the new version back into WebSim 39:50 - Story test: You are a genetically enhanced soldier in a mystical forest, where you must retrieve forbidden knowledge. However, an evil sorcerer who controls time stands in your way, and everyone is connected telepathically. 41:33 - Summary of what you can do with Server Commander 42:50 - Setting up Server Commander on your machine: Node.js, checkout code, and install dependencies 45:16 - Running for the first time: choosing port, local or server, and domain name 47:33 - Setting up Custom GPT 50:50 - Publishing and testing the new Custom GPT 53:46 - Thank you! Like and subscribe for the follow-up video, ask questions in the comments, further plans
excellent work ! was looking for a while for something like open interpreter to run through custom actions in a custom GPT. This is the first project that comes close to that, even better than what imagined in some ways ! Thank you for sharing this!
this also shows how much potential was there for custom GPTs, but OpenAI developers failed to properly realize it - they did not add the possibility to have personalized (per user) settings/ configuration.
@@llucis-v yeah... I think openai failed too. Custom gpt store did not become an app store. There could be also reasons where one would prefer api and control over their app. My feeling is that openai is not interested in pushing this further.
@@EduardsRuzga day job is bioinformatics/genomics research, software engineer formation and avid coder; like you I want to use LLMs for exactly this kind of diy stuff you showed here, for work automation and general productivity boosting etc. Been wanting to use custom GPTs for exactly stuff like this, give them access to my local environment through custom actions etc. Did not find time to implement it, you DID it!
I do not have access to Devin. I do test things like Aider/Cursor/GPTEngineer/Websim/Claude/Create.xyz, getlazy.ai, Replit Agent As far as code editing goes. To me, it feels comparable or slightly better to state in which Replit and getLazy are. GPTEngineer/Aider/Cursor are better but you need to pay separately from ChatGPT for them and there are some ways in which ChatGPT Server Commander is better too. On the other hand, this is a hobby project, it's free, it's open source, and I am working on improving it, other can fork and improve it too. I do like some parts of it better than Aider or Cursor. GPTEngineer is pretty cool but its frontend only for now + cost/utility is hmm bit lacking. But it probably has a good future ahead of it. I am thinking of the next video to compare Replit/GPTEngineer/GetLazy and this one head-to-head.
@@EduardsRuzga That video sounds like a great idea! We need realistic, honest and especially practical USABILITY reviews of such projects. There is so much hype out there and youtubers churning out AI videos just for quantity/views, not quality and practical usability of the content.. E.g. videos reviewing/testing latest local model (through ollama) for coding (e.g. through ContinueDev) despite the fact that most of these models are just not capable of handling even basic coding tasks - so what's the point in making videos showing how bad they are (or sometimes just how to install them, despite being practically useless..).
@@EduardsRuzga I would also be interested to see a video of how YOU practically use aider, and maybe even compare it to claude-dev? (enough Cursor videos out there, not buying into that). They both use prompt-caching so the cost shouldn't be that high.. I actually sniffed around aider's internals as I was also considering making a GPT with custom actions out of it as well.. So far I did some dry runs to see the prompts, and exposed the repo map function (the one based on tree-sitter), I think it could be workable and even improved with a custom/personal GPT pointing back to a dynamic codebase server doing even better code indexing..
but of course I might be wrong (not enough time to test them..), if you found any solid practical & reliable uses for small local LLMs, please do make a video about any such findings - they are sorely needed in the youtube space imo
@@llucis-v thanks for super valuable feedback! I am less focusing on tools like Aider and ClaudeDev as they are local, I am more excited and want to carve a niche in the cloud tools like GPTEngineer and WebSim Aider is cool, I did check the first versions of ClaudeDev and did not like it, but now heard they improved considerably. I think Aider founder should take it in the direction of becoming a plugin that anyone can add to their coding projects. Imagine it being an npm or pipy package anyone can add to their project. Open source building block for Cursors. Any startup can pick it up and add capability to their product. Let's connect somewhere, do you use Discord or any other messaging platform? Telegram, Linkedin etc?
Links:
ChatGPT Server Commander GitHub Repo: github.com/wonderwhy-er/ChatGPTServerCommander
WebSim IDE Plugin: ?plugin=@xech/ide
New WebSim adventure: websim.ai/c/t8etqIvkNTppQQ2Uz
Chapters:
00:00 - Intro
00:47 - Jump-in joke test with RunwayML
01:00 - Last video recap: WebSim limitations
02:44 - Plan for today
03:07 - Review of WebSim Text Adventure app and what we will be changing with ChatGPT
05:48 - Download WebSim code, using Voice In with ChatGPT Server Commander
06:07 - Creating folders, moving files, renaming file, and opening Finder window in the new folder
08:12 - Asking ChatGPT to explain what the app is and how it works
10:48 - Asking ChatGPT to extract, run, and test the initial story generation feature in isolation
13:36 - Brainstorming with ChatGPT on how to improve story generation
19:43 - Asking ChatGPT to edit and run the code
21:19 - ChatGPT makes a syntax error, its detected and reported, ChatGPT iterating on it by itself
24:09 - Testing and iterating on new, sometimes hilarious, story setup generation: almost 40k unique combinations
35:10 - Integrating and reviewing the new story generator back into WebSim
37:51 - Uploading the new version back into WebSim
39:50 - Story test: You are a genetically enhanced soldier in a mystical forest, where you must retrieve forbidden knowledge. However, an evil sorcerer who controls time stands in your way, and everyone is connected telepathically.
41:33 - Summary of what you can do with Server Commander
42:50 - Setting up Server Commander on your machine: Node.js, checkout code, and install dependencies
45:16 - Running for the first time: choosing port, local or server, and domain name
47:33 - Setting up Custom GPT
50:50 - Publishing and testing the new Custom GPT
53:46 - Thank you! Like and subscribe for the follow-up video, ask questions in the comments, further plans
excellent work ! was looking for a while for something like open interpreter to run through custom actions in a custom GPT.
This is the first project that comes close to that, even better than what imagined in some ways ! Thank you for sharing this!
And open source. Feel free to fork :)
@@EduardsRuzga did it, thanks! Do you have online forums that you visit/post? Found you on medium, good stuff!
this also shows how much potential was there for custom GPTs, but OpenAI developers failed to properly realize it - they did not add the possibility to have personalized (per user) settings/ configuration.
@@llucis-v yeah... I think openai failed too. Custom gpt store did not become an app store.
There could be also reasons where one would prefer api and control over their app.
My feeling is that openai is not interested in pushing this further.
@@llucis-v what do you do your self?
@@EduardsRuzga day job is bioinformatics/genomics research, software engineer formation and avid coder; like you I want to use LLMs for exactly this kind of diy stuff you showed here, for work automation and general productivity boosting etc.
Been wanting to use custom GPTs for exactly stuff like this, give them access to my local environment through custom actions etc. Did not find time to implement it, you DID it!
good content : )
This is cool - how does it compare to Devin and other s/w dev tools?
I do not have access to Devin.
I do test things like Aider/Cursor/GPTEngineer/Websim/Claude/Create.xyz, getlazy.ai, Replit Agent
As far as code editing goes.
To me, it feels comparable or slightly better to state in which Replit and getLazy are.
GPTEngineer/Aider/Cursor are better but you need to pay separately from ChatGPT for them and there are some ways in which ChatGPT Server Commander is better too.
On the other hand, this is a hobby project, it's free, it's open source, and I am working on improving it, other can fork and improve it too.
I do like some parts of it better than Aider or Cursor.
GPTEngineer is pretty cool but its frontend only for now + cost/utility is hmm bit lacking. But it probably has a good future ahead of it.
I am thinking of the next video to compare Replit/GPTEngineer/GetLazy and this one head-to-head.
@@EduardsRuzga That video sounds like a great idea! We need realistic, honest and especially practical USABILITY reviews of such projects.
There is so much hype out there and youtubers churning out AI videos just for quantity/views, not quality and practical usability of the content.. E.g. videos reviewing/testing latest local model (through ollama) for coding (e.g. through ContinueDev) despite the fact that most of these models are just not capable of handling even basic coding tasks - so what's the point in making videos showing how bad they are (or sometimes just how to install them, despite being practically useless..).
@@EduardsRuzga I would also be interested to see a video of how YOU practically use aider, and maybe even compare it to claude-dev? (enough Cursor videos out there, not buying into that). They both use prompt-caching so the cost shouldn't be that high.. I actually sniffed around aider's internals as I was also considering making a GPT with custom actions out of it as well.. So far I did some dry runs to see the prompts, and exposed the repo map function (the one based on tree-sitter), I think it could be workable and even improved with a custom/personal GPT pointing back to a dynamic codebase server doing even better code indexing..
but of course I might be wrong (not enough time to test them..), if you found any solid practical & reliable uses for small local LLMs, please do make a video about any such findings - they are sorely needed in the youtube space imo
@@llucis-v thanks for super valuable feedback!
I am less focusing on tools like Aider and ClaudeDev as they are local, I am more excited and want to carve a niche in the cloud tools like GPTEngineer and WebSim
Aider is cool, I did check the first versions of ClaudeDev and did not like it, but now heard they improved considerably.
I think Aider founder should take it in the direction of becoming a plugin that anyone can add to their coding projects.
Imagine it being an npm or pipy package anyone can add to their project. Open source building block for Cursors.
Any startup can pick it up and add capability to their product.
Let's connect somewhere, do you use Discord or any other messaging platform? Telegram, Linkedin etc?
By
hm?