I'm curious as to what unique value add is being done on top on the core chatgpt (or claude or whatever) website and chat interface.
Now they have very long context windows, ability to upload and parse through files, custom system prompts, can work with images, remember user history across chats...
So what's the unique value adds in other products that are using chatbots?
My 'hot take' currently is that the core product of chatgpt and similar bots is so good that there's very little value to 'wrapper' apps/products. I'm a heavy tech user and all of my use cases and problems I can solve with the core chatgpt right now.
The only real limitations and problems I can think of are:
- privacy, though there seem to be modes for not using/storing your data, depending on your level of trust of the companies, - reliance on internet, I suppose one day I'd need a local chatbot with gpt4 capabilities somewhere remote, - maybe file uploads (and RAG capabilities) are limited on the core gpt4/claude sites, though I'm not sure this will remain the case for any length of time.
So yeah I'm not sure what unique value 'wrapper' products are providing, but I'd love to see examples and be proven wrong and hopefully learn something new.
I think the first and most important part is the context. In ChatGPT I need to feed in the context every time. An AI agent sitting in my spreadsheet already has access to all my sheets and workbooks. Therefore it can help me with the task right then and there.
Similarly, an AI on a writing platform such as Overleaf can suggest better sentences because it already has the context of the entire essay.
The second important part is actions. ChatGPT can output a table, but an AI agent in my spreadsheet app can directly make changes to the tables. Similarly, the AI agent can summarize an online meeting, and email / slack the team members their respective tasks as discussed.
For one thing I like agents that live inside my IDE and can directly make patches that affect various files. If it is patching things and then I tell it "I still see a squggle on line 75" I feel like I am collaborating with it. If I am having to cut and paste code out of Copilot and then fix up the imports and then I have to tell it the code is wrong I feel stupid.
I use cline and cursor, not sure they can be though of as GPT wrappers
If it's not a wrapper for GPT it is a wrapper for something else that's similar. The one I enjoy the most now is Windsurf and it is such a blast to have it tell me it is having trouble with its tools and ask for help. ("You're running on Microsoft Windows, you write C: not /C: and use the backslash instead of the slash in paths")
I am building llmpad.com yeah you can think of it as a GPT wrapper I think differently. I started building it 18 months ago, actually because I wanted to customise the system prompt and keep my data. Then I wanted to use other AI's Anthropic, Groq etc. Then I wanted to generate images. Then I wanted to work with tools for search, or UI cards. Then I wanted to work with realtime audio api's. Then I wanted to create react components kinda like V0. this week I wanted to work on tasks, that run and chain scheduled prompts, its turning into a little agent. All the time using one UI that looks the same, is tweaked to my liking and has some nice power user features like Shift Arrow on input for history, switching models etc.
I am also sure tomorrow there will be another new thing.
I think of a little like Notepad++ for llm's instead of just Notepad.
I use it hourly, in the morning I have a summery of HN from overnight along with news, and reminders.
I host it on Digital Ocean, it uses about 15-20 API's from AI companies but also web search, document and kv stores etc.
Interested in trying it let me know.