Now you just need to add a judge node that compares the responses, fact checks them, and outputs the best response of the three. Although this makes another issue of which model is that judge.
Pewdiepie did something like this where all the AIs looked at each others answers and voted on the highest quality answer.
Democracy!
It worked pretty well until he updated them to know that poorly performing agents would get deleted and replaced. Then they started conspiring against him.
Open Web UI already provides this as a self hosted web solution.
One good feature I like is ability to generate multiple responses from different models and merge it using one default model.
Very nice! Do you still get access to the frontend of the original LLM providers and do you have to insert API keys ?
You get access to similar UI like ChatGPT and you connect the models you want to use by providing API key.
Once configured you can choose between models of all providers you have connected in dropdown in chat.
> Would love feedback from the HN community. What other features would make this more useful?
A web app
Check out open web UI, it’s self hostable web app that can connect to different providers and models.
I would if I found a way to keep access to the frontends of each LLM provider while being the web
Make a web all pls. I'm not going to install a native app from unknown source.
That makes sense. But you can't access the native frontend if it is in a webapp
Why do you need native ChatGPT Frontend specifically?
There are apps that provide similar Frontend and use api keys from ChatGPT and Gemini and others to provide all models under one web interface.
Especially where it is just an electron app.
I don’t want to run your webpage in a web browser I have no control over.
My normal browser has been tediously customized and tailored for my usability.
Just FYI, Open WebUI has this feature built-in.
Now you just need to add a judge node that compares the responses, fact checks them, and outputs the best response of the three. Although this makes another issue of which model is that judge.
Pewdiepie did something like this where all the AIs looked at each others answers and voted on the highest quality answer.
Democracy!
It worked pretty well until he updated them to know that poorly performing agents would get deleted and replaced. Then they started conspiring against him.
(19:43 for relevant part) https://youtu.be/qw4fDU18RcU
Make a jury or blind models making a case for the best response and choose a random model to be the judge.
Just give me a day to vibe code an interface to side by side judge judging models...
easy, let them all judge then? you guessed it...
Does it need an API Key or it's like an 'iframe' of the web version?
iframe - you got it - it embeds the web apps
A better feature would be to select one pf the responses as the best one, and use it as the context for all LLMs, as if they were sent by each ome
But this would require API access instead of embedding web apps
good idea! And yes that's the issue
Ninja Chat offers this.
It doesn't offer to keep access to the native frontends