Would the idea be to have many specialist models (forms, wikipedia, final cut, etc) and have a parent model choose which is best (potentially also a specialist model for selecting specialist models) with the idea an llm would give a larger goal and trigger this cascade of specialists to quickly do the task?
I think this is the logical next step in AI. There will be a few large "smart" AI models, but mostly smaller more hyper-specialized. In large part because they are faster to build....and faster/cheaper to run. As the model this post is about shows - you don't need to be able to write the works of shakespear to be able to run a computer.
The frontier models are already getting STUPID expensive. More than the average person certainly can reasonably afford for any use case. And even the business users are having a hard time justifying the prices for anything other than the most bleeding edge, "big picture" things (like creating a huge project plan).
I mean this is what MoE is the first step of. Slice your intelligence to make it more efficient without losing much capability.
CUA seems to disaggregate. Train specialist models and make delegation explicit. But I think the future for capable but efficient models involves doing it all in one package. I have no idea how that architecture would look like or how to train it.
Like a human would, I can walk and talk, without thinking hard about walking and still not stumbling. Specialized sub-circuits that interact. Not via a text protocol, but by directly interacting.
Did you see the Astra plays Portal 2 videos? You can see how separate viewing a scene, understanding the scene and acting in the scene still are. And in this model of interaction, a very capable model still makes stupid mistakes.
It is remarkable, that it works at all. But it is neither effective nor efficient. And in my opinion the real revolution comes not by capability, but by integration.
So to tie the loop: I think CUA makes an interesting first step towards specialization, but I am not sure if that is the logical conclusion of how AI should work.
It's a specialized classifier for actions on a form. So.. not totally. If you are building a bot system, which can fill form quickly without having to reason too much about the form, this probably can help.
Please give us a hyper general cookie consent popup dismissal add-on :)
Would the idea be to have many specialist models (forms, wikipedia, final cut, etc) and have a parent model choose which is best (potentially also a specialist model for selecting specialist models) with the idea an llm would give a larger goal and trigger this cascade of specialists to quickly do the task?
I think this is the logical next step in AI. There will be a few large "smart" AI models, but mostly smaller more hyper-specialized. In large part because they are faster to build....and faster/cheaper to run. As the model this post is about shows - you don't need to be able to write the works of shakespear to be able to run a computer.
The frontier models are already getting STUPID expensive. More than the average person certainly can reasonably afford for any use case. And even the business users are having a hard time justifying the prices for anything other than the most bleeding edge, "big picture" things (like creating a huge project plan).
Conscious thought seems to follow this too. No one thinks to breath, we often get in a groove and do deep thinking during a rote task, etc.
I mean this is what MoE is the first step of. Slice your intelligence to make it more efficient without losing much capability.
CUA seems to disaggregate. Train specialist models and make delegation explicit. But I think the future for capable but efficient models involves doing it all in one package. I have no idea how that architecture would look like or how to train it.
Like a human would, I can walk and talk, without thinking hard about walking and still not stumbling. Specialized sub-circuits that interact. Not via a text protocol, but by directly interacting.
Did you see the Astra plays Portal 2 videos? You can see how separate viewing a scene, understanding the scene and acting in the scene still are. And in this model of interaction, a very capable model still makes stupid mistakes.
It is remarkable, that it works at all. But it is neither effective nor efficient. And in my opinion the real revolution comes not by capability, but by integration.
So to tie the loop: I think CUA makes an interesting first step towards specialization, but I am not sure if that is the logical conclusion of how AI should work.
Interesting approach focusing on decision scoring rather than text generation.
I read the nod to Typesafe Jev. Are you implying use of RLCD or still RLHF?
Nobody outside TypeSafe even knows what RLCD is, so that’s impossible to answer.
Are we able to use CUA to build a system similar to Grok Bot?
It's a specialized classifier for actions on a form. So.. not totally. If you are building a bot system, which can fill form quickly without having to reason too much about the form, this probably can help.