I don't know. I know that AI is a "Large Language Model" (LLM), But it's already outdated, as time keeps going on, and then you have to update the information or data sets to be newer versions.
I feel like they really need to stop training AI as a Large Language Model, where it is trained on all data, and instead break it down into resources and components. Imagine like a Cognitive ability part, then have separate parts that are tacked on in regards to specific subject matter.
So for instance if a user does a search, the basic "cognitive part" thinks to itself "here are the search terms" How do I do this function in Rust Language?", i'll search for an answer using the resources "programming language", "functions", "Rust Language" etc. This way they should be able to cheapen the training and have more up to date information.
Because I think as it is now, they gather as much up to date information that they can, train the model on it, test it to make sure it's working correctly, send it off to be used by humans after 6 months or so of testing.
I'm a bit bewildered with ChatGPT. I have used ChatGPT 4.0 in functions using Rust language for the Bevy game engine. Bevy is currently on version 11.2. ChatGPT had training data up to version 0.9 i think. Then it went down to 0.5. It knows about the 11.2. or 0.10. But it can't give you functions that would work. It's all out-dated. You have to change things around. It would be best for a LLM to be up to date. I feel that it would be best to compartmentalize the data for certain things into smaller components or resources. That way you might be able to train a small model on that data, and maybe somehow find a way to integrate that model with another model as some kind of off-shoot.
It's kind of like how Visual Studio Code is really a glorified text editor, but you add plugins and extensions to give it more functionality as a programmers tool for the specific uses you are using it for.
I feel this somehow needs to happen with AI.
It reminds me of the Terminator franchise and story. In the future John Connor re-programs the terminators. At the start he is having his tech guys look into it, go through the code bit by bit. There's a file for opening a hand, another for closing it, more files for moving specific fingers. All those files are fine. No need to re-write the complete code, just change the orders.
Likewise, the majority of the Rust language is fine. A few tiny things changed from version 1.72 to 1.73. No need to learn the whole language again, just learn the changes, train on the changes. There shouldn't be a need or want to train a model on the complete information known to human kind. That's a ridiculous amount of training. Break things down, and somehow make the updates be the things re-trained on, or learnt.
