Can I train an AI using data I input? (ie. can I use openai/etc. and provide data it must follow?) And keep building that model with more and more?

[contract]

I've been using several AI programs open ai, Gemini, etc. which are great, but here's what I'm struggling with. Every new chat clears history, or every chat has so much history that it crashes my browser.

1. Open a chat.
2. Input data, historical text, etc. etc. Would love to upload whole books, etc. 100s of texts.
3. Have all responses based on open Ai's standard database + PLUS my data input.
4. Be able to go back to this model that has learned all this, without having to re-upload over and over.
5. (side note: possible to get AI to stop refusing to say, I can't provide false information, I can't agree if I don't agree, etc. it's super annoying to fight AI when I am commanding it to agree. Gemini is the worst on this.)

Is there a way to build custom models/learning AI's? Maybe I am just overthinking; overlooking something.

 

[chubes]

Check out the concept of "fine-tuning" using the API. It allows you to add custom training data for the models.

https://platform.openai.com/docs/guides/supervised-fine-tuning

This might be what you are looking to do.

 

[secretagentdad]

You can run deep seek or gwen on a beefy ass baremetal server or even just a desktop but its slow.
The GPU cost basis is really unfavorable to small scale operations for weight training.
I'm messing with training shit on deliberately skewed sub sets of old common crawls and keywords sets sampled and indexed strategically.

Specialized indexing you can almost get competitive with providers like tier.net if you stack coupons and clearance listings.

 

[eliquid]

I've been using several AI programs open ai, Gemini, etc. which are great, but here's what I'm struggling with. Every new chat clears history, or every chat has so much history that it crashes my browser.

1. Open a chat.
2. Input data, historical text, etc. etc. Would love to upload whole books, etc. 100s of texts.
3. Have all responses based on open Ai's standard database + PLUS my data input.
4. Be able to go back to this model that has learned all this, without having to re-upload over and over.
5. (side note: possible to get AI to stop refusing to say, I can't provide false information, I can't agree if I don't agree, etc. it's super annoying to fight AI when I am commanding it to agree. Gemini is the worst on this.)

Is there a way to build custom models/learning AI's? Maybe I am just overthinking; overlooking something.

Yes and I do it all the time.

It's called Projects in ChatGPT AND in Claude
Or
Gems in Gemini

These features have been around a while now, start using them.

 

[zelch]

have you tried building custom gpts inside chatgpt? It wont solve all of your problems you listed but possibly will solve most of them