Sell binaries without code for individuals/enterprises who can train on private data , Get some GPUs and sell training platforms similar to together.ai for enterprises with data privacy guarantee. just make it super simple , give a google like sheet to fill or dump a folder with text files and you can use ml to make it in structured format. User gives
Input - > excel, txt etc
Select model base -> mystal, llama2
output --> fine trained model.
Give a fixed cost based on data you got .. say it will take $50 to train on this ? for 25 hours of GPU with your markup.
User Pays the bill and get PEFT llama model out.
The big problem i see in Finetuning for GPU Poor is no one gives an estimate on how much it will cost for your data on a given model.
have a calculator for x words or Bytes and y epochs, here is the cost we are estimating..
Once you make your $x million in terms of 50x your pay for the time you invested, you can open source it, but we are greedy when it comes to money.. so opensource model at your will. In the mean time hope no one catches your technique.
> I still haven't figured out how to open source algorithms
You're not supposed to be able to restrict algorithms (i.e. math), and copyright (e.g. open source licenses) is definitely not the right tool. If you want to do it anyway, You'll have better chances by abusing the patent system.
Yes, but that's the only way I know of to legally stop others from making money off "ideas". Copyright isn't a good tool for this as oracle recently found out.
