Such a complex topic is difficult to know where to start with. Perhaps a fundamental question is a good place to start: Do you have a decent PC, and how much are you willing to learn? There are basically two approaches here.
- Your own computer: Learning to use it properly takes a significant amount of time, as well as the right hardware and software. Thereās no easy way, but youāll learn a lot.
- You can rent the service or server for a few weeks at a time, at a cost of around $5 to $10. Although billing is based on usage, this doesnāt concern most people. This option is ideal if you just want to try it out, but you wonāt learn much that way.
Now, Iām going to talk about using your own PC. What Iām saying here isnāt necessarily the absolute truth or the best way to do things. Rather, Iām trying to make my advice as general and universal as possible. What do you need mostly:
- A decent graphics card is useful because it is faster at AI calculations than the CPU. This applies to all types of AI, whether text-based, image-based, or otherwise.
- A GitHub account and the GitHub app to download āprogramsā (repositories).
- The Python programming language and its software (PyTorch). While you donāt need to know the language itself, your operating system must be able to run it. The GitHub repository is the best place to check which version is required. Rest assured that you will be notified if a required version is missing, at the latest when you open the file.
Of course, all of this can vary depending on your requirements. Some software runs better on AMD processors, some on NVIDIA processors, and some on both. Donāt worry about it. It sounds more complicated than it is. You might ask why we need Python and GitHub? Not every programmer wants to write a graphical user interface, let alone one for every operating system. This way, however, it runs on almost every operating system. The developer can focus on the AI functions rather than which button to press. Thatās why weāre taking this detour ā to ensure accessibility regardless of the operating system.
Now the path forks again:
- Do you really want to train your own LLM? It takes a lot of time - weāre talking at least a day or weeks. Itās difficult to be more specific than that since it depends on factors such as the hardware and the amount of data.
- Or do you create a LoRA? What exactly is that? You can think of it as an add-on for a game, like DLC. You can use someone elseās free LLM and create a LoRA based on Bambu tutorials, for example. This way, you can teach the AI new things without occupying your computer for days
Personally, I recommend creating a LoRA at least the first time around to get familiar with the topic, test whether it works, and avoid finding out days later that it doesnāt work at all. A Lora can be created in just a few hours.
Most people would probably have fallen asleep by now if they had read all this text. It sounds complicated - thereās just so much to read! Itās not actually complicated, but there are a few steps that most PC users donāt usually have to deal with.
Please also bear the following in mind: If youāre developing your own LLM model, it wonāt be as large or complex as those created by companies. Technically, it should be referred to as a āsmall modelā. Unless youāre willing to let your computer run for weeks - which some people actually do - you should plan ahead.
So, if you were expecting a one-click solution, Iām afraid Iām going to have to disappoint you ā you might as well stop reading now and opt for the AI training rental service.. However, if youāre willing to use the above programs, nothing stands in your way.
I need to split this comment up, though. Iāll continue in the next one.
. I think we will hit the limit to what AI can do (or what we are willing to let it do very soon).