How do you make the AI tool
Hello everyone, I have seen lots of indie developers coming up with amazing AI tools for copyrighting, logo creation, avatar creation, etc. but after some research about making tools with AI and it seems that it requires thousands of dollars of computational / GPU power, but I am not really getting the point how these indie developers being able to make tools like copy.ai, jenni.ai, ai with logo, etc?
I am sure they didn't have to invest thousands of dollars in order to train their models. And thus I don't have money in order to create an AI tool, does that mean I should quit thinking about AI and all?
I was very excited about building something with AI for Images, but after asking someone about the possibilities of it, they said that making AI models require thousands of dollars and if you don't have that stop thinking about AI.
Does anyone have any advice related to this?
It's actually pretty simple. They're using AI APIs.
For writing apps, most of them use GPT-3, which is connected to an API by OpenAI, and for Stable Diffusion, they're also using them.
With an API, someone else handles all the cloud computing and storage, all they have to do is make the interface user friendly and they're good to go.
As for training, doing it from scratch is of course very expensive, but most of these companies are doing something called "fine tuning", which is taking an existing AI and then essentially doing "training lite" to put it in simple terms.
This is not expensive at all.
If you're curious, I'm developing an AI API myself. When we launch, we'll have Stable Diffusion and BLOOM (GPT-3 alternative that's less restrictive) on the cloud.
We're mostly trying to make open source AI more accessible, since they're too large to host on a computer. For how we're doing it, we're using AWS. Purchasing your own servers is indeed very expensive.
I have a discord for my API where we post updates: https://discord.com/invite/DsJXxcMmkC
Feel free to join and ask your questions. We also plan to add more AIs in the future, as well as different stable diffusion models and ability for users to train their own AI (fine tuning).
Thank you so much Richard for the answer, I have seen two AI tools which claim to make logo design through AI. According to my research Stable Diffusion, Dall E etc are there to create Art through AI, but I didn't really find anything which can create AI for logos or design. How are these indie developers being able to create these logo generator through AI without having any API available for the same? Did they created their own dataset on top of Stable Diffusion?
They're able to create logo generators and the like via what I mentioned before: fine tuning. It is exactly as you said, they create their own model on top of stable diffusion.
This process is called "Dreambooth", and requires a hefty GPU, so most people do it on the cloud (unless you've got $ to drop on your own server lol).
It's actually not prohibitively expensive, considering your dataset only needs 5+ images.
The text generator AIs like GPT-3 are also fine tunable. Once we launch, we plan to allow users to fine tune BLOOM and dreambooth stable diffusion.
For the record, there are many open source models that other people have trained, this one for example:
https://huggingface.co/nitrosocke/redshift-diffusion
So there might be one that's for logos as well, but with some searching, I am not able to find one. So it might be something you can train yourself.
Feel free to ask more questions through my twitter DM or here. My prelaunch website is going live soon so you can stay tuned for that!
https://twitter.com/TheRealEtch/status/1601405291723395073
Let's say I want to fine-tune a model that makes the logo on top of stable diffusion, but I don't have a very high end GPU, how much would it gonna cost to rent out the server on cloud? like the actual cost, if I am generating 5 logo on one go?
If you're setting up all the infrastructure yourself, then the price is probably a few dollars at most.
You can also do it for free if you look for a dreambooth google collab. Those are easier and less time intentive than to fine tune a model with than renting cloud space and setting the dreambooth architecture yourself.
Generating 5 logos in one go would be what you do with your model once it's created.
To clarify any misconceptions, you dreambooth to produce a model, and you get to keep that forever if you like, completely free to use.
So once you create a model, you've created a permanent logo generator.
The costs will come from running the generator on the cloud.
I'm planning on adding dreambooth to my API once we get Stable Diffusion on there.