Question:
How do I make generative AI?
Basically something like GPTs.
And I mean AI, not chatbots.
Whatās the difference?
Chatbots need to manually learn stuff but AI can learn itself.
So does creating ChatGPT or DALL-E count?
I want to make AI, not make a chat thingy using an AI model.
I want to create AI. And btw I donāt have any cycles.
Iām asking how to make AI
Ok, elaborate. This sounds interesting.
@SnakeyKing You can try using TensorFlow (use with Python) to do machine learning and AI creation.
Can you explain how to use it?
And btw I prefer to use the js
version
nvm I think I found the tutorials
Actually IDK but freecodecamp might have something.
Ok here:
TensorFlow
good luck, tensorflow is hard lol
TensorFlow is one way to do it with JS (and I think the most popular one).
Brain.js is another way.
Creating an AI is an extremaly and difficult task, it takes a lot of time and can be frustrating (as one who had a hell of a time trying to create one)
My AI had the solely task to play a simple game to achiev the highest score and oh boy, never been so frustrated trying to create the thing. Read the docs, learn from it, test it and try again. Itās an arduous journey!
Remember that Machine Learning learn from data. So you have to prepare your data before anything else.
The basic steps is to create a model, compile it and train it.
The answer is that you spend an astronomical amount of money. Iām not being flippant: generative AI is prohibitively expensive to produce, thatās why the entire product-ecosystem is tightly controlled by a few extremely rich monopolies.
That being said, you can train small, specialized models on consumer hardware, sometimes with fairly good results. But part of that depends on what your definition of āconsumerā is, and what youāre trying to do. For small/medium sized models you need somewhere around 40GB of VRAM to train them, so for example an NVIDIA Tesla A100 with 40GB of VRAM is selling on amazon (used) right now for about $7,000. Large models require distributed computing at a vast scale (think hundreds of TB of VRAM).
The problem is, at least with the current state of technology, the entire model has to be loaded into volatile memory in order to perform the types of mathematical operations on the data that produce meaningful inference (this is especially true with language models, as language is like this giant sea with only certain very particular pathways from any one single point to another and very little opportunity to compress the complexity of the space (this is true when generating language, however once itās generated it becomes extremely compressible, which is a great illustration of entropy at work, but I digress). Keep in mind also that this is far from general programming, it requires knowledge of esoteric, highly quantitative approaches to computing and without a strong background in advanced mathematics the learning curve can be punishingly steep.
Again, all of this kind of varies by what exactly youāre trying to do. But really the short answer is you just donāt, you fork over some cash to a rich corporation and get an API key. This is of course a matter of opinion, but the current AI fad is all sort of stupid and exploitative, and its capabilities are only kind of impressive, I think generally speaking people are just getting taken for a ride, itās just another opportunity for a few small groups of very wealthy and powerful people to pick regular folks up by our ankles and shake the change out of our pockets, that being said Iām not bashing the underlying technology which can be a useful tool but is extremely expensive to produce or even useābut basically everything Iāve said, its relevance is dependent on knowing what weāre even talking about when we say āAIā, which is far from clear, I mean the topic is nuanced to say the least, even between two iterations of the same model the distinctions can be exceptionally subtle and refined, often to the point of being evanescent.
How? Study, and study. Then invest lot time learning how to train and evaluate, then money to be able to train big enough models.
Please avoid JS with AI. Learn with python, build with C, infer with whatever
I didnāt use JS to do AI. I prefer to use Python or C++. Iāve been looking at R as a complementary language too.
But since OP said he prefers to use JS I just list the libraries available.
Itās a machine learning tool (less of Generative AI or LLM and more of an input/output type thing I believe).
You may consider looking into LangchainJS: js.langchain.com