Making an AI in python
Posted by GrassWinter4767@reddit | learnprogramming | View on Reddit | 10 comments
So recently I have been seeing a bunch of videos of people who: “Trained AI to drive” or something and I think that is just the coolest thing in the world. BUT one problem. I have absolutely no idea how to do it. If there is a guide or tutorial or course you could recommend or just general advice that would be great. Thanks in advance!
SaltAssault@reddit
When people say "AI" nowadays, they mostly mean Large Language Models (LLMs). Creating one from scratch is a massive undertaking and something for large, well-funded teams. What you see online is mostly people using AIs already developed by these teams, which they do by integrating them into their own software. You do this via their respective App Programming Interfaces (APIs). An API is like an official window through which your software can communicate with someone else's software. You usually have to register with the company for this to get a unique key that you use every time you use their API, because that makes sure the window is closed for people who shouldn't have access.
I would reccommend looking at using OpenAI in your projects, it's a pretty decent LLM with not too outlandish prices and with a lot of documentation and tutorials online. It's important to know that you usually (always?) pay for LLMs per query. I.e., every single time you ask their AI something via an API, you're charged a bit of money. This can easily get out of hand if you're not mindful.
It's an advanced topic for a beginner, but it's good to encourage one's curiosity. Just don't get too discouraged if it turns out to be a bit too tricky for now.
Soggy_Struggle_963@reddit
There are also plenty of models you can run locally for free if you have the hardware for it.
SuspiciousDepth5924@reddit
Pretty simplified you sort of have 3 "parts" to an AI application.
* A model, either something "off-the-shelf" or something you've trained* (training your own non-trivial model takes a lot of time and money)
* Some low level stuff that takes the model and runs stuff on your GPU
* Some "glue code"/"user interface" to tell the low level stuff what model to use and what to do with it.
Python is generally only used for the glue/user interface bit as python code is way to slow to do the low level AI stuff in any reasonable time.
You can get a _very_ simplified view of what the AI actually does here: https://en.wikipedia.org/wiki/Neural_network_(machine_learning)
In short its a shitload of
if(input*weight > threshold) { send_to_next_layer(1)} else { send_to_next_layer(0) }
With the training fine tuning the weights and thresholds for every "neuron" to better match the training data.
Wingedchestnut@reddit
Pretty sure it's not ML and DL but Reinforcement Learning
SuspiciousDepth5924@reddit
I'm not a 100% up to speed on the current state of the art, but IIRC yes, they use Reinforcement Learning which is based on Transformers which is based on Neural Networks, which is based on a bunch of vector/matrix math etc.
What you want to call it mostly depends on how far you want to peel the onion ("Magic AI wizard box" <-> "GPU shader math").
Wingedchestnut@reddit
It's called Reinforcement Learning, I forgot the name but there are some basic python 'games' you can try like balancing a pole which are known exercises. You can try this but Imo this is already a more advanced field of data science that's mainly done for academic research, it isn't applied that much in real life, thinking out loud would be training robot, gamedevelopment..
sleggerthorn1909@reddit
So at first you should learn what kind of algorithms there are and whats the difference. Reinforcement Learning, Unsupervised and Supervised. From there on go on and learn the different Algorithms that are standard. Then take one and build your first project.
bravopapa99@reddit
Be wary, a lot of those videos are edited bullshit; the result of months of work to get it working, it creates a false impression that the subject matter is easy, like a lot of tutorial videos.
This is a good place to go: I did it myself a long time ago, never got to complete it because... life etc.
https://developers.google.com/machine-learning/crash-course
Soggy_Struggle_963@reddit
Do you have any background in software development? I feel jumping straight into machine learning is a bit of a steep learning curve.
GrassWinter4767@reddit (OP)
I have made a couple simple python projects before but no nothing like this. This is kinda my end goal and I need a route to achieve it.