I've played around with several open source libraries for various tasks. One thing that is critical: Have a use case in mind! I've found it really tough to wrap my head around some of this without a clear idea of a problem I could be solving. It doesn't have to be "real", but it should be concrete in your mind — otherwise it's hard to tell if the thing is working.

Generally I'd also say: Have an idea of how much data you need. Some tools require a lot of data, some actually work on relatively little.

I've now deployed production ML models and tapped into some of the OpenAI endpoints for certain uses — and I think I'm still just a "web dev", however, I do have some upgrades skills in terms of using and understanding these technologies.

Like — I didn't have to write any of this from scratch, I tapped into really useful dependencies I pulled into a project and learned to make the most of. Like anything, it takes some deep focus to get the hang of, but we are definitely in "standing on the shoulders of giants" territory.

Literally doing certain AI stuff "from scratch" is the work of world-leading-experts, but in a lot of ways my own work with these technologies is about as deep as my work with databases, where I really just get to know the abstractions and domain-specific languages I am provided.

I would like to build my own AI like Jarvis 🔥

I've built and recently launched AmjadGPT last weekend.

As for tools/technologies, I used Langchain (js lib) and Next.js as the frontend UI.

AI is definitely a hobby, but I'm slowly intertwining it into my career and using it more at work.

I built a browser extension that turned emoji art into accessible captions using Tensorflow. The ML inference worked well, but it would slow down the website a lot.

Back in the late 80s I made some character recognition neural networks from scratch in ZX Spectrum BASIC 🔥

I would like to build my own AI like Siri🤩