As a developer, my first step was to take an online class on how to prompt for chat gpt et el.

My second class is on to go agentic AI (calling AI from a program).

I recently attended a short presentation on RAG (ask Chat GPT). This filled in a lot of holes in my brain about LLMs.

Pls clarify if you wish to develop AI such as new models and writing code that interacts with GPU or develop FOR ai such as GPT wrapper

Are you trying to build AI applications or research AI itself? Completely different paths.

If it's the former - skip the math and start calling APIs. OpenAI, Anthropic, or open-source models via Replicate. Spend a week building something real: add a chatbot to your product, build a document Q&A system, whatever solves an actual problem.

The deep learning theory? You can learn that later if you actually need to fine-tune models or optimize inference. Most developers never do. Don't let the AI hype convince you that you need a PhD to ship useful AI features.

Are you trying to build AI applications or research AI itself? Completely different paths.

If it's the former - skip the math and start calling APIs. OpenAI, Anthropic, or open-source models via Replicate. Spend a week building something real: add a chatbot to your product, build a document Q&A system, whatever solves an actual problem.

Focus on prompt engineering, handling token limits, streaming responses, managing costs, error handling. These are the 80% of "AI development" for application builders.

The deep learning theory? You can learn that later if you actually need to fine-tune models or optimize inference. Most developers never do. Don't let the AI hype convince you that you need a PhD to ship useful AI features.

It's basically Typescript / and / or Python engineering, working with streaming data, calling APIs as others mentioned. Try to pick apart an agent framework, such as Mastra or any of the million others. If you're a solid engineer you have all the AI building skills you need, just go build.

The best way in my opinion is learning from an open source project that is trying to do something similar to what you want.

If one is "experienced software developer", I'd say one then can guide oneself to right direction for learning by starting asking the most basic/obvious/childish questions, e.g. what libraries/languages/datasets/data formats/concepts is needed to understand and map/chart the territory (yes, that is what OP is basically doing now, asking but not just barreling into messy action) with a goal to create a minimal prototype.

True experienced developers have learned how to solve any programming problem. This related reading was actually yesterday mentioned at HN: https://beej.us/guide/bglcs/