Don't let the dizzying pace of advancements in the ChatGPTera overwhelm. Be inspired by it, accelerate your learning & build cool stuff from shoulders of communities.
Youtube lectures. Find a youtuber (or a blogger) you enjoy learning from. There are great content-creators for almost everything. However, as this is my advice to my alternate-universe-self, I would strongly urge Andrej Karpathy's channel for what I think provides super cool hands-on learning by pulling (and reconstructing) ML stuff apart, decorated with wonder, and grounded with some math intuition.
Build! This is non-negotiable. At least, tweak tutorials towards applications that interest you. Collaborate and build with others, makes it fun & purposeful!
Follow your wonder. I recently stumbled upon 3Blue1Brown's channel. What a treat! I revisited information theory with wordle
Keep your intuition of math and data strong, especially since this speaks to you.
Wisely choose a structured online course with assignments if you have something very particular to learn - new frameworks, general foundational math/applications for neural networks (deeplearning.ai), niche use-cases (Biomedical AI), etc
Github - Let your code live (and die) in public (I still largely fail)
Write - personal notes. Document your learning for yourself
Write - publicly, online. Stackoverflow, possibly blogs, etc. (Let it be useful. Don't parakeet a tutorial or blog piece. Write to capture your aha moments, tips, or to explain/teach concepts, or to share a cool application you made)
Whilst I'd want my alternate self to learn the modern techniques that are state of the art for a reason, I would also insist on strong foundations and first principles learning with a healthy balance of math+building. By means of a bad analogy - the study of concepts behind the first electric motors did not cease to be relevant to an aspiring electrical engineer today. To be fair, I am biased due to my academic training, where physics concepts were timeless. On an extreme, I even vouch for the lab sessions where we'd would program 8085 microprocessors and CROs from 1980s.