SOTA (The Hype vs Confusion)

The problem isn’t that AI is too big.
The problem is that most people never stay in one place long enough to understand it.

Why actually most of the public is confused in AI and feeling overwhelming is because they're not on a single lane and didn't get to understand what really is currently getting build and what they can further contribute as there're so many lanes in AI if Someone tell me to categorize here are from my side:

• LLM reasoning (e.g., chain-of-thought, agents)

• Multimodal models (vision + text + audio)

• Efficient models (small, fast, on-device)

• Retrieval / RAG systems

• Robotics / embodied AI

• Scientific AI (biology, materials)

If you don't pick a lane you'll surely feel lost, so pick one and become enough good at that so that if anyone asks you something related to that field you must be able to not just tell them what it is but also being able to implement.

NOTE: NOT SAYING THE OTHER LANE IS A TRAP, BUT MULTIPLE LANES AS YOUR CORE IS A TRAP, SO FOCUS ON ONE JUST KNOW ABOUT OTHERS.

Apart from this, read research papers but related to domain specific not flying here and there, and build things.

Try to find out what're issues in current pipeline, just fixing it'll be something which will be considered as SOTA and which will help you as individual to run a sprint with the speed of marathon. Let's try to understand with an example:

“Take RAG systems. Most people build basic pipelines. But real issues are:

• retrieval returns irrelevant chunks

• context overload reduces accuracy

• latency kills usability
  Fixing these is where actual progress happens.”

If you worked on any of this(not limited scope on this example just) this will give us real edge over the random AI folks.

From generalist I mean just scratching the land and moving further, if you go enough deep into a field that you get either water (personal usecase) or best case is oil(something new and useful in commercial way), and then switching onto something is not someone who's normal generalist.

Just doing things in a single direction is the only best thing as of now is remaining else if we try to become generalist that'll just give us a platform where no train will ever come to onboard you. As AI is something which resolves anyone's problem of basic idea, so any founder, any engineer, any technologists, if they require someone for basic support they'll prefer AI due to recruitment process, and only experts will be more required. and that's where we'll lag and left behind thinking why even after doing the hard things I'm still gone nowhere.

So just do the work as having a single road for the final destination, else multi-paths will lead to nowhere.

(Not any expert but someone who's been through the same road so sharing my experience)