Stop telling beginners to learn neural networks first

Someone asked a simple question on Threads recently: "I don't have any knowledge of AI. Where should I start? And how would I apply my knowledge in real life work?"

The first reply was the one you'd expect: start with the basics, learn how neural networks work, train your first model on some example data.

It's well-meaning advice. It's also exactly backwards for almost everyone who asks that question.

The muscle fiber problem

Imagine someone tells you they want to start exercising. Maybe they want more energy, maybe they want to lose weight, maybe they just want to stop being winded on stairs. And your response is: "First, let's cover the different types of muscle fibers in the human body - slow-twitch, fast-twitch, how each responds to training stimulus."

Technically correct. Completely useless. That person is going to nod politely and never work out again.

The advice fails because it optimises for the wrong thing. It treats understanding the system deeply as a prerequisite for getting a benefit from the system, when in reality those are two different goals that call for two different paths. Most people don't need to understand fiber types to get stronger, run further, or feel better. They need to walk, then jog, then find a form of movement they don't dread.

AI advice has the same failure mode. 'Learn how neural nets work, train your first model' is the muscle-fiber answer. It's real knowledge, but it's front-loaded onto people who haven't yet decided this is something they want to go deep on - and most of them won't, because most people don't need to.

Two different questions, wearing the same outfit

'How do I learn AI?' is actually two separate questions that get asked with the same four words.

'How do I build AI systems?' - this is the researcher and engineer path. You do need the fundamentals here: linear algebra, how gradient descent works, what a transformer actually does under the hood, how to train and fine-tune a model. This is a real discipline with a real floor of technical depth, and there's no shortcut around it if your goal is to design or improve the systems themselves.

'How do I use AI to do my job better?' - this is what almost everyone asking is actually asking, whether they realise it or not. A marketer wanting to speed up campaign research, a lawyer wanting to draft first-pass contract summaries, a founder wanting to prototype faster, a student wanting a better study partner - none of these people need to know what a loss function is. They need to know how to write a good prompt, how to verify AI output instead of trusting it blindly, and which tools fit which tasks.

Answering question two with the training playbook for question one is why so many beginners bounce off AI entirely. They go looking for how to use ChatGPT to write better emails and get handed a linear algebra syllabus. Most people, reasonably, close the tab.

But you'll never really understand it without the fundamentals

This is the pushback, and it's not wrong - it's just answering a different question than the one that was asked. Yes, if you want to eventually train your own models, fine-tune an open-source LLM, or work as an ML engineer, the fundamentals aren't optional. Nobody is arguing you can skip linear algebra and become a research scientist.

But 'you'll never train your own model without basics' is true and irrelevant for someone who never wanted to train a model in the first place. They wanted to stop spending four hours a week writing status reports. Telling them they need to understand backpropagation first is like telling the guy who wants to jog around the block that he'll never deadlift 400 pounds without learning about fast-twitch fibers. Correct, and not what he asked.

Depth without direction just becomes a wall between someone and the thing they actually wanted.

The fundamentals matter enormously - for the 5% of people whose actual goal is building the systems. For everyone else, the wall is the whole outcome.

So where should a beginner actually start?

If you're the person who asked the original question, here's a more useful version of the advice.

1. Start with a problem you already have, not a technology you're supposed to learn. Don't open a tutorial titled 'Introduction to Machine Learning'. Open ChatGPT, Claude, or whatever tool is in front of you, and try to solve something real: draft an email you've been putting off, summarise a document you don't want to read, brainstorm a plan for something you're stuck on.
2. Learn by using, not by studying. The skill that transfers to real work isn't knowing how attention mechanisms work - it's knowing how to ask good questions, how to iterate on a bad answer, and how to spot when the model is confidently wrong. You build that by doing it fifty times, not by reading about it once.
3. Let curiosity pull you deeper, don't push yourself there. If, six months in, you find yourself wanting to know why the model does what it does, or you want to build something off-the-shelf tools can't do, that's the moment the fundamentals become worth your time - because now you have a reason to want them. Motivation that comes from a real question sticks. Motivation borrowed from someone else's roadmap doesn't.
4. Ignore the implication that using AI without training one isn't 'really' AI skill. It's a different skill, not a lesser one. The person who can get reliable, useful output out of a model for real work is doing something plenty of PhDs can't do well.

The actual answer

Not everyone needs to know how to train a model, the same way not everyone doing squats needs to know their fiber-type ratio. What almost everyone needs is a real problem, a tool in hand, and permission to start small.

If the fundamentals ever become the bottleneck between you and something you're trying to build, you'll know - and you'll learn them fast, because by then you'll actually want to.