Learning Machine Learning & GenAI as a beginner – what surprised me the most

[abhinyaay]

Body

I recently started learning Machine Learning and Generative AI,
and honestly, it has been both exciting and overwhelming.

What surprised me the most is that:

• ML is less about fancy models and more about understanding data
• Simple techniques like TF-IDF or Logistic Regression still matter a lot
• GenAI is powerful, but without fundamentals, it feels like magic without control

At first, I thought I needed to learn everything at once:
deep learning, transformers, fine-tuning, RAG, etc.
But slowly I realized that building small things and understanding why
something works is more important than rushing.

For other beginners:

• Don’t skip ML fundamentals
• Try to implement small projects
• Confusion is normal (and part of the process)

I’d love to hear from others:

• Did you start with ML or GenAI?
• What concept confused you the most in the beginning?

Guidelines

• I have read and understood this category's guidelines before making this post.

[ohhaus]

You're already thinking correctly. The excitement is for the toy, the overwhelming feeling is when you realize it's a real tool. You've discovered the most important thing: without the fundamentals, you're just using a magic box you don't understand. That's how you get nonsense outputs and can't fix them.

Starting with ML first is the only sane path. Jumping straight to GenAI is like trying to build a race car when you don't know how an engine works. The concept that confuses everyone in the beginning is that the model itself is the least important part. The data is everything. Cleaning it, understanding its shape, its biases—that's 90% of the real work.

Your plan is right. Build small things. Implement linear regression from scratch with NumPy. Then do logistic regression. Then a decision tree. You will understand more from making a stupid classifier work on the Titanic dataset than from copying a fancy transformer tutorial.

The math isn't optional. You don't need a PhD, but you must understand why a gradient is calculated, what a loss function actually measures, and why your data is split three ways. This is what turns the magic into a controlled, debuggable engineering process.

[wang900115]

In 2022, I dabbled in machine learning and learned about its underlying principles. It involved a lot of partial differential equations, engineering mathematics, and linear algebra, as well as some technical terminology. I found it quite interesting at the time.

The most challenging part was understanding the mathematical notation and viewing the models from a system-wide perspective.

[abhinyaay]

Thanks for reading!
Would love to hear how others started their ML or GenAI journey.

[UtsavKash19]

Body

I recently started learning Machine Learning and Generative AI, and honestly, it has been both exciting and overwhelming.

What surprised me the most is that:

• ML is less about fancy models and more about understanding data
• Simple techniques like TF-IDF or Logistic Regression still matter a lot
• GenAI is powerful, but without fundamentals, it feels like magic without control

At first, I thought I needed to learn everything at once: deep learning, transformers, fine-tuning, RAG, etc. But slowly I realized that building small things and understanding why something works is more important than rushing.

For other beginners:

• Don’t skip ML fundamentals
• Try to implement small projects
• Confusion is normal (and part of the process)

I’d love to hear from others:

• Did you start with ML or GenAI?
• What concept confused you the most in the beginning?

Guidelines

• I have read and understood this category's guidelines before making this post.

Great reflection — and honestly, a very accurate realization for someone just starting out in Machine Learning and GenAI.

What surprised me the most when learning this field was also how important the basics are compared to the hype around advanced models.

---

MY EXPERIENCE STARTING WITH ML VS GENAI

I started with traditional Machine Learning first, and in hindsight, that made GenAI much easier to understand later.

Key reasons:

• Understanding data preprocessing, features, and evaluation builds real intuition.
• Models like Linear/Logistic Regression, Decision Trees, and TF-IDF teach why predictions work.
• Without this foundation, GenAI can feel like a black box instead of a controllable system.

So your realization is actually a very strong learning signal — you’re on the right path.

---

CONCEPTS THAT WERE CONFUSING AT THE BEGINNING

Some things that commonly confuse beginners (including me):

1. Bias vs Variance

   • Hard to visualize at first.
   • Becomes clear only after experimenting with real datasets.

2. Overfitting

   • Models performing perfectly on training data but failing in reality felt counter-intuitive.

3. Embeddings in GenAI

   • Understanding that text becomes vectors representing meaning, not just numbers, takes time.

4. Transformers & Attention

   • The math and architecture look scary, but conceptually it’s about learning relationships in sequences.

---

ADVICE FOR BEGINNERS

Your points are already excellent. I’d add:

• Build tiny projects (spam classifier, sentiment analysis, simple RAG app).
• Focus on intuition before complexity.
• Learn evaluation metrics (accuracy, precision, recall, F1).
• Accept that confusion is part of progress, not failure.

Most importantly:

➡ You don’t need to learn everything.
➡ You just need to keep building and understanding one step at a time.

---

SHORT ANSWER TO YOUR QUESTIONS

Did you start with ML or GenAI?
Started with traditional ML → helped massively with GenAI later.

Most confusing concept?
Bias–variance tradeoff and embeddings.

---

You’re thinking in exactly the right direction.
That shift from “learn everything fast” to “understand deeply and build small” is usually the moment real progress begins.

[ganapathijahnavi]

If you’re surprised that ML/GenAI isn’t just “build a model and done,” that’s normal but it’s also the reality of the field.

Machine Learning is not about models first. It’s about:

1. Understanding the problem
2. Understanding the data
3. Choosing the right approach
4. Evaluating properly

Most beginners waste time jumping into:

1. Transformers
2. LLM APIs
3. “Build your own ChatGPT” tutorials

Without understanding:

1. Bias vs variance
2. Overfitting
3. Evaluation metrics
4. Feature engineering
5. Why a model fails

If you want a solid path, here’s a better order:

Step 1 – Math basics (not advanced, just enough)

1. Linear algebra basics (vectors, matrices)
2. Probability basics
3. Gradient intuition

Step 2 – Classical ML first

1. Linear regression
2. Logistic regression
3. Decision trees
4. Random forests
5. KNN
   Implement at least one model from scratch. It forces you to understand what’s happening.

Step 3 – Work with messy datasets
Real learning happens when:

1. Data is incomplete
2. Labels are noisy
3. Accuracy is low
4. Your model fails
   That’s actual ML.

Step 4 – Then move to Deep Learning
Only after you understand why simpler models fail.

Step 5 – GenAI
GenAI is built on:

1. Transformers
2. Optimization
3. Representation learning

If you skip fundamentals, you’ll only know how to call APIs not how things work.

Also, don’t confuse:
. “Using AI tools”
with
. “Understanding AI systems”
Very different skill levels.

->If your goal is:
Research → go deeper into theory
Industry → focus on data pipelines, deployment, evaluation
Just GenAI apps → learn APIs + prompt engineering + system design

Be clear about your goal.

That’s it. No hype. Build fundamentals → build small projects → break things → debug → repeat.

[rikream]

Interesting reflection. Many people jump straight into GenAI tools now, but having ML fundamentals really helps make sense of what’s happening under the hood.

I personally started with traditional ML first (regression, classification, basic NLP) before looking into generative models. The biggest early confusion for me was understanding the difference between model performance metrics and real-world usefulness — a model can score well but still fail in practical scenarios because of data quality or bias.

Your point about building small projects is very true. Implementing even simple pipelines (data cleaning → feature extraction → model → evaluation) teaches more than just reading theory. GenAI becomes much easier to understand once those basics are solid.

Curious to know — are you focusing more on NLP/GenAI now or still exploring general ML first?

[rohmantr]

I can relate to this a lot.

When I started, I also assumed machine learning meant jumping straight into deep learning and transformers. It took time to understand that most real progress comes from understanding data: how it’s collected, cleaned, represented, and evaluated. Models are often the smallest part of the system.

You’re absolutely right about fundamentals. Techniques like TF-IDF and Logistic Regression look simple, but they teach core ideas: feature engineering, linear decision boundaries, bias–variance trade-off, and proper evaluation. Those concepts transfer directly to more advanced models.

With Generative AI, it’s easy to feel like everything is magic. Large models abstract away so much complexity that you can get results without understanding what’s happening underneath. But without grounding in probability, optimization, embeddings, and evaluation, it’s hard to reason about failure cases or improve systems intentionally.

I started with classical ML before moving into generative models. The most confusing concepts early on were:

1. Overfitting vs underfitting in practical terms
2. Regularization and why it works
3. Backpropagation (not the math itself, but the intuition behind gradient flow)
4. Evaluation metrics and when accuracy is misleading

Building small projects was what made things click. Even simple pipelines—clean data → vectorize → train → evaluate. help build intuition that scales to transformers and RAG systems later.

Your approach is solid: depth over speed, understanding over trend-chasing.

[Randa-Lakab]

I really relate to this. I also started feeling overwhelmed because everything in ML and GenAI looks advanced and fast-moving.

For me, I started with traditional Machine Learning before diving into Generative AI. And honestly, that helped a lot.

The concept that confused me most at the beginning was:

1️⃣ The difference between models and data quality
I kept thinking the model was everything. Later I realized that clean data, feature engineering, and problem understanding matter more than using a “fancy” model.

2️⃣ Overfitting vs Underfitting
It sounded simple in theory, but understanding how it appears in real projects (especially when validation accuracy behaves strangely) was confusing at first.

3️⃣ Why simple models still work
Like you said, Logistic Regression, TF-IDF, or even basic feature scaling can outperform complex deep models on structured data. That was surprising.

My Advice to Beginners

1. Start with:

• Linear Regression

• Logistic Regression

• Basic probability & statistics

• How training/validation works

2. Build tiny projects (spam classifier, sentiment analysis, simple prediction model).

3. Only then move to transformers and LLMs.

GenAI feels magical, but once you understand:

• embeddings

• tokenization

• attention

• optimization

…it stops feeling like magic and starts feeling like engineering.