Ask HN: How to get started with machine learning?

When someone posts "Ask HN: How to get started with machine learning?" on Hacker News, they're not asking for a textbook recommendation. They're asking for permission to skip the PhD. And honestly? They should. The old gatekeeping around ML — that you need a math degree and three years of grad school — is crumbling. I've watched self-taught engineers land ML roles in 18 months. I've also watched people with master's degrees freeze up when asked to deploy a model. The difference isn't intelligence. It's how they learned.

This guide is the one I send to friends who DM me that exact question. No fluff. No "learn linear algebra first" scare tactics. Just a path that works.

What "Getting Started With Machine Learning" Actually Means in 2025

Five years ago, getting started meant implementing gradient descent from scratch in NumPy. Today? That's like learning to forge your own hammer before building a house. The field has matured. According to Stanford's 2024 AI Index Report, the number of open-source models on Hugging Face grew 400% in two years. You don't need to build from scratch anymore. You need to know what to use, when to use it, and how to debug when it breaks.

Related: I've explored this before in Machine Learning and Ketosis.

I break ML newcomers into two camps. Camp A wants to build things — products, apps, automations. Camp B wants to understand things — research, theory, pushing boundaries. This guide is for Camp A. If you're Camp B, you'll still benefit from the practical parts, but you'll eventually need the math I'm deliberately skipping here.

Here's what "getting started" actually requires: Python fluency, data intuition, and the ability to read a paper and implement the key idea. That's it. The rest is noise.

Related: This connects to what I wrote about Gemini Robotics 2 Brings Google's AI Into the Physical World.

The 4-Week Roadmap That Actually Works

I've tested this sequence with three career-switchers. All three shipped working projects by week four. Not perfect projects. Working ones. There's a difference.

Week 1: Python and Pandas — But Only the Parts You'll Use

Skip the full Python course. You need five things: variables, functions, loops, list comprehensions, and error handling. Then jump straight into Pandas. Why Pandas first? Because 80% of ML work is data wrangling, and nothing teaches data intuition faster than cleaning messy datasets with your own hands.

Related: For more on this, see ai content generator for linkedin.

Here's my actual week-one exercise I give people: download the Titanic dataset from Kaggle. Load it. Find the column with the most missing values. Fill those gaps with the median. Create a new column that combines age and fare into a single feature. Plot it. This sounds trivial. It's not. I've interviewed candidates who can explain transformers but can't handle missing data. Don't be that person.

Use Kaggle's free micro-courses for Pandas and data cleaning. They're interactive, browser-based, and take maybe 8 hours total. No setup headaches.

Week 2: Scikit-Learn and the "Hello World" Models

Start with three algorithms: linear regression, decision trees, and k-means clustering. That's it. Don't touch neural networks yet. Each of these teaches a fundamental concept — regression teaches prediction, trees teach feature importance, clustering teaches unsupervised learning.

I recommend the official Scikit-Learn tutorial for this. It's dry but correct. Supplement it with StatQuest videos on YouTube — Josh Starmer explains concepts visually in a way that actually sticks. His decision tree video alone is worth more than most paid courses.

By the end of week two, you should be able to load any CSV, train a model on 80% of it, and predict the remaining 20%. If you can do that, you're already ahead of where I was at month three.

Week 3: Your First Real Project (Not a Tutorial)

This is where most people quit. Tutorials feel productive. They're not. Following along with someone else's code teaches you to type, not to think. You need a project where nobody has written the solution.

Some ideas that worked for people I've mentored:

• Scrape Craigslist apartment listings and predict rent prices based on square footage, bedrooms, and neighborhood
• Download your own Spotify listening history and cluster songs by audio features
• Use a public health dataset to predict diabetes risk from basic biomarkers

The key rule: pick something you care about. I once spent three weeks building a model to predict NBA game outcomes. It was terrible. I learned more from that failure than from any course. Caring about the problem makes you push through the debugging sessions that actually teach you.

Week 4: Deploy Something. Anything.

A model on your laptop is a hobby. A model someone else can use is a portfolio piece. Use Gradio or Streamlit — both let you wrap a model in a web interface with about 10 lines of Python. I prefer Gradio for quick demos; Streamlit for anything I want to customize.

Deploy it to Hugging Face Spaces (free) or Render (also free tier). Send the link to three friends. Watch them break it. Fix the edge cases. Congratulations — you've now done more than 90% of ML course graduates.

3 Mistakes That Keep Beginners Stuck for Years

I've made all of these. So has everyone I know who actually works in ML.

Mistake 1: Waiting Until You "Know Enough Math"

You need exactly enough math to debug why your loss isn't decreasing. That's it. For linear regression, you need to understand what a derivative means conceptually. For neural networks, you need to understand chain rule. You do not need to derive backpropagation by hand before training your first model. I've shipped models to production that I couldn't derive from scratch. The math comes naturally when you're curious about why something isn't working.

According to a 2023 survey by Kaggle, only 28% of professional data scientists have a master's degree or higher. The rest learned on the job. Let that sink in.

Mistake 2: Chasing State-of-the-Art Before Learning the Basics

Everyone wants to fine-tune Llama. Almost nobody needs to. I've seen beginners spend weeks trying to get a transformer working when a random forest would have solved their problem in an afternoon. Start simple. When simple breaks, upgrade. This isn't just practical advice — it's how you develop the diagnostic intuition that separates senior engineers from juniors.

Mistake 3: Tutorial Hell

You know the cycle. Finish one tutorial, feel accomplished, start another. Six months later, you've completed 40 courses and can't build anything from scratch. The fix is uncomfortable: build something ugly. Ship something that barely works. The discomfort of not knowing what to do next is the signal that you're actually learning. Tutorials remove that discomfort. That's why they're addictive. That's also why they don't work.

What I Actually Use Day-to-Day (And What I Don't)

My daily stack is less glamorous than you'd think. Python, Pandas, Scikit-Learn, and Jupyter notebooks cover 70% of my work. For deep learning, PyTorch. For quick experiments, Google Colab with a T4 GPU. For deployment, FastAPI plus Docker.

What don't I use? AutoML tools for anything serious. They're fine for prototyping, but I've never seen one handle edge cases well. And I rarely use JAX or TensorFlow anymore — PyTorch won the ecosystem war for most practical work.

For learning resources, fast.ai remains the best free course I've found. Jeremy Howard teaches top-down: build something cool first, understand the theory later. It's the opposite of academia and it works. The Hugging Face NLP course is also excellent once you've got the basics down.

One workflow I've settled into: when I'm researching a new technique, I'll read the abstract and conclusion of the paper first. If it still seems relevant, I'll skim the method section. Then I'll find the official implementation on GitHub and read the code. Papers explain the ideal. Code reveals the compromises. You learn more from the compromises.

How to Stay Current Without Burning Out

ML moves fast. Papers drop daily. New models every week. If you try to keep up, you'll burn out in three months. Here's my rule: follow five people whose judgment you trust, and ignore everything else until they signal it matters.

My personal list: Andrej Karpathy (practical deep learning insights), Simon Willison (LLMs and tooling), Lilian Weng (technical depth on OpenAI's blog), Chip Huyen (ML systems design), and the Hugging Face blog. That's it. Five sources. If something is genuinely important, at least two of them will mention it within a week.

For structured learning, I set aside two hours every Friday morning to read one paper and implement one small piece of it. Consistency beats intensity. Every single time.

Of course, there's a faster way to get hands-on with AI without drowning in setup. Tools like AI-Mind let you generate content, analyze text, and experiment with AI outputs without writing a single line of prompt engineering. It's not a replacement for learning ML — nothing is — but it's useful for understanding what modern AI can and can't do before you invest months in training your own models. The first 30 generations are free, so there's no reason not to poke around and see what the output quality looks like across different content types. Sometimes the best way to understand a tool is to use a polished version of it first, then work backward to understand the mechanics.

Key Takeaways

• Start with Python and Pandas, not math. Data wrangling is 80% of the work and teaches intuition faster than theory.
• Build a real project by week three. Tutorials feel productive but don't teach you to solve novel problems.
• Deploy something by week four using Gradio or Streamlit. A live project is worth more than 10 completed courses.
• Follow five trusted sources, not the entire ML firehose. Consistency beats intensity for long-term learning.
• Simple models solve most problems. Don't reach for transformers when a random forest will do the job.

Sources

• Stanford HAI, AI Index Report 2024, 2024. Annual comprehensive report tracking AI trends, including open-source model growth on Hugging Face.
• Kaggle, State of Data Science and Machine Learning Survey, 2023. Industry-wide survey of 30,000+ practitioners on education, tools, and workflows.
• fast.ai, Practical Deep Learning for Coders, 2024. Free top-down ML course emphasizing project-based learning over theory-first approaches.
• Scikit-Learn, Official Tutorial Documentation, 2024. Core library tutorials covering regression, classification, and clustering fundamentals.