You will likely need about €200 for API credits or a bit of compute and about 200 hours of stubborn practice; in return you get the know-how from the world-class, public learning paths and tools with no gates! But you must choose to do the work. The simple plan is to pick one course, build one meaningful project, and loop hard on evaluation and iteration until the results improve. If this sounds interesting, keep reading!
The secret: there are no gates
Start with fast.ai – Practical Deep Learning for Coders and follow it end to end. It is pragmatic, free, and pairs well with public notebooks. Add Google’s Machine Learning Crash Course if you want clear visuals and core concepts.
Set up without headaches
Use Mamba as a fast, drop-in replacement for conda to create clean environments. Then install PyTorch using the official selector. If you prefer the cloud, run notebooks on Kaggle or Google Colab with free GPUs.
Your first project
Pick something a human can easily override or cancel so it is safe, but where mistakes are visible enough that you care to improve. Examples include personal email triage, flagging blurry photos, or a tiny tabular prediction at work. Ship a very small end-to-end slice and write down what went wrong and what changed.
Take a look here for a guideline on how professionals pick their AI projects.
The four pillars
Iteration
Change one thing at a time, track results, and repeat. That feedback loop is the craft.
Data and features
Most wins come from better data quality and representation, not fancy tricks.
Modeling
Spinning up a decent model is often trivial with today’s libraries; your edge rarely comes from a flashier architecture. fast.ai drives this point home in Practical Deep Learning for Coders.
Evaluation
Choose metrics that reflect reality. For classical ML, the scikit-learn evaluation guide is a solid reference. For current generative AI work, read Hamel Husain on building task-specific evals and testing loops.
Iteration
Change one thing at a time, track results, and repeat. That feedback loop is the craft.
Practical costs
APIs and hosted runtimes are pay-as-you-go. Skim pricing before you scale prompts or training time. Start small, log usage, and upgrade only when you see clear benefits.
When you want to go deeper
- Kaggle Learn micro-courses for quick skill gaps in Python, Pandas, and intro ML.
- tinygrad to peek under the hood of a minimal, hackable deep learning framework.
Bonus for career edge
Technical skills are the baseline. The unfair advantage is linking real business needs to the strengths of AI and choosing projects that move a metric the company cares about. When capabilities are equal, the person who can translate between business goals and feasible ML work brings more value.