Video summary
Lộ trình học AI toàn diện từ A đến Z năm 2025
Main summary
Key takeaways
Main ideas & lessons
- Purpose of the video: Provide a comprehensive AI learning roadmap (A to Z) for 2025, built from the creator’s 10 years of experience as an AI engineer.
- Who it’s for:
- People who want to learn AI from scratch
- People already learning AI but without a clear plan/direction
- Non-IT backgrounds who want to transition into AI
- High-level structure:
- IT learners: 6 steps
- Non-IT learners: 7 steps (includes an extra foundational IT step)
Roadmap (step-by-step)
Step 0 (non-IT / unrelated fields only): Equip basic IT fundamentals
Core idea: AI is a sub-branch of CS. Without foundations, you’ll constantly struggle. You don’t need to take every CS course, but you should cover three key subjects:
- Basic Programming
- Learn programming fundamentals: variables, operators, loops, conditionals.
- Data Structures + Algorithms
- Learn common data structures: linked lists, stacks, queues, trees, graphs.
- Learn algorithm complexity (efficiency in time and resources).
- Database
- Understand how AI/ML depends on data for training/testing/deployment.
- Learn to collect, store, and query data efficiently and professionally.
Suggested resources mentioned:
- CS50 (Harvard University): watch Lecture 0 to Lecture 7 (“Scratch to SQL”)—skip or review parts as needed.
- FreeCodeCamp: “Introduction to Programming and Computer Sense”.
Step 1: Learn the right level of mathematics
Core idea: For most industry roles (data scientist / ML engineer / AI engineer), you mainly need basic math understanding, not advanced academic math.
Math goal: be able to fine-tune and interpret models, and help with interview readiness.
Key topics:
- Statistics (most important)
- Probability for accurate data handling and processing.
- How to treat outliers (e.g., super-rich values in income data).
- Statistical evaluation to judge reliability/performance.
- Example: R² / coefficient of determination
- Example meaning: R² = 0.85 ⇒ explains 85% of variation.
- Linear Algebra
- Operations with vectors and matrices (how ML models use training data).
- Eigenvalues/eigenvectors for dimensionality reduction concepts.
- Calculus
- Only a small amount is needed—mainly understanding an extended version of derivatives (e.g., gradient and related ideas).
- Trigonometry (limited use)
- Mainly cosine similarity (similarity between vectors).
- Sometimes used in tasks like image rotation.
Suggested resources mentioned:
- Khan Academy (English): math supplements.
- 3Blue1Brown (spelled as “Blue One Brown” in the subtitles): YouTube videos for intuitive linear algebra/calculus.
Step 2: Learn Python
Core idea: Python is essential for AI/data engineering due to:
- Clear syntax
- Rich libraries
- Strong community support
Learning guidance:
- If new: start with basics first (syntax and fundamentals).
- Avoid jumping directly into pre-built libraries—don’t become dependent and fail when something breaks.
Suggested tools:
- Jupyter Notebook (beginner-friendly)
- PyCharm / VS Code (more professional IDEs)
Learning resources mentioned (two options):
- Harvard University Python programming course
- w3schools: Python tutorial
Step 3: Learn AI/ML libraries and ecosystem (practical tooling)
Core idea: Python’s success in AI comes from a large ecosystem of libraries.
Libraries/frameworks mentioned:
- NumPy (spoken as “nPAI”)
- Foundations for multidimensional arrays and numerical computation.
- Pandas
- Data manipulation and analysis (tables, time series).
- Data visualization tools (names unclear in subtitles; intent includes tools such as Plotly)
Learning approach:
- Use tutorials from each library’s homepage (or via w3schools).
- Learn with examples (NumPy, pandas, and visualization tutorials).
Then study ML fundamentals:
- Machine learning = algorithms that learn patterns from data rather than relying on manually coded rules.
- Common mistake: only using model/library APIs without understanding the underlying algorithms and why predictions happen.
Focus areas within ML:
- Supervised learning: trained on labeled data
- Unsupervised learning: trained on unlabeled data
ML libraries/frameworks mentioned:
- scikit-learn (spoken incorrectly as “Pilon”)
- For building/training/evaluating models; supports supervised and unsupervised learning.
- A second boosting/GPU-focused library mentioned with subtitle errors (intended as a faster training/boosting library using GPU and parallel threads).
Outcome of Step 3:
- Understand ML algorithm theory
- Practice training models on datasets
- Learn optimization for better performance
ML learning resources mentioned (three):
- “Machine Learning for Everybody” (FreeCodeCamp, ~4 hours)
- Machine Learning Specialization (Coursera; platform name unclear in subtitles)
- CS229 Machine Learning (Stanford)
Step 4: Learn Deep Learning
Core idea: Deep learning uses artificial neural networks to learn from complex data (images, audio, text, video) and is central to modern AI.
Applications mentioned:
- Computer Vision
- facial recognition, self-driving cars, smart healthcare
- Natural Language Processing (NLP)
- chatbots for translation, virtual assistants
Deep learning learning sequence:
- Understand how neural networks are formed
- Learn advanced network types:
- Convolutional networks for vision and regression tasks
- Learn Transformers architecture for NLP (emphasized as widely important recently)
Deep learning libraries mentioned:
- PyTorch (spoken as “PyThos”)
- TensorFlow
- Keras
Recommendation given: prioritize PyTorch (described as combining Keras-style simplicity with TensorFlow flexibility).
Deep learning course resources mentioned (plus extras):
- FreeCodeCamp: “Deep Learning (for Beginners)” (course name slightly corrupted in subtitles)
- FreeCodeCamp: “PyTorch for Deep Learning and Machine Learning” (~26 hours)
- “Deep Learning Specialization” (Coursera; platform unclear in subtitles)
- Stanford: CS230/CS231/CS23 Deep Learning (title partially corrupted)
Extras mentioned:
- CS231n: Deep Learning for Computer Vision (Stanford)
- CS224n: Natural Language Processing with Deep Learning (Stanford)
Step 5: Build real-world projects (beyond toy datasets)
Core idea: After ML + DL basics, build projects to convert knowledge into practical capability and recruiter-visible proof.
Project guidance:
- Early practice often uses toy datasets with curated guidance.
- Real-world projects require you to handle:
- what to do next
- end-to-end workflow without mentor intervention
- Projects are described as a transition from “book knowledge” to “personal knowledge”.
How to get real datasets (mentioned):
- Use Kaggle (implied as “the world’s largest data and machine learning platform”).
Technologies to learn in this stage:
- Docker
- Cloud platforms: AWS, Google Cloud, Microsoft Azure
Portfolio/recruiter advice:
- Publish work on GitHub (subtitles mis-transcribed; intended as GitHub)
- Engage with the community (likes/stars/comments can help)
Additional promotional content (courses + tools)
The speaker promotes upcoming courses, including:
- No-code AI course
- build/train/evaluate/deploy without programming
- Basic course for non-IT beginners
- basic Python + AI/data science libraries + AI applications
- Advanced data science & machine learning course
- includes NLP + Computer Vision
- uses a private dataset from the speaker’s previous company (recruitment domain)
- Deep Learning for Computer Vision (basic) and Advanced CV
- topics include object detection, segmentation, GANs, etc.
- advanced course includes Docker
- Math for AI course
- probability, statistics, linear algebra, calculus
- emphasizes practical application (not only theory)
- Contact method: “Contact me via Zalo” (number shown in an image; not included in subtitles)
Speakers / sources featured (as named or clearly referenced)
Speaker / creator
- Unnamed video creator
- Speaks as “myself”; described as an AI engineer with 10 years of experience.
Courses, channels, and organizations referenced
- Harvard University – CS50
- Lecture 0 to Lecture 7 (“Scratch to SQL”)
- FreeCodeCamp
- “Introduction to Programming and Computer Sense”
- “Machine Learning for Everybody”
- Deep learning courses (multiple referenced)
- w3schools
- Python tutorial
- library tutorials (NumPy/pandas/visualization)
- Khan Academy
- math resources (English mentioned)
- 3Blue1Brown (“Blue One Brown” in subtitles)
- math/physics explanation channel
- Stanford University
- CS229 (Machine Learning)
- CS231n (Computer Vision)
- CS224n (NLP with Deep Learning)
- another deep learning course mentioned with corrupted subtitle title
- Coursera (subtitles show “CERA” due to transcription issues)
- Machine Learning Specialization
- Deep Learning Specialization
- Kaggle (implied as the data platform)
- GitHub (subtitles mis-transcribed as “Kitap”)
- Cloud platforms: AWS, Google Cloud, Microsoft Azure
- Docker
Libraries/frameworks referenced
- NumPy
- Pandas
- Visualization tools (auto-subtitle errors; intended plotting/visualization libraries such as Plotly)
- scikit-learn (spoken incorrectly as “Pilon”)
- Boosting/GPU-focused library (subtitle errors; intended for fast boosting/training)
- PyTorch (spoken as “PyThos”)
- TensorFlow
- Keras
- Jupyter Notebook
- PyCharm / VS Code
Contact platform
- Zalo