Six ways to learn AI, depending on what you want to build
Karpathy, 3Blue1Brown, fast.ai, Hugging Face, Andrew Ng, and Doug Does AI’s interactive courses: what each is good for, what you need first, and where to go next.
Last checked: October 3, 2026.
Start with your goal
- I want to see why neural networks work: start with 3Blue1Brown.
- I want to implement the machinery myself: choose Karpathy’s Neural Networks: Zero to Hero.
- I already code and want to build a useful ML application: choose fast.ai.
- I want to use and fine-tune models in the Hugging Face ecosystem: choose the Hugging Face LLM Course after an introductory deep-learning course.
- I want a structured foundation across machine learning: choose Andrew Ng’s Machine Learning Specialization.
- I want to explore one idea and explain it back in a short session: try a dougdoes.ai lesson, then use a longer course when you want more depth.
There is no overall winner here. A visual explanation, a code-along implementation, a graded course, and a small interactive experiment solve different problems. Several of these resources work well together.
1. Karpathy: Neural Networks: Zero to Hero
Choose it for: understanding neural networks by implementing them.
The sequence builds from backpropagation through character-level language models to a GPT-style transformer. Expect Python, tensors, debugging, and sustained code-along work. The official prerequisites are solid Python and introductory mathematics. The syllabus lists micrograd at 2 hours 25 minutes and the GPT build at 1 hour 56 minutes; those are viewing times, before exercises or troubleshooting. Official syllabus
Access and practice: public videos, reference notebooks, and exercises. The course points learners toward a Discord community. Code outputs and reference implementations help you check your work; this is not a promise of individual instructor review. Notebooks and exercise directions
Tradeoff: choose Karpathy when you want to implement autodiff and investigate neural-network internals in depth. It asks for sustained coding practice. Short browser experiments can help you revisit a calculation between coding sessions.
2. 3Blue1Brown: neural-network and transformer explanations
Choose it for: visual intuition before, or alongside, implementation.
Start with the neural-network and gradient-descent explanations, then move to transformers and attention. The selected material explains the ideas visually rather than walking you through a complete model-training project. You can start without a coding environment; vectors, matrices, and derivatives become increasingly relevant. Neural-network lesson · Attention lesson
Access and practice: public videos and lesson pages. Some pages include interactive examples, reveal prompts, and answer-check questions. This is useful feedback on a specific question, rather than review of an independently built model.
Time: the current playlist lists the introductory neural-network video at 18:40 and attention at 26:10. These are individual video runtimes, not a complete-course estimate. Official playlist
Tradeoff: a strong choice when an equation or attention diagram has not clicked. Add a coding course if your goal is to implement, train, and debug a model yourself.
3. fast.ai: Practical Deep Learning for Coders, Part 1
Choose it for: getting a working application first, then understanding its parts.
Part 1 covers vision, NLP, tabular data, recommendation, and deployment. It assumes coding experience, preferably Python, and high-school mathematics. The provider describes nine lessons of around 90 minutes each, with reading and practical work beyond that. This comparison covers the 2022 Part 1 course; Part 2 is a separate, longer sequence. Course overview
Access and practice: the course and online book are free. Lessons use notebooks, projects, and code experiments, with community forums and linked solutions to book questions. Hosted notebooks can require an account; compute availability and provider limits are separate from free course access. Getting started and exercises
Tradeoff: choose fast.ai when your goal is a broader applied-ML project, especially outside language modeling. It is a larger commitment than a single browser exercise. Its older recordings also make current library documentation worth keeping nearby.
4. Hugging Face: LLM Course
Choose it for: working with pretrained models, datasets, tokenizers, and the surrounding tools.
The course is free and ad-free. It asks for good Python and recommends an introductory deep-learning course first. Its published study guideline is approximately 6–8 hours per chapter, spread over a week; that is a workload estimate, not video length. It combines explanations, runnable notebooks, quizzes, and community forums. Some features require a Hugging Face account. Introduction and prerequisites
The fine-tuning material covers both the Trainer API and a custom training loop. The later SFT chapter includes chat templates, LoRA, and evaluation. Fine-tuning chapter · SFT and LoRA
Tradeoff: choose Hugging Face when you want broader familiarity with its ecosystem and more tasks. Doug Does AI’s SQL sequence explores a narrower question: did one adapter improve answers, and what does its score fail to establish?
5. Andrew Ng: Machine Learning Specialization
Choose it for: a structured introduction to machine learning beyond LLMs.
The three-course sequence includes regression, classification, neural networks, trees, clustering, recommenders, and reinforcement learning. The entry requirements are basic programming and high-school mathematics. Practice notebooks, quizzes, and graded assignments make it more structured than a video playlist. Official program and syllabus
Access and time: self-paced, multi-week study. The provider’s live catalog and FAQ show different duration estimates, so a single exact total is a poor study-time estimate. On DeepLearning.AI, videos and community access are free; Pro adds hands-on labs and certificates. The listed price is US$30 monthly or US$300 billed annually, before applicable taxes. Check the current membership terms before paying. Coursera is a separate enrollment option with its own terms.
Tradeoff: choose the Machine Learning Specialization when you want a broad ML curriculum and formal assignment progression. It is not the most direct route here to building a small GPT or running a modern LoRA experiment.
6. Doug Does AI’s interactive courses
Choose it for: a small concept you can manipulate, explain, and revisit.
Doug Does AI has three interactive course tracks:
- Building Language Models from Scratch: six lessons, from count-based predictions to a small Shakespeare transformer, plus a downloadable CPU Python project.
- Fundamentals of Machine Learning: six lessons on a neuron, loss, gradients, optimizers, batches, and softmax. Basic algebra is enough to start.
- Fine-tuning a Model for SQL: five lessons on a WikiSQL experiment, SFT, LoRA, and evaluation. The optional local lab requires Apple Silicon and Python 3.12.
The language-model and SQL tracks assume basic Python. You can read the lessons and explore prepared examples without installing it. Some controls replay recorded model results or illustrate a calculation; they do not train a model in your browser. Run the downloadable projects to do the training yourself.
Access and feedback: public lessons, prepared examples, written hints, and local explanation drafts work without an account. AI feedback on your explanation and page chat require sign-in, eligible access, and available usage. Authored checks and hints remain useful when AI feedback is unavailable. AI feedback can be wrong; compare it with the worked explanation and actual code results. Course hub
Time: try a 15–30-minute study session on one idea. This is a suggested way to use the material, not a measured completion time or a promise that every lesson fits. Coding, setup, and further experiments can take longer.
Tradeoff: each lesson is deliberately smaller in scope than a long technical lecture. The tracks overlap with some topics in the courses above, but the comparison does not establish equivalent coverage or learning outcomes. Choose a longer course when you want its breadth, detailed derivations, or sustained implementation practice.
What the topic coverage actually looks like
This is a selective map of the reviewed material, not a claim that similarly named topics receive equal depth. “Not listed” means the reviewed materials did not establish that topic in the named course or track; the provider may teach it elsewhere.
| Course or track | Training fundamentals | Language models and transformers | Pretrained-model adaptation | Wider scope or important boundary |
|---|---|---|---|---|
| Karpathy | Implemented in code; manual gradients | Character models and GPT build | SFT/LoRA not listed in this syllabus | Detailed internals and a separate tokenizer build |
| 3Blue1Brown | Visual/conceptual explanations | Visual transformer and attention explanations | No fine-tuning lab established in these selected lessons | Concept checks; no end-to-end coding project established |
| fast.ai Part 1 | Applied, with notebooks | NLP applications; a from-scratch GPT sequence not established | Transfer learning; not presented here as an LLM SFT/LoRA course | Vision, tabular data, recommenders, deployment |
| Hugging Face LLM Course | Introductory deep learning recommended first | Architectures and practical model use | Fine-tuning, SFT, LoRA, evaluation | Datasets, tokenizers, NLP tasks, demos |
| Machine Learning Specialization | Structured foundations and assignments | A GPT build is not listed | LLM SFT/LoRA not listed | Broad classical ML and neural-network foundation |
| Doug Does AI: ML fundamentals | Small inspectable calculations | Separate track | Loss connects to SFT; no fine-tuning project here | Six focused topics, with optional Python practice |
| Doug Does AI: Shakespeare track | Training a small model | Counts, representations, attention, character transformer | Separate track | CPU project; does not implement a production BPE tokenizer |
| Doug Does AI: SQL track | Applied training loop; basics assumed | Uses a pretrained model | One SFT/LoRA experiment and its evaluation | Narrow single-table task; no production-readiness claim |
Combine resources without taking six courses at once
These are suggested study routes, not validated equivalence paths. Pick one main course and use the others to resolve a specific gap.
Route A: I need the ideas to click
Start with 3Blue1Brown’s neural-network and gradient-descent lessons. Then use Doug Does AI’s neuron, loss, and gradient modules to predict what should happen before moving a control. If you want a broad syllabus next, continue with the Machine Learning Specialization. If you want to implement differentiation, move to Karpathy’s micrograd lesson.
Route B: I want to build and understand a language model
Make Karpathy your main coding course. Use Doug Does AI’s counts lesson when conditional probabilities feel abstract, and the fundamentals modules when you want a smaller calculation to check. Doug Does AI’s Shakespeare project is another compact model to train and inspect. It does not replace Karpathy’s manual-backpropagation, BatchNorm, or tokenizer work. Keep those when implementation depth is your goal.
Route C: I want to adapt an existing model
If deep learning is new, begin with an introductory course such as fast.ai. Work through Hugging Face’s early chapters on models and fine-tuning, then its SFT/LoRA material. Use Doug Does AI’s SQL track as a bounded case study in separating valid output from a correct answer. Its prepared results are accessible without the Apple Silicon lab. If your goal is ecosystem fluency, keep Hugging Face as the main path.
Route D: I have 20 minutes today
Choose one question, not a course-completion target. Watch part of an explanation or open one interactive lesson. Predict an outcome, test it, and write two sentences about what changed. At the next session, try to reproduce the explanation before reopening your notes. Short sessions can support sustained study; they do not make the underlying syllabus disappear.
What feedback means in this comparison
- An answer check tells you whether a specific choice or calculation matches the authored result. Some 3Blue1Brown and Doug Does AI lessons provide these.
- Code and tests show what an implementation does. Karpathy and fast.ai provide practical work, not just something to watch.
- Graded assignments and quizzes add structured checks. Availability depends on the course and access tier.
- Community help can give you a human perspective, but response times and individual review are not guaranteed.
- Doug Does AI’s feedback responds to your short explanation in the lesson context. It is usage-limited assistance, not certification or guaranteed expert review. Other providers also offer forms of interactive help.
Before you commit
Try one representative lesson and one exercise. Check whether you can explain the result without replaying the solution. Choose the longer external option when it matches the work you want to do, and use a smaller interactive lesson when a focused experiment helps you keep going.
The comparison draws on official syllabuses, public lesson pages, and linked repositories checked on October 3, 2026. It does not include a controlled learning-outcomes study or testing of every paid feature. Prices, platform features, and course contents can change; the linked provider pages are the place to confirm current access.
For checkpoints alongside implementation work, see the Karpathy study companion.