Lesson 6 — Recap and FAQ
The five lessons in one page
Lesson 1 — the problem. Some tasks cannot be written as rules, because we cannot state the rules we ourselves use. Machine learning sidesteps this by deriving the rules from labelled examples. It only became practical around 2012, when large datasets and graphics processors finally arrived together.
Lesson 2 — the definition. AI is any technique letting a machine derive its behaviour from data instead of hand-written instructions. Everything deployed today is narrow AI: excellent at one task, useless outside it. No model understands what it does, which is fine when a mistake is cheap and visible and dangerous when it is not.
Lesson 3 — the three circles. AI is the goal, machine learning is the dominant method, deep learning is the branch that handles raw data by discovering its own features. Use deep learning on images, sound and text; start with a tree ensemble on tables, where it usually wins anyway.
Lesson 4 — the mechanism. Training scores the model's mistakes as a number and nudges every parameter downhill against that score, millions of times. Too few examples let the model memorise instead of generalise, and that failure is only visible if you hold back a test set.
Lesson 5 — reality. AI pays off on narrow, high-volume tasks with reliable labels. Generative AI works as assistance with review, not autonomy. Data availability, tolerance for probabilistic answers and a measurable target decide a project before any algorithm does.
Glossary of terms you will meet everywhere
| Term | What it means, plainly |
|---|---|
| Model | a function with adjustable numbers, tuned by training to map inputs to outputs |
| Parameters / weights | those adjustable numbers; a large language model has hundreds of billions |
| Training | the loop that adjusts the parameters to reduce mistakes on the examples |
| Inference | using the trained model to get a prediction; cheap and fast |
| Feature | one input variable the model reads, such as "customer tenure in months" |
| Label | the correct answer attached to a training example |
| Loss | the number that scores how wrong a prediction is |
| Epoch | one full pass through the whole training dataset |
| Overfitting | memorising the examples instead of learning the pattern |
| Drift | the world changing after training, so the frozen model gets steadily worse |
| Hallucination | a generative model producing fluent, confident, false content |
| Fine-tuning | taking a model trained by someone else and adapting it to your data |
| Transformer | the architecture behind modern language models, introduced in 2017 |
| GPU | the parallel processor that makes training deep networks feasible |
The 14 questions people actually ask
1. Do I need a PhD to work in AI?
No. A PhD is close to necessary for research — inventing new architectures, publishing. The overwhelming majority of AI jobs are not research: they are applying known methods to a company's data, which demands solid engineering, good data handling and domain understanding. Many productive practitioners came from software development or data analysis.
2. How much mathematics do I really need?
Less than the internet implies, more than zero. To use libraries competently you need intuition for three areas: linear algebra (a vector is a list of numbers, a matrix transforms them), probability (a prediction is a distribution, not a verdict) and calculus (a derivative is a slope, which is how gradient descent works). You need to manipulate proofs only if you intend to do research. The Mathematics for AI course teaches exactly this intuition.
3. Which language should I learn?
Python, without hesitation. Not because it is the most elegant language, but because the entire ecosystem is built in it: NumPy, pandas, scikit-learn, PyTorch, TensorFlow. Choosing another language means reimplementing tools that already exist and losing access to every tutorial and paper implementation. Python for AI starts there.
4. Will AI replace my job?
It will more likely change its content than remove the role. The tasks most exposed are repetitive and well-documented: standard transcription, first-draft writing, routine boilerplate code. What resists is deciding what should be done, arbitrating between conflicting constraints, and being accountable for the result. The practical response is to learn to work with these tools rather than compete with them at what they are good at.
5. Is AI dangerous?
The realistic risks are not the cinematic ones. They are, in rough order of how much damage they currently cause: automated discrimination by models that faithfully reproduce historical bias; decisions nobody can explain or contest; mass surveillance made cheap by facial recognition; large-scale disinformation from cheap synthetic media; and misplaced confidence in systems whose failure mode is a confident wrong answer. All five are present-tense problems, and all five are governance problems as much as technical ones. The AI Ethics course covers them.
6. Does ChatGPT learn from my conversations?
Not during the conversation itself. The model's parameters are frozen at inference: it does not update as you talk. What creates the impression of learning is the context window, which holds the current conversation and lets the model refer back to it. Once the conversation ends, that context is gone. Providers may separately use conversations to train future versions, which is a privacy question rather than a live-learning one.
7. Why does AI invent facts?
Because a language model is optimised to produce likely text, not true text. It has no internal database to check against and no representation of its own uncertainty. When the information is missing, the most likely continuation is a well-formed sentence, and a well-formed sentence containing a fabricated citation is more likely than an admission of ignorance. Retrieval techniques reduce this substantially by supplying real source documents, without eliminating it.
8. What is the difference between AI and automation?
Automation executes a process a human specified: if this happens, do that. It is entirely predictable and it never improves on its own. AI derives the process from data, handles cases nobody enumerated, and returns probabilities rather than certainties. A rule that blocks any transaction over ten thousand is automation. A model that scores a transaction as 3% likely to be fraudulent based on its shape is AI.
9. How long to become employable?
With serious, consistent effort, six to twelve months is realistic to reach junior level: Python, data handling, classical machine learning, two or three finished projects you can explain end to end. Portfolio projects carry more weight than course certificates, because they show you handled real data and made decisions.
10. Free tools or paid ones?
Everything you need to learn is free. Python, scikit-learn, PyTorch and TensorFlow are open source. Google Colab lends you a GPU in a browser at no cost. Hugging Face hosts hundreds of thousands of pre-trained models. Nothing needs to be bought before you are training models large enough to justify renting hardware.
11. Can AI be creative?
It recombines what it has seen in ways that are frequently novel and sometimes genuinely surprising. Whether that is creativity depends on your definition. What it demonstrably lacks is intention: it does not decide to make something, does not know why one variation is better, and has no stake in the result. It is a formidable instrument in the hands of someone who does.
12. Why do models need such enormous energy?
Training a large model means performing an astronomical number of operations across thousands of processors for weeks. That is a one-off cost. Inference is cheap per request but happens billions of times, so at scale the total often exceeds the training cost. Smaller specialised models, covered in the premium catalogue, are frequently the sensible answer: a compact model fine-tuned for one task can match a giant one on that task at a fraction of the cost.
13. What is the difference between open-source and proprietary models?
Proprietary models are accessed through an API: no infrastructure to run, usually the best raw performance, and your data leaves your premises. Open-weight models can be downloaded and run on your own hardware: full data control, no per-request cost, and you carry the operational burden. The choice is usually driven by confidentiality and volume rather than by benchmark scores.
14. Where should I go after this course?
In this order: Python for AI to get the tooling, Mathematics for AI for the intuition, then Machine Learning to build your first real models. After that, follow your interest — Deep Learning, NLP, Computer Vision or LLMs.
This discovery course gave you the mental model. The premium AI catalogue gives you the practice: notebooks, real datasets, projects, and a verifiable certificate after a 40-question examination. Included in every paid plan.
Last step — Take the quiz and see your attestation →