Five Lines is a free learning site that explains technical concepts in five lines of simple English. A track is one whole field worked out in order: broken into modules, every concept in reading order, a graded test at the end of each module. The AI and machine learning track is written and free to read now. Web Development, Cybersecurity, Data Science and Systems are next.
Attention lets a model decide which parts of the input matter for each output.
Instead of one fixed summary it computes a weighted blend of all positions.
Positions that are relevant get a high weight, the rest get a low one.
This removed the fixed-vector bottleneck of old seq2seq models.
It is the single idea the entire transformer architecture is built around.
Most material fails in one of two directions: it drowns you before you know what a loss function is, or it hand-waves so hard you finish unable to say what a transformer changed. A track is the third option: the whole field, in the order the ideas actually depend on each other.
A track is every concept in its field, ordered so each one stands on the last. AI / ML is 356 of them across 17 modules, from what a dataset is to reproducing a research paper. Name a term anywhere on the site and there is a card that teaches it. The build fails otherwise.
Five lines per concept, read in under a minute. You are never handed a forty-page chapter and asked to work out which paragraph mattered.
Every module ends in a graded test and every answer is explained, right or wrong. The report names the modules you actually know rather than the ones you have read. It is per track, so finishing one tells you nothing about the next.
A track is a whole technical field broken into modules, each module a list of concepts, each concept five lines. AI / ML is written and free to read now; the rest are in the same format and are not written yet.
From what a dataset is to reproducing research papers.
The browser, the server, and everything that has to agree between them.
How systems are attacked, and what actually stops it.
Getting a defensible answer out of messy data.
What the machine is really doing underneath the language you write in.
This is one track, free and unlocked, end to end. Every module ends in a graded test; three cumulative checkpoints and a final sit across all of them. Every track that follows is built to this shape.
How machines learn from data, the main algorithm families, and how to measure whether any of it worked.
Neural networks from a single neuron up to the training tricks that make deep models actually converge.
Getting machines to read images: convolution, CNN architectures, detection, segmentation and vision transformers.
Turning language into numbers, and the recurrent architectures that handled sequences before transformers arrived.
The attention mechanism piece by piece, and how the encoder, decoder and decoder-only stacks are assembled from it.
What a large language model actually does at each step, and the sampling knobs that shape what comes out.
Structuring what you send a model so the output is reliable, parseable and hard to hijack.
Representing meaning as vectors, and the index structures that find the nearest ones fast enough to matter.
Grounding a model in your own documents: the full pipeline from raw files through to a cited answer.
Changing the weights rather than the prompt, and knowing when that is actually the right lever.
Models that take actions in a loop: tools, planning, memory, and knowing when to hand control back to a person.
Models that create rather than classify: autoencoders, GANs, diffusion, and the multimodal systems built on them.
Deciding whether the thing you built actually works, especially when there is no single correct answer.
Everything between a working notebook and a service other people depend on: shipping, tracking, watching and paying for it.
How AI systems get attacked through their inputs, their data and their tools, and what actually stops it.
Reinforcement learning through to RLHF, plus the compression and systems tricks that make frontier models runnable.
Moving from consuming the literature to producing it: reading, reproducing, experimenting and publishing.
Every card carries a level, and the picker in the app narrows both the deck and the tests to match. It is a filter you can change at any time. All three are free, and the advanced cards are not held back for anybody.
The high-level basics. Read only these, end to end, and you still come away with a correct mental model of the whole field.
Adds how things actually work and the trade-offs you meet in practice. This is the one most people want.
Everything, including internals, optimisation, failure modes and the research-level material. Plus the cumulative checkpoint tests.
Five lines, then the next five. Read a handful and you will know within a few minutes whether this suits how you learn. Nothing asks you to sign in to find out.