https://abxda.com/academy/en/ Free courses to really understand technology. en CC BY-NC 4.0 · Abel Coronado academy@abxda.com (abxda academy) Sat, 17 Oct 2026 12:00:00 GMT https://abxda.com/academy/en/generative-ai/lessons/programming-vs-learning/ https://abxda.com/academy/en/generative-ai/lessons/programming-vs-learning/ Sun, 04 Oct 2026 12:00:00 GMT Learning from data What changes when a machine learns instead of following rules? Programming is writing rules: rules + data → answers. Learning flips the flow: data + answers → rules. (12 min) https://abxda.com/academy/en/generative-ai/lessons/the-world-as-numbers/ https://abxda.com/academy/en/generative-ai/lessons/the-world-as-numbers/ Sun, 04 Oct 2026 12:00:00 GMT Learning from data How does a machine see an image, a sound, or a text? A model only receives numbers. A grayscale image is a table of brightness values from 0 (black) to 255 (white); in color it is three tables, one per channel (red, green, blue). (10 min) https://abxda.com/academy/en/generative-ai/lessons/a-model-and-its-error/ https://abxda.com/academy/en/generative-ai/lessons/a-model-and-its-error/ Sun, 04 Oct 2026 12:00:00 GMT Learning from data How do you measure how wrong a model is? A straight line rises (or falls) at a constant pace. Its weight w, the slope, is how much the prediction rises per unit of input; its bias b is the height when the input is zero. The prediction is w·x + b. (11 min) https://abxda.com/academy/en/generative-ai/lessons/gradient-descent/ https://abxda.com/academy/en/generative-ai/lessons/gradient-descent/ Sun, 04 Oct 2026 12:00:00 GMT Learning from data How does a machine find the right numbers without being told? Training means searching for the weight that leaves the error lowest. The machine cannot see the whole curve: it only knows the error and the slope where it stands. (11 min) https://abxda.com/academy/en/generative-ai/lessons/the-artificial-neuron/ https://abxda.com/academy/en/generative-ai/lessons/the-artificial-neuron/ Sun, 04 Oct 2026 12:00:00 GMT Neural networks What exactly does an artificial neuron do? An artificial neuron computes y = f(w₁x₁ + w₂x₂ + b): it multiplies each input by its weight, adds them, adds the bias, and applies an activation function. (11 min) https://abxda.com/academy/en/generative-ai/lessons/what-a-neuron-cannot-do/ https://abxda.com/academy/en/generative-ai/lessons/what-a-neuron-cannot-do/ Sun, 04 Oct 2026 12:00:00 GMT Neural networks Why is a single neuron not enough? XOR is not linearly separable: its two diagonals cross, so no line gets all four corners right. (6 min) https://abxda.com/academy/en/generative-ai/lessons/layers-and-activations/ https://abxda.com/academy/en/generative-ai/lessons/layers-and-activations/ Sun, 04 Oct 2026 12:00:00 GMT Neural networks How do networks gain power by stacking layers? A layer is a group of neurons that look at the same inputs; the forward pass computes layer by layer, from input to output. (10 min) https://abxda.com/academy/en/generative-ai/lessons/backpropagation/ https://abxda.com/academy/en/generative-ai/lessons/backpropagation/ Sun, 04 Oct 2026 12:00:00 GMT Neural networks How is the blame for an error shared among thousands of weights? A network’s final error is shared among thousands of weights hidden in layers. Knowing how much blame each one carries is the credit assignment problem. (9 min) https://abxda.com/academy/en/generative-ai/lessons/training-well/ https://abxda.com/academy/en/generative-ai/lessons/training-well/ Sun, 04 Oct 2026 12:00:00 GMT Neural networks Which training decisions separate a network that learns from one that does not? The optimizer turns the slope into a step. SGD uses it as is; momentum adds inertia; Adam also gives each weight its own step size. Each one wants its own rate. (12 min) https://abxda.com/academy/en/generative-ai/lessons/generalizing-and-evaluating/ https://abxda.com/academy/en/generative-ai/lessons/generalizing-and-evaluating/ Sun, 04 Oct 2026 12:00:00 GMT Neural networks How do you know whether a model really works? Lowering training error is not enough: what matters is error on data the model has not seen. (9 min) https://abxda.com/academy/en/generative-ai/lessons/convolutions-and-vision/ https://abxda.com/academy/en/generative-ai/lessons/convolutions-and-vision/ Sun, 04 Oct 2026 12:00:00 GMT Deep learning How does a network learn to see edges, shapes, and objects? Connecting every pixel to every neuron does not scale: a small image already needs millions of weights. (6 min) https://abxda.com/academy/en/generative-ai/lessons/sequences-and-memory/ https://abxda.com/academy/en/generative-ai/lessons/sequences-and-memory/ Sun, 04 Oct 2026 12:00:00 GMT Deep learning How does a network process something that happens in order, like a sentence? A sequence is data whose order matters and whose length varies. A recurrent network reads it step by step with the same cell and passes a hidden state, its memory, from one step to the next. (10 min) https://abxda.com/academy/en/generative-ai/lessons/the-deep-learning-big-bang/ https://abxda.com/academy/en/generative-ai/lessons/the-deep-learning-big-bang/ Sun, 04 Oct 2026 12:00:00 GMT Deep learning Why did in 2012 what did not work in 1990? The takeoff of deep learning did not come from a new idea, but from three ingredients that coincided: massive data, compute, and training tricks. (10 min) https://abxda.com/academy/en/generative-ai/lessons/embeddings-meaning-as-geometry/ https://abxda.com/academy/en/generative-ai/lessons/embeddings-meaning-as-geometry/ Sun, 04 Oct 2026 12:00:00 GMT Deep learning How does a machine turn meaning into distances? An embedding is a dense, learned vector: a row of a table tuned by data, instead of a fixed slot per word. (11 min) https://abxda.com/academy/en/generative-ai/lessons/generating-is-learning-a-distribution/ https://abxda.com/academy/en/generative-ai/lessons/generating-is-learning-a-distribution/ Mon, 12 Oct 2026 12:00:00 GMT Generating images What does it mean for a network to "generate" something new? Classifying means drawing a boundary between groups. Generating asks for more: knowing the distribution of the data, where it crowds and how far it reaches, to draw new samples. (10 min) https://abxda.com/academy/en/generative-ai/lessons/generative-adversarial-networks/ https://abxda.com/academy/en/generative-ai/lessons/generative-adversarial-networks/ Sun, 11 Oct 2026 12:00:00 GMT Generating images How does a generator learn by competing against a critic? A GAN is two competing networks: the generator turns a randomly drawn number into a fake data point; the discriminator estimates how likely a data point is to be real. (11 min) https://abxda.com/academy/en/generative-ai/lessons/diffusion-models/ https://abxda.com/academy/en/generative-ai/lessons/diffusion-models/ Mon, 12 Oct 2026 12:00:00 GMT Generating images How is an image created by removing noise step by step? The forward process mixes an image with noise, step by step, on a fixed schedule, until only static is left. It learns nothing, and in the end every image ends up as the same kind of static, which is easy to make. (12 min) https://abxda.com/academy/en/generative-ai/lessons/from-text-to-image/ https://abxda.com/academy/en/generative-ai/lessons/from-text-to-image/ Sun, 11 Oct 2026 12:00:00 GMT Generating images How does the model understand what you ask for in words? CLIP trains an image encoder and a text encoder together. Both output vectors of the same size, compared with cosine similarity. (13 min) https://abxda.com/academy/en/generative-ai/lessons/predicting-the-next-word/ https://abxda.com/academy/en/generative-ai/lessons/predicting-the-next-word/ Sun, 11 Oct 2026 12:00:00 GMT Transformers and language How can something that only predicts words write an essay? A language model gives, for a context, a probability to every word (or piece) that can come next. (11 min) https://abxda.com/academy/en/generative-ai/lessons/tokens/ https://abxda.com/academy/en/generative-ai/lessons/tokens/ Sun, 11 Oct 2026 12:00:00 GMT Transformers and language What does a language model actually read, and why does it matter? A language model reads neither letters nor words: a tokenizer cuts text into tokens, pieces from a fixed vocabulary, and only each piece’s number reaches the model. (11 min) https://abxda.com/academy/en/generative-ai/lessons/attention/ https://abxda.com/academy/en/generative-ai/lessons/attention/ Mon, 12 Oct 2026 12:00:00 GMT Transformers and language How does a word decide which other words to look at? A word can mean different things (“bank”), yet it enters the model with a single vector. To understand it you have to look at the others: that is what attention does. (12 min) https://abxda.com/academy/en/generative-ai/lessons/the-transformer/ https://abxda.com/academy/en/generative-ai/lessons/the-transformer/ Mon, 12 Oct 2026 12:00:00 GMT Transformers and language What is inside an LLM, block by block? A transformer always follows the same path: tokens, numbers, embeddings plus position, a stack of blocks, and, from the last token’s list, one score per word that softmax turns into probabilities. (12 min) https://abxda.com/academy/en/generative-ai/lessons/pretraining/ https://abxda.com/academy/en/generative-ai/lessons/pretraining/ Mon, 12 Oct 2026 12:00:00 GMT Scale What does a model learn from reading half the internet? Pretraining repeats a single exercise: cover the next piece of a text and guess it. Since the answer is already in the text. (13 min) https://abxda.com/academy/en/generative-ai/lessons/scaling-laws-and-costs/ https://abxda.com/academy/en/generative-ai/lessons/scaling-laws-and-costs/ Mon, 12 Oct 2026 12:00:00 GMT Scale How much does a model cost and what do you get by making it bigger? Scaling a model means turning three knobs: more parameters (N), more training tokens (D), and more compute (C). (13 min) https://abxda.com/academy/en/generative-ai/lessons/gpus-memory-and-quantization/ https://abxda.com/academy/en/generative-ai/lessons/gpus-memory-and-quantization/ Tue, 13 Oct 2026 12:00:00 GMT Scale Why does GPU memory rule over everything else? A model works out of the GPU’s memory, its VRAM. It takes parameters × bytes per parameter: 70 billion in FP16 is 140 GB, more than an H100’s 80. (13 min) https://abxda.com/academy/en/generative-ai/lessons/inference-how-a-model-answers/ https://abxda.com/academy/en/generative-ai/lessons/inference-how-a-model-answers/ Mon, 12 Oct 2026 12:00:00 GMT Scale What happens, and what does it cost, between typing and the answer appearing? An answer has two phases. Prefill processes your whole message at once and sets the wait until the first token; decode writes one token per turn and sets how long the rest takes. (13 min) https://abxda.com/academy/en/generative-ai/lessons/fine-tuning-and-lora/ https://abxda.com/academy/en/generative-ai/lessons/fine-tuning-and-lora/ Mon, 12 Oct 2026 12:00:00 GMT From model to assistant When should you adapt a model, and how is it done cheaply? Fine-tuning keeps training an already trained model on pairs of input and desired answer: it is the same loop as in lesson 8 (answer, measure the error, backpropagate, adjust), with far less data. (11 min) https://abxda.com/academy/en/generative-ai/lessons/aligning-assistants/ https://abxda.com/academy/en/generative-ai/lessons/aligning-assistants/ Mon, 12 Oct 2026 12:00:00 GMT From model to assistant What turns a text predictor into a useful assistant? A base model only continues text: it was trained to guess the next piece, not to serve a person. Aligning it means getting it to do what the person intends: be helpful, honest, and harmless. (14 min) https://abxda.com/academy/en/generative-ai/lessons/reasoning-models/ https://abxda.com/academy/en/generative-ai/lessons/reasoning-models/ Mon, 12 Oct 2026 12:00:00 GMT From model to assistant Why does thinking longer before answering improve the answer? A language model writes one token per pass. If it answers right away, the whole calculation must happen inside; if it writes out the steps (a chain of thought), each result stays as context and each token adds a pass of computation. (12 min) https://abxda.com/academy/en/generative-ai/lessons/context-and-memory/ https://abxda.com/academy/en/generative-ai/lessons/context-and-memory/ Mon, 12 Oct 2026 12:00:00 GMT From model to assistant Does the model remember what you said? Does it learn while you chat? The model is stateless: on every turn the app resends it the whole conversation. Its weights do not change while you chat. (12 min) https://abxda.com/academy/en/generative-ai/lessons/rag-read-before-answering/ https://abxda.com/academy/en/generative-ai/lessons/rag-read-before-answering/ Mon, 12 Oct 2026 12:00:00 GMT From model to assistant How do you give a model your own knowledge without retraining it? RAG does not retrain the model: with every question it looks up the chunks of your documents that are most similar and pastes them into the prompt, along with the question. (13 min) https://abxda.com/academy/en/generative-ai/lessons/agents-and-tools/ https://abxda.com/academy/en/generative-ai/lessons/agents-and-tools/ Mon, 12 Oct 2026 12:00:00 GMT From model to assistant What changes when a model can act, not just talk? A model only writes text. To act, it is given tool descriptions and writes function calls (a name and data); the program around it runs them and pastes the result into its context. (13 min) https://abxda.com/academy/en/generative-ai/lessons/hallucinations-and-limits/ https://abxda.com/academy/en/generative-ai/lessons/hallucinations-and-limits/ Mon, 12 Oct 2026 12:00:00 GMT Using AI with judgment Why is it wrong with such confidence, and what can be done? A hallucination is a fluent, confident answer that is false or unsupported. It is not an occasional breakdown: it follows from how a language model is trained. (12 min) https://abxda.com/academy/en/generative-ai/lessons/evaluating-ai-systems/ https://abxda.com/academy/en/generative-ai/lessons/evaluating-ai-systems/ Tue, 13 Oct 2026 12:00:00 GMT Using AI with judgment How do you check that an AI system works for your case? A benchmark is a fixed public exam. It is useful for comparing models in general, but it saturates and can be contaminated: its score does not tell you whether the system works for your case. (12 min) https://abxda.com/academy/en/generative-ai/lessons/risks-and-governance/ https://abxda.com/academy/en/generative-ai/lessons/risks-and-governance/ Mon, 12 Oct 2026 12:00:00 GMT Using AI with judgment What can go wrong, and who is accountable when it does? Prompt injection has no complete fix. It does harm when private data, content from strangers, and a way out come together: remove one of the three or add human approval before acting. (13 min) https://abxda.com/academy/en/generative-ai/lessons/strategy-for-organizations/ https://abxda.com/academy/en/generative-ai/lessons/strategy-for-organizations/ Mon, 12 Oct 2026 12:00:00 GMT Using AI with judgment Where to apply generative AI in your organization, and how to start? A use case is a specific task, not “use AI.” It is ranked with three questions: how often it happens, what each one is worth, and how risky it is. Something small and frequent can be worth more than something big and rare. (12 min) https://abxda.com/academy/en/generative-ai/lessons/the-complete-map/ https://abxda.com/academy/en/generative-ai/lessons/the-complete-map/ Mon, 12 Oct 2026 12:00:00 GMT Wrap-up How do all the pieces fit together, from neuron to chatbot? Each lesson grew out of another’s limit: XOR called for hidden layers; the short memory of networks that read in order called for attention; the base model, which only continues text, called for alignment. (11 min)