From tokens and attention to tuned models you can run yourself
Generative AI turns text into a programmable surface: given the right tokens, attention weights and training stages, a neural network can summarise, classify, retrieve, converse and look at images. This collection walks that machinery end to end, from tokenisation and embeddings through the transformer block, pretraining, supervised and parameter-efficient fine-tuning, prompt engineering, retrieval-augmented generation and multimodal models. It is laid out as 63 short self-contained pieces, so you can begin at retrieval or at preference tuning instead of working through everything in order. By the end you can load a model, prompt it deliberately, ground it in your own documents, tune it on your own data and judge whether the output is worth trusting.
Coming soon
Where these models came from, what they are good at, and how to reach one.
Where these models came from, what they are good at, and how to reach one.
The pieces a language model is made of, from raw text to a finished forward pass.
The pieces a language model is made of, from raw text to a finished forward pass.
Classifier and clustering work you can do with a model you never train.
Classifier and clustering work you can do with a model you never train.
Getting reliable text out of a model, and wiring it into something larger.
Getting reliable text out of a model, and wiring it into something larger.
Search, retrieval and images — stretching a text model beyond text.
Search, retrieval and images — stretching a text model beyond text.
Training a vector space that fits your own documents.
Training a vector space that fits your own documents.
Turning an encoder into a task-specific tagger or classifier.
Turning an encoder into a task-specific tagger or classifier.
Supervised tuning, parameter-efficient tuning and preference alignment.
Supervised tuning, parameter-efficient tuning and preference alignment.
“The clearest on-ramp from using language models to understanding them.”
It keeps the rigour of a full course text and drops the semester-shaped padding. Each article stands alone, so a reader can start at retrieval or at fine-tuning without working through everything before it. The result is a route into generative AI that rewards curiosity rather than endurance.
Nano College takes the standard university textbook for a course and turns it into short, self-contained pieces you can actually finish. We keep the ground the textbook covers and the rigour it insists on; what we leave out is the padding that only exists to fill a semester.