About the course
Neural networks are already part of work and life. Yet almost nobody knows what is inside. We show it in plain language, then you build your own model from scratch and stop depending on a single service.
Why it matters
Everything you spent months telling the neural network about yourself and your work goes with it. Someone who understands how models work calmly moves to another one.
You will see why a neural network is confidently wrong and when its answer cannot be trusted.
Every technology company is looking for people who can build and fine-tune models. The course takes you to interview level.
Outcome
Not a certificate but skills you can show: code, a model, and answers to interview questions.
Weeks 5–8
How text becomes tokens, what attention is, and why the model writes the way it does.
Weeks 8–12
A small language model from an empty file: you write the code, tests check that it is correct.
Weeks 15–17
Teach an open model to answer in the style you need and on your data.
Week 21
An assistant that answers from your files and says honestly where it found the answer.
Weeks 9, 13, 14
How much memory, how many GPUs and how much money it takes to train and run a model.
Weeks 23–24
A question trainer, answer reviews, and preparation for the salary conversation.
The app
Real screens from the web version, in the same theme as the site. The interface supports six languages. The iPhone version is being prepared for release. Shown here with the Russian interface.
Method
Every topic takes three steps. Without the derivation the code makes no sense, and without the code an interview answer falls apart at the first follow-up question.
01
You derive every formula yourself, on paper. Hard ideas start with a plain walkthrough on small numbers, then the formula. There are 130 such walkthroughs and 23 diagrams.
02
A language model from scratch: autograd, tokenizer, transformer, KV cache, LoRA, RL. You do the exercises without AI, and they are checked by the same tests as the reference implementation.
Why divide q·k by √d in attention?
03
Interview-level questions with a timer and answer checking, rubrics for six interview formats, a question bank, and a tutor grounded in the course. The tutor asks questions instead of handing out answers.
No sign-up
Three questions from the entry diagnostic, one each on Python, tensor shapes and derivatives. The app has 24, and your answers tell it which week to start from. Nothing you answer leaves this page.
The questions on this page need JavaScript. The full diagnostic is in the app: app.snulya.com/diagnostic
24 weeks
Six phases, 20–25 hours a week. If you only have 10 hours, stretch it to 48 weeks and keep the order. Module 0 comes first for those who need the foundations, and a Capstone comes after.
Hover over a week or tap it to see the topic.
Module 0 based on the diagnostic
1 Foundations
2 Transformer
3 Inference, GPUs
4 Post-training
5 Beyond
6 Interviews
Capstone 4–6 weeks
Внимание
checkpoint: test yourself at the end of a phase
Inside
Tutor
The tutor answers from the course materials and points to the right week. In Socratic mode it guides you with questions until you reach the answer yourself. It does not write exercise solutions: you write them, and the tests check them.
tutor · week 7 · Socratic mode
Pricing
The course text is open under CC BY-NC-SA, nanolm under Apache-2.0. You pay for the tutor and answer checking on our key, spaced repetition of problems, and the certificate.
$0
Everything for self-study
790 ₽ per month
Or $149 a year
Payment for Pro is not connected yet. Free once you sign in: 30 tutor messages and 5 answer checks
$79 one-time
Six weeks before your interviews
Payment is not connected yet
Members join with an invite code. In the dashboard, leads see who is on which week, which skills have reached the target level, and how the checkpoints went.
$39 per seat per month
10–50 seats, billed yearly. Pro for everyone, a group progress dashboard, a shared quota.
$2,500 per cohort
Up to 30 students for 6 months. Pro for everyone, an instructor dashboard: weeks, skills, checkpoints.
from $15,000 per year
Seats by agreement, no limit on tutor messages, and you can connect your own model key.
Prices in US dollars. Payment is not available yet. The course, the problems and the tutor on your own key are free, and once you sign in you can try the tutor on our key for free.
The core needs derivatives of composite functions, matrices and basic probability: from week one you compute gradients by hand. If that has faded, Module 0 covers linear algebra, calculus and probability from scratch, with small-number examples.
The core is planned for 20–25 hours a week, Module 0 for 15–20. If you only have 10 hours, stretch the core to 48 weeks. Keep the order of topics: each week builds on the previous ones.
Answer all 24 questions. Starting at week 1 requires at least 80% overall and 75% in every block. At 50% or above, work on weak Module 0 blocks; below 50%, complete F1–F6. An unfinished diagnostic does not establish readiness for the core.
Yes, and that is the free plan. Paste an OpenRouter or Anthropic key in settings: tutor requests go through your key and your account. OpenRouter also has free models. Everything else in the app works with no key at all.
The whole course is available in English and Russian: the program, Module 0, the tracks, the exercises and the tutor. The app interface comes in six languages: Russian, English, German, Spanish, Portuguese and Chinese. The tutor replies in the language you ask in. Terms like attention, backprop and KV cache stay in English, as in interviews.
No. The nanolm exercises are done without AI, otherwise you never learn how much you lean on hints. The tutor explains the topic and asks a leading question, and the tests show whether your solution is right.
Payment is not connected yet. Materials, problems, local progress and the tutor on your own key are free. Once you sign in, you can try 30 tutor messages and 5 answer checks on our key. Extended reviews and the certificate belong to Pro and are not available for purchase yet.
