Never trained a model? Start here. One road takes you from how Python runs your code to training, fine-tuning and grounding your own GPT: the skills AI engineering roles ask for, built from first principles.
239 lessons · 928 tested code examples · foundations free, no card
Pick a path
Prefer a goal-based route? Each path is the same lessons, ordered around the role you want.
1. Where are you today?
2. Where do you want to be?
01 · 3 paths
Start here if you are new to coding, data or the maths.
02 · 3 paths
How models learn, from classic ML to deep learning and vision.
03 · 3 paths
What frontier labs build: LLMs, RL and research.
04 · 5 paths
Put AI into products and keep it fast, reliable and cheap.
Looking for one topic? Search the Library
Or build one project · Build Atlas
Nine parts, in order. Each adds a working piece to Atlas, an AI assistant you build from scratch.
Beyond the courses
Show, don't tell
Every example in every lesson is run before it's published, and the numbers in the text are the numbers it prints. No pseudo-code, no “left as an exercise”.
Fine-tune an 8B model by training half a percent of it.
d, ffn, kv, r = 4096, 14336, 1024, 16 # Llama-3-8B
shapes = [(d, d), (d, kv), (d, kv), (d, d),
(d, ffn), (d, ffn), (ffn, d)]
lora = 32 * sum(r * (i + o) for i, o in shapes)
share = lora / 8.03e9
print(f"trains {lora/1e6:.0f}M of 8,030M = {share:.2%}")trains 42M of 8,030M = 0.52%
Why a single modest GPU can customise a large model.
Inside a lesson
A real section from Build an LLM from Scratch.
What should the loss be at step 0, and why does it matter?
An untrained model should spread its bets evenly over the vocabulary, so the starting loss should be ln V. If it's far higher, the model starts confidently wrong.
for std in [1.0, 0.02]:
emb = torch.randn(V, d) * std
logits = emb[tokens] @ emb.T
loss = F.cross_entropy(logits.view(-1, V),
targets.view(-1))
print(f"init std {std}: starting loss "
f"{loss:.2f} | ln(V) = {math.log(V):.2f}")init std 1.0: starting loss 128.26 | ln(V) = 4.53 init std 0.02: starting loss 4.53 | ln(V) = 4.53
Write overfit_one_batch(model, x, y).
A healthy model must memorize a single batch. If it can't, the plumbing is broken. The worked solution drives the loss from 3.90 to 0.003.
Open this lessonBuilt from first principles by AI engineers · About the platform
Plans
The foundations are free forever, no card. Pro unlocks every other lesson, plus each new lesson as it's published.
Secure payments by Razorpay · prices in INR, GST included where applicable · plan details
FAQ
Software engineers who want to understand and build AI systems, not just call an API. You should be comfortable programming; Domain 1 starts from how Python actually runs your code, so no ML background is needed.
No. Every code example runs on an ordinary laptop CPU in seconds to a couple of minutes. Lessons explain where GPUs change the picture.
Most lessons are a 10–25 minute read, plus time for the self-check questions and the exercise. Work through them in order: each lesson lists what it builds on.
All of the foundations: the full Python track plus the CS and systems lessons, with no card and no sign-in required to read. Unrolled Pro unlocks everything after them, including the rest of Build Atlas.
Not yet: 239 lessons are written, covering the foundations and Parts 1–7 of Build Atlas. Agents and serving (Parts 8 and 9) are being written now. Pro includes every new lesson as it is published, at no extra cost.
Yes, anytime directly with one click and your plan won't renew. You keep access until the end of the period you've paid for.
One road from arrays of numbers to a GPT you train, fine-tune and ground in documents, plus paths for specific goals. Python and the foundations are free, with no card.