In 6 weeks you'll read
arXiv
on release day and tell signal
from noise yourself.
Next cohort: September 15, 2026. 15 seats. Enrollment open.
ML engineer, author of Seeallochnaya: technical takes on the AI industry, no fluff.
This course is Seeallochnaya in cohort format: 6 weeks of taking LLMs apart, piece by piece, live.
fine-tune, build integrations,
but don't understand why it works.
arXiv —
and feel you missed every other paper.
Understand how LLMs actually work: transformer,
attention,
scaling laws — at the engineering level.
Tell a working agent from a slick demo.
Get how reasoning models work
and why o1 isn't magic.
See the 2026 industry picture: what Anthropic, OpenAI, and Google are doing — and why.
Architectures of GPT-2, Llama, DeepSeek — broken apart, atom by atom: tokens, training, inference.
Test your LLM choices against 15 practicing engineers in the cohort.
Next cohort: September 15, 2026. 15 seats. Enrollment open.
Join the cohort6 modules over 6 weeks
Transformer, attention, exactly how token
prediction works. Breaking down GPT-2 as a minimum-viable
architecture.
Scaling laws, emergent abilities,
RLHF, instruction tuning.
What separates each generation of models.
o1, o3, how chain-of-thought works
at training time, test-time compute,
where it actually differs from regular LLMs.
How models learn to use tools:
function calling, MCP,
where this still breaks down right now.
What's actually open and what isn't, how to read technical reports from Meta/Mistral/DeepSeek, what to look for in benchmarks — and why you shouldn't trust them.
OpenAI/Anthropic/Google capex, GPU economics,
multimodality, what “AGI” means in 2026,
what to expect in 2027.
arXiv, your choice.
3 plans — pick yours
$750
“Self-directed”
For those who learn solo. No personal homework review — only model answers. Questions go to office hours.
Includes:
$1,200
“With feedback”
Igor reviews your homework and the final project personally.
Everything in Basic, plus:
$2,200
“With personal work”
Work with Igor directly. Four 45-minute calls, a review of your AI project at work, plus a one-on-one final defense.
Everything in Standard, plus:
Yes. If you've done fine-tuning or written inference code — come in. If you've only used the ChatGPT API and never opened HuggingFace — too early, start with YSDA or karpov.
The cohort starts September 15. Everyone moves together, on one schedule. 6 weeks, 2 live sessions per week. After each — homework with a deadline, reviewed in the next session. Shared Telegram chat.
4–6 hours per week on homework plus 3 hours of sessions — 7–9 hours total. Reasonable to do alongside a job.
YSDA and Karpov teach ML from scratch — how to train models. This course is for those already working with LLMs who want to understand how they work inside and why.
Free lectures give you information — there's plenty of that already. On the Basic plan, you pay for live sessions and a 15-person cohort where you can ask Igor. From there: Standard adds homework review, Premium adds a review of your project.
Yes. All sessions are recorded and stay forever — watch later if you missed it. But homework deadlines stay on the cohort schedule — everyone moves together.
Only Igor. All 12 sessions, all office hours, homework review on Standard and Premium — Igor himself. No assistants — that's why the cohort is capped at 15.
15 seats per cohort. Starts September 15, 2026.
Application received
We'll reach out within 24 hours via your contact. We'll help you pick a plan if you're unsure.
If Telegram didn't open — message the bot @kotenkov_llm_bot and send the application: