Coding LLMs from the Ground Up: A Complete Course

Sebastian Raschka just released one of the more generous things you'll see from an independent AI researcher this year: a roughly 15-hour, seven-part video series that walks through coding a large language model from the ground up. No shortcuts, no black-box libraries doing the heavy lifting. You set up the Python environment yourself with uv, build tokenization and data loaders, hand-code self-attention and multi-head attention, assemble the actual transformer architecture, pretrain it, and then fine-tune it twice — once for classification with a spam detector, once for instruction following.

The series grew out of supplementary material for his book Build a Large Language Model (From Scratch), but Raschka says it now works as a standalone course, and it's about five times longer than the abbreviated workshop he put out last year that people apparently loved. His reasoning for why this matters is a car analogy, and it's a decent one: you don't learn how engines work by starting with a Formula 1 car. You start with a go-kart. Schumacher and Senna both tinkered with go-karts before they ever sat in anything fast, and that hands-on feel for steering and torque is exactly what made them useful to their pit crews later. Building a small LLM from scratch is the go-kart. You're not training the next GPT-5, you're learning why attention layers exist and what a loss curve actually looks like when pretraining starts to work.

There's a real-life backstory here too. Raschka mentions, almost in passing, that he's dealing with a serious neck injury and hasn't been able to work at a computer for three weeks while he tries conservative treatment before possibly facing surgery. The timing, he says, is brutal — he'd just gotten back into a writing rhythm after a run of reasoning-model articles. So this release is partly recycled footage he'd already recorded, repackaged as something useful to share while he can't produce new written work. As a bonus for paid subscribers, he's also throwing in a 2.5-hour non-coding talk recorded right after the Llama 4 launch, comparing the state of LLMs in 2025 to where things stood back at GPT-2 in 2018.

What stands out is how unglamorous the pitch is. This isn't "build a startup with LLMs" content. It's closer to a electronics kit for adults who want to actually understand the wiring instead of just plugging things in.