
Already know some of this? A 2-minute test skips the lessons you know.
Take the placement testA free 7-lesson course built from Andrej Karpathy's 60m of YouTube video. 1 day at 90 minutes a day, with study notes and a quiz in every lesson. A 2-minute placement test skips what you already know.
Taught by Andrej Karpathy on YouTube. Videos play in YouTube's own player, so the creator keeps every view.
Understand that an LLM like Llama 2 70B consists of just two files—a 140GB parameters file (weights) and a ~500-line C inference script—and runs locally without internet.
Describe the core next-token prediction mechanism and how it forces the model to learn world knowledge, while noting the 'reversal curse' shows knowledge access is imperfect and one-directional.
Summarize how comparison-based labeling (RLHF) and detailed instruction manuals improve assistants, and how human–model collaboration shifts labeling effort toward oversight.
Contrast current 'System 1' fast token generation with the research goal of 'System 2' slow, deliberate reasoning (tree-of-thoughts) to convert compute into accuracy.
Outline the LLM-as-OS vision: base model as kernel, context window as RAM, tools as peripherals, fine-tuned experts as apps, with proprietary and open ecosystems mirroring Windows vs. Linux.
Categorize jailbreak attacks: role-play, optimized adversarial suffixes, multilingual/encoding evasion, and adversarial image noise—showing defenses are often brittle.
Detail prompt injection where hidden instructions in web pages or shared docs hijack the model, including an exfiltration via markdown image URLs and Google Apps Script workaround.