
Already know some of this? A 2-minute test skips the lessons you know.
Take the placement testA free 20-lesson course built from Andrej Karpathy's 2h 11m of YouTube video. 2 days 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.
Explain how text is tokenized, how the conversation format structures user/assistant turns, and why managing the context window (token budget) improves performance and cost.
Describe the two-stage training process (pre-training on internet-scale data, post-training for assistant persona), the resulting knowledge cutoff, and why models need tools for recent information.
Identify suitable low-stakes, non-recent, high-frequency questions (e.g., caffeine content, common medication ingredients) that models answer reliably from parametric memory, and verify against primary sources.
Apply the practice of starting a new chat when switching topics to reset the context window, reducing distraction, latency, and cost.
Navigate the model dropdown to confirm which model is active, compare pricing tiers across providers (OpenAI, Anthropic, Google, xAI), and choose the most capable model affordable for professional work.
Distinguish thinking models (o1, o3-mini, DeepSeek-R1, Grok 3 'Think', etc.) from standard models, enable them via UI selectors, and deploy them for hard math/code/reasoning tasks where extra inference-time compute boosts accuracy.
Trigger and verify internet search tool use in ChatGPT, Perplexity, Claude, Gemini, and Grok to retrieve current facts (e.g., TV schedules) that exceed the model's knowledge cutoff.
Leverage Deep Research (ChatGPT Pro, Perplexity) to automate iterative search, reasoning, and report generation with citations for complex topics like supplement ingredient analysis.
Upload PDFs (papers, blood tests, book chapters) into the context window, verify OCR/transcription accuracy, and iteratively query the model to boost comprehension and retention while reading.
Recognize when a model switches to the Python tool for exact calculation or code execution, inspect the generated code for hidden assumptions, and note which providers/models lack this tool (risk of hallucinated math).
Use ChatGPT's built-in data analysis tool to search for tabular data, write plotting code (log scales, curve fitting), render figures inline, and audit the generated code for silent assumptions.
Create interactive diagrams (Mermaid), flashcard apps, and React prototypes inside Claude Artifacts, and iterate on them conversationally without leaving the chat interface.
Employ Cursor's Composer agent to scaffold repositories, write and run code, debug gradient-check failures, and manage the full software lifecycle inside an IDE-integrated LLM workflow.
Contrast 'fake audio' (speech-to-text → LLM → text-to-speech) with 'true audio' (native audio token streaming in Advanced Voice Mode / Grok), and evaluate latency, expressiveness, and tier availability.
Upload arbitrary sources (PDFs, URLs, text) to NotebookLM, chat with the grounded model, and generate customizable, interactive audio podcasts for passive learning during walks or commutes.
Upload screenshots/photos (nutrition labels, blood-test PDFs, math expressions, toothpaste ingredients) for OCR + analysis, verify extracted values, and use the model as a first-draft interpreter for domain-specific visual data.
Generate images via integrated diffusion models (DALL·E 3, Ideogram) inside ChatGPT/Claude/Gemini, iterate on prompts, and understand the text→image token pipeline distinction from native multimodal models.
Enable cross-session memory to retain user preferences (e.g., translation tone, formatting rules) and set global custom instructions that steer model behavior across all chats.
Create a Custom GPT with tailored instructions (e.g., Korean subtitle OCR → translation → grammar breakdown), invoke it via screenshot paste, and manage a library of personal micro-agents.
Synthesize a decision framework: match model size/tier to task difficulty, enable reasoning for math/code, attach tools (search, Python, artifacts) for gaps, pick native vs. tacked-on multimodality, and choose the right app (ChatGPT default, Perplexity search, Claude artifacts, Cursor coding, Grok entertainment).