Andrej Karpathy
Former OpenAI founding member and Tesla AI director, now an educator; coined "vibe coding" (vibe-coding). Karpathy's "Deep Dive into LLMs like ChatGPT" (Feb 2025, 9.7M views, 3.5h) is a general-audience explainer built on three training stages (pretraining โ SFT โ RL) mapped to how children learn from textbooks (exposition โ worked examples โ practice problems), delivered as a single-take Excalidraw + live-browser screencast with zero hook engineering.
Signature mental models from the lecture: base model as lossy-compression-of-the-internet; assistant as labeler-simulation ("you're not talking to a magical AI, you're talking to an average labeler"); params-vs-context-memory; models-need-tokens-to-think; "RLHF is not RL" (gameable reward models vs indefinite verifiable-domain RL); Swiss-cheese capabilities. Reproduced GPT-2 for $600/24h (lm.c) vs the original $40k.
Karpathy's closing advice โ "use them as tools in the toolbox, check their work, and own the product of your work" โ is the same conclusion matt-pocock and dhh reach from the practitioner side: the human owns taste and verification.
Practical sequel, How I use LLMs (Feb 2025, 2.7M views, 2h11m): a tour of ChatGPT, Claude, Gemini, Grok and Perplexity organised by what goes into the context window: tool-selection-by-freshness, thinking models only when the fast answer fails, and verify-like-a-junior-analyst. Personal stack at the time: ChatGPT Pro ($200/mo) as default, Perplexity for search, Claude Artifacts for diagrams, Cursor Composer for code, SuperWhisper on F5 for ~half of all queries, and few-shot custom GPTs for learning Korean. Karpathy reads books alongside an LLM ("don't read books alone").
Linked from
AlphaGoLLM training pipelineLossy compression of the internetModel parameters (weights)Models need tokens to thinkMove 37Open weightsParameters vs context memoryRLHF is not RLSleep-time computeSmart zone / dumb zoneSuperhuman blind spotsTool selection by freshnessVerify like a junior analystVibe coding