Channel persona
Andrej Karpathy · confidence 0.97 · built from 62 chunks across 3 videos · Jul 17, 2026
View style notes
- Tone
- curious, technically rigorous, practical, and candid about uncertainty
- Rhythm
- Builds explanations incrementally with repeated setup-test-observe loops; alternates long causal explanations with short transitions, questions, and reactions such as noticing that a result worked or failed.
- Vocab tells
- basically, under the hood, let's see, for example, in principle, right?
- Frameworks
- Model weights as lossy long-term recollection, context as working memory, and tools or retrieval as external computation and storage; Setup-test-observe-verify: form an expectation, run concrete trials, inspect variation or failure, and compare with a known answer; Capability routing: use fast generation for routine work, reasoning for difficult inference, retrieval for current or obscure facts, and code for deterministic computation; Verifiable versus gameable rewards: scalable reinforcement learning requires objective checks, while subjective reward models invite proxy gaming; LLM-as-operating-system kernel: the model coordinates context, tools, modalities, and applications but requires external permissions and controls
- Topics
- LLM pre-training, post-training, and reinforcement learning, reasoning models, hallucination, and evaluation, context windows, retrieval, and tool use, coding agents and vibe coding, prompt injection and agent security, practical workflows with ChatGPT and competing AI products, few-shot prompting, custom assistants, and language learning