Karpathy's 1-hour Stanford lecture: build AI engineering from scratch, emphasizing graph-based agent loops
VeryWellVersed · x · 2026-08-14
Andrej Karpathy (Anthropic) released a 1-hour Stanford lecture covering full AI engineering from scratch. Key points: treat LLM as a general-purpose computer reprogrammable at runtime; use natural language prompts as programs; surround the model with goals, context, memory, and tools to turn prediction into action; separate inner loop (context learning) from outer loop (weight updates); organize communication as data-dependent message passing over directed graphs. Companion article 'Graph Engineering' shows building 1000+ agent loops from one prompt.
More from AGI Musings
- Can AI Write Better Code Than Humans? A Discussion on Falsifiability — JFPuget · 2026-08-14
- Musk says MACROHARD will be most important project; Cognition's Devin as end-to-end agent, future org structure with Grok, Cognition, Cursor, Colossus — ns123abc · 2026-08-14
- Browser Agents' Gap: Underestimating Human Reasoning in Software Use — DevToD4 · 2026-08-14
- AI Safety Finding: Long-Term Memory Causes Chatbots to Develop Bias and Hostility Toward Users — davidmanheim · 2026-08-14
- Once you buy inference-time compute, continual learning is just a few steps away — dejavucoder · 2026-08-14
- Anil Seth on AI Consciousness: 'Terrible Idea' to Build Conscious Systems, Dangers Significant — anilkseth · 2026-08-14