Karpathy's 1-Hour Stanford Lecture Distills AI Engineering Into 5 Layers
weichiuma · x · 2026-10-05
Andrej Karpathy delivered a 1-hour Stanford lecture on AI engineering from scratch, boiling the field down to a layered progression:
- 10% LLM: treat GPT like a general-purpose computer reprogrammable at runtime
- 30% Prompt: a program for that computer written in natural language
- 30% Agent: wrap the model with goals, context, memory, and tools to turn prediction into action
- 20% Loop: separate the inner loop (learning from context) from the outer loop (training updates weights)
- 10% Graph: organize communication as data-dependent message passing over directed graphs
The quoted tweet also points to an article on "Harness Engineering," arguing that when agents fail, rewriting prompts is the wrong layer — reliability comes from adding a harness layer with context, safe tools, durable state, verification, and recovery.
More from AGI Musings
- AI safety's disconnect from systems safety engineering questioned; FMF cites early INPO ties — chrismeserole · 2026-10-05
- LLMs are pure map with no territory: why you can't make a model of a model — inductionheads · 2026-10-05
- LLM Progress Called Incremental, Not Exponential, as Narrative Debate Heats Up — inductionheads · 2026-10-05
- Gary Marcus backs WSJ skepticism: 9%-of-GDP AI spending faces a natural ceiling — GaryMarcus · 2026-10-05
- Timnit Gebru wins 'Alternative Nobel', calls AI existential-risk talk a ploy to dodge regulation — PolarBearby · 2026-10-05
- Would COVID vaccines ship faster in 2026 with OpenAI and Anthropic in the loop? — flowersslop · 2026-10-05