AI Is Not Magic: The Physics, Math, and Hardware Behind It
summary
Explains that LLMs are linear algebra plus optimization running on specialized silicon: words live as embeddings in 768 dimensions, similarity search uses vector databases, learning is gradient descent that reduces error stepwise, and transformers use attention over token vectors. Points out GPUs (not CPUs) and tensor cores do heavy parallel 4x4 matrix work — a single clock can do a 4x4 matrix per clock — and that scaling is limited by cooling and megawatts of power, which drives current engineering tradeoffs.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:07AI is math, physics, hardware
- 3:58Embeddings and vector DBs
- 5:50Neural networks as math pipelines
- 7:19AI learning with gradient descent
- 8:42Transformers revolutionize AI
- 10:31Hardware behind AI power
- 11:36Why GPUs matter for AI
- 13:40Tensor cores accelerate AI
- 15:07Physics limits and AI future
related talks
41mUnveiling the simple math that powers AI
12mKeynote: How LLMs Have Changed the Way Organizations Do AI- 12mKeynote: How LLMs Have Changed the Way Organizations Do AI
30mLLM Inference Arithmetics: the Theory behind Model Serving
36mLLMs and AI agents demystified
16mAI: Common Misperceptions and Uncommon Perceptions
51mAI Is Not Magic
23mGhost in the Machine: AI, Spirituality and a New Humanism