Sean Carroll's Mindscape: Science, Society, Philosophy, Culture, Arts, and Ideas 230 | Raphaël Millière on How Artificial Intelligence Thinks
47 snips
Mar 20, 2023 
Raphaël Millière, a philosopher and cognitive scientist at Columbia University, dives into the intricacies of artificial intelligence and its thought processes. He distinguishes between artificial and biological intelligence, highlighting the simplicity of AI's neural networks compared to human brains. Millière discusses the evolution of language models and their learning mechanisms while emphasizing their limitations and the implications for AI safety. The conversation also probes ethical questions surrounding AI's perceived rights and the significance of embodied experiences in human cognition.
AI Snips
Chapters
Transcript
Episode notes
Neural Networks vs. Brains
- Artificial neural networks are simpler than biological neurons, with each node performing basic math.
- The largest current models, like GPT-3, have significantly fewer parameters than the human brain.
Neural Network Training
- Neural networks are initialized with random parameters or weights, which are then adjusted during training.
- Language models predict the next word or token in a sequence, learning grammar and structure implicitly.
Language Model Function
- Language models predict the next word, not entire sentences, using tokenization for efficiency.
- Despite this, they implicitly learn grammar and sentence structure through training.