Radiolab The Alien in the Room
151 snips
Dec 12, 2025 The podcast delves into the complex world of artificial intelligence, tackling the question of what it really is. Early ambitions of AI are explored, with a shift from rule-based systems to learning machines that mimic brain functions. Discover how neural networks learn from experience and how training on vast datasets has unleashed unexpected capabilities. Insights into language models reveal the challenge of understanding context over long texts. AI's creative outputs are framed as predictions, showcasing how technology continues to evolve and shape our understanding of intelligence.
AI Snips
Chapters
Books
Transcript
Episode notes
NetTalk’s Rapid Learning
- Terry Sejnowski and colleagues taught a program called NetTalk to pronounce English by feeding it transcripts and correct phonemes.
- The system started as noise but learned quickly and generalized to pronounce novel sentences after training.
How Neural Nets Learn Features
- Neural nets are layers of simple units connected by weighted links that get tuned by math to minimize error.
- The middle layers act as a black box that discovers helpful features without explicit labeling.
Generalize Through Many Examples
- Train models on diverse examples and average learned weights to generalize to new inputs.
- Use many varied examples so learned connections capture the broad concept, not just a single instance.
