Closer To Truth Blaise Agüera y Arcas: What is Intelligence in an AI Future?
Dec 3, 2025
Blaise Agüera y Arcas, a VP at Google and expert in AI and computational neuroscience, dives deep into the essence of intelligence. He explores how evolution and AI are intertwined, revealing that prediction is core to both. Blaise discusses the breakthroughs of transformer models, the limitations of AI's self-improvement claims, and the importance of human feedback in AI systems. He highlights the collective nature of intelligence, suggesting a future where human and AI minds collaborate within networks, shaping a new understanding of consciousness and cooperation.
AI Snips
Chapters
Books
Transcript
Episode notes
Generality Comes From Prediction
- Blaise defines general intelligence as broad ability to generalize beyond training, not an ineffable soul or agency.
- He argues modern large unsupervised predictive models demonstrate true generality by solving novel tasks.
Engineers Expected A Trick
- Blaise recounts that teams predicted a 'trick' or fairy dust was needed for general intelligence but were surprised by scaling next-token models.
- He says engineers were surprised when simple scaling produced systems that solved word problems and more.
Transformers Enabled Scalable Prediction
- The transformer architecture enabled scalable, parallel training and long-context prediction, fueling modern progress.
- Blaise emphasizes prediction over specific architectures as the central ingredient of recent AI advances.
