Microsoft Research Podcast Will machines ever be intelligent?
26 snips
Mar 23, 2026 
Subutai Ahmad, a computer scientist translating neocortex ideas into AI, and Nicolò Fusi, a researcher on transformer and LLM architectures, debate machine intelligence. They compare transformer attention to cortical columns, discuss continual learning vs memorization, energy and sparsity trade-offs, and whether many small, parallel modules could recreate brain-like intelligence.
AI Snips
Chapters
Transcript
Episode notes
Why Transformers Scaled Where RNNs Struggled
- Nicolò explained transformers separate attention (mapping token relations) and feedforward (storing knowledge), enabling parallel, bottleneck-free computation.
- This design removes sequential state carry bottlenecks of RNNs and unlocked massive-scale training on parallel hardware.
Encoder Layers Lift Relevant Information
- Nicolò described encoder layers as boosting relevant information prominence rather than successive summarization.
- The model highlights intent-relevant tokens until a latent space representing task intent and necessary knowledge emerges.
Thousand Brains Build Parallel World Models
- Subutai described the thousand brains theory where many cortical columns each build complete local world models and vote to reach a consistent interpretation.
- Each cortical column (~50k neurons) operates in parallel, forming distributed, sensory-motor predictions rather than a single latent space.