Will machines ever be intelligent?

2026-03-23 · Source: Microsoft Research · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Brain-Inspired AI · Depth: Expert, extended

Summary

The Microsoft Research podcast "Will machines ever be intelligent?" features Doug Burger, Nicolò Fusi (Microsoft Research), and Subutai Ahmad (Numenta) discussing the fundamental differences between transformer-based large language models (LLMs) and the human brain's architecture. Published March 23, 2026, this episode explores whether current AI systems are truly intelligent by comparing their efficiency, representation, and sensory-motor grounding. The discussion highlights LLMs' brute-force, compute-intensive approach to learning, contrasting it with the brain's highly efficient, distributed, and continuously learning cortical columns. Key differences include the brain's sparse, parallel processing at 10-12 watts versus LLMs' sequential, n-squared attention mechanisms requiring megawatts, and the brain's continuous, localized learning compared to LLMs' teacher-forced, static dataset training.

Key takeaway

For AI Scientists and Research Scientists designing future intelligent systems, recognize that current monolithic LLMs, while powerful, are fundamentally inefficient compared to biological brains. Consider shifting research towards distributed, sensory-motor-grounded architectures that prioritize continuous, sparse learning, potentially by designing small, efficient digital cortical columns. This approach could lead to more energy-efficient and adaptable AI, bridging the gap between current AI capabilities and true general intelligence.

Key insights

Human brains and current LLMs employ fundamentally different architectures and learning mechanisms for intelligence.

Principles

• Intelligence can be viewed as compression.
• Biological systems prioritize energy efficiency.
• Continuous learning is fundamental to biological intelligence.

Method

The human brain utilizes approximately 100,000 cortical columns, each acting as a complete sensory-motor processing system, operating in parallel and continuously updating world models through sparse, localized learning.

In practice

• Explore sparse representations for AI efficiency.
• Integrate sensory-motor loops into AI architectures.
• Investigate continuous, localized learning paradigms.

Topics

• Large Language Models
• Transformer Architecture
• Neuroscience
• Cortical Columns
• AI Efficiency

Best for: AI Scientist, Research Scientist, AI Researcher, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• I don’t think we are close to “AI scientists” — Current LLM-based AI agents lack continuous implicit learning, limiting their ability to achieve human-level scientific or complex knowledge work.
• AI Wearable Watches You Back — AI's future hinges on co-designed, continuously running agentic systems requiring vastly increased, energy-efficient compute.
• AI Cartoon of The Year — and five (re)readings for 2026 — The future of AI may shift from pure LLMs to more robust, neurosymbolic approaches addressing current limitations.
• Scaling LLMs won't get us to AGI. Here's why. — Current LLMs are pattern matchers that interpolate, but cannot extrapolate to genuinely novel structures required for AGI.
• IEEE Transactions on Artificial Intelligence, Volume 7, Issue 2, February 2026 — The volume showcases diverse AI advancements, from LLM applications in autonomous driving to specialized models for protein language processing.
• The Predictive Mind: How Human Brains and AI Speak the Same Language — Human brains and large language models both operate on predictive processing, continuously generating expectations and updating based on new information.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Microsoft Research.