Adam Marblestone on Why AI is Missing Something Fundamental About the Brain

Brain's Unique Learning Mechanisms

• 💡 The brain achieves high learning efficiency from minimal data, unlike Large Language Models (LLMs) which require vast amounts of data.
• 🧠 A key difference lies in the brain's complex, evolved loss functions and reward signals, which are highly specific and dynamically applied across different brain areas and developmental stages.
• 🎯 The cortex excels at omnidirectional inference, predicting any subset of variables from any other, in contrast to LLMs primarily focused on next-token prediction.

Learning and Steering Subsystems

• 🔑 Steve Byrnes' theory proposes a Learning Subsystem (cortex) that models a Steering Subsystem (innate responses from lower brain areas like hypothalamus/brainstem).
• 🌱 The Learning Subsystem generalizes abstract concepts (e.g., the word "spider") to trigger innate responses from the Steering Subsystem, effectively wiring learned features to innate reward functions.
• 🔬 Genomic analysis shows the Steering Subsystem has more diverse and bespoke cell types for specific, innate functions, while the Learning Subsystem has more uniform cells for general learning.

Biological vs. Digital Intelligence

• ⚡ Biological hardware presents trade-offs, offering energy efficiency and co-located memory/compute, potentially leveraging neuronal stochasticity for probabilistic inference.
• ⚠️ A significant limitation of biological brains is the inability to be copied or easily modified at the synaptic level, unlike digital AI models.
• 🚀 Future AI development should consider co-designing algorithms with hardware that mimics biological advantages, such as low-voltage switches and integrated memory/compute.

Impact of Neuroscience on AI

• 📈 Connectomics, the detailed mapping of brain connections, is crucial for understanding the brain's architecture, learning rules, and initializations to inform AI.
• 📊 "Brain-data-augmented" AI could use human brain activity patterns as auxiliary loss functions during training, potentially leading to richer representations and better generalization.
• ⏳ Significant investment (hundreds of millions to low billions) in neuroscience technology development is needed to make comprehensive brain mapping feasible and practical for AI insights.

Automating Formal Mathematics

• ✅ AI is well-positioned to automate formal mathematics through systems like Lean, which allow for mechanically verifiable proofs.
• 🛠️ This approach transforms math proving into an RL (Reinforcement Learning) task with a clear, verifiable signal, similar to AlphaGo.
• 💡 Automating mechanical proof steps can accelerate mathematical progress and enable applications like formally verified, unhackable software, shifting human effort to conceptual organization and conjecture.