Interview, Podcast
Adam Marblestone – AI is missing something fundamental about the brain
- Predicts the biological question of how the brain functions could represent a quadrillion-dollar opportunity and the most significant scientific inquiry, contingent on technological empowerment of the field rather than pure intellectual effort.
- Hypothesizes that the cortex operates via omni-directional inference and probabilistic AI, where every area predicts subsets of variables from others, potentially utilizing energy-based models to capture joint distributions.
- Outlines a cortical architecture separating a "steering subsystem" with innate, pre-wired responses and sensory capabilities (such as face detection) from a "learning subsystem" capable of generalizing abstract concepts from diverse inputs.
- Suggests evolutionary pressures encoded complex loss functions and learning curricula into the genome, wiring learned features to innate reward systems to drive social instincts and cortical expansion with relatively few genetic changes.
- Anticipates that hardware evolution will co-locate memory and compute while utilizing lower voltages and natural stochasticity to mimic neural efficiency, potentially reducing costs through optical connectomics paradigms similar to the Human Genome Project's million-fold cost reduction.
- Projects that AI timelines for transformative events remain in the 5-to-10-year range, with the capability to sequence a human brain at a cost of hundreds of millions to low billions dollars requiring significant technological push and concerted effort.
- Identifies a scaling paradigm shift where fixed training compute budgets may yield smarter outcomes when split among multiple agents, and where inference-time compute used to elicit capabilities will be increasingly distilled into model training.
- Foresees a surge in formally verified software using tools like Lean for cybersecurity and proof automation, though conceptual math organization and generating new theorems remain challenges requiring potential external innovation or computational assistance.
- Notes that while current AI resembles limited reinforcement learning, future systems may integrate model-free and model-based approaches akin to basal ganglia and cortical functions, potentially requiring minimal drives like curiosity for competence.
- Warns of risks in "moonshot" biology and neuroscience projects, including the potential for failure in executing complex connectome generation and the "specification problem" in formal methods, while advocating for a research portfolio combining AI-informed reverse engineering with bottom-up biological discovery.
- Suggests that future AI may require auxiliary loss functions to predict neural activity patterns, distilling visual cortex data into networks, and that the gap map of fundamental capabilities could serve as the basis for numerous deep-tech startups.
- Maintains uncertainty regarding the nature of subjective experience, which might involve new physics, and acknowledges that the brain's world model remains ambiguous between hidden neural states and symbolic languages.
- Proposes that continual learning likely involves architectural mechanisms such as multiple time scales of plasticity, hippocampal-cortical consolidation, and thalamic gating or constraint satisfaction, though details remain unresolved.
- Indicates that while AI models have driven recent computational neuroscience theories, there is a push to identify actual neural primitives rather than imposing AI vocabulary, with progress expected through describing systems via architecture and learning rules.
- Anticipates that the "specification problem" for formal methods may be overcome in the coming years, allowing for interpretable world models specified in equations and a return to symbolic methods for provable safety properties.