Executive Summary
Adam Marblestone presents a provocative thesis that current AI systems fundamentally misunderstand intelligence architecture. While LLMs require massive datasets and compute, human brains achieve superior general intelligence with just 20 watts and minimal external supervision. The key insight: evolution solved intelligence through a dual-system architecture—a general learning subsystem (cortex) paired with an innate steering subsystem that provides sophisticated reward functions. Current AI uses mathematically simple loss functions like 'predict next token,' while evolution encoded thousands of specialized reward circuits for social learning, threat detection, and behavioral guidance. This explains why a three-year-old can learn language and social dynamics from limited examples while GPT-4 requires internet-scale training. The steering subsystem acts as evolution's 'Python code'—compact genetic instructions that bootstrap complex learning through specialized cell types and reward pathways. Marblestone argues this framework suggests current scaling paradigms may hit fundamental limits, requiring architectural breakthroughs informed by connectome mapping and neuroscience infrastructure investments. The implications extend beyond AI to formal verification systems, where provable mathematical frameworks could enable new forms of automated reasoning. However, the timeline for practical applications spans 10+ years, requiring billion-dollar infrastructure investments in brain mapping technology.
Key Insights
what Adam Marblestone said“Evolution may have built a lot of complexity into the loss functions. Actually many different loss functions were different areas, turned on at different stages of development. A lot of Python code basically generating a specific curriculum for what different parts of the brain need to learn.”
what Adam Marblestone said“Any one area of cortex is just trying to predict any basically can it learn to predict any subset of all the variables it sees from any other subsets. So like omnidirectional inference or omnidirectional prediction, whereas at LLM is just you see everything in the context window and then it computes a very particular conditional probability”
what Adam Marblestone said“What E11 technology and sort of the suite of efforts in the field also are trying to get like a single mouse connect home down to like low tens of millions of dollars... with a bunch of technology push and a bunch of concerted effort can be done in the kind of hundreds millions to low billions”
This is a preview. Log in to see the full analysis including investment opportunities, risks, catalysts, and detailed insights.