The Next Step Toward Understanding The Nature Of Intelligence - EP 49 Sebastian Seung
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
Sebastian Seung, the world's leading connectome researcher from Princeton, sits down to discuss his groundbreaking work mapping the complete neural wiring of a fly brain—a decade-long scientific milestone published in Nature in 2024. (00:47) The conversation covers his journey from theoretical physics to neuroscience, the painstaking process of creating connectomes (detailed maps of neural connections), and the profound insights gained from understanding how 140,000 neurons and millions of synapses create intelligence in a fruit fly.
- Main Theme: The episode explores how connectome research is transitioning from pure academic pursuit to practical applications through brain emulation technology, with Seung announcing his new startup Memazing that aims to reverse-engineer animal brains to create more efficient AI systems and eventually enable human brain emulation.
Speakers
Sebastian Seung
Sebastian Seung is a professor of computer science and neuroscience at Princeton University and one of the world's foremost experts on connectomics. Originally trained as a theoretical physicist at Harvard and MIT, he transitioned to neuroscience and led the team that completed the first full connectome of a fruit fly brain—a groundbreaking achievement published in Nature in 2024. He previously served as head of research at Samsung and has written a bestselling book about the connectome, establishing himself as a pioneer in understanding how neural wiring creates intelligence.
Key Takeaways
Brain Structure Reveals Function in Remarkable Ways
The fly connectome revealed that neural architecture directly reflects function, challenging skeptics who believed dead brain maps wouldn't reveal how living brains work. (21:27) Seung discovered neurons that literally gesture toward their function—motion-detecting neurons have branches that point in the direction of motion they detect, like "pantomimers" that have been wildly gesticulating for 200 million years. This form-function relationship demonstrates that connectomes aren't just static maps but functional blueprints that can be read and understood, providing a foundation for building brain emulations.
AI Acceleration is Creating a "Connectome Spring"
After the "connectome winter" that followed the 1986 worm brain mapping, AI breakthroughs have revolutionized the field by making previously impossible analysis feasible. (11:48) What would have taken humans 100,000 years to trace manually, AI can now accomplish in months, though human experts still need to correct AI errors. This acceleration means neuroscience adjacent to AI will progress exponentially, while other areas may lag behind. The convergence of AI and connectomics is creating unprecedented opportunities to understand and replicate brain function.
The Information Gap Requires Strategic Assumptions
While connectomes provide unprecedented detail about neural wiring, significant information gaps remain that must be bridged through educated assumptions and hybrid approaches. (34:54) Missing parameters include the relative strength of excitatory versus inhibitory synapses, time scales of neural responses, and dynamic properties that can't be read directly from static connectomes. Scientists are using techniques like backpropagation combined with connectome constraints to tune these parameters and create faithful brain emulations.
Energy Efficiency is the Next AI Frontier
The human brain operates on just 10 watts while NVIDIA's DGX H100 requires 10 kilowatts, and a fruit fly brain uses only one microwatt while performing remarkable computations. (37:57) This massive efficiency gap represents a major opportunity for creating "intelligence too cheap to meter"—AI systems that could run everywhere rather than being centralized in massive data centers controlled by a few corporations. Understanding how biological brains achieve such efficiency could slash AI costs and democratize artificial intelligence.
Brain Emulation Offers a Dark Horse Path to AGI
While companies like OpenAI pursue large language models as a path to artificial general intelligence, brain emulation represents an alternative approach that could achieve human-level AI by directly replicating brain structure and function. (39:34) By creating digital versions of biological brains with one-to-one correspondence between simulated and real neurons and synapses, this approach could bypass current AI limitations and provide a more direct route to understanding and replicating intelligence.
Statistics & Facts
- The fruit fly brain contains 140,000 neurons and tens of millions of synapses, with the total length of neural wires equivalent to about a football field's worth inside the tiny brain. (10:19)
- The fly connectome project began in 2014 and took about a decade to complete, requiring a fly brain to be cut into 7,000 slices, each a thousand times thinner than a human hair using an atomically sharp diamond knife. (09:12)
- If connectome costs drop by a factor of two every year following Moore's law, a human connectome would cost a million dollars in 20 years, or a billion dollars if wanted in 10 years, demonstrating the exponential cost-time tradeoff. (13:14)
