This page collects my (work-in-progress) research notes on designing artificial general intelligence (AGI). By writing and sharing these, I want to:
- Study the technical problems associated with designing intelligence
- Debunk unscientific claims and hype to save others' time and effort
- Receive feedback and correct errors in my reasoning
- Questions and comments welcome!
- Best way to reach me is via Twitter
The information presented here spans topics in math, computer science, and biology, and in places involves a good deal of speculation. Statements that I think are true, but have not been experimentally or mathematically verified, are followed by a red asterisk*Be skeptical! as a reminder not to believe everything I say blindly. I’ve tried to provide a thorough set of references, so you can and should come to your own conclusions.
We must not believe those, who today, with philosophical bearing and deliberative tone, prophesy the fall of culture and accept the ignorabimus. For us there is no ignorabimus, and in my opinion none whatever in natural science. In opposition to the foolish ignorabimus our slogan shall be:
We must know — we will know!
What the brain is not
Before attempting to understand the mechanism of intelligence, we need to be sure that excessively weird things aren’t happening.
- Turing-equivalence: the brain is exactly as computationally capable as standard artificial computers.
- Not quantum: the brain does not directly use quantum mechanics for computation.
- Not hardcoded: the organization of the brain is not specified at the level of synapses, so there must be an algorithm describing the topological details of neural circuits.
- Reservoir computing suggests that neural networks can be largely random and fixed, but optimized networks outperform random ones.
- Continuous-time computation: although time in physical reality is continuous, discrete time is sufficient to describe any physical computing system.
Computation in the brain
Where and how is data stored? How is information transmitted and routed?
- The binding problem concerns the routing, association, and storage of information in neural networks that are fixed over rapid processing timescales.