Digital Intelligence and Coordination
It is the differences between human intelligence and artificial intelligence out of which we construct the narratives about how AI will change the future. Many of our narratives are constructed from observed or anticipated differences in the following traits/dimensions:
- Capability
- Cost
- Agency
- Corrigibility
- Consciousness
For instance, narratives of economic disruption run on projections about cost/capability/agency. P(doom) narratives add in things like corrigibility, and e/acc mixes in elements of consciousness.
This project considers another interesting trait, which rarely makes its way into the mainstream narratives: that of digital implementation. AIs are digital, unlike any other known form of equivalent intelligence. Digital systems provide a large collection of interesting informational primitives which are highly relevant for problems of coordination. We’re only just beginning to explore the space of applications of this mixture of intelligence and digital abstraction.
Without discounting the importance of other narratives around AI, this project focuses narrowly on understanding what we might be missing by paying less attention to the power of digital intelligence for coordination.
Status: Active
Linked Posts:
Coherence
This project aims to understand how parts form coherent wholes, and what things like computational and informational restrictions have to do with this process.
Status: Active
Linked Posts:
Structure of Intelligence
Intelligence is a solution to the problem of adapting across diverse environments. Fields like learning theory, computational complexity theory, game theory, and social choice theory each study what is theoretically possible for intelligent agents in certain contexts under specific sets of constraints (information, computation, trust, etc.).
But some rich structures of intelligent systems only emerge when these constraints are combined. I believe there are low-hanging fruits to be found in taking underconstrained and the overly-abstracted models often used in these settings, and trying to give them flesh.
Questions this project seeks to answer:
- What is the proper theoretic framework for jointly addressing the computational and informational limitations of intelligent systems?
- What are toy problems which capture the need for inference-time computation/search for a jointly informationally and computationally efficient learner?
- How to intelligent systems bias their search processes toward solutions with verifiable structure?
- When does verifiability stand at cross purposes to capability?
Status: Active
Linked Posts:
The Physics of Emergence
We should expect that concrete intelligent systems will arise through dynamical optimization and selection processes. A priori, it’s quite surprising that many of the simple dynamics that we see at play in nature and in AI training would be able to find any of the complex structures that we observe or hypothesize.
Ideas from different parts of physics (statistical mechanics, non-equilibrium thermodynamics, etc.) as well as the general study of complex systems can help shed light on some of what is happening here. Singular Learning Theory (SLT) and the Free Energy Principle (FEP) take this approach.
Status: Active
Linked Posts:
Learning Theory Sequence
A sequence of posts about learning theory and generalization.
Status: Complete
Linked Posts: