Macheng Shen
I think about theories of intelligence, distributed cognition, and the physical limits of learning systems.
I think about intelligent systems as finite physical systems that must build control-sufficient internal structure while paying for memory, communication, delay, synchronization, and adaptation. That framing connects biological multiscale competency, current AI systems, and infrastructure-scale coordination problems.
Harness Engineering and the Physical Instantiation of Intelligence
A pedagogical essay on why AI capability is a property of the larger physical-computational system: model, memory, tools, harness, and world coupling.
Meta-Control, Information Gain, and the Architecture of Autonomous Learning
A pedagogical research note on how prediction error, anticipated information gain, learning progress, and pragmatic value may fit together inside a true control plane for autonomous learning.
Why Distributed Memory Matters for Lifelong Agents
Long-term memory is not passive storage — it is a policy for preserving and updating state so that future prediction, control, and identity maintenance improve without recomputing everything from scratch. Connects distributed engrams, complementary learning systems, Nested Learning, and the stability–plasticity problem.
Where Objectives Come From — and Why Solutions Become Strategic Assets
Why reward is often only a representation of goals, why objectives may emerge from joint agent–world dynamics, and why some solutions behave like strategic assets that transform later games.
Line Loss for Intelligence: Thermodynamics, Topology, and Cognitive Cones for Planet-Scale AI Infrastructure
As AI becomes infrastructure, intelligence will not just be computed; it will be transported across space. The bottleneck may be transport, synchronization, and control-relevant communication rather than raw compute.
↳ Adversarial Review: A Stress Test
↳ Debate: Is Information More Fundamental Than Energy?
From Mutual Information to Endogenous Viability
Why mutual information is only a starting point for understanding intelligence, and why a deeper theory may need endogenous viability, relevance, and distributed multiscale organization.
↳ Adversarial Review: A Stress Test from Three Directions
↳ Debate: From Persistence to Goals — The Category Jump
Bridge to the research branch
The wave / credit-transport notes are best read as a subproblem inside this larger agenda. They ask how update-relevant information moves through real learning systems. The bigger thesis still needs additional layers: relevance, task formation, endogenous objectives, multiscale control, and physical viability.