ideasthesia

Sign in Subscribe

Ryan Collison

I write about cool science and share it here.

Humanoid robot wireframe with luminous sensorimotor pathways and prediction-error flows

Robotics and Embodied Active Inference

Most robot control systems separate sensing from acting. Active inference collapses that distinction: a robot doesn't just react to the world, it continuously predicts it — and acts to make its predictions come true. The result is behavior that looks uncannily purposeful, even on simple hardware.

Software architecture visualization showing discrete and continuous computational pathways

PyMDP and RxInfer: The Active Inference Software Stack

Active inference has the math. PyMDP and RxInfer have the implementations. PyMDP is the researcher's sandbox — clear, Pythonic, built for learning the framework. RxInfer is for production-scale agents that need to run in real time. This is how they differ, and when to use each.

Factor graph visualization with variable and factor nodes connected by flowing message pulses

Message Passing and Belief Propagation in Active Inference

The free energy principle is elegant in theory but computationally demanding. Message passing on factor graphs is how you implement it in practice — turning the brain's update equations into something an algorithm can run.

Layered probabilistic world model visualization showing state-observation-action relationships

The Generative Model: Building World Models for Active Inference Agents

In active inference, behavior is driven by a generative model: a probabilistic description of how the world works and how observations arise. Building that model — specifying priors, likelihoods, and transition dynamics — is half the work of designing an intelligent agent.

Synthesis of inference scaling insights about thinking

Synthesis: What Inference Scaling Teaches About the Nature of Thinking

Give a language model more time to think and it produces better answers. Inference scaling suggests that thinking is search: navigating a space of possible coherent states to find the most meaningful one. Which raises the question: is that what human reasoning is too?

Test-time compute as active inference free energy minimization

Test-Time Compute Meets Active Inference: Reasoning as Free Energy Minimization

Scaling inference compute and minimizing free energy turn out to describe the same process. AI reasoning and biological cognition may be solving the exact same optimization problem from different angles.

Pay-per-intelligence business model visualization

The Economics of Inference: Pay-Per-Intelligence Business Models

Intelligence is becoming a utility you dial up or down by the problem. Need to solve something hard? Rent more compute for that inference run. The economics reshape who can afford sophisticated AI, which problems get solved, and what companies are worth building.