ideasthesia

Sign in Subscribe

Ryan Collison

I write about cool science and share it here.

Knowledge graph with semantic relationships enabling structured retrieval and multi-hop reasoning

Graph RAG

Vector databases were supposed to solve AI’s knowledge problem. Just embed your documents, store them, retrieve relevant chunks, and feed them to your language model. Simple, right? Except it doesn’t work for complex questions. Vector retrieval fails at multi-hop reasoning, misses contextual relationships, and can’t connect

Spiking neural network chip architecture with event-driven computation and temporal coding

Neuromorphic Computing

Your laptop uses about 65 watts to run. Your brain uses about 20. And your brain is doing far more: integrating sensory streams, maintaining autobiographical continuity, running predictive models of the world, coordinating movement, processing language, and managing thousands of concurrent processes—all on the energy budget of a dim

High-dimensional vector space with holographic distributed representations binding symbols to patterns

Hyperdimensional Computing

What if we’ve been computing in the wrong number of dimensions? Most neural networks operate in spaces of hundreds or thousands of dimensions. Hyperdimensional computing operates in 10,000 dimensions or more. And in those vast spaces, something magical happens: vectors become almost orthogonal to each other by default,

Neural network internals with circuit motifs and feature decomposition revealing computational structure

Mechanistic Interpretability

Neural networks work. They generate coherent text, recognize faces, translate languages, and write code. But nobody—including the people who built them—knows exactly how they do it. Until recently, AI systems were black boxes: weights and activations arranged in patterns we couldn’t decode. We could measure performance, but

Category theory diagrams with functors and morphisms mapping compositional structure across domains

Applied Category Theory

What if there was a mathematical language that could describe neural networks, quantum physics, linguistic composition, and biological systems all at once? What if the same formal tools could illuminate how sentences compose meaning, how proteins fold, and how machine learning architectures preserve structure? That language exists. It’s called

Neural organoid with emergent activity patterns—biological computing substrate self-organizing

Organoid Intelligence

What if the most powerful computer you could build wasn’t made of silicon, but of neurons? Brain organoids—miniature self-organizing neural tissue grown from stem cells—are already being trained to play Pong, control robots, and recognize patterns. They’re not conscious (probably), but they are computational. And

Bioelectric field patterns across cellular collective—voltage gradients directing morphogenetic computation

Basal Cognition

Your cells are thinking. Not metaphorically. Actually thinking. They’re making decisions, storing memories, solving problems, and coordinating with their neighbors to build and maintain the impossibly complex thing you call a body. They do this without neurons, without brains, without anything resembling what we typically call “cognitive architecture.” And