ideasthesia

Sign in Subscribe

Ryan Collison

I write about cool science and share it here.

Semiconductor chips designed for hyperdimensional computing

Intel and IBM Bet on Hyperdimensional: Industry Applications

Intel and IBM are investing in hyperdimensional computing — an architecture that encodes information in high-dimensional vectors to enable fast, energy-efficient learning. It's a genuine alternative to neural nets, not a refinement.

Edge device running hyperdimensional computing efficiently

Hyperdimensional Computing Beats Transformers (On Edge Devices)

Transformers dominate benchmarks but consume enormous power. Hyperdimensional computing sacrifices some accuracy for radical efficiency — making real-time AI feasible on tiny embedded devices. The tradeoffs show where intelligence needs to live.

High-dimensional geometric space with surprising properties

Why High Dimensions Are Magic: The Geometry of Hypervectors

High-dimensional spaces break your intuition in useful ways. Random vectors become nearly orthogonal, making it possible to encode, bundle, and bind symbolic structures into single dense vectors — without losing information.

Hypervector operations binding and bundling in high-dimensional space

The Algebra of Hypervectors: Binding, Bundling, and Permutation

You can multiply two 10,000-dimensional vectors to bind concepts, add them to bundle a set, and permute to encode sequence. Together, these three operations form a surprisingly powerful algebra for thought.

Sparse distributed memory architecture visualization

Pentti Kanerva and the Origins of Hyperdimensional Computing

Long before 'HDC' was a field, Pentti Kanerva was asking how a brain-like memory could work with noisy, high-dimensional patterns. His sparse distributed memory model is where the revolution quietly started.

High-dimensional vector space with holographic distributed representations

Computing in 10,000 Dimensions: The Hyperdimensional Revolution

What if cognition works by bundling 10,000-dimensional random vectors? Hyperdimensional computing is a bold bet — and it's already beating deep learning on some hardware-constrained tasks.

Synthesis of interpretability insights revealing coherence geometry

Synthesis: What Neural Network Internals Teach Us About Coherence

Interpretability tools are cracking open trained neural networks and finding unexpected structure: linear representations, modular circuits, emergent concepts. These findings offer a new lens on coherence, using artificial systems as a model for biological originals.