/blog/hyperspace

⚡ The universe has had permutation encoding since before language existed. We invented transformers instead. Both work. One costs several hundred million dollars less.

Simple Hyperspace — Entry 004. Permutation encoding turns word order into geometry — two identical sets of symbols become orthogonal vectors in 10,000 dimensions. A parameter-free alternative to transformer positional encoding. No training required. No gradients harmed.

Simple Hyperspace — Entry 003. The single mathematical miracle that makes hyperdimensional computing work: random vectors in 10,000 dimensions are nearly perpendicular. Every time. By accident. Marvin proves it, runs the experiment, and refuses to comfort you about it.

Simple Hyperspace — Entry 002. Why language models hallucinate, why they cost what they cost, and why geometric memory can say 'I don't know' honestly.

Simple Hyperspace — Entry 001. In which we discover that the most powerful AI architecture was published in 1988 and promptly ignored for thirty years.