Category:Machine Learning

The Model Already Knew the Answer — It Just Couldn't Stop Talking Long Enough to Say It

A look at a provocative new finding: two random vectors, no training, and a small open-weight model's accuracy nearly doubles.

The Chip That Learned to Think in Parallel

Why the hardware running today's AI was built to draw video game triangles, and what happens when you point a few thousand of them at the same problem.

Shrinking a 2.8-Trillion -Parameter Model

Kimi K3 shipped as the largest open-weight model anyone has released. This piece looks at what the compression crowd has actually managed to do about that, two days in.

You Built the Whole Chain

Building Blocks · Part 13

Article 1 opened with a single weight, multiplying a single number. Nothing since has introduced anything more magical than that.

The Same Model, Two Very Different Jobs

Building Blocks · Part 9

The exact same matrix multiplications run whether a model is learning or answering. So what could possibly be different enough to need its own dedicated hardware, memory budget, and math?

Tokens, Revisited: how text becomes math

Building Blocks · Part 6

Before a single vector or tensor can exist, raw text has to be chopped into pieces a model can count on one hand.

Tensors: the shape data takes inside AI

Building Blocks · Part 5

Vectors were one row of numbers. Tensors are what you get when a model needs to stack rows into grids, and grids into stacks of grids.

Dimensions: why AI thinks in thousands of directions at once

Building Blocks · Part 4

Every entry in an embedding is a separate axis of meaning. Here's what that actually buys a model, and what it costs.

Embeddings: where similar ideas learn to live near each other

Building Blocks · Part 3

Vectors gave every idea an address. Embeddings are what happens when a model chooses those addresses on purpose, so that meaning has geography.