Stories by driftsumi-e

The Unreasonable Effectiveness of Linear Prediction as a Perceptual Metric

Residual Matrix Transformers: Scaling the Size of the Residual Stream

Skate 3 in the Browser

Discrete Fourier Transform faster than n log n [pdf]

Integer multiplication below n log n

Dust: Pretraining Transformers Without Backpropagation

Single cortical neurons as deep artificial neural networks

SoftServe: A Scalable Quasi-Newton Method for Deep Learning

Chess, but it looks like doodles on paper

Self-Play Pretraining with Zero Data

Learning to Discover Interesting Mathematics

Log-Depth Recurrent Language Modeling

Realtime Pathtracing in Octane

94% on AIME with 1B Params

86" 3D Display

Fluid Simulation in One Tweet

Solving rubiks cubes "without search"

Jev might be Qwen finetune

DeepSeek's Parent Company

Replicating Jev with Qwen-2.5

1 2 3 4 5 6 7 8 9 10