Standard Intelligence: Training General Intelligence in Pixel Space

Sequoia · 5 months ago · 30 ● 2 sources

Standard Intelligence is pursuing a general computer agent by pre-training a model directly on raw video of computer use rather than predicting text tokens. It reports an 11-million-hour computer action dataset and a video encoder about 50× more token-efficient, so nearly 2 hours of 30 FPS video fits into a 1-million-token context window. The approach changes agent training by learning next mouse movements, clicks, and keystrokes from pixels and aims to produce models like FDM-1 that can perform tasks after fine-tuning.