Digest
This week in open AI: synthetic data, diffusion models, and hardware
Synthetic desktop trajectories lift OSWorld scores 18 points, a 7B diffusion model trains on 1.5T tokens, Midjourney enters ultrasound hardware, and Anthropic shuts Fable after a US government order.
ByAlex Sokoloff·
This week saw breakthroughs in model training—from 3.1 million synthetic desktop trajectories that reversed negative transfer in UI agents to a from-scratch 7B diffusion model trained on 1.5 trillion tokens. Meanwhile, Midjourney announced a full-body ultrasound scanner, Anthropic shuttered Fable under government pressure, and Noam Shazeer moved from Google to OpenAI. Across the board, researchers tackled real-time reasoning, annotation errors, and edge deployment.
In this digest
- ProCUA-SFT dataset pushes UI-TARS 7B to 45% OSWorld success with 3.1M synthetic desktop trajectories — Synthetic desktop trajectories just did what human data couldn't: a 3.1M-sample dataset lifted UI-TARS 7B's OSWorld score 18.7 points above baseline.