OpenMOSS releases MOSS-Image open diffusion transformer

A discussion thread asks why symbolic AI structures keep resurfacing in ML research despite fading after neural nets rose to dominance. The poster notes symbolic AI's 1970s heyday and argues more research should revisit it.

Reddit user aziib generated PS2-style images with FastH3-V2 at 4 steps per 15 seconds, paired with a custom "modern ps2 style" LoRA published on Civitai.

OpenAI engineer Rennie Song uses a dot to collect feedback from social media and user reports, update a live tracker in ChatGPT Space, and investigate and fix bugs with Codex. OpenAI is pitching the workflow as "using dots to build better dots."

Hugging Face's Clem Delangue said preorders passed 20,000 units after Nvidia's CEO showed the $399 robots at GTC Berlin; Thomas Wolf counted over 4,000 sold in 24 hours versus ~3,000 on launch day. Deliveries are slated before Christmas.

A study finds an agent's harness configuration can triple its cost without improving accuracy. The scaffolding around the model, not just the model itself, drives spend.

Built from a prompt to Claude Opus 5.5 High, the artifact repaints an uploaded photo in a few seconds using a canvas element and simulated paintbrush. It relies on Sobel edge detection with no external libraries beyond Google Fonts.