AI labs race on inference speed: GPT 5.6 Sol hits 750 tok/s

LLaDA-Image, a unified 6B-parameter image generation and editing model, is now available on GitHub, arXiv, and Hugging Face.

Reality Defender CTO Alex Lisle says AI menus look like "an alien trying to make a pizza," trained on a narrow, "pleasing" aesthetic from corpora like 2015 Chili's menus. Customers viscerally sense something is off, driving demand for AI-detection tools.

A pilot study found coding agents often stick with grep over LSP-backed semantic navigation, and forcing the semantic path sometimes reduced task success. The author argues tool LLM-friendliness—context and output shape—matters as much as precision.

Gimlet, which helps divide AI tasks among different chips, raised $300 million in a new round, valuing the company at $3 billion.

A Reddit user shared a video made with MiniMax that turns Anakin and Obi-Wan into dancers, with music from Suno. The clip is a creative test using movie screenshots.