AI news for Nov 6, 2026

Ai2 has released ToyBlocks, the first of what it calls FoundationWorlds: simulated settings built so researchers can study robot learning under controlled conditions. ToyBlocks centers on a robot arm and blocks on a tabletop, and generates training demonstrations automatically rather than depending on people teleoperating the robot. Skill training runs in Nvidia's Isaac Lab, and a task-and-motion planner sequences those skills into complete demonstrations. The code, data and models are published openly.

Hugging Face published patch release v5.22.2 of its Transformers library, described in the release notes as an in-between patch for security and distributed fixes. The release contains two changes: hardening of the library's Jinja sandboxing, credited to Rocketknight1, and a distributed fix that sets device_id only for NCCL, credited to michaelbenayoun and 3outeille. A full changelog comparing v5.22.1 to v5.22.2 is linked from the release page.

Salesforce has introduced Koa Decisions, a new decision model that reads a request and a list of categories and returns one choice, a yes-or-no answer, or a score in a single step. It is the next version of Salesforce’s HyperClassifier and was built by applying new training techniques to Google’s Gemma, adding support for images and audio. Salesforce says the model is aimed at routing and other quick choices inside agent workflows, while leaving larger language models for writing and reasoning. People can try a demo through a free Trailhead account. Salesforce plans to extend Koa Decisions to flows, Apex, prompt templates and a developer API.