What’s New This Week? AI Model, MR Glasses, TypeScript 4.2, Go Generics & More

1. Wenhui Large‑Scale Pretrained Model

China's Zhiyuan Institute released the 11.3‑billion‑parameter “Wenhui” model, aiming to give self‑supervised models cognitive abilities by integrating user behavior, common‑sense knowledge and reasoning. It is the largest Chinese pretrained model after DALL·E, matching international state‑of‑the‑art.

2. Rokid Vision 2 MR Glasses

Rokid unveiled the dual‑camera mixed‑reality glasses Vision 2, which fold like normal glasses, support 4G/5G, and combine AR with AI to boost productivity and deliver immersive experiences.

3. TypeScript 4.2 Beta

TypeScript 4.2 beta introduces stricter checks for the in operator, relaxes rules between optional properties and string index signatures, and adds missing helper functions. It also makes template literal types the default for template string expressions.

4. Go Generics Proposal

Ian Lance Taylor announced the submission of a backward‑compatible generics proposal for Go. If accepted, a beta implementation is expected in Go 1.18, allowing developers to experiment with generic code.

5. Docker’s 2021 Priorities

Docker highlighted three focus areas for 2021: distributed collaboration, increased micro‑service adoption, and accelerating AI/ML usage in applications.

6. Industrial Internet Action Plan (2021‑2023)

The Ministry of Industry and Information Technology issued a plan to improve new‑type infrastructure, build 5G‑connected factories in ten key industries, and establish a national industrial‑internet big‑data centre network.

7. ICLR 2021 – Robust DARTS

Researchers from Meituan and Shanghai Jiao Tong University presented RobustDARTS, a method that uses supernet feature roots to detect collapse and adds an auxiliary skip connection to improve DARTS stability without extra hyper‑parameters.

8. TextBox – An Easy, Efficient Text Generation Library

The RBOX team at Renmin University released TextBox, an open‑source library that combines modular components with ready‑to‑use baselines, enabling researchers to build custom models or quickly run state‑of‑the‑art baselines.