Developer Community Tech Week is back with us again, so check out this week’s important news for us developers.

  • Zhiyuan Research Institute released a new super-large pre-training model “Wenhui” with a parameter magnitude of 11.3 billion

  • AI company Rokid has released the Rokid Vision 2: Compatible with 4G and 5G networks

  • TypeScript 4.2 Beta release

  • The Go generic proposal has been submitted and the Go 1.18 beta is expected to be piloted

  • Docker released the development team’s three preferred directions for 2021

  • Ministry of Industry and Information Technology issues action Plan for Industrial Internet Innovation and Development (2021-2023)

  • ICLR 2021: Robust Differentiable Neural Network search for DARTS-

  • Smart pen: an easy-to-use, efficient open source text generation library

News of the industry

Industry   News

1. Zhiyuan Research Institute released a new super-large pre-training model “Wenhui” with a parameter magnitude of 11.3 billion

China News Network Beijing news on January 12 “Wenhui”, a new super-large pre-training model, was released recently to explore and solve the problem that the current large-scale self-supervised pre-training model does not have cognitive ability. The Wenhui model uses a data-driven approach to build a pre-training model that connects user behavior with common knowledge and cognition, allowing the AI to “learn” and create. Data show that the parameter scale of “Wenhui” model reaches 11.3 billion, second only to the 12 billion parameter number of DALL·E model. It is the largest pre-training model in China at present, and has realized parallel running with the international leading pre-training technology.

2, AI company Rokid released binocular MR glasses Rokid Vision 2: compatible with 4G and 5G networks

Today, artificial intelligence company Rokid released its latest binocular mixed reality (MR) glasses, the Rokid Vision 2. The Rokid Vision 2 has a thin, thin look that folds like regular eyeglasses, and it extends existing user interfaces to include phones, laptops, game consoles and tablets, and is compatible with 4G and 5G networks. In addition, the glasses combine AR and AI capabilities to combine mixed reality with the real world, helping to boost productivity and provide users with an immersive experience.

3. TypeScript 4.2 Beta is released

TypeScript version 4.1 introduces a new type: Template literal types. This type can model specific patterns of strings.

But in 4.1, there are some strange inconsistencies between template string types and template string expressions. This is because template string expressions are not compatible with the new template string type.

So in TypeScript 4.2, template string expressions now always start with template literal types.

Perform more stringent checks for the IN operator

Relax the rules between optional attributes and string index signatures

4. The Go generic proposal has been submitted and the Go 1.18 beta is expected to be tested

Ian Lance Taylor, a key member of the Go team, announced that he had submitted a Proposal to add generics to Go. “The language change to add generics to Go is fully backward compatible, and the existing Go program will continue to work as it does now,” Ian said on his blog. Ian also said that if the proposal is approved, the Go team will release a full but possibly unoptimized implementation for developers to try out before the end of the year, perhaps in the Go 1.18 beta.

5. Docker released the development team’s three preferred directions for 2021

Docker team SCOTT JOHNSTON has released a list of major trends related to development teams in 2021, picking three directions, including distributed collaboration, more microservices, and accelerated AI/ML adoption.

The first is open, distributed collaboration. Its team’s survey of how thousands of Docker developers worked in 2020 revealed that 51% preferred working remotely. Deployment of remote development work will be completed in 2021.

The second is more microservices and less complexity. Microservices applications are proving their value in 2020, especially by squeezing the time from design to production and responding quickly to demand. Of the Docker developer community, 65 percent said their organization already uses microservices.

The third is to accelerate the popularization of AI/ML. The Docker team expects the use of AI/ML tools by development teams in their applications to increase in 2021.

6, The Ministry of Industry and Information Technology issued the Industrial Internet Innovation and Development Action Plan (2021-2023)

January 13, the Ministry of Industry and Information Technology (MIIT) today issued a notice on printing and distributing the Industrial Internet Innovation and Development Action Plan (2021-2023). IT House learned that the notice noted that the new infrastructure was further improved. An industrial Internet network infrastructure covering all regions and industries has been initially completed, and 30 5G-connected factories have been built in 10 key industries. A national system of industrial Internet big data centers will be basically completed, and 20 sub-centers at the regional level and 10 sub-centers at the industry level will be built.

The academic frontier

Academic News

1, ICLR 2021 | Meituan, Shanghai University, etc. : Robust Differentiable Neural Network search DARTS-

Differentiable methods (DARTS) are the most popular methods in neural network architecture search (NAS). There are many methods to solve the problem of performance collapse and improve its stability. RobustDARTS[1] proposed to use the characteristic roots of the hypernet as a marker to determine whether a collapse is entering or not. In this paper, a counterexample was found in the experiment, that is, the performance of the model searched was still improving even when the characteristic roots were increasing. Inspired by the unfair competitive advantage of jump links in FairDARTS[2], this paper uses an additional skip, which simply and effectively improves the robustness of DARTS without the use of additional hyperparameters and indicators.

* Links to papers: openreview.net/forum?id=KL…

2, “magic pen” shenghua: an easy to use, efficient text generation open source library

Text generation, as a hot field in natural language processing, has attracted extensive attention in academia and industry in recent years. With the deepening of the research, the sub-tasks and corresponding models in the field of text generation are more and more abundant, and some excellent open source frameworks have emerged. Existing open source text generation libraries can be broadly divided into two categories. One category tends to be modular driven, providing various components for researchers to build different models, but has the disadvantage that many specific baseline cannot be achieved through modules. The other one prefers the baseline driver, which is convenient for users to call the baseline quickly, but the disadvantage is that the overall structure is not unified enough and the scalability is limited.

Therefore, in order to take into account the advantages of both, the AI BOX team of Renmin University of China launched a new open source text generation library TextBox, which provides many interfaces of the new Baseline while ensuring modularization of components, enabling researchers to build their own models based on modules on the one hand. Alternatively, you can quickly call the existing baseline to complete the experiment.

* Links to papers: arxiv.org/pdf/2101.02…

* The above information comes from the network, edited by the public number “Jingdong Zhilian Cloud developer”,

Does not represent jd zhilian cloud position *

Recommended reading:

  • Pan jianwei’s team went to Nature again to build the world’s first integrated quantum communication network with a total length of 4,600 kilometers

  • The Turing Test is obsolete, AI needs a new benchmark; No Flash, No Adobe player

  • 2020H1 Growth of China’s AI cloud service market far exceeds expectations; The C++20 standard is officially released

Welcome to [JD Zhilian Cloud] to learn about the developer community

More wonderful technical practice and exclusive dry goods analysis

Welcome to [JD Zhilian cloud developer] public account