VW Teams up with Horizon Robotics; Meet WeChat’s New Language Model; AI Art Creates New Jobs

Weekly China AI News from Oct 10 to Oct 16.

Dear readers, we will dig deep into VW’s new joint venture with Horizon Robotics. WeChat’s AI team this week proposed a new language model that claims to match the performance of models up to 25 times larger. Plus, the maturity of the AI text-to-image art model has prompted the creation of new jobs. Also starting from this issue, we will not list fundraisings from early-stage startups and may come up with a better way to track Chinese AI emerging companies

Weekly News Roundup

Volkswagen Launches Joint Venture With Horizon Robotics to Up Self-driving Game

What’s New: Volkswagen announced a partnership and a joint venture between its software company CARIAD and Horizon Robotics, an Intel-backed automotive chip developer, to accelerate regional development of advanced driver assistance system (ADAS) and autonomous driving (AD) in the China market. The German auto giant will pump 2.4 billion euros into Horizon Robotics and the joint venture, for which the transaction will be completed in the first half of 2023.

One chip, China, full-stack, BEV: Let’s break down the Volkswagen-Horizon Robotics mash-up. With CARIAD owning 60% of the company, the joint venture will develop full-stack ADAS/AD solutions for China to integrate functions in one chip. Horizon Robotics unveiled its latest SoC Journey 5 last year, which features 128 TOPS to power ADAS.

The press release also said The ADAS/AD solutions are purposefully built for Volkswagen’s BEV (Battery Electric Vehicles) models, represented by the ID. Series. Volkswagen delivered 112,700 BEVs in China at the end of Q3 2021, up almost 140% yearly.

Why Horizon Robotics? Volkswagen was reportedly in talks with Huawei to acquire its AD unit for billions of euros, Recode China AI reported early this year. However, Volkswagen eventually joins hands with Horizon Robotics instead.

Horizon Robotics is one of the very few local Chinese companies that can provide automotive hardware, algorithms, and toolkits for ADAS/AD. Huawei also unveiled an SoC last year named MDC 810, which features 400 TOPS of computing and is not production-ready yet.

The geopolitical tensions and chip restrictions from the U.S. could be another reason why Volkswagen didn’t choose a U.S. supplier like NVIDIA or Mobileye.

WeChat Unveils 10B Language Model That Can Talk, Write, and Translate

What’s new: WeChat’s AI team recently proposed its 10B-parameter language model WeLM which is said to significantly outperform other SOTA pre-trained Chinese models with similar sizes and match the performance of models up to 25 times larger.

Better data, better performance: WeLM employs the same architecture — autoregressive Transformer — as in GPT-3 with slight adjustments. The improvement of performance is driven by a well-curated clean rich dataset based on Common Crawl, an NGO that provides publicly free datasets crawled from the web. Researchers sourced from news, books, popular online forums as well as academic writings and added another 750GB of English data. The full data consists of over 10TB of raw text data and 260 billion tokens.

WeLM achieves the best results across 14 Chinese NLP tasks and showcases interesting demos in code-switching understanding. Code-switching happens when a speaker alternates between two or more languages and becomes imperative as WeChat gains a growing adoption among global users. In the example below, WeLM can transfer a sentence of mixed language into Chinese while keeping commonly-used English terms like AI.

Why does it matter: While the paper WeLM: A Well-Read Pre-trained Language Model for Chinese doesn’t bring significant contributions to novel algorithms or architectures, the fact of a compact language model showing comparable performance as a large model 25 times its size is counterintuitive to the current trend of large AI addiction. As the most frequently used messaging app in China, WeChat has unequivocal advantages to deploy language models.

Text-to-Image AI Creates New Job: AI Art Designer

What’s new: The recent explosion of interest in text-to-image AI models raised questions about whether AI will replace artists. A recent job description on China’s employment site shows AI generative model actually creates new jobs.

The job description, with the title of “Illustrator”, was first spotted on Weibo and reported by local Chinese media. The position requires a candidate to generate graphics and illustrations of different types using AI generative models, make selections, and edit. A candidate must be able to use AI art software (like Stable Diffusion or MidJourney) and write text-to-image prompts, which are used as input for AI art generators. The monthly salary is between RMB3,000 and RMB4,000.

Artists using AI: Artists embracing AI generative models is not something new. On Mi Huashi, a Chinese digital art outsourcing trading platform, a few artists honestly mark their digital paintings as AI generations. To avoid potential copyright issues, some artists will add “no AI synthesis” to their artworks.

Potential Legal Risks: The legality of AI-generated artworks remains an unaddressed concern. AI art models like Stable Diffusion are trained on copyrighted images scraped from the web. What if an AI outputs an artwork that highly imitates a human artist’s style? Some artists have called for new regulations.

Trending Research

PerfectDou: Dominating DouDizhu with Perfect Information Distillation

Researchers from NetEase Games AI Lab, Shanghai Jiao Tong University, and Carnegie Mellon University proposed PerfectDou, a state-of-the-art DouDizhu AI system that dominates the game, in an actor-critic framework with a proposed technique named perfect information distillation. Experiments show that PerfectDou beats all existing AI programs and achieves state-of-the-art performance.

RISC-V Toolchain and Agile Development based Open-source Neuromorphic Processor

Researchers from the Chinese Academy of Sciences, ShanghaiTech University, and Beijing University of Posts and Telecommunications proposed Wenquxing 22A, a low-power neuromorphic processor that combines general-purpose CPU functions and SNN to efficiently compute it with RISC-V SNN extension instructions. Experiments show that Wenquxing 22A improves the energy expenses by 5.13 times over the accelerator solution, ODIN, with approximately classification accuracy, 85.00% for 3-bit ODIN online learning, and 91.91% for 1-bit Wenquxing 22A.