SenseTime IPO Hit By US Investment Ban; Baidu Releases 260-Billion Parameter NLP Model Enhanced By Knowledge Graphs; Alibaba-Backed Startup Unveils $10,000 L4 Robotaxi Kit

China’s AI news in the week of December 12, 2021

Chinese AI Giant SenseTime Delays IPO Over US Investment Ban Threat

Just a week before SenseTime to go public in Hong Kong, the Chinese leading AI company was placed on a blacklist by the US government that would ban American investors from investing in the company.

Last Friday, the US Treasury Department added multiple Chinese companies — including SenseTime — to the list, citing their ties with Chinese militaries. SenseTime was projected to announce its pricing of IPO on the Hong Kong Exchange on the same day. Specializing in facial recognition technology that can distinguish certain races based on facial characteristics, SenseTime was accused of assisting the Chinese government to recognize Muslim ethnic Uyghurs.

SenseTime will delay its IPO, and its IPO prospect is clouded, Chinese media Caijing reported citing people familiar with the matter. Also unknown is how the US investors behind SenseTime will deal with this new regulatory change. The company has raised over a total of $2.6 billion.

In response to the US government’s move, SenseTime said in a statement that “We strongly oppose the designation and accusations that have been made in connection with it. The accusations are unfounded and reflect a fundamental misperception of our company. We regret to have been caught in the middle of geopolitical tension.”

In 2019, one subsidiary of SenseTime, Beijing SenseTime, was put on a so-called “entity list” by the Commerce Department’s Bureau of Industry that imposes additional license requirements on any exports, reexports, or transfers of all items. However, in its Hong Kong IPO prospectus, SenseTime explained that the 2019 ban would have a limited impact on the company.

In August, we reported that SenseTime filed for Hong Kong IPO to raise $2 billion, but it halved its goal to $768 million over concerns on regulatory crackdowns. Founded in 2014, the company reported revenue of RMB3.446 billion in 2020 from sales of its AI products and services in enterprise software, smart city, mobile devices, and automobiles. Its adjusted net loss in 2020 is RMB878.4 million.

As of June 30, 2021, SenseTime has RMB8.93 billion as cash and cash equivalents and RMB7.94 billion in time deposits.

Baidu Beefs Up NLP Model ERNIE 3.0 With 260 Billion Parameters

Last Thursday, Chinese tech giant Baidu released its largest natural language processing model, ERNIE 3.0 Titan (or PCL-BAIDU Wenxin), in a research collaboration with Shenzhen-based scientific research institution Peng Cheng Laboratory. With 260 billion parameters, the model is said to be the world’s largest monolithic Chinese NLP model.

ERNIE is a series of pre-training language models Baidu created in 2019 to compete with Google’s BERT, Microsoft’s Megatron, and OpenAI’s GPT. Unlike most language models trained on unstructured data, Baidu researchers devised an approach to unify knowledge graphs, an old-fashioned network to represent relationships between knowledge subjects, and massive text corpora into one representation. That’s why ERNIE 3.0 Titan is also crowned as the world’s first knowledge-enhanced 100-billion-scale model. The advantage, as Baidu said, is to improve the model’s understanding and reasoning ability.

ERNIE 3.0 Titan is an upgraded version of ERNIE 3.0, an NLP model Baidu revealed this August that topped the language understanding benchmark SuperGLUE. While the paper is not available yet, here’s what we know.

ERNIE Titan 3.0 has achieved state-of-the-art results in more than 60 tasks, including machine reading comprehension, text classification, and semantic similarity, as well as in over 30 few-shot and zero-shot benchmarks.

The model was developed through a collaboration between Peng Cheng Laboratory’s self-developed computing system Peng Cheng Cloud Brain II and Baidu’s deep learning platform PaddlePaddle.

Baidu researchers came up with a controllable algorithm and a trustable algorithm to ensure the modal’s language generation is bound with human instruction.

Researchers applied model compression technology to simplify the ERNIE 3.0 Titan model for real-world scenarios. The compressed model retains only 0.02% of the original size but can achieve comparable performance.

Alibaba-Backed Robotaxi Startup Releases $10,000 Self-Driving Kit

DeepRoute.ai, a Shenzhen-based self-driving company, recently released an AI driver solution that costs less than $10,000 to turn a passenger vehicle into an L4 robot vehicle.

The self-driving kit, coined DeepRoute-Driver 2.0, packages five solid-state lidars and eight cameras into a top-mounted sensor rack, in which lidars account for roughly half of the expense. The company also demonstrated how its AI driver navigated the Central Business District of Shenzhen for an hour without intervention, handling complex road conditions crowded with various automobiles, cyclists, and countless pedestrians.

DeepRoute.ai plans to sell its production-ready self-driving kit to OEMs in 2022 and expects to see L4 self-driving passenger vehicles in 2024. However, the company hasn’t disclosed any interested customers.

DeepRoute.ai is stealing the show from another local startup AutoX. In September, the startup announced a $300 million Series B round from Alibaba, Jeneration Capital, and Geely.

Another Chinese self-driving upstart Momenta joins hands with SAIC Mobility to start public trials of robotaxi services in Shanghai.

Investment News

• Anshi Biotechnology, a Beijing-based biotech company, has raised over $200 million in its Series A funding round led by Vivo Capital, Bain Capital, and Chunhua Capital. The company aims to build a new-generation targeted therapy platform.
• Alva Systems, a Beijing-based AI & AR solutions provider, has raised hundreds of millions of RMB in its Series B funding round led by SoftBank China Venture Capital. Founded in 2011, the company develops AI-enabled augmented reality solutions and heterogeneous computing for industrial applications.
• EverPro Technologies, a Wuhan-based developer of active optical cable (AOC) products, has raised RMB300 million in its Series B funding round led by Yunfeng Fund and joined by Longzhu Fund and Firstred Capital. EverPro boasts a series of products and solutions, such as active optical cable related modules, optical fibers and cables, active optical cable products, USB3.0 SuperTT Dongle, cloud terminal products, as well as ultra high-speed connections for cloud terminal, machine vision, audio and video entertainment, video conferencing and digital signage based on USB3.0.