Beijing Issues First-Ever Driverless Robotaxi Permits; Huawei Anticipates AI Compute to Jump 500 Times by 2030; CogView2 Challenges DALL-E-2 With Better Results

Weekly China AI News from April 24 to May 1

Dear readers, last week was a bonanza for autonomous driving enthusiasts, from Baidu and Pony.ai securing first-ever driverless ride-hailing permits, to Guangzhou-based Ruqi Mobility raking in one billion yuan from autonomous driving companies. Meanwhile, BAAI, the “Chinese OpenAI”, just released their second-gen text-to-image generation CogView2 as a Chinese equivalent of DALL-E-2. Read more below.

News of the Week

Baidu, Pony.ai Secure China’s First-Ever Driverless Permit to Offer Passenger Rides

Chinese search engine giant Baidu and autonomous driving upstart Pony.ai received a batch of permits from Beijing authority that allow them to deploy driverless robotaxis for ride-hailing. For the first time in China, residents can hail a driverless robotaxi on public streets without pre-screening.

How it works: Permits indicate the removal of drivers from the seat behind the steering wheel. However, a safety operator still must be present in the front passenger seat. Baidu exec estimated a fully driverless robotaxi will pass the regulation to hit the street by the end of 2022. 10 licensed driverless robotaxis from Baidu and four from Pony.ai will provide services in an area of 60 square kilometres in Yizhuang, a suburban town of southeast Beijing. Without a commercial driverless permission, the pilot program is free now. Users can hail a Baidu robotaxi ride from 10:00 to 16:00 and a Pony.ai ride from 9:00 to 17:00.

Why it matters: The driverless approval in Beijing represents a regulatory tipping point in China’s autonomous driving industry that transits from a driverless testing stage to a driverless operation stage. On the other hands, Beijing has always been the benchmark city in China in the field of autonomous driving policy so the regulation will prompt other cities to follow the steps.

China’s catching up: The permit is the latest addition to China’s fast-pace autonomous driving policemaking with the aim to catch up with the U.S. Waymo, the Alphabet’s autonomous vehicle venture, has been providing fully driverless robotaxis in a small town of Phoenix since 2019. Waymo and GM’s Cruise are also granted to offer fully driverless rides in San Francisco, with pre-screening needed. Earlier this year, the Federal vehicle safety regulators have cleared the way for the production and deployment of driverless vehicles that do not include manual controls such as steering wheels or pedals.

Pony.ai last week also received the first-ever taxi license in China that authorizes the company to operate 100 autonomous vehicles as traditional taxis in Nansha, Guangzhou. The robotaxi still retains a safety driver.

AI Compute to Grow 500 Times by 2030, Huawei Chairman Says

Powerful ML/AI applications require a large amount of computational power. At Huawei’s annual global analyst summit held in Shenzhen, Huawei rotating chairman Ken Hu said the AI computing power will increase by a factor of 500, to more than 100 ZFLOPS.

In numbers, Huawei predicts that by 2030, there will be 3.3 ZFLOPS of general computing power (FP32) available, a 10-fold increase over 2020; and 105 ZFLOPS of AI computing power (FP16), a 500-fold increase over 2020.

ZFLOPS? FLOPS is short for floating point operations per second. 1 exaFLOPS (EFLOPS) = 10¹⁸ FLOPS. 1 zettaFLOPS (ZFLOPS) = 10²¹ FLOPS

Methodology: The conclusion was originated from Huawei’s Computing 2030 report and first made headlines in 2021, but Huawei didn’t explain its methodologies. Today’s fastest supercomputer is the Japanese Fugaku which reaches 442.01 petaFLOPS since June, 2020. Meta’s upcoming supercomputer RSC will also be the fastest AI supercomputer in the world, performing at nearly 5 exaflops of mixed precision compute. Intel said they will produce a machine with one ZFLOPS in 2027.

Why it matters: Predicting the global general and AI computing capacity is difficult. Usually the amount of compute is correlated to annual hardware spending. But the anticipation signifies the growing demand of AI compute and strategic advantages of AI companies that have invested into AI compute centers. Tesla for example has unveiled its supercomputer Dojo would perform 112.32 PetaFLOPS of computing power in FP16.

Inspur, China’s largest server company, also made its prediction: by 2025, the global AI compute capacity will reach 6.8 ZFlops, a 30-fold increase over 2020, which is in line with Huawei’s prediction.

Tsinghua U, BAAI Reveal Text-to-Image Generation Model CogView 2

Researchers from Tsinghua University, Beijing Academy of Artificial Intelligence (BAAI), last week introduced the second generation of its text-to-image generation model CogView 2, following OpenAI’s release of DALL-E-2.

What is CogView? CogView is a 4-billion-parameter Transformer combined with VQ-VAE for general-domain text-to-image generation. The model can also be finetuned for various downstream tasks, e.g. style learning, super-resolution, text-image ranking and fashion design. The paper was accepted to NeurIPS 2021.

The CogView2 made an improvement to address slow generation and complexity for high-resolution images, using hierarchical transformers and local parallel auto-regressive generation. CogView2 is a 6B-parameter transformer with a simple and flexible self-supervised task, Cross-modal general language model (CogLM), and finetune it for fast super-resolution. You can read the paper CogView2: Faster and Better Text-to-Image Generation via Hierarchical Transformers for more.

CogView vs DALL-E-2: Chinese researchers boasted the new text-to-image system, shows very competitive generation compared to concurrent state-of-the-art DALL-E-2. The system is said to create better results with Chinese text inputs. You can play with the demo here.

OpenAI hasn’t released a publicly available demo or open-sourced the codes but a controlled demo. Below is an image generation result from New York Times.

We input the same-meaning Chinese text description into the CogView2 system and waited for about three minutes. Below is the closest one I chose.

Papers & Projects

Tencent’s Go-Playing AI Defeats a Human Champion in 1-on-1 Mahjong

Fine Arts is a Go-playing computer program created by Chinese social media and gaming giant Tencent. The AI system once defeated human Go Champion Ke Jie. Now its creators eye on Mahjong. In the paper Actor-Critic Policy Optimization in a Large-Scale Imperfect-Information Game, researchers from Tencent and other elite Chinese universities extended the actor-critic algorithm framework in deep reinforcement learning to tackle a large-scale 2-player zero-sum imperfect- information game, 1-on-1 Mahjong, whose information set size and game length are much larger than poker. The proposed algorithm, named Actor-Critic Hedge (ACH), modifies the policy optimization objective from originally maximizing the discounted returns to minimizing a type of weighted cumulative counterfactual regret.

First Explainable Natural Language Analogical Reasoning Dataset

“Raising children is like gardening — nurture them and be patient.” Humans often use analogies to better explain something using a similar comparison. In the paper E-KAR: A Benchmark for Rationalizing Natural Language Analogical Reasoning, researchers from Fudan University, ByteDance AI Lab, South China University of Technology, and University of California Santa Barbara, proposed a first-of-its-kind Explainable Knowledge-intensive Analogical Reasoning benchmark (E-KAR). The benchmark consists of 1,655 (in Chinese) and 1,251 (in English) problems sourced from the Civil Service Exams, which require intensive background knowledge to solve. Empirical results suggest that this benchmark is very challenging for some state-of-the-art models for both explanation generation and analogical question answering tasks, which invites further research in this area.

Rising Startups

Ruqi Mobility, a ride-hailing platform backed by Tencent and GAC Group, has raised over RMB1 billion yuan ($157 million) in its Series A funding round. Founded in 2018, the Guangzhou-based company operates fuel-efficient, hybrid electric vehicles with well-trained drivers. What’s noteworthy is Pony.ai and WeRide also participate in the funding, signaling their willingness to add robotaxis to commercial ride-hailing platforms.

Han’s Robot, an intelligent robot startup built by Han’s Laser Technology Industry Group, has raised nearly RMB200 million yuan ($30.3 million) in its Series B+ funding round. Founded in 2017, the Shenzhen-based company has developed collaborative robot for automated production lines and multi-sensing autonomous vehicles for deliveries.