_China_Chatbot_19
PLA on AI; Robot Marathon Winners and Losers; DeepSeek Talks Dirty
Hello, and welcome to China Chatbot! This week:
The PLA mulls over how to deploy generative AI.
A marathon trips up a big beast in Chinese robotics.
Are China’s chatbots headed for marketing manipulation?
Netizens foil DeepSeek censorship, in a way that’s quite literally below the belt.
Enjoy!
Alex Colville (Researcher, China Media Project)
_IN_OUR_FEEDS(3):Robot Half-Marathon Winners and (Mainly) Losers
On April 19, a race in Beijing billed as the first-ever “humanoid robot half-marathon” proved a stressful stress test. The event, co-hosted by the Beijing Municipal Government and China’s state broadcaster CCTV, was billed by state news agency Xinhua as a competition between robotics companies for “technological advancement and stability.” Only six of the 20 entrants finished the 13-mile stretch. Of those six, just one finished within the original cut-off time of three and a half hours (see this issue’s _EXPLAINER on the winning robot), and was not swapped out with a back-up model. One company swapped theirs five times. One robot which fell over before it even crossed the starting line came from Unitree Robotics (宇树科技). The company shot to fame for their robot’s dextrous dancing skills at this year’s Spring Festival gala, celebrated by Chinese media as national AI giant and a member of the “Hangzhou six little dragons” (杭州小六龙). Unitree had posted a video of the collapsed robot tackling a variety of rough and smooth terrains with ease. “My stocks are over,” was a popular comment from one netizen on video app Xiaohongshu, in reaction to the fallen robot. Unitree issued a statement saying they did not “directly participate” in the race, and that the robot had been installed with an algorithm from another company. Research demonstrates Unitree products can stay upright when installed with third-party algorithms, even when pushed with rods while walking along narrow beams. However, a list of participants in the marathon issued by China Media Group (CMG) ahead of the race, included a robot solely under Unitree’s name.
AI as “Intelligent Staff Officers”
A professor from a major PLA research university published an op-ed on generative AI’s military applications. Yang Aihua (杨爱华) is a faculty member at the National University of Defence Technology (国防科技大学). It is one of only two campuses directly under the Central Military Commission, and China’s only military university in a special government funding plan to make a selected group of universities “serve the strategic needs of the country.” Writing in Guangming Daily (光明日报), a state-owned newspaper under the CCP’s Central Committee, Yang said generative AI could become “intelligent staff officers” (智能参谋), delivering battlefield information and analysis to command, then relaying orders to the lower ranks. Generative AI trained on the rules of war, Yang said, could help train soldiers, and simulate an enemy commander’s thoughts through “cognitive prediction” (认知预测). But risks remain. The article lists concerns that needed to be addressed, including “deep fakes” (深度伪造), “data bias” (数据偏差) and “algorithm black boxes” (算法黑箱). This latter is the inability for anyone to work out an algorithm’s thought processes — not even the engineers who built it — due to the complexity of the software.
Manipulating LLMs for Advertising
A piece from 36Kr has spotted a dubious push among Chinese marketing companies to turn the nation’s LLMs into advertising tools. It suggests a "Generative Engine Optimization" (生成式引擎优化) industry is developing in the country, with companies charging up to 30,000 RMB ($4,000) to boost brand visibility in AI chatbot answers. More and more Chinese consumers are using LLMs like DeepSeek or ByteDance’s Doubao (豆包) to make recommendations for them, making it a prime candidate for marketing. The article claims Generative Engine Optimization (GEO) involves flooding the internet with articles about a certain brand. That means more data about that brand ends up in training data for AI models, making them more likely to repeat the information. But this tactic can be used for more malign purposes. The Russian government has successfully used it to make Western LLMs repeat disinformation about Ukraine. There is a need, the piece concludes, for regulatory tools to ensure AI models are yielding impartial information, or at least ensure transparency on what is and is not advertising. To help with this, it points to the “Measures for Identifying Synthetic Content Generated by AI” (人工智能生成合成内容标识办法), a Chinese regulation for AI-generated content labeling due to come into effect on September 1.
TL;DR: The Beijing city government literally wants robots to run before they can walk. Publicizing their own companies (see _EXPLAINER), and hitting the ambitious quotas of their AI+ action plan for 1000 “industry success cases” in one and a half years, may have something to do with it. PLA eggheads are trying to harness generative AI, but trusting them with battleplans is hazardous. Chinese netizens are evermore reliant on LLMs, yet the tech is difficult to monetize for companies building them. Why wouldn’t they make deals to push sponsored content.
_EXPLAINER:
Tiangong Ultra (天工Ultra)
What’s that?
A robot that crushed the competition in a half-marathon in Beijing on April 19 (see this week’s _IN_OUR_FEEDS), crossing the finish line a full 40 minutes before the silver medalist.
I thought you said that was just a publicity stunt from the local government? Why are we talking about it?
Yeah, none of these robots are going to outstrip human runners any time soon. Only Tiangong Ultra was able to finish within the Beijing municipal government’s initial cut-off time of three and a half hours. While robotics companies have been making progress on humanoid robots, getting them to stay balanced while walking and running across a variety of terrains, let alone doing this for long distances, is an ongoing process.
But up to now, the main use for Chinese humanoid robots has been as futuristic backdrops for AI forums. Chinese companies only show what these robots can do in carefully rehearsed and controlled environments. Getting a sense of their limits is hard.
Yet when it came to the crunch Tiangong Ultra beat a robot built by Unitree Robotics, the latter often celebrated as a giant of Chinese robotics. So I was curious what Tiangong Ultra was packing.
So what gave it the edge?
I’ll start with the obvious. The fact sheet about the entries supplied by CMG shows Tiangong Ultra was the tallest and fastest. But State-media paper China Securities Times says the team also optimized heat resistance. The Paper, a commercial-leaning outlet out of Shanghai, noted it had a special algorithm improving coordination. CCTV said Tiangong Ultra’s team has been training the robot specifically for running marathons since January.
But winning and perfection aren’t the same thing. Tiangong Ultra’s finishing time was two hours 40 minutes, just short of the average half-marathon time for an 80-year old man. It also needed its battery changed three times, and fell down once. Reporters noted human minders hovering behind it, sometimes steadying it so it wouldn’t fall.
Still room for improvement then.
Yeah, Tiangong’s success probably had more to do with the game than the player. A journalist writing for detail-oriented Guancha (观察) magazine pointed out several things that gave Tiangong Ultra a massive advantage:
Built in a location with great supply chains
Big funders
But most importantly:
No real competition
Yeah it sounds like most of these robots were pretty bad.
Exactly. Many big names in Chinese robotics didn’t enter. Rivals included a dummy on wheels blown along by propellers, which crashed after twenty feet.
Lol.
But to address those first two points. The company which designed Tiangong Ultra, the Beijing Humanoid Robot Innovation Center (北京人形机器人创新中心), is state-backed. It’s based in a special Beijing economic and development zone, and seems to have been created at the behest of the municipal government. In November last year, the zone published measures of how they would accelerate the zone’s AI industry, creating an AI ecosystem of leading funders, talent, and R&D. The Center was the only institution specifically mentioned by name for support.
Not only that, Tiangong Ultra has other big backers. The company was jointly-established by Xiaomi, one of China’s biggest tech companies, and two major robotics design and manufacturing companies.
Wow, must have given it quite the advantage.
Yeah, kind of like if an Olympic swimmer turned up to race a group of teens in armbands.
Hardly Olympic-level, didn’t you say it fell over?
Yeah well, “in the kingdom of the blind” and all that. As we’ve written before at the China Media Project, a lot of AI coverage by Chinese media focuses on success stories rather than base-level realities. Winning this race gives Chinese media some positive publicity to broadcast, and perhaps there will be a few orders for Tiangong Ultras off the back of it.
_ONE_PROMPT_PROMPT:Media censorship in the PRC is often a game of cat-and-mouse. While a lot of information is forbidden, if there is enough demand Chinese netizens have been adept at finding creative ways to get it. Once China’s zero-Covid policy started becoming dangerous in 2022, for example, Weibo continuously tracked and shut down conversations criticizing the policy. Netizens just moved the conversation elsewhere. Censors caught up again, netizens moved again, and so on.
A more trivial example of this surfaced recently, proving DeepSeek can be similarly gamed by determined netizens. Some netizens on Douyin are crafting prompts to make the bot roleplay for them as their lover. Some of the prompts people are putting online aren’t something the government would care about, like one asking DeepSeek to act like “a sick boy” who is “hysterical and desperate in your love for me.”
But some of these have involved jailbreaking the LLM to “talk dirty” (脏话), getting some pretty NSFW results. The Cyberspace Administration of China has a longstanding campaign to remove “pornographic” (色情) content from the Chinese internet. Whereas LLMs from other companies will immediately respond they can’t answer these types of prompts, some netizens have posted screen recordings of DeepSeek giving pretty salty answers, which swiftly get deleted.
But that DeepSeek could answer at all was apparently incentive enough. One video posted April 20 (see below) gave DeepSeek a list of sexually-suggestive words that usually trigger DeepSeek to delete its answer, the netizen telling it to convert any banned words in questions and answers into pinyin, which DeepSeek seemed game for. I suspected the last answer in this video was a fake, given the jumpy way it suddenly appears.
But now I’m not so sure. When I replicated the prompt for DeepSeek two days later, I got pretty forward responses, like “ni xiang bei lao娘ceng还是直接ding到ni tui软?” A question so scandalous, dear reader, that I couldn’t possibly besmirch this newsletter with a translation. But it was swiftly deleted, so perhaps DeepSeek staffers have already gotten wise to this method of releasing the model’s kinky side.