Wednesday brought a major shake-up in the automotive community: Volkswagen has acquired a stake in the Chinese EV startup Xpeng. Both partners are planning to jointly develop new electric models. It is well known that VW and other German automakers are lagging behind in China’s electric segment. Now, it seems that the Wolfsburg-based company aims to catch up with the help of its Chinese competitors, as Felix Lee analyzes. Even premium subsidiary Audi is set to form a closer alliance with the Chinese joint venture partner SAIC.
In our interview, we delve into intelligent language models: “In China, there are already more than 80 language models the size of ChatGPT,” says Hans Uszkoreit in conversation with Frank Sieren. The 73-year-old is one of Europe’s leading AI researchers. With his start-up Nyonic, he aims to make Germany a world-class player in AI, working closely with a team from Shanghai. Uszkoreit has a strong connection to China, having served as Chief Scientist of the Artificial Intelligence Technology Center (AITC) in Beijing. His wife, Xu Feiyu, is also a distinguished AI researcher and was recently the head of AI at SAP. Uszkoreit discusses Europe’s prospects in the competition with China and highlights the weaknesses and dangers of applications like ChatGPT.
Meanwhile, regarding the replacement of Foreign Minister Qin Gang in Beijing, there continues to be a chilling silence while his predecessor-successor, Wang Yi, travels the world. On Wednesday, he held his first meetings in his reappointed role as Foreign Minister in Ankara, meeting with his counterpart and President Erdogan. We may never learn what misstep cost Qin his position.
Most automotive market experts had suspected that the Volkswagen Group was facing a tough situation due to poor sales figures. However, this news is likely to make a big impact. Europe’s largest automaker intends to close the gap, particularly in electromobility, by collaborating with Chinese competitors to halt its decline in the world’s largest passenger car market.
The main VW brand will now work together with Chinese EV manufacturer Xpeng on electromobility, software and autonomous driving. The alliance will be strengthened by a nearly 5 percent stake, for which Volkswagen will pay 700 million dollars, as announced by the company after a supervisory board meeting in Wolfsburg on Wednesday.
The initial plan is to jointly develop two electric models for VW’s mid-range segment. These models will be built in Volkswagen’s relatively new factory in Hefei, Anhui Province, in eastern China and are expected to enter the Chinese market in 2026. VW is currently transforming the Hefei site into a new production and development center.
In addition to this, the group revealed that its subsidiary Audi has signed a strategic agreement with joint venture partner SAIC. Initially, the collaboration will involve Audi adopting SAIC’s electric platform for the A3 and A4 models. By 2027, Audi aims to have jointly developed an E-platform with SAIC.
One of the issues Audi currently faces is that the EVs based on Audi’s MEB modular electric drive matrix are perceived as poorly connected by Chinese customers. They also charge too slowly and have less power compared to Chinese competitors. Therefore, SAIC will provide support in the coming years before Audi ventures into the Chinese market with a new platform.
The SAIC platform is intended for mid-range models, while Audi plans to build upper-class electric vehicles with its second partner, FAW. It will be the first time that the VW Group sources the core technology of an electric vehicle from Chinese partners.
Both decisions demonstrate how far Chinese EV manufacturers have outpaced their German competitors. And more significantly, it appears that Chinese competition is now unbeatable without assistance from the People’s Republic.
SAIC has been one of Volkswagen’s two major joint venture partners in China since the beginning of their business in the country. However, the Chinese automaker had looked up to the Germans for nearly four decades to learn from them. Now, this relationship has reversed remarkably quickly.
VW is still trying to downplay its own weaknesses. “We are not buying technology; we are jointly developing vehicles,” asserts Ralf Brandstaetter, head of Volkswagen China, in a conference call with journalists. However, at least until 2027, buying technology is precisely the plan. It is still uncertain whether the Volkswagen Group will genuinely collaborate with Chinese partners to develop electric platforms in the future. The discussions are ongoing.
Xpeng is one of the many newcomers in the Chinese EV segment, and it is not even the most innovative. Established in 2014, the private company, supported by several internet giants and based in Guangzhou, ranks rather low among automakers. According to the China Passenger Car Association (CPCA), it was only twelfth in terms of EV sales in the first quarter. The company relies more on clever marketing concepts than on its size.
Recently, BYD has taken the lead by far. The former battery manufacturer has overtaken Volkswagen as the number one player in the overall market, including conventional combustion engine cars. VW’s strong second position is solely due to its combustion engine cars. In electromobility, VW and Audi are niche players.
In the first half of this year, VW’s EV sales dropped by 1.6 percent to 62,400 cars compared to the same period last year. BYD sells ten times more EVs than VW. As for Audi, it has completely lost ground in the Chinese market for EVs, having sold just over 3,000 battery-powered vehicles in the first quarter.
You recently founded a German ChatGPT start-up with a presence in China. What is the objective?
We established the start-up Nyonic in Berlin, with a research team in Shanghai. Our aim is to address two weaknesses of ChatGPT. Current language models are predominantly focused on English and perhaps Chinese. We intend to change that. Additionally, we want to transform ChatGPT into a reliable work tool for specific industries, which it currently isn’t.
What is missing for that transformation?
ChatGPT needs to provide precise and up-to-date answers, for instance, in aiding automobile development and manufacturing. The current ChatGPT is still at a high school or bachelor’s level. We must elevate it to the standard of automotive engineers. The machine needs to achieve both technical and business reliability; otherwise, it will remain a mere toy.
But why different languages? English is a global language, and Chinese is heading in that direction.
Much European knowledge is not yet available in English, but only in various European languages. Additionally, AI systems should bring the same benefits to users worldwide. Language diversity is a significant step for ChatGPT and crucial for a multipolar world. The majority of the world is multilingual. The US can afford Anglo-centric technology, and China can afford China-focused technology. However, the majority of the world, including Southeast Asia, Africa, and Europe, relies on covering numerous languages and cultures.
Why do you take on this effort? At 73 years old, you are already at retirement age. You are and remain a pioneer in European AI. You could relax.
Everyone on our team, regardless of age, wants to be part of the biggest AI revolution of our lifetime. Additionally, we wish for Europe not to be left behind by the US and China this time. Europe possesses many intelligent individuals, and numerous innovations have come from here. But then we repeatedly failed. China already has more than 80 language models similar to ChatGPT, while the EU only has three. In Germany, there’s only one, which is not yet among the best. We aim to change that.
What about the Chinese who work on this with you?
Most of them, like our CEO Dong Han, are bilingual or multilingual, thus have a natural interest in multilingual AI. Many Chinese models are designed solely for use in China. Some of our experts studied in Germany or other European countries, where they learned about multilingual language technology. But we also recruited top engineers from Google because they want to contribute to our project.
The topic of artificial intelligence is not new. You have been working on it for over 30 years. What’s different about ChatGPT that has sparked so much excitement now?
The system has achieved something we call understanding, similar to humans. It’s defined differently for the machine since it lacks our holistic worldview and doesn’t adapt its knowledge based on user input. But the effect is the same: The machine responds as if it truly comprehends the inputs.
But wasn’t that already the case before?
Yes and no. There were systems that could play chess, enable autonomous driving, do demand forecasting and calculations, translate, compose music and more. Each system needed to be trained with data from its specific domain. It learned from hundreds of thousands, sometimes millions of examples, on how to behave. It’s a bit like training a dog, rewarding it for eating what it should and correcting it when it shouldn’t. The dog eventually learned without knowing what meat, cake or bones are…
… but it has only handled small, trivial tasks.
That’s not much different in previous machine learning. For language AI, it means learning many individual facts, words and terms. Now, with the Transformer technology on which the new AI is based, it can better understand contexts and learn actual meaning from them. It’s like learning about the world outside in a dark hut without windows through thousands of books and the internet. Initially, you only learn words like sun, tree, flowers or car. Over time, you gather so much information that you can depict the outside world more precisely. The sun rises and sets, clouds can cover it, casting shadows. Some flowers follow the sun. The Transformer technology learns to navigate in this world, which increasingly resembles ours but isn’t exactly the same. As a result, the machine can now solve tasks in this world it was never trained for. That’s the significant step of the Transformer technology on which ChatGPT is based, as the name suggests: Generative Pretrained Transformer.
Can you explain this with an example?
I can ask the machine to summarize text about the functions of a car for a ten-year-old. It has never done this before, but it knows what summarizing means and that ten-year-olds use a different language than forty-year-olds. The machine has already written many texts for ten-year-olds, gradually learning how to construct sentences, which words to use, and which ones are too difficult. It describes how a car functions using words a ten-year-old would understand.
Even the average adult would find that challenging.
Yes, although they instantly understand what a car is and the machines have only approximated the image of a car in many small steps. The wonder is that the machine can do it, even though it learned only from plain texts. It learned words, then phrases, then whole sentences, and then complete texts we feed it. All in context, allowing it to create entirely new, meaningful and even sophisticated texts from words and phrases.
So, progress now comes from training the language models rather than new algorithms.
Absolutely. Recent advancements came from data and training, not from new algorithms. Google and Baidu employ more complex algorithms than OpenAI. I try to explain this to policymakers repeatedly. They keep urging us to program algorithms more precisely to prevent “wrong” opinions from forming. However, that’s not the fault of algorithms. The machine must deal with the versatility and contradictions of what it has learned from a vast array of text data. As a result, it sometimes arrives at different conclusions than desired. Often, these conclusions are closer to reality than what some wish for and, thus, considered “wrong”. It’s now about the quantity of text data, the mix, and the way I train the system, but no longer about algorithms.
What are ChatGPT’s remaining weaknesses?
Although ChatGPT is exceptionally intelligent, it still exhibits surprising knowledge gaps that it occasionally fills with imagination. We cannot fully rely on its knowledge and judgment yet.
Will the next version fix these issues?
You might be surprised to learn that there’s currently no reliable method to train the machine to get exactly what I want. It’s all still trial and error, more of an art than a science. There’s no curriculum; I simply feed data into giant tubes and see what happens. We don’t truly know what the machine is doing in the process. When the results come, we say: Well learned, but dear machine, you shouldn’t believe everything you’ve read. And don’t say everything you believe. Then we specifically feed data to correct or improve the machine’s behavior and see what happens.
Is it that straightforward?
Training the machine, feeding it data, is incredibly costly. It costs several million each time. Therefore, we can’t simply try out ten variations. As a result, I don’t really know how good my method truly is because I have relatively few comparisons.
So, it’s possible that we’re blindly following a winding path uphill because we can’t find the cable car.
Yes, and this trial-and-error process isn’t easy to accept for someone seeking reliable, ever-improving methods. Part of what we’re doing is still alchemy. I’m concocting the magic potion. The magic potion works. I can pass on this knowledge, but I don’t know why, only how it works. I only partly understand the side effects. At the same time, we’re talking about machines that could pass a bar exam and work as certified lawyers.
What has surprised you the most in the current development?
Many of my colleagues and I were convinced that the Transformer model would also need to learn incrementally, like humans, and slowly improve through more and better training data. We were amazed when researchers employed brute force, setting more than 100 billion parameters and fed the machine hundreds of billions of words. The system didn’t collapse, as we had feared, nor did it produce nonsense only. We were all baffled.
The machine is already so powerful that it can process very long texts in one step. This is entirely different from humans, who process text incrementally, word by word and sentence by sentence. Often, in the middle of a sentence, humans can predict how it will continue.
Hans Uszkoreit, 73, has worked in Germany, the US and China as a university professor, research manager, industry advisor, and co-founder of several start-ups. He is considered one of Europe’s leading AI researchers. Uszkoreit is the scientific director at the German Research Center for Artificial Intelligence (DFKI). He initiated and coordinated German and international research collaborations and led several renowned European projects. Uszkoreit has authored over 250 international publications. Recently, he co-founded the Berlin-based ChatGPT-like start-up “nyonic” together with a research team in Shanghai.
Uszkoreit has strong ties to China: He served as the Chief Scientist of the Artificial Intelligence Technology Center (AITC) in Beijing. His wife, Xu Feiyu, is also a prominent AI researcher and was previously the head of AI at SAP.
Read the second part of the interview here, which discusses whether a malevolent AI could become dangerous or if the greatest risk still comes from humans.
Taiwan’s armed forces have conducted a maneuver to fend off air attacks on the country’s largest international airport. The exercise on Wednesday is part of the annual Han Kuang military exercises. According to the Financial Times, these annual military drills have recently faced criticism from military analysts, who view them as mere demonstrations and call for more authentic tests involving real attack scenarios. Wednesday’s exercises mark a step towards this transformation, the newspaper says.
The drills were intended to test the military’s cross-sector coordination and emergency response capabilities under the pressure of a potential Chinese invasion, as reported by CNN, citing Taiwan’s Ministry of National Defense.
During the exercise, a formation of Black Hawk helicopters landed at Taiwan Taoyuan Airport, located about 30 km west of the capital, Taipei. Each helicopter deployed six soldiers with machine guns, simulating a Chinese invasion force. The Black Hawks were escorted by a group of Apache attack helicopters that engaged ground-based defenders. After a half-hour battle, the defending forces gained the upper hand and restored control. Firefighters also practiced extinguishing simulated fires.
In other parts of the country, Taiwan’s military canceled some of the multi-day Han Kuang exercises due to the approaching Typhoon Doksuri. The first rainbands from the typhoon reached Taiwan on Wednesday. ck
At Bremen University of Applied Sciences, Prof. Dr. Sandra Heep holds a unique position as the sole professor specializing in China. For several years, she has been leading the program “Applied Business Languages and International Business Management” with a focus on China, as well as overseeing the China Center on campus. Her expertise covers a broad range of subjects, including China’s economic system, its role in the global economy, its political system and its foreign policy.
Heep is content with this diverse role as the head of the China Center, where she advises stakeholders from politics and business on China-specific issues. “Having a broad understanding helps me keep the bigger picture in mind, rather than just focusing on specialized knowledge,” she explains.
However, as the program director, Heep faces the challenge of educating students to become experts on a country that is becoming increasingly repressive internally and more aggressive externally. “It’s not easy to inspire young people to pursue a China-focused study despite the political developments,” she admits. One solution she sees is to promote China savviness outside of sinology, for example, in political science and economics, as well as through accessible programs in schools. “Germany already has a shortage of China experts, and this problem will worsen in the coming years. We urgently need to find new ways to ignite young people’s interest in engaging with China and learning the Chinese language.”
Heep knows the hesitant approach to China from her own biography. The Chinese language had always fascinated her. At the beginning of her philosophy studies at the University of Bonn, she sat in on sinology lectures. But after a few weeks, she dropped it and concentrated on her minor subjects of political science and psychology. “At that time, I couldn’t imagine going to China for a longer period of time,” she recalls. And why learn a language she wouldn’t use?
However, China never completely left her mind. With a scholarship, Heep went to the University of Nanjing after her basic studies, learned Chinese and traveled the country. Back in Germany, she completed her philosophy studies in Berlin. In her political science dissertation, she examined whether China’s state-guided financial system was compatible with exercising power in the international financial system.
She subsequently worked as a guest researcher and scientific staff, including at the Institute of World Economics and Politics of the Chinese Academy of Social Sciences in Beijing and held a position as head of the program “Economic Policy and Financial System” at the Merics Institute. Before being appointed to Bremen in 2018, Heep spent two years advising the German Ministry of Finance in the G20 project for cooperation with China: “In retrospect, it was one of the most crucial periods in my career because it allowed me to become familiar with the workings of the federal government and gain valuable insights for policy advising.”
Influenced by her philosophy studies, Heep places particular importance on conceptual clarity in her research on China. Currently, she is investigating the increasing politicization of the Chinese economic system and whether, given the Communist Party’s growing focus on national security, political control and ideology, this system can still be understood as capitalist. Svenja Napp
Under strict supervision, Yuan Meng is escorted to his aircraft in a transport box. The first panda born in France is leaving Paris, where he was raised at the Beauval Zoo, heading for China, the home of his parents. Hundreds of visitors and zoo staff bid him “Bon Voyage” and wish him the best of luck for his next mission: In Chengdu, in the southwestern Chinese province of Sichuan, Yuan Meng will be taken to a reproduction center where he will contribute to the conservation of his species.
Wednesday brought a major shake-up in the automotive community: Volkswagen has acquired a stake in the Chinese EV startup Xpeng. Both partners are planning to jointly develop new electric models. It is well known that VW and other German automakers are lagging behind in China’s electric segment. Now, it seems that the Wolfsburg-based company aims to catch up with the help of its Chinese competitors, as Felix Lee analyzes. Even premium subsidiary Audi is set to form a closer alliance with the Chinese joint venture partner SAIC.
In our interview, we delve into intelligent language models: “In China, there are already more than 80 language models the size of ChatGPT,” says Hans Uszkoreit in conversation with Frank Sieren. The 73-year-old is one of Europe’s leading AI researchers. With his start-up Nyonic, he aims to make Germany a world-class player in AI, working closely with a team from Shanghai. Uszkoreit has a strong connection to China, having served as Chief Scientist of the Artificial Intelligence Technology Center (AITC) in Beijing. His wife, Xu Feiyu, is also a distinguished AI researcher and was recently the head of AI at SAP. Uszkoreit discusses Europe’s prospects in the competition with China and highlights the weaknesses and dangers of applications like ChatGPT.
Meanwhile, regarding the replacement of Foreign Minister Qin Gang in Beijing, there continues to be a chilling silence while his predecessor-successor, Wang Yi, travels the world. On Wednesday, he held his first meetings in his reappointed role as Foreign Minister in Ankara, meeting with his counterpart and President Erdogan. We may never learn what misstep cost Qin his position.
Most automotive market experts had suspected that the Volkswagen Group was facing a tough situation due to poor sales figures. However, this news is likely to make a big impact. Europe’s largest automaker intends to close the gap, particularly in electromobility, by collaborating with Chinese competitors to halt its decline in the world’s largest passenger car market.
The main VW brand will now work together with Chinese EV manufacturer Xpeng on electromobility, software and autonomous driving. The alliance will be strengthened by a nearly 5 percent stake, for which Volkswagen will pay 700 million dollars, as announced by the company after a supervisory board meeting in Wolfsburg on Wednesday.
The initial plan is to jointly develop two electric models for VW’s mid-range segment. These models will be built in Volkswagen’s relatively new factory in Hefei, Anhui Province, in eastern China and are expected to enter the Chinese market in 2026. VW is currently transforming the Hefei site into a new production and development center.
In addition to this, the group revealed that its subsidiary Audi has signed a strategic agreement with joint venture partner SAIC. Initially, the collaboration will involve Audi adopting SAIC’s electric platform for the A3 and A4 models. By 2027, Audi aims to have jointly developed an E-platform with SAIC.
One of the issues Audi currently faces is that the EVs based on Audi’s MEB modular electric drive matrix are perceived as poorly connected by Chinese customers. They also charge too slowly and have less power compared to Chinese competitors. Therefore, SAIC will provide support in the coming years before Audi ventures into the Chinese market with a new platform.
The SAIC platform is intended for mid-range models, while Audi plans to build upper-class electric vehicles with its second partner, FAW. It will be the first time that the VW Group sources the core technology of an electric vehicle from Chinese partners.
Both decisions demonstrate how far Chinese EV manufacturers have outpaced their German competitors. And more significantly, it appears that Chinese competition is now unbeatable without assistance from the People’s Republic.
SAIC has been one of Volkswagen’s two major joint venture partners in China since the beginning of their business in the country. However, the Chinese automaker had looked up to the Germans for nearly four decades to learn from them. Now, this relationship has reversed remarkably quickly.
VW is still trying to downplay its own weaknesses. “We are not buying technology; we are jointly developing vehicles,” asserts Ralf Brandstaetter, head of Volkswagen China, in a conference call with journalists. However, at least until 2027, buying technology is precisely the plan. It is still uncertain whether the Volkswagen Group will genuinely collaborate with Chinese partners to develop electric platforms in the future. The discussions are ongoing.
Xpeng is one of the many newcomers in the Chinese EV segment, and it is not even the most innovative. Established in 2014, the private company, supported by several internet giants and based in Guangzhou, ranks rather low among automakers. According to the China Passenger Car Association (CPCA), it was only twelfth in terms of EV sales in the first quarter. The company relies more on clever marketing concepts than on its size.
Recently, BYD has taken the lead by far. The former battery manufacturer has overtaken Volkswagen as the number one player in the overall market, including conventional combustion engine cars. VW’s strong second position is solely due to its combustion engine cars. In electromobility, VW and Audi are niche players.
In the first half of this year, VW’s EV sales dropped by 1.6 percent to 62,400 cars compared to the same period last year. BYD sells ten times more EVs than VW. As for Audi, it has completely lost ground in the Chinese market for EVs, having sold just over 3,000 battery-powered vehicles in the first quarter.
You recently founded a German ChatGPT start-up with a presence in China. What is the objective?
We established the start-up Nyonic in Berlin, with a research team in Shanghai. Our aim is to address two weaknesses of ChatGPT. Current language models are predominantly focused on English and perhaps Chinese. We intend to change that. Additionally, we want to transform ChatGPT into a reliable work tool for specific industries, which it currently isn’t.
What is missing for that transformation?
ChatGPT needs to provide precise and up-to-date answers, for instance, in aiding automobile development and manufacturing. The current ChatGPT is still at a high school or bachelor’s level. We must elevate it to the standard of automotive engineers. The machine needs to achieve both technical and business reliability; otherwise, it will remain a mere toy.
But why different languages? English is a global language, and Chinese is heading in that direction.
Much European knowledge is not yet available in English, but only in various European languages. Additionally, AI systems should bring the same benefits to users worldwide. Language diversity is a significant step for ChatGPT and crucial for a multipolar world. The majority of the world is multilingual. The US can afford Anglo-centric technology, and China can afford China-focused technology. However, the majority of the world, including Southeast Asia, Africa, and Europe, relies on covering numerous languages and cultures.
Why do you take on this effort? At 73 years old, you are already at retirement age. You are and remain a pioneer in European AI. You could relax.
Everyone on our team, regardless of age, wants to be part of the biggest AI revolution of our lifetime. Additionally, we wish for Europe not to be left behind by the US and China this time. Europe possesses many intelligent individuals, and numerous innovations have come from here. But then we repeatedly failed. China already has more than 80 language models similar to ChatGPT, while the EU only has three. In Germany, there’s only one, which is not yet among the best. We aim to change that.
What about the Chinese who work on this with you?
Most of them, like our CEO Dong Han, are bilingual or multilingual, thus have a natural interest in multilingual AI. Many Chinese models are designed solely for use in China. Some of our experts studied in Germany or other European countries, where they learned about multilingual language technology. But we also recruited top engineers from Google because they want to contribute to our project.
The topic of artificial intelligence is not new. You have been working on it for over 30 years. What’s different about ChatGPT that has sparked so much excitement now?
The system has achieved something we call understanding, similar to humans. It’s defined differently for the machine since it lacks our holistic worldview and doesn’t adapt its knowledge based on user input. But the effect is the same: The machine responds as if it truly comprehends the inputs.
But wasn’t that already the case before?
Yes and no. There were systems that could play chess, enable autonomous driving, do demand forecasting and calculations, translate, compose music and more. Each system needed to be trained with data from its specific domain. It learned from hundreds of thousands, sometimes millions of examples, on how to behave. It’s a bit like training a dog, rewarding it for eating what it should and correcting it when it shouldn’t. The dog eventually learned without knowing what meat, cake or bones are…
… but it has only handled small, trivial tasks.
That’s not much different in previous machine learning. For language AI, it means learning many individual facts, words and terms. Now, with the Transformer technology on which the new AI is based, it can better understand contexts and learn actual meaning from them. It’s like learning about the world outside in a dark hut without windows through thousands of books and the internet. Initially, you only learn words like sun, tree, flowers or car. Over time, you gather so much information that you can depict the outside world more precisely. The sun rises and sets, clouds can cover it, casting shadows. Some flowers follow the sun. The Transformer technology learns to navigate in this world, which increasingly resembles ours but isn’t exactly the same. As a result, the machine can now solve tasks in this world it was never trained for. That’s the significant step of the Transformer technology on which ChatGPT is based, as the name suggests: Generative Pretrained Transformer.
Can you explain this with an example?
I can ask the machine to summarize text about the functions of a car for a ten-year-old. It has never done this before, but it knows what summarizing means and that ten-year-olds use a different language than forty-year-olds. The machine has already written many texts for ten-year-olds, gradually learning how to construct sentences, which words to use, and which ones are too difficult. It describes how a car functions using words a ten-year-old would understand.
Even the average adult would find that challenging.
Yes, although they instantly understand what a car is and the machines have only approximated the image of a car in many small steps. The wonder is that the machine can do it, even though it learned only from plain texts. It learned words, then phrases, then whole sentences, and then complete texts we feed it. All in context, allowing it to create entirely new, meaningful and even sophisticated texts from words and phrases.
So, progress now comes from training the language models rather than new algorithms.
Absolutely. Recent advancements came from data and training, not from new algorithms. Google and Baidu employ more complex algorithms than OpenAI. I try to explain this to policymakers repeatedly. They keep urging us to program algorithms more precisely to prevent “wrong” opinions from forming. However, that’s not the fault of algorithms. The machine must deal with the versatility and contradictions of what it has learned from a vast array of text data. As a result, it sometimes arrives at different conclusions than desired. Often, these conclusions are closer to reality than what some wish for and, thus, considered “wrong”. It’s now about the quantity of text data, the mix, and the way I train the system, but no longer about algorithms.
What are ChatGPT’s remaining weaknesses?
Although ChatGPT is exceptionally intelligent, it still exhibits surprising knowledge gaps that it occasionally fills with imagination. We cannot fully rely on its knowledge and judgment yet.
Will the next version fix these issues?
You might be surprised to learn that there’s currently no reliable method to train the machine to get exactly what I want. It’s all still trial and error, more of an art than a science. There’s no curriculum; I simply feed data into giant tubes and see what happens. We don’t truly know what the machine is doing in the process. When the results come, we say: Well learned, but dear machine, you shouldn’t believe everything you’ve read. And don’t say everything you believe. Then we specifically feed data to correct or improve the machine’s behavior and see what happens.
Is it that straightforward?
Training the machine, feeding it data, is incredibly costly. It costs several million each time. Therefore, we can’t simply try out ten variations. As a result, I don’t really know how good my method truly is because I have relatively few comparisons.
So, it’s possible that we’re blindly following a winding path uphill because we can’t find the cable car.
Yes, and this trial-and-error process isn’t easy to accept for someone seeking reliable, ever-improving methods. Part of what we’re doing is still alchemy. I’m concocting the magic potion. The magic potion works. I can pass on this knowledge, but I don’t know why, only how it works. I only partly understand the side effects. At the same time, we’re talking about machines that could pass a bar exam and work as certified lawyers.
What has surprised you the most in the current development?
Many of my colleagues and I were convinced that the Transformer model would also need to learn incrementally, like humans, and slowly improve through more and better training data. We were amazed when researchers employed brute force, setting more than 100 billion parameters and fed the machine hundreds of billions of words. The system didn’t collapse, as we had feared, nor did it produce nonsense only. We were all baffled.
The machine is already so powerful that it can process very long texts in one step. This is entirely different from humans, who process text incrementally, word by word and sentence by sentence. Often, in the middle of a sentence, humans can predict how it will continue.
Hans Uszkoreit, 73, has worked in Germany, the US and China as a university professor, research manager, industry advisor, and co-founder of several start-ups. He is considered one of Europe’s leading AI researchers. Uszkoreit is the scientific director at the German Research Center for Artificial Intelligence (DFKI). He initiated and coordinated German and international research collaborations and led several renowned European projects. Uszkoreit has authored over 250 international publications. Recently, he co-founded the Berlin-based ChatGPT-like start-up “nyonic” together with a research team in Shanghai.
Uszkoreit has strong ties to China: He served as the Chief Scientist of the Artificial Intelligence Technology Center (AITC) in Beijing. His wife, Xu Feiyu, is also a prominent AI researcher and was previously the head of AI at SAP.
Read the second part of the interview here, which discusses whether a malevolent AI could become dangerous or if the greatest risk still comes from humans.
Taiwan’s armed forces have conducted a maneuver to fend off air attacks on the country’s largest international airport. The exercise on Wednesday is part of the annual Han Kuang military exercises. According to the Financial Times, these annual military drills have recently faced criticism from military analysts, who view them as mere demonstrations and call for more authentic tests involving real attack scenarios. Wednesday’s exercises mark a step towards this transformation, the newspaper says.
The drills were intended to test the military’s cross-sector coordination and emergency response capabilities under the pressure of a potential Chinese invasion, as reported by CNN, citing Taiwan’s Ministry of National Defense.
During the exercise, a formation of Black Hawk helicopters landed at Taiwan Taoyuan Airport, located about 30 km west of the capital, Taipei. Each helicopter deployed six soldiers with machine guns, simulating a Chinese invasion force. The Black Hawks were escorted by a group of Apache attack helicopters that engaged ground-based defenders. After a half-hour battle, the defending forces gained the upper hand and restored control. Firefighters also practiced extinguishing simulated fires.
In other parts of the country, Taiwan’s military canceled some of the multi-day Han Kuang exercises due to the approaching Typhoon Doksuri. The first rainbands from the typhoon reached Taiwan on Wednesday. ck
At Bremen University of Applied Sciences, Prof. Dr. Sandra Heep holds a unique position as the sole professor specializing in China. For several years, she has been leading the program “Applied Business Languages and International Business Management” with a focus on China, as well as overseeing the China Center on campus. Her expertise covers a broad range of subjects, including China’s economic system, its role in the global economy, its political system and its foreign policy.
Heep is content with this diverse role as the head of the China Center, where she advises stakeholders from politics and business on China-specific issues. “Having a broad understanding helps me keep the bigger picture in mind, rather than just focusing on specialized knowledge,” she explains.
However, as the program director, Heep faces the challenge of educating students to become experts on a country that is becoming increasingly repressive internally and more aggressive externally. “It’s not easy to inspire young people to pursue a China-focused study despite the political developments,” she admits. One solution she sees is to promote China savviness outside of sinology, for example, in political science and economics, as well as through accessible programs in schools. “Germany already has a shortage of China experts, and this problem will worsen in the coming years. We urgently need to find new ways to ignite young people’s interest in engaging with China and learning the Chinese language.”
Heep knows the hesitant approach to China from her own biography. The Chinese language had always fascinated her. At the beginning of her philosophy studies at the University of Bonn, she sat in on sinology lectures. But after a few weeks, she dropped it and concentrated on her minor subjects of political science and psychology. “At that time, I couldn’t imagine going to China for a longer period of time,” she recalls. And why learn a language she wouldn’t use?
However, China never completely left her mind. With a scholarship, Heep went to the University of Nanjing after her basic studies, learned Chinese and traveled the country. Back in Germany, she completed her philosophy studies in Berlin. In her political science dissertation, she examined whether China’s state-guided financial system was compatible with exercising power in the international financial system.
She subsequently worked as a guest researcher and scientific staff, including at the Institute of World Economics and Politics of the Chinese Academy of Social Sciences in Beijing and held a position as head of the program “Economic Policy and Financial System” at the Merics Institute. Before being appointed to Bremen in 2018, Heep spent two years advising the German Ministry of Finance in the G20 project for cooperation with China: “In retrospect, it was one of the most crucial periods in my career because it allowed me to become familiar with the workings of the federal government and gain valuable insights for policy advising.”
Influenced by her philosophy studies, Heep places particular importance on conceptual clarity in her research on China. Currently, she is investigating the increasing politicization of the Chinese economic system and whether, given the Communist Party’s growing focus on national security, political control and ideology, this system can still be understood as capitalist. Svenja Napp
Under strict supervision, Yuan Meng is escorted to his aircraft in a transport box. The first panda born in France is leaving Paris, where he was raised at the Beauval Zoo, heading for China, the home of his parents. Hundreds of visitors and zoo staff bid him “Bon Voyage” and wish him the best of luck for his next mission: In Chengdu, in the southwestern Chinese province of Sichuan, Yuan Meng will be taken to a reproduction center where he will contribute to the conservation of his species.
