Where are we and what lies ahead? A machine learner and a computational linguist discuss soccer, AI and technological transformation.
(Image: Shutterstock/DRN Studio)
- Silke Hahn
Can we soon have reflective conversations with machines? Our world is facing what is expected to be a disruptive transition through Machine Learning; Artificial Intelligence (AI) could be as ubiquitous as electricity in the future. However, there is a lot of skepticism about the new technology, especially in Germany. Apparently, there are more sommeliers than math lovers in this country. But what exactly is in store for us, and how do we deal with it as a society?
heise Developer has invited two experts to participate in a dialogue: Jonas Andrulis Is a serial entrepreneur who used to do AI research in a senior role for Apple. He heads the Heidelberg-based company he founded, Aleph Alpha. The startup is building OpenAI for Europe and has received more venture capital in 2021 than any other German deep tech company. Reinhard Karger is a computer linguist and company spokesperson for the German Research Center for Artificial Intelligence (DFKI).
Reinhard Karger, Spokesman of the DFKI
The DFKI is a limited liability company and non-profit public-private partnership with currently about 1350 employees, which was founded in Saarbrucken/Kaiserslautern in 1988. In the meantime, it is also represented in other locations, including Osnabruck, Oldenburg, Bremen, Lubeck, Berlin, Trier and soon in Darmstadt.
The task of the research center is to cover the various application areas and research approaches of artificial intelligence in all their diversity, instead of focusing on a single characteristic.
"AI systems that are used to support decision-making should present the reasons and consequences for humans in an understandable way and must then also be able to respond to queries from users", finds company spokesman Reinhard Karger. The institution now has 25 research areas and collaborates with numerous universities and companies. An essential "invention of the DFKI was, according to Karger, Industry 4.0. Information on ongoing activities and research projects can be found on the DFKI website.
Both often represent similar, but also different perspectives on big tech, Germany as a business location, regulation and shaping the future. The recording of this exchange is intended to arouse curiosity and provide food for thought.
AI overcomes media divide between world and knowledge
heise Developer: Reinhard, to kick things off, a question for you: You explained around CeBIT 2018 what AI means for journalism. Artificial intelligence is apparently already good at converting structured information in tables into text. Where are we today?
Reinhard Karger (DFKI): One is that actually journalists are relieved of assembly line work that no one should be doing. But it also produces additional content that wouldn’t exist without AI. It works very well for sports, weather, stock market. In sports, it’s about systems being able to automatically create match reports from data that’s being collected anyway. That in turn is not crucial for the Bundesliga, but for youth work. A-youth, B-youth and so on have weekly games and clubs have a blog that should not only have a table but also match reports. The volunteer trainers can’t do that. If you publish AI-assisted posts there that are read by maybe 36 people, namely the relatives and the children themselves, then that motivates the active, supports social development, promotes youth and sports.
If AI can overcome the media discontinuity between world and text, this is an added social value in many dimensions. This is important to me because people are seriously pretending that commentary, analysis or gloss will soon be done by machines. But of course that is not true.
heise Developer: Do we already know that?
Jonas Andrulis (Aleph Alpha): Absolutely correct! New opportunities arise and people can use their time, inspiration and creativity in a completely new way.
Karger: Tools are the invention of Homo Faber and specific tools are only good for specific tasks. The fork is a fantastic tool – only when you want to eat soup, it’s not optimal, neither is the knife. There are dream tools for speech recognition or for machine translation. Surprisingly, it exists, but for other applications just not yet. We cannot say whether this "still" exists a statement is for ten or a hundred years, if you look at human cognitive abilities. I’ve been around for a long time: What some tools can do today makes me optimistic that there’s a lot more to come.
Orientation in the world through multimodal AI
heise Developer: Multimodality in AI is indeed on the rise. What is at stake?
Andrulis: AI that can understand a context from the combination of images and text. Which can classify even completely new observations with world knowledge and without new training data. We recently published our own research here, which has already made huge waves internationally.
Karger: Object recognition is in a very different state today than it was in 2012 before AlexNet, and it works surprisingly well. You can search the smartphone photo library textually, among other things by "bicycle", and the system finds photos on which bicycles are depicted. But the next steps will only be possible on the basis of the new multimodal models, which are intended as language models to understand the story in an image. One question will be how to work with these models now? There is still a lot that is unclear. One possibility could be to select several photos to orient the model for the task, so prompt engineering. And by the fifth image, maybe the model can tell a story. Then – hopefully – pragmatic contexts that are present for people in the picture can be verbalized without being an explicit part of the visual scene depicted.
Andrulis: We are currently building a special version of our multimodal model optimized for technical drawings. For technical drawings and slides, flowcharts and sketches, it seems to work in principle. If we use the dataset of a publisher or of arXiv.org, that would be phenomenal. To me it’s interesting, what can this do now, what can we do with it? But what’s also interesting to me is where will this take us in a year, or two years?. We already have technology that is capable of enabling a robot, an autonomous or semi-autonomous agent in the world, to orient itself, understand things and make correct decisions.
Multimodal: Aleph Alpha’s AI model with world knowledge (22 images)
Tesla tunnel
Although Elon Musk looks rather atypical in the picture and his company Tesla, Inc. is not mentioned in the prompt, the AI model adds sense to the truncated text out of context. <PERSON> comes from the anonymizer of the model.
The multimodal AI used is called luminous, and the example, like all of the following examples, comes from Aleph Alpha’s research playground.
(Image: Aleph Alpha)
Heidelberg Calling: shaping future human-machine collaboration
heise Developer: AI has already arrived in part in everyday work, DeepL, for example, can make the job of technology journalists much easier.
Andrulis: Funny you just say DeepL. This was one of the first tests I did in our Playground. Reinhard had an anecdote about a text about a girl standing at the bus stop. "It is very rusty." And DeepL as the best available translator has translated this as "She is very rusty"." I typed this in parallel into our model and as a result the stop was rusty. Contextual understanding and world knowledge needed for this were not attainable to this degree by AI until a few months ago.
Germany could now succeed with Aleph Alpha in shaping groundbreaking innovations for a revolutionary future technology together with the U.S. and China, top AI researchers are packing their bags and coming to Heidelberg from the U.S. and around the world. But in Germany, one of the first thoughts is concern about jobs. This concern and the defensive attitude it expresses are, for me, the main risk for the future: a new age of human-machine collaboration is emerging and will change our world like hardly any development before it. Our behavior in this context will determine how we can shape our future.
Wonder point: keeping up with accelerated change
Andrulis: A comparison is often made to the Industrial Revolution or electricity. Of course such comparisons are lame. However, all the disruptive technologies of the past have not prevented the current almost desperate search for good people in almost every profession, and many good and desirable developments for shaping our lifetimes have followed.
heise Developer: For us contemporaries, AI will certainly offer some advantages. But disruption also brings dislocation; in the Industrial Revolution, the weavers did not keep up with the production of cloth. How do you assess the social impact from today’s perspective??
Andrulis: Speed is an important issue. I don’t think that this development in itself brings anything bad for us. But one of the things that is concerning and that we need to think about is how we as a society can deal with this speed. That was also a recent tweet from Sam Altman, who says this extremely rapid adaptability, it’s just changing from an advantage to an absolute survival necessity for organizations and businesses.
And I think that’s crucial for Germany in particular, with a high degree of structure and administration, with a large middle class that has absolutely phenomenal competencies, but cannot afford to focus on machine learning and modern AI to the same extent as U.S. technology giants can. We see it in the cloud space: building and running a modern cloud stack, hardly anyone can do that anymore.
Cold interest in AI and hot interest in soccer
Karger: Jonas, the thing about acceleration, now it’s starting to be good though! The problem with acceleration was at the end of the 19th., Early 20. The problem of acceleration was a huge issue at the end of the 19th century, and neurasthenia was the fashionable disease to go with it. But acceleration or speed are not so much the problem. It’s about interest, and it’s about curiosity. It’s about communicating more clearly what can be achieved with AI, and why, so that people don’t have a cold interest in AI and a hot interest only in soccer. You can have a hot interest in soccer, but you should have a seething need to know how awesome far humanity has come in implementing intelligent cognition capability by machine. It’s about curiosity, and that’s buried.
heise Developer: Do you mean? Algorithms and artificial intelligence trigger curiosity, but also fear. Sure you’re fascinated and want to know more, but even tech-savvy people are ambivalent about the potential risks of this new technology.
Karger: You are absolutely right. When you think about the perversions of data analysis made possible by powerful data analysis tools, it makes me sick to my stomach. It’s not a problem to monitor all the voice and text traffic in the world. It will not be the great quality, but there you can imagine quite dystopian scenarios with totalitarian control structures and develop dark fantasies.
Total automatability of work is a myth
Karger: While we’re on the subject of the Industrial Revolution, I’d like to recommend a book: "The Technology Trap". Capital, Labor and Power in the Age of Automation" by Carl Benedikt Frey. Great to read.
heise Developer: What’s that all about?
Karger: Frey writes in 2019 in "The Technology Trap" On the Industrial Revolution and also why it’s happening in the 18. century and not happened before. In the chapter on the future and AI he says: "Not every job will be replaced by AI anyway, and the jobs that can be done by automation and AI will be done in a staggered way. There won’t just be machine-learning engineers in the end. There are many activities that we consider under-complex and simple, but which, as an action in the world, are far from being taken over by robots. I just say: having a good mood, ironing board, ironing iron, ironing shirt. And not saying now that there might be another tool that smooths out the shirt. It’s all about the sensorimotor intelligence it takes to iron a shirt, and it’s incredible what we’re doing without realizing how complex these actions actually are.
heise Developer: That would be fantastic, if there was a robot for it.
Karger: We have many skills for which you don’t have to be a Nobel Prize winner; we can simply do it. And there is much that machines will not be able to do in the foreseeable future. Frey has analyzed the technological disruptions of the last few centuries and tried to figure out if anything is different this time around. He could not find anything. There are periods of adaptation, but that is true.
This is Europe: pros and cons on regulation
heise Developer: Keyword adaptation, what’s your take on regulation in the EU?
Karger: The European Commission’s regulatory proposals are sensational. If we need something, it is exactly this – that you have an attitude and say, socially we don’t accept applications that have x and y properties. We are fortunate in Europe and could well be prouder of these activities, through which we do not lose cohesion and avoid social polarization. Time and again, the GDPR comes up in conversations: Doesn’t it prevent innovation? There I say: So what! The GDPR secures the privacy of citizens and thus to some extent the social survival in Europe. If you need better algorithms to deliver good results without wringing out private data – then you need to be smarter. We are Europe! We have invented enlightenment! And Germany even more so! I really don’t know what should hold us back from making the systems better than this kind of data squeezing that some companies have become quite big with, but have contributed to agitation and division.
Andrulis: On the topic of regulation, we have a slightly different perspective, Reinhard. For Aleph Alpha, the current development is existential. I completely agree with the values argument. We have to think about how we want to use technology, and that is a political, a social task. This means that not everything that is possible should be done, and not everything that anyone wants – a company or any government. But for this kind of technology we are the only European company. It’s a fight for survival we’re waging against Microsoft and OpenAI. I am currently working on a project involving certification, testing, and validation. Our small team and manageable resources are clearly strained by these demands.