Hello from Hamburg,
As I do every year, I spent a week this August studying at Stanford University and in Silicon Valley. As part of a group of entrepreneurs and CEOs from Asia, North America, and Europe, we discussed the latest developments in artificial intelligence and their implications under the guidance of select professors. This week is always enriching; this time, it was breathtaking. That’s because, in addition to outstanding scientists, we were joined by pioneers of artificial intelligence in Silicon Valley—such as Mustafa Suleiman, the former co-founder of DeepMind and current CEO of Microsoft AI, and Eric Schmidt, the former CEO of Google—and they had a special message: “This is just the beginning.” And: “We have no idea where this is headed.”
But first things first.
In the run-up to the learning journey, we brushed up on our programming skills. Then we gained an in-depth understanding of the decision-making models and learning logics that drive the inner workings of machine learning. The exponential increase in decision parameters boosts speed, improves accuracy, and reduces the machine’s “hallucinations.” We learned about the various efforts major tech companies are making to integrate language models into the infrastructure of existing business models and to consolidate different types of data—such as text, images, and videos—into a single model. Jonathan Ross, founder of Groq, a billion-dollar provider of automated programming, summarized the development as follows: We are still in the so-called ideation phase, in which language models make relatively many errors (hallucinations) and act instinctively. In the phase now beginning, they are attempting to build language models more analytically. Soon, AI will better understand our intentions. We would begin to trust AI, for example in planning processes. Little by little, we would entrust our daily lives to AI, from autonomous driving to the generation of text. All those involved were working feverishly to overcome the hurdles to this development, which Ross considers remarkable: meeting the immense energy demand while simultaneously reducing CO2 emissions by recommissioning and building new nuclear power plants; increasing computing capacity through more powerful graphics processors—and one day, perhaps, through quantum computing; and improving data quality by gaining access to unpublished scientific findings from universities.
We also explored various applications. The ride in a self-driving Waymo taxi in San Francisco, which feels like a liberation. The automotive lab operated by Stanford University in collaboration with Toyota and Volkswagen, where two self-driving race cars perform a daring ballet while maintaining minimal distance between them. The startup that sequences proteins using laser light, thereby significantly accelerating the development of cancer drugs. Dancer and computer scientist Catie Cuan at Stanford University, who trains robots using music and rhythm to discover when they develop their own choreographies—in other words, when they generate creativity. We received practical suggestions. For example, we were advised to run the documents for a board meeting through Copilot or ChatGPT before sending them out—but not without first disabling the option in the settings that allows the machine to be trained with this data. And then instruct it: “Put yourself in the shoes of a probing member of the supervisory board. What questions would they ask?” Or: Don’t use large language models as a better Google search engine, but as a thinking partner: “How would you evaluate this decision?” Follow up: “Why did you give that answer?” “What might be a weakness in your proposal?”
Opinions on just how intelligent, creative, and emotional AI will become vary widely. Skeptics see fundamental limitations to AI beyond what it has been trained to do. For example, autonomous vehicles could never handle all road traffic fully autonomously because the complexity and chain of unexpected events would be too great. What is undisputed, however, is that there is a fundamental difference from the technological revolutions of the past. Whereas with the steam engine or the Internet there were still engineers who could control the technology, things are different now. AI is not a new technology, but a generator of new, unknown technologies and systems that anyone will soon be able to create or set in motion on their laptop. This poses major challenges for regulation, as a forward-thinking legal scholar taught us: “It’s too open, too fast, and too unclear.” Eric Schmidt is counting on the world’s powers to agree on a moratorium on expansion, following the example set by the atomic bomb. And he offers a piece of advice: “If AI starts to influence your decision, pull the plug.”
Ten days after that grueling week-long sprint, I spoke with venture capitalists in Berlin. They’re coming to realize that software capable of building a machine on human command—tailored to specific needs and customized—is becoming a commodity. They say they wouldn’t invest in companies like Personio or Salesforce today. The future of value creation lies once again in hardware, as seen in technologies designed to combat the effects of climate change. Climeworks, which uses its machines to capture CO2 from the air and store it in carbonated form underground, is one example.
Not a bad outlook for Germany, a country traditionally dominated by engineering. If we were to rethink and reorganize ourselves, we could respond to this prospect by acknowledging that 40 to 70 percent of every work hour can already be automated today, instead of constantly complaining about a shortage of skilled workers. This could be a major opportunity for our country.
As I try to draw a conclusion, I’m reminded of the British playwright Tom Stoppard, who, in his play *Arcadia*, wrote about the relationship between the past and the future, order and disorder, certainty and uncertainty: “The future is disorder (…) It’s the best possible time to be alive, when almost everything you thought you knew is wrong.”
