We throw computing at our ignorance. And it works.
Daniel Dobos of Swisscom on AI’s energy demands, lessons from CERN, radiation risks in space and teaching children to question technology.
Discuss this episode and meet other Swiss tech builders in the TechDrive WhatsApp group.
Key insights: executive summary
- More computing does not mean we fully understand learning. Daniel argues that today’s AI progress relies heavily on large amounts of computing and energy. He expects further advances to come from understanding learning better and building models more closely grounded in the physical world.
- The human brain gives him reason to expect more efficient AI. Daniel uses the brain’s low energy requirements as inspiration for neuromorphic computing and other approaches informed by physics and neuroscience. His point is that intelligence need not always require the energy demands of today’s large systems.
- AI in space comes with physical risks. Drawing on his experience with electronics at CERN, Daniel warns that radiation can alter stored or processed information. He questions how reliably an AI system’s output could be checked when its hardware is exposed to radiation and solar storms.
- AI literacy should develop critical thinking. When Daniel introduced his children to generative AI, they eventually recognised its potential for misuse themselves. He describes that moment as a powerful opening for discussion, with a parent or teacher present to help them think it through.
- Technology creates impact through people working together. At THE Port humanitarian hackathons, research tools met problems brought by humanitarian organisations. Daniel’s examples include adapting detector-packaging technology to help identify disaster victims and inviting a magician to explain encryption.
- New technology can feel ordinary surprisingly quickly. Daniel recalls his son becoming comfortable with a robotaxi in Shanghai after about 15 minutes. He expects robotics adoption to differ across places and use cases, with efficiency and the experience of using a robot both influencing its role.
“We solve all our ignorance with throwing enormous amount of computing, enormous amount of energy on solving our ignorance. And surprisingly, it works.”
“I think AI and digital technologies have their role in education, but it has to be well thought through in the entire education process.”
What does Daniel mean by throwing computing at our ignorance?
Daniel sees the impressive output of current AI systems alongside an incomplete understanding of how learning works. In his view, large amounts of computing and energy compensate for some of that missing understanding.
He includes himself in that criticism. His interest is in what comes next: approaches that draw more deeply on physics, neuroscience and models of the world, potentially producing useful intelligence with much less energy.
It is an argument for further scientific progress. He expects better understanding to change how systems are built, rather than treating ever-larger computing requirements as inevitable.
How did working at CERN lead him into AI?
Daniel spent much of his career at CERN developing silicon and diamond detectors. The volume and speed of experimental data made advanced analysis essential.
In the conversation, he describes the ATLAS work he was involved in as producing around 80 terabytes of data per second, with systems performing around 40 million inferences per second. He explains that the team used specialised hardware, including FPGAs and digital signal processors, to process data in real time.
That experience shaped his interest in AI as a tool for a specific scientific problem, and his attention to the hardware and constraints behind it.
What does technology for good look like in practice?
After the Higgs boson discovery, Daniel and colleagues asked how their research could affect people’s everyday lives. That question helped lead to THE Port humanitarian hackathons, bringing researchers and other specialists together around challenges from humanitarian organisations.
One example involved adapting packaging technology used for CERN detectors to give forensic teams more time to identify people who had died in disasters. Daniel describes its value in human terms: helping families learn what happened to someone they loved.
AI was sometimes part of the solution, but he stresses the importance of the team. People who had not worked together could combine their skills to solve problems they could not tackle alone.
Why did a humanitarian hackathon need a magician?
Daniel recalls a team working on homomorphic encryption, which allows calculations on encrypted information without first decrypting it. A magician helped explain the idea using a box, folded paper and a number.
The demonstration made a difficult technical concept easier for others to understand. For Daniel, it showed why multidisciplinary teams should include people whose contribution is not obvious at first.
Communication and practical knowledge can determine whether an idea reaches the people who need it. He also gives the example of logistics expertise identifying where an otherwise digital project might fail.
How can children learn to question AI?
Daniel describes sitting with his children as they explored tools that generated images and text. After experimenting creatively, they began asking whether the same tools could be used to do something harmful to a teacher, friend or relative.
The next realisation was personal: someone else could do the same to them. Because the questions came from the children, Daniel felt they were ready for a more meaningful discussion about consequences.
His approach combines exploration with an adult who is available to discuss what happens. He wants AI in education to support creativity and critical thinking, with careful attention to when and how the tools are introduced.
Why is Daniel cautious about running AI in space?
Daniel’s experience at CERN includes operating electronics in environments with high radiation. He sees related challenges for computing in space: radiation can flip a bit from zero to one or the other way around.
His concern is how an altered calculation might affect an AI recommendation and how the system would detect that something had gone wrong. He also raises the additional disruption that solar storms could cause.
He recognises possible advantages, including access to solar energy, but wants the engineering risks to receive equal attention, especially when AI supports critical decisions.
Could the brain and dreaming inspire more efficient AI?
Daniel points to the human brain as evidence that useful intelligence can operate with limited energy. He cites a range of roughly 20 to 50 watts to explain his optimism about finding more efficient approaches. This is an illustration of his argument, rather than a like-for-like benchmark against a data centre.
He is interested in neuromorphic computing and approaches informed by physics and neuroscience. Watching his children learn also makes him question what current AI methods may be missing.
He describes explaining a concept in the evening, then finding that a child could explain it the following morning. That leads him to ask what an equivalent of dreaming, reorganising or relearning concepts might mean for AI. He presents it as a research question, not an established explanation.
What did a robotaxi ride in Shanghai reveal about adoption?
Daniel recalls taking his 11-year-old son in an autonomous car in Shanghai. After about 15 minutes, the experience had become normal to his son.
He also saw humanoid robots serving coffee or selling products. He acknowledges that some tasks could be completed more efficiently by machines without a human shape, while the interaction itself could add something to the experience.
His expectation is that adoption will vary. Different places and businesses will experiment with humanoid and non-humanoid machines, then discover where each makes sense.
Who is Daniel Dobos?
Daniel Dobos is Research Director at Swisscom and a particle physicist who spent much of his career at CERN, working on detectors and the analysis of large experimental datasets. He helped establish THE Port humanitarian hackathons and is interested in bringing research into practical use. In this conversation, he also speaks as a father of three who wants technology to be safe and useful for the next generation.
Episode show notes
Daniel Dobos joins Justinas Mišeikis for a drive through Zürich that connects particle physics with AI, humanitarian work and everyday life. From processing experimental data at CERN to helping forensic teams identify disaster victims, he explains why the people around a technology matter as much as the technology itself.
The conversation explores children discovering AI’s risks, radiation challenges for computing in space, the brain’s energy efficiency and what learning during sleep might suggest for future research. Daniel also shares his experience of robotaxis and humanoid robots in Shanghai, before returning to the episode’s central question: how much better could AI become if we understood learning more deeply?
More conversations from the Swiss tech ecosystem
- Nobody Has Left a Robot Working Overnight | Yulia Sandamirskaya, ZHAWYulia Sandamirskaya, ZHAW Zurich University of Applied Sciences
- Physical AI Is Ready, Buildings Are Not | Kateryna Portmann, ANYboticsKateryna Portmann, ANYbotics
- Why Robots Can Backflip but Struggle With Real Work | Enzo WälchliEnzo Wälchli, Operator Circle
Follow the Swiss tech ecosystem, one drive at a time
New in-car interviews with founders, investors and researchers building Zurich's deep tech, AI and robotics scene. Want to discuss this episode, ask follow-up questions and meet other builders?