We Throw Computing at Our Ignorance | Daniel Dobos on TechDrive Zürich

At a rooftop barbecue at CERN, a group of researchers asked each other what any of their work changed for ordinary people. One answer that came out of those evenings involved body bags, and the packaging used to protect particle detectors, and giving forensic teams enough time to return names to families.

Daniel Dobos trained as a particle physicist and spent most of his career at CERN building silicon and diamond detectors. The data rates forced him into machine learning long before the current wave: the ATLAS experiment produced 80 terabytes a second, far too much to store, so the analysis had to happen in real time. He went on to co-found THE PORT, a set of humanitarian hackathons run with challenges from the United Nations and the International Red Cross.

The technical heart of this conversation is a claim about understanding rather than cost. Daniel argues that we do not know how learning works, so we substitute scale for the missing knowledge, and that scale turns out to work surprisingly well. Good enough for fluent text, not good enough to call the problem solved. He points at reasoning benchmarks where a ten year old still beats the frontier models, and at the human brain running on 20 to 50 watts, as evidence that a far more efficient answer exists and we have not found it yet.

In this episode:
• Why 80 terabytes a second forced CERN into real time inference on FPGAs, not GPUs
• Forty million inferences a second, years before the AI hype
• The rooftop barbecue that moved him from pure research into the ecosystem
• THE PORT, and how detector packaging ended up giving families back their dead
• Why a hackathon team working on encryption recruited a magician
• The Harvard study suggesting one generation may not be smarter than the last
• What social media should have taught us about scaling a technology
• Why he let his children play with AI image generators before saying anything
• The question they arrived at on their own, and why that mattered more than a warning
• Fire as the oldest example of a technology that runs both ways
• Radiation in orbit, and why a flipped bit is worse for a transformer than for deterministic code
• The brain on 20 to 50 watts, and what neuromorphic and energy based approaches might change
• Learning overnight, and what dreaming could mean for a language model
• Fifteen minutes in a robotaxi in Shanghai, and how quickly an eleven year old stopped noticing

About the guest: Daniel Dobos is a particle physicist who spent most of his career at CERN working on silicon and diamond detectors and on real time machine learning for the ATLAS experiment. He is a co-founder of THE PORT humanitarian hackathons and a father of three.

About the show: TechDrive Zürich is a series of candid, in-car conversations with the people building technology in and around Zürich, hosted by Justinas Mišeikis.

🔗 Connect
• Guest: Daniel Dobos - linkedin.com/in/danieldobos
• TechDrive Zürich, hosted by Justinas Mišeikis - techdrive.tv
• WhatsApp Group: chat.whatsapp.com/JgGidD8SqsuFzohXvZmo5z
• Follow the LinkedIn newsletter for the whole series: linkedin.com/newsletters/techdrive-z%C3%BCrich…
• TechDrive Zurich website: techdrive.tv
• YouTube channel: youtube.com/@TechDrive-Zurich