LLM scaling, agent reliability and Europe’s AI stack ft. Daniel Gillblad

About this episode

No news roundup this time, and in English. I met Daniel at CyberChess in Riga, where we ended up on stage together, and the conversation afterwards was good enough that I asked him on the show a week later. He has spent 25 years in AI, first in academia and later running research labs at the Swedish Institute of Computer Science and the Chalmers AI Research Centre. For the past three years he has been Chief of AI at Recorded Future. I wanted to know from someone who was doing this long before the LLM boom whether the past few years were a genuine scientific breakthrough or just very good scaling. His answer surprised me, and it reframed how I think about where the remaining headroom actually sits. We also get into what benchmarks are still worth, and why the stochastic-parrot debate says more about us than about the models. Then agents, where my own experience is that the demos look great and reliability falls apart on anything longer than ten minutes. Daniel explains what is really holding these systems back, and why the fix has less to do with AI than most people expect. From there to the Hugging Face incident: we both think the industry latched onto the wrong part of it, and t