← ALL NEWS

INGENIØREN · 21 Nov 2026 · DENMARK

Do we get smarter when AI solves the big problems? Teknisk set

This is an episode of Teknisk set, the weekly podcast from Danish engineering outlet Ingeniøren, hosted by Liv Bjerg Lillevang and Henrik Heide. It asks whether people get smarter when AI solves the biggest problems, or whether we should simply outsource the hardest challenges to a machine without understanding how it works.

The starting point is the Millennium Problems, a list of seven mathematical puzzles with a million-dollar prize for each. Only one had ever been solved by a human, and the remaining six had stood unsolved since 2010. According to the episode, OpenAI's language model then took on the Navier–Stokes equation, which describes the motion of gases and liquids and underpins calculations for weather forecasts, ocean currents, airflow around a wing and blood flow in the body. The model is said to have cracked it in 88 hours. The world's leading mathematicians were left in shock. They could see that the model had created thousands of agents working in groups on different parts of the problem, but they could not work out how it reached the solution.

That gap is the heart of the discussion. Top mathematicians have worked through long chains of proofs, shifting levels of abstraction as they go. If future mathematicians hand the complex parts of mathematical language to an AI, the episode asks whether they remain mathematicians or become professionals who are merely good at prompting.

Two DTU professors join the discussion: Thomas Bolander (logic and mathematics) and Søren Hauberg (Mathematics and Computer Science, DTU Compute). The same two were interviewed three years ago, when language models had just taken off but were still flawed and unimaginative. They reflect on what the development means for identity, learning and reflection, ask whether AI can be a good gardener, and speculate on how history would have looked if ancient Greek mathematicians had had a language model. They also predict where language models will stand in three years.

One reader comment on the page notes that the 88-hour figure means little without knowing how much hardware and how many models were used.

Read the original ↗