
13 years ago I left academia to become a founder. My last research paper on language models ended with an open conjecture that I never found the time to pursue. tud.qucosa.de/landing-page/?… Today I gave that paper to ChatGPT. What surprised me was that it quickly connected my conjecture to areas of mathematics I had never thought to combine: branching processes, Perron–Frobenius theory, and analytic combinatorics. The problem had been sitting at the intersection of several fields. Those connections turned out to be the key. They led to a proof of the conjecture and a much deeper understanding of why it is true. We proved that these models undergo a remarkable phase transition: every non-critical model makes very long sentences exponentially unlikely, but exactly at the critical boundary that exponential decay disappears and becomes a universal power law. github.com/TorstenStueber/Yi… The irony is hard to miss: my last academic paper was about language models. Fourteen years later, it was finally completed with the help of one. The recent breakthroughs of AI in mathematics suddenly feel very personal.















