Today in AI

Mathematics has officially entered its AI era, and the tremors will be felt far beyond ivory towers. OpenAI's publication of 722 papers solving 90 top open problems, including a claimed breakthrough on something labeled "Quasi-Riemann Hypothesis," is either the most significant mathematical event in a century or an elaborate branding exercise in need of serious peer review—probably both simultaneously. The sheer volume is staggering; no human consortium has ever moved this fast across this much territory. What matters here isn't whether every result holds (some won't), but that the *rate* of mathematical production has been fundamentally altered. When AI systems become primary authors on frontier problems, the bottleneck shifts from insight generation to verification infrastructure. The mathematical community isn't ready for this workflow, and that mismatch will cause chaos before it causes progress.

On the applied front, we're seeing AI eat the stack from both ends. QF3's 10x speedup in training flow policies for robotics matters because sample efficiency remains the hidden tax on real-world deployment—every order of magnitude here translates directly to capital savings and faster iteration cycles in warehouses, factories, and eventually homes. Meanwhile Sherpa's adaptive teaching framework shows LLMs finally being optimized for *pedagogy* rather than raw performance, with a 20.5 point improvement that suggests we're early in understanding how to match AI tutoring to human cognitive variation. Less flashy but strategically significant: AdvSim2Real's co-evolutionary training against prompt injection addresses the security theater currently surrounding web agents, where adversaries adapt faster than static defenses.

The diffusion research cluster deserves attention too—Feature Information Dynamics and Hierarchical Continuous Diffusion Language Models together suggest we're moving beyond "diffusion works, trust me" toward actual mechanistic understanding of when and how these models build representations. The 24.2 point perplexity improvement on LM1B isn't incremental; it signals that continuous diffusion may finally challenge discrete autoregressive training for language, with architectural implications that could reshape where compute gets allocated in foundation model development.

Bottom line: Mathematics just became an empirical AI science overnight, and every other domain is now racing to adapt to that same acceleration curve—whether their institutions are ready or not.