AI · 7h ago
GPT-5.6 Solves 30-Year Math Problem in Convex Optimization
On July 17, 2026, GPT-5.6 Sol, guided by a researcher's prompt, proved a lower bound of Omega(d^2) function evaluations for convex optimization, matching a 30-year-old upper bound. The proof took 148 minutes of sustained reasoning. This settles the complexity of a fundamental problem underlying machine learning algorithms.
Meridian48 take
The achievement is genuine, but the headline's 'nobody noticed' is overstated — the HN thread had 513 points; the real story is how AI is now tackling lower bounds, which are harder than finding algorithms.
gpt-5.6convex-optimization