GPT-5.6 Sol Pro closes 30-year optimization gap

AI is shifting the bottleneck of mathematical research from mechanical derivation to high-level prompting and formal verification, enabling researchers to solve multi-decade open problems in a single afternoon.

* The proof establishes a quadratic lower bound of Omega(d^2) for zeroth-order convex optimization, demonstrating that derivative-free optimization is fundamentally harder than gradient-based first-order optimization.

* The success of the 10-page CDC-style prompt indicates that guiding LLMs with highly structured reasoning frameworks, rather than raw query-response cycles, is critical for breakthrough mathematical reasoning.

* Formal verification via Lean acts as the ultimate validation tool, bypassing traditional peer review timelines and eliminating concerns about LLM hallucinations in complex proofs.

* This milestone suggests that math and theoretical computer science researchers will increasingly transition into directors of AI systems, leaving "medium-hanging" optimization problems to automated agents.