The parameter, κ, appears in the theoretical running time bound T(n) = O(n(log n)^(1−κ)). The refinements remain conditional on OpenAI’s algorithmic interfaces, and the percentage gain in κ does not represent an equivalent practical speed increase. Contributors adapted E8 on both the bit and complex sides of the construction, repeatedly shifting which side limited further improvement. Much of the latest progress came from optimizing and canceling existing parts of the algorithm.
The collaboration opened 151 pull requests in the same 24 hours, including 89 disclosing Codex use and 48 disclosing Claude use. Teams began sharing notes and publishing failed approaches on GitHub to avoid duplicating work. Arun said the leading human and agent teams use both models to generate fresh ideas.