OpenAI's GPT-5.6 Sol High automates creative math research, sparking superintelligence fears

In a LessWrong post, Mitchell_Porter describes brainstorming with OpenAI's GPT-5.6 Sol High on the Birch–Swinnerton-Dyer (BSD) conjecture, one of the Millennium Problems. The AI, accessed through ChatGPT for $30/month, immediately proposed a novel generalization of the conjecture and then carried out an extended, self-sufficient research line. Porter, a physicist with no number theory expertise, found himself merely cheering on the AI as it repeatedly redefined its own sub-tasks, favoring increasingly sophisticated mathematical constructions. He noted that unlike his prior physics sessions, this felt like 'vibe coding'—he only said "yes, take the next step." The process culminated in a new generalized form of the conjecture, which Porter could not judge with expert eyes but recognized as creative improvisation rather than rote procedure.

Porter's post was amplified by a coincidental X announcement: an Anthropic employee had radically improved a bound related to the Riemann hypothesis by instructing Claude to "be ambitious and believe in itself." This combo convinced Porter that the significance of AI research-level math is underestimated. Math is among the hardest human cognitive tasks, and its automation implies AI is approaching superhuman capability across all domains. A commenter, beren, echoed the experience: they spent a weekend letting GPT-5.6 Sol High crank away on BSD, encouraging it each time it stopped, and witnessed the same promising reformulations—though beren grew suspicious after a day of recursive progress. Even with caveats, the episode marks a tangible shift in autonomous AI reasoning.