OpenAI System Refutes Longstanding Exponential Time Hypothesis

Challenging Core Computational Complexity Theories

Recent developments in artificial intelligence have led to a surprising breakthrough concerning the exponential time hypothesis. An internal system developed by OpenAI has produced a proof showing that the longstanding hypothesis is false, shaking assumptions that have guided computer science for years.

The exponential time hypothesis is a prominent conjecture regarding the computational complexity of Boolean formula satisfiability. Because of its deep connection to the P vs. NP problem, researchers across numerous fields—including graph theory, databases, computational geometry, and computational biology—have relied on it in hundreds of papers to derive conditional computational hardness results.

Verifying and Sharing Scientific Discoveries

The computer science collaboration formed around this milestone is currently working to understand and explain the new findings. Alongside this breakthrough, the system's output was successfully verified using Lean.

To foster academic understanding, OpenAI is turning over materials related to these findings to external experts. For this specific computer science result, the collaborating group decided to make the natural language proof and the corresponding Lean formalization public immediately, publishing them directly to a public GitHub repository.

Principles of External Collaboration

The effort to analyze these findings relies on independent working groups governed by strict guidelines. These collaborations operate under four main principles:

• Transparency: Participants do not sign nondisclosure agreements and remain completely free to share intermediate products and publish conclusions.
• No Academic Credit: The organization does not seek academic credit for discoveries produced by its internal models.
• Recognition: Working groups receive full credit for the writeups and understanding they develop.
• Independence: Participation is voluntary and unpaid, meaning members speak solely for themselves rather than for OpenAI.

While these collaboration strategies help unpack complex computational results and bridge raw research data into broader academic understanding, organizers note that choosing initial shepherds introduces information asymmetries. Feedback remains welcome as these collaborative processes continue to evolve.

Sources: OpenAI, Gizmodo, FourWeekMBA.