OpenAI Retracts Three Mathematical Findings

In a recent update shared via X (formerly Twitter), Dan Roberts—a scientist at OpenAI and co-founder of Diffeo (which was acquired by Salesforce)—announced several changes to the organization's mathematical repository on GitHub.

Roberts, who has a background in gravity studies and co-authored The Principles of Deep Learning Theory, indicated that the openai/math repo is a living project. The team intends to continuously refine the content, adding new formalizations and correcting any errata as they are discovered.

📊 Summary of Repository Changes

The latest update to the math/history.md file reflects a significant amount of churn in the formalization process:

---

⚠️ The Withdrawn Manuscripts

While the repo grew in some areas, three specific results were pulled back. These complex topics involve advanced algebraic geometry, specifically focusing on $\text{K3}$ surfaces and abelian varieties.

The following items have been removed from the active record:

• Algebraicity of Weil classes on split abelian eightfolds
• Algebraicity of Kuga–Satake Correspondences for $\text{K3}$ Surfaces
• The rational Hodge conjecture for products of $\text{K3}$ surfaces

"To save you a click: As a result, we have withdrawn the following three manuscripts..." — @SmokeAwayyy

---

🔄 The "Software-Style" Proof Cycle

The community reacted with a mix of fascination and skepticism. Some observers noted that mathematical proofs are now being treated like iterative software releases.

Community Perspectives

The reaction on X highlighted a shift in how formal mathematics is being published:

• Anderson (@ItsGoharr): Suggested that proofs are shipping like software now.
• Bicycle (@BicycleSocials): Expressed a more critical view, asking, "Respectfully why did you not check these before releasing them?"

This approach mirrors a git workflow more than a traditional academic journal submission:

git commit -m "Fix errata in K3 surface proof"
git push origin main

---

🖼️ Key Figures and Context

Below are the profiles of the individuals involved in the discussion:

Dan Roberts (@danintheory)

Community Contributors

Reference Material:

Post Date: October 8, 2026