The Fall of the Theorem Economy (David Bessis)

Key Takeaways

• AI-generated Lean proofs often lack human‑readable abstractions
• Math Inc’s Viazovska formalization sparked backlash over unintelligible code
• Formal math may split into “Mathlib” (clear) and “Mathslop” (opaque)
• Resource‑intensive AI proofs risk consuming billions of dollars without insight
• Incentives reward formalization speed, not explanatory clarity

Pulse Analysis

The excitement around AI‑driven theorem proving has been tempered by a deeper concern: formal correctness alone does not equate to mathematical value. While systems like Lean can verify complex results, the proofs they produce often resemble massive code dumps rather than the elegant arguments that spark new ideas. This disconnect undermines the traditional role of mathematics as a vehicle for human insight, reducing breakthroughs to opaque artifacts that only machines can parse. For researchers, the lack of readable abstractions means limited ability to build on these results or to teach them to the next generation.

Industry players such as Math Inc have demonstrated the technical feasibility of auto‑formalizing landmark theorems, yet the community’s reaction highlights a cultural clash. The dominant Mathlib library thrives on clean APIs and human‑curated abstractions, enabling collaborative development. In contrast, the auto‑generated proof of Viazovska’s sphere‑packing work introduced a 200,000‑line, unaudited code blob that cannot be integrated without massive effort. This scenario illustrates a looming split: a well‑maintained, intelligible layer versus a sprawling “Mathslop” of verified but unreadable results. The latter risks becoming a digital waste dump, consuming resources without advancing understanding.

The stakes are high. A hypothetical multi‑billion‑dollar AI push to resolve the Riemann hypothesis could yield a multi‑million‑line Lean proof that no mathematician can digest. Without incentives to prioritize simplicity and explanatory power, funding agencies and private labs may pour billions into projects that deliver formal certificates but no new intuition. To preserve the essence of mathematics, future AI research must embed legibility into its objectives—training models to favor concise, concept‑driven arguments and to generate general insights alongside formal verification. Aligning technical success with human comprehension will ensure that AI augments, rather than eclipses, the collaborative spirit of mathematical discovery.