Do OpenAI's internal models actually solve fundamental problems, or are they merely another marketing tool designed to justify staggering compute costs?
OpenAI has announced the generation of mathematical, computer science, and physics proofs using its internal models. Reports claim these include the refutation of a central computer science assumption and a partial proof of a Millennium Prize Problem. Amid relentless pressure from investors demanding economic returns on multi-billion-dollar infrastructure investments, the company is shifting its focus from traditional marketing benchmarks to the generation of scientific proofs. Nevertheless, the company itself acknowledges that these manuscripts are merely raw data requiring external audit.
OpenAI has enlisted external experts, including a Nobel laureate, to verify the generated mathematical and physics proofs. These experts act as shepherds for peer review, with validation tailored to each specific discipline. In computer science and mathematics, proofs are additionally translated into Lean, a specialized programming language that enables a computer to check every logical step. Still, specialists note that this does not guarantee complete independence from the model, as Lean only verifies compliance with formal conditions while leaving the interpretation of meaning to human experts.
Revision of Basic Algorithmic Constraints
An inflection point in computer science came with OpenAI's claim that its system refuted the Exponential Time Hypothesis (ETH). This hypothesis concerns the Boolean satisfiability problem (SAT), which underpins computational complexity theory. With each new variable, the search space doubles in overall complexity, creating an exponential barrier for computing resources. According to the hypothesis, efficient short paths for the worst-case scenario do not exist—an assumption backed by hundreds of research papers across various fields.
Infrastructure Risks and the Economics of Proofs
Any change to foundational algorithms triggers a recalculation of inference costs and the operation of specialized software like SAT solvers used in chip and software verification. The practical impact of this theoretical breakthrough remains uncertain, since real-world applied problems are often simpler than the worst-case scenarios posited in ETH. Meanwhile, the mathematical component also touched upon the local smoothing conjecture, originally proposed by mathematician Christopher Sogge in 1991 and partially solved in two dimensions in 2019 by Guth, Wang, and Zhang, whereas OpenAI's models have now produced a proof for all dimensions.
The long-term viability of AI giants' business models depends on their ability to convert raw compute into verifiable applied outcomes. If the verification of complex physical and mathematical derivations requires Nobel laureates and manual audits by world-class scientists, scaling these solutions to the corporate sector runs up against severe staffing and timeline constraints. Infrastructure demands daily multi-billion-dollar infusions, yet monetization relies on trust in raw outputs filtered by independent experts.
The transition to machine-generated science redefines how AI capabilities are measured while simultaneously exposing the limits of automation in foundational disciplines. When machine synthesis requires personal validation by the original authors of hypotheses, the economic model of artificial intelligence proves constrained by scarce human capital, returning us to the fundamental question of whether these results represent genuine utility or costly administrative overhead.
Editorial analysis
We have written about this before: July 1 — OpenAI Uses GPT-4 to Explain AI Interpretability; August 1 — according to OpenAI, generating proofs for all ten problems cost approximately $2,000 at the API rates of the Sol model. OpenAI announced that its internal models generated proofs across physics, computer science, and mathematics, including partial progress on a Millennium Prize problem. Pure academic marketing designed to justify massive computing bills.
We covered this: July 1, August 1