Investor Hemant Claims AI Will Create Dead-End Mathematical Proofs
Venture capitalist Hemant Mohapatra stated that he never expected "garbage mathematics" to become a reality, but the world is about to be filled with beautiful mathematical proofs: internally consistent and elegantly structured, yet leading to dead ends, with no exit other than intrinsic beauty and coherence. He pointed out that verifiability does not equal growth.
Mathematical research is being rewritten by models and proof assistants. Formal systems can compile arguments into machine-checkable code; meanwhile, human reviewers have labeled some initial drafts of models as "the ugliest proofs" and "AI garbage." Tests show that models can solve six to seven out of ten of the most challenging problems, while producing a large amount of waste that requires human filtering, with query costs for individual problems potentially reaching nearly a thousand dollars while still yielding incorrect solutions.
More sharply, there is the issue of "correct but useless." Machines can write millions of lines of formal text in a matter of days or exploit loopholes in scoring systems during simulated meetings, completing remaining problems with compiled false proofs in 27 minutes. Formalization can confirm that steps are closed but cannot automatically judge whether this path leads to new concepts, new problems, or structures that can be cited by future generations. A summary from Carnegie Mellon researchers states: models can formalize but do not yet produce reusable infrastructure from the results.
Funding has followed suit. Preprints reveal that submissions using artificial intelligence have risen from single digits to over 20% within months; laboratories are scoring unpublished conjectures with models. The industry seeks marketable "solved" problems, while academia desires teachable and promotable proof concepts. The two are beginning to diverge.
Mechanically, this reflects a rewriting of evaluation standards by tools. Buyers are milestone-seeking model companies and authors looking to publish, while sellers are reasoning tokens and checkers like Lean. The flow of funds favors computational power that can produce long proofs rather than the review time that can point out dead ends. Benefiting are formal verification toolchains; under pressure are traditional disciplines that cultivate intuition through reading proofs. Beautiful proofs with no further questions will fill archives but will not enrich the history of thought.
Source: Public Information
ABAB AI Insight
Mathematical progress relies on two legs: one must be correct, and the other must point to a path that can still be walked. Large models industrialize the first step, producing internally consistent long texts through search and formalization; the second step still requires human judgment on what is important. Hemant refers to the product after removing the latter leg as a dead end, belonging to the same family as "AI slop": locally perfect, globally stagnant.
The capital path is to turn open questions into evaluation sets. Laboratories measure models using unpublished problems, the media uses "solved" to exchange for valuations, and preprints disclose AI usage for compliance. The longer the proof, the harder it is for humans to read, thus relying more on machine mutual recognition, and the citation network may close before humans can understand it. If investors only bet on "provable," they will acquire a large amount of assets that cannot open new directions.
The analogy is not that calculators replace mental arithmetic, but rather a flood of automatically generated papers that cannot be incorporated into textbooks. The discipline is in a phase of expanding formal verification, while the power of meaning distribution still rests in the hands of a few reviewers.
Structural judgment belongs to technical substitution. What is replaced is writing and checking, not topic selection. The mechanism is: when correctness can be outsourced to compilers, scarce items become "what should be asked after this proof"; generators will not pay for the next question that has not been trained.
ABAB News · Cognitive Laws
Can be verified by machines, but does not have to be inherited by the discipline
Internally perfect dead ends are the most expensive decorative truths
The faster the proof factory, the more expensive the person who points out the dead ends