A Reddit-driven claim that GPT-5.4 produced a new analytic number theory method solving an open Erdős problem on primitive sets is circulating with reported comments from Terence Tao and Jared Duker Lichtman, and how the AI research community handles the next seventy-two hours will say more about AI's scientific credibility than the claim itself.
The specific problem at the center of the claim is the Erdős conjecture on primitive sets, a longstanding open question in analytic number theory concerning the sum of reciprocals of integers in primitive sets weighted by a logarithmic factor. Jared Duker Lichtman proved a significant result related to this conjecture in 2022, establishing that the primes maximize a particular quantity over all primitive sets, which represented one of the more celebrated number theory results of that year. The claim now circulating on Reddit is that GPT-5.4 found a new method that extends or completes work in this area, and that both Lichtman and Terence Tao, arguably the most prominent living mathematician, have commented on the result.
Before treating this as confirmed scientific news, several things need to be established that the Reddit thread alone cannot establish. First, where exactly did Tao and Lichtman comment, and what precisely did they say? A comment noting that a model's output is interesting or worth examining is categorically different from a comment affirming that a proof is correct and novel. Tao maintains an active blog and has previously engaged publicly with AI-assisted mathematics, including AlphaProof's work on IMO problems in 2024, and his bar for endorsing a result is extremely high. Second, has the purported proof been written up in a form that can be checked line by line by independent mathematicians, or does the claim rest on a model output that has not yet been subjected to formal verification? Third, and most critically for assessing the nature of the achievement: did GPT-5.4 produce the key insight autonomously, or did it accelerate or complete a human-directed proof search where the conceptual framework was already substantially in place?
The framing of AI mathematical achievement has significant consequences beyond academic credit assignment, and this is where the story becomes directly relevant for founders and investors. If a frontier model can produce genuinely novel proof strategies for open problems in analytic number theory without substantial human scaffolding, the commercial implications for research infrastructure are considerable. Pharmaceutical companies, materials science firms, cryptography labs, and financial modeling operations all employ research mathematicians and theoretical scientists whose core function is producing novel structured reasoning under uncertainty. A model that can contribute at that level is research infrastructure in a way that changes headcount calculations and institutional value chains.
If, on the other hand, GPT-5.4 accelerated a proof that Lichtman or another mathematician was close to completing anyway, the achievement is still meaningful as a demonstration of AI-assisted research but does not imply the same autonomous frontier capability. The distinction matters because much of the current investor narrative around frontier AI value rests on the stronger claim. Labs and their backers are arguing, with increasing confidence, that models are approaching the ability to conduct genuine scientific research rather than merely assisting it. A viral Reddit thread that gets treated as confirmation of that argument without rigorous verification is benchmark theater, not science, regardless of how impressive the underlying output might be.
The mathematics community has better verification infrastructure than most fields for exactly this kind of claim. Proofs are either valid or they are not, and the community of people capable of checking a result in analytic number theory, while small, is highly motivated to do so quickly when a significant claim surfaces. If the GPT-5.4 result is real, formal confirmation will come within days to weeks through the standard channels: a preprint on arXiv, commentary from Tao's blog, or a formal write-up from Lichtman himself. The absence of those things after a reasonable window is itself informative.
The moat question for startups building on AI research claims
Assuming some version of this story holds up under scrutiny, the commercial question it raises for startups is one the industry is not yet equipped to answer cleanly: who captures value when AI models contribute to frontier research before institutions know how to validate, credit, or monetize that contribution? Academic publishing moves slowly. Patent systems were not designed for AI-generated mathematical methods. And the models producing these outputs are owned by labs whose commercial interests do not necessarily align with making the outputs freely available to the research community.
Startups building in AI-for-research verticals, whether in drug discovery, materials science, formal verification, or pure mathematics, are operating in a credit and IP environment that has not caught up with what the models can do. The defensible position in that environment is not owning the model capability, which commoditizes faster than any startup can maintain a moat around it. It is owning the domain validation infrastructure: the evals, the expert networks, the verification pipelines, and the institutional relationships that can distinguish a genuine research contribution from a plausible-looking hallucination. That infrastructure is scarce, takes time to build, and is not something a frontier lab will build for every specialized domain.
The Erdős claim deserves to be taken seriously and verified rigorously, in that order. If it confirms, it will mark a meaningful moment in the case for AI as research infrastructure rather than research assistant. If it does not confirm, it will be a useful reminder that viral scientific claims require the same skepticism as viral financial ones, and that the gap between a model producing impressive mathematical text and a model producing correct mathematical proofs is still real, still consequential, and still worth measuring carefully before building a business case around it.
Also read: Gemini Omni rumors are circulating ahead of Google I/O and what founders do with unverified platform leaks says more than the leaks themselves • AI models are hitting a data quality wall and the open web is the reason why • Derrick Downey built a number one App Store hit with Claude and no coding experience and the template he used is sitting there for anyone willing to try it
