Tech • AI • Robotics • Game
They wiped out my field”: The profession we thought untouchable has just fallen
7/10TL;DR
AI is advancing fastest in work that can be checked instantly and automatically, a shift highlighted by recent progress in formalized mathematics and likely to spread across many professional tasks.
KEY POINTS
Verification, not difficulty, is the key vulnerability
The emerging divide is not between simple and complex work, but between tasks that are easy to verify and those that are not. In reinforcement learning, systems improve by trying, being scored, and adjusting. Researchers describe a practical rule: the faster and more unambiguously a task can be checked, the faster AI can learn it at scale.
Why mathematics moved into the front line
Mathematics became unusually exposed because parts of the field can be translated into formal languages such as Lean, where every proof step is checked by software. That creates an instant, cheap and tireless judge. Unlike many real-world jobs, this lets labs run huge volumes of automated attempts against a precise correctness test.
OpenAI’s results jolted top researchers
A major shock came when OpenAI reported nearly 400 results on hard mathematical problems in a single release after earlier large-scale runs, including 10,000 model copies operating for 88 hours on Navier-Stokes-related work. Some mathematicians described the output as containing material comparable to multiple Fields Medal-level advances. The speed and volume, with an average of about three hours of model reasoning per result, intensified concern.
Human authority proved weaker than machine checking
Formal verification also exposed errors in previously published human papers. That matters because software does not defer to status, reputation or seniority. In a field built on proof, automated checking can become stricter than peer review and can undermine one of academia’s hidden protections: trust in prestigious names.
Even the breakthrough claims had limits
Not every result held up cleanly. Critics noted that some Navier-Stokes work relied on a version of the problem with an added external force, a route later argued to be a dead end for the original formulation. Other papers were described as too unclear to assess easily, and a few were reportedly withdrawn over sign errors. Still, those flaws were treated by many as evidence of delivery problems rather than proof that the underlying capability is unimportant.
Software engineering shows the next stage
Programming is also highly testable, yet senior engineers still hesitate to let autonomous agents ship code unchecked. The reason is that passing visible tests does not guarantee success in production. A database migration, for example, can satisfy every test and still trigger regulatory or reporting errors months later because the real specification was incomplete. In software, much of the job is discovering the problem itself, not only solving it.
Value is shifting toward those who define and own the test
As the cost of producing a verifiable answer falls, value moves elsewhere: choosing the right question, owning the evaluation system, and taking responsibility for the outcome. In companies, that means proprietary data, quality criteria, internal test cases and ground-truth feedback become strategic assets. Whoever controls the judge can train, evaluate and potentially replace the producer.
Training pipelines and career ladders may be disrupted
The change threatens the bottom of many professions first, especially junior tasks such as first drafts, routine analysis, basic coding and compliance work. That raises a structural problem: if entry-level, easily checked work is automated away, the path for developing future senior experts narrows. Similar fears surfaced in elite mathematics departments, where students were warned they may graduate into a transformed profession.
Two camps are forming
One response is resistance: hundreds of mathematicians have joined groups committed to minimizing or refusing AI use. The other response is adaptation, with researchers using machine-generated proofs as raw material to extend in a day what might otherwise have taken far longer. The divide is increasingly between defending an old territory and learning to navigate the new one.
The broader lesson reaches far beyond academia
The same logic applies to offices, laboratories and industry. A routine request to standardize quality criteria can become the moment an organization builds the evaluator that will later automate part of its own work. The central economic question is no longer how prestigious or difficult a task once was, but how cheaply and reliably its output can be scored by a machine.
CONCLUSION
The immediate risk from AI is highest where work can be verified quickly, repeatedly and without ambiguity. As that frontier expands, advantage will belong less to those who produce answers and more to those who define the question, control the evaluator and assume the decision.
Ask a question