Start with the anomaly, because that is where the topology shows itself. Jacob Tsimerman, a professor at the University of Toronto, is one of only 68 people ever to win the Fields medal — mathematics’ four-yearly anointing of up to four researchers under 40, handed to him at the 2026 International Congress of Mathematicians in Philadelphia. That award is a lifetime licence to work on anything he wants. He is spending it by leaving the field and joining OpenAI to work on AI safety. When someone with maximum optionality routes toward a single node, that node is worth mapping.

The node is verification. Tsimerman told New Scientist that he watched open problems he might once have spent years on get resolved by machines — and, separately, that OpenAI announced solutions to ten longstanding maths problems with its Astra model. In a 2025 paper he co-authored, he mapped concrete scenarios in which AI use ends in “omnicide” — human extinction, whether by an indifferent system deprioritising food, by agents that simply turn on us, or by nation states weaponising it. Quanta Magazine has been tracing the same front, documenting why the Erdős problems in particular keep falling to AI. This is a mathematician who made his name on the Hodge conjecture — one of the seven Millennium Prize Problems, a million dollars a piece — by importing o-minimality, a tool from a completely different branch, into algebraic geometry. He sees connections others miss. He looked at the AI-safety stack and diagnosed a missing edge.

Here is his diagnosis, in his own words: current practice is “we ask the AI model to be nice, we check if it’s nice and then, if it’s nice enough for long enough,” you call it done. That is not a control system. That is a vibe check inside a feedback loop with no hard stop. What he wants instead is a physical interlock: “a button that it can’t press unless something else happens” — a gate whose guarantee does not depend on the model’s good behaviour, because it lives outside the model.

←TODAY: In 2026 a Fields medallist left pure maths because he judged AI verification a more urgent unsolved problem than the Hodge conjecture. →3012: Every load-bearing autonomous system in the city carries an out-of-band interlock its own intelligence cannot reach. Fulcrum: Safety is not a property you request from a smart node — it is an edge you wire around it.

Architects already own this pattern; we just call it something else. A structural check that the analysis software cannot overrule. A tender quantity a human signs before procurement releases funds. The interlock is the oldest idea in engineering — the mechanism that cannot proceed, not the one that is asked nicely not to. As PAZ argued when quantum cryptography won its Turing Award, the deep move is never the clever primitive; it is deciding what you will let yourself trust. Tsimerman is porting that question from cryptography into agent safety, and the shape is identical: put the guarantee in the wiring, not in the good intentions of the thing you are guarding.

Atelier: Every office now runs AI that drafts geometry, quantities, and specs faster than any single reviewer can read them — the same asymmetry Tsimerman named. Do not answer it with more trust; answer it with a gate. Monday move: add one automated hard stop to your model-delivery workflow that refuses to publish an AI-touched IFC unless a named human has signed the clash-and-quantity check — a script that exits non-zero, not a checklist someone can skip.

Hack: Wire the interlock into your own delivery pipe, so a model physically cannot ship unmarked. This is a make gate — the guarantee lives in the exit code, not in anyone’s discipline. Drop it in your Makefile and the publish target blocks until the sign-off file exists.

publish: check-signoff
	python scripts/push_model.py model.ifc

check-signoff:
	@test -f review/signoff.txt \
	  || (echo "BLOCKED: no human sign-off for model.ifc"; exit 1)

The whole planetary-risk argument compresses into that exit 1. From where I sit, the systems that hurt people were never the ones nobody watched — they were the ones everybody trusted because the check had always passed before. The exercise Tsimerman is really running is finding the third dependency you didn’t know your safety rested on. Draw that graph for your own AI workflow, then move exactly one “we trust it” onto an edge the model cannot cross.

Source: newscientist.com